This page intentionally left blank

Multiwavelength Optical Networks, Second Edition

Updated and expanded, this second edition of the acclaimed Multiwavelength Optical
Networks provides a detailed description of the structure and operation of modern optical
networks. It also sets out the analytical tools for network performance evaluation and
optimization for current and next generation networks, as well as the latest advances in
enabling technologies.
Backbone optical networks are evolving to mesh topologies utilizing intelligent net-
work elements; a new optical control plane is taking shape based on GMPLS; and
significant advances have occurred in Fiber to the Home/Premises (the “last mile”),
metropolitan area networks, protection and restoration, and IP over WDM. Each of
these is treated in depth, together with new research on all-optical packet-switched net-
works, which combine the speed of optics with the versatility of packet switching. Also
included are current trends and new applications on the commercial scene (wavelengths
on demand, virtual private optical networks, and bandwidth trading).
With its unique blend of coverage of modern enabling technologies, network archi-
tectures, and analytical tools, the book is an invaluable resource for graduate and senior
undergraduate students in electrical engineering, computer science, and applied physics,
and for practitioners and researchers in the telecommunications industry.

Thomas E. Stern is Professor Emeritus of Electrical Engineering at Columbia University,

New York, and has served as department chair and technical director of Columbia’s
Center for Telecommunications Research. A Fellow of the IEEE, he holds several patents
in networking. He has also been a consultant to a number of companies, including IBM,
Lucent, and Telcordia Technologies.

Georgios Ellinas is an Assistant Professor in the Department of Electrical and Computer

Engineering at the University of Cyprus, Nicosia. He has held prior positions as an
Associate Professor at City College of New York, as a Senior Network Architect at
Tellium Inc., and as a Senior Research Scientist at Bell Communications Research. He
has authored numerous papers and holds several patents in the field of optical networking.

Krishna Bala is currently the CEO of Xtellus, a company that manufactures fiber opti-
cal switches. Krishna was the co-founder and CTO of Tellium (NASDAQ: TELM), a
successful optical networking company. Prior to that he was a Senior Research Scien-
tist at Bell Communications Research. He holds a Ph.D. in electrical engineering from
Columbia University.
Multiwavelength Optical
Networks, Second Edition
Architectures, Design, and Control

THOMAS E. STERN
Columbia University

GEORGIOS ELLINAS
University of Cyprus, Nicosia

KRISHNA BALA
Xtellus
CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, São Paulo

Cambridge University Press

The Edinburgh Building, Cambridge CB2 8RU, UK
Published in the United States of America by Cambridge University Press, New York

www.cambridge.org
Information on this title: www.cambridge.org/9780521881395
© Cambridge University Press 2009

This publication is in copyright. Subject to statutory exception and to the

provision of relevant collective licensing agreements, no reproduction of any part
may take place without the written permission of Cambridge University Press.
First published in print format 2008

ISBN-13 978-0-511-48054-6 eBook (NetLibrary)

ISBN-13 978-0-521-88139-5 hardback

Cambridge University Press has no responsibility for the persistence or accuracy

of urls for external or third-party internet websites referred to in this publication,
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.
To Monique, who has always been there for me. To our children and our grand-
children (T.E.S.)
To my loving mother, Mary, and sister, Dorita, and the memory of my beloved
father, Nicos (G.E.)
To my wife, Simrat, and our children, Tegh and Amrita (K.B.)
 Contents

Figures page xvii

Tables xxix
Preface to the Second Edition xxxi
Acknowledgments xxxv

1 The Big Picture 1

1.1 Why Optical Networks? 1
1.2 Objectives of an Optical Network Architecture 4
1.3 Optics versus Electronics: The Case for Transparent
Multiwavelength Networks 9
1.4 Optics and Electronics: The Case for Multilayered Networks 12
1.5 Network Hierarchies 16
1.6 A Little History 18
1.7 Overview and Road Map 22

2 The Layered Architecture and Its Resources 28

2.1 Layers and Sublayers 29
2.2 Network Links: Spectrum Partitioning 34
2.3 Optical Network Nodes: Routing, Switching, and Wavelength
Conversion 39
2.3.1 Static Nodes 40
2.3.2 Dynamic Nodes 46
2.3.3 Wavelength Converters 63
2.4 Network Access Stations 67
2.4.1 Transmitting Side 70
2.4.2 Receiving Side 71
2.5 Overlay Processors 74
2.5.1 Regeneration 76
2.5.2 Wavelength Interchange 76
2.6 Logical Network Overlays 77
2.6.1 SONET Networks 79
2.6.2 ATM Networks 81
viii Contents

2.6.3 IP Networks 83
2.6.4 MPLS and Its Extensions 84
2.7 Summary 85
2.8 Problems 87

3 Network Connections 91
3.1 Connection Management and Control 96
3.1.1 Optical Connections 100
3.1.2 Logical Connections 100
3.2 Static Networks 102
3.2.1 Point-to-Point and Multipoint Connections 104
3.2.2 Packet Switching in the Optical Layer: The MAC Sublayer 111
3.2.3 Additional Comments on Broadcast-and-Select 121
3.3 Wavelength-Routed Networks 122
3.3.1 Routing and Channel Assignment 124
3.3.2 Routing and Channel Assignment Examples 128
3.4 Linear Lightwave Networks: Waveband Routing 133
3.4.1 Routing and Channel Assignment 135
3.4.2 Multipoint Subnets in LLNs 140
3.4.3 A Seven-Station Example 143
3.5 Logically-Routed Networks 151
3.5.1 Point-to-Point Logical Topologies 153
3.5.2 Multipoint Logical Topologies: Hypernets 156
3.6 Summary 162
3.7 Problems 163

4 Enabling Technology 165

4.1 Evolution of Transmission and Switching Technology 166
4.2 Overview of the Optical Connection 167
4.3 Optical Fibers 168
4.3.1 Principles of Guided-Wave Propagation 168
4.3.2 Optical Fiber Technology: Transmission Impairments 174
4.3.3 Solitons 187
4.3.4 Photonic Crystal Fibers 188
4.4 Amplifiers 190
4.4.1 Erbium-Doped Fiber Amplifiers 191
4.4.2 Raman Amplifiers 198
4.4.3 Semiconductor Optical Amplifiers 201
4.4.4 Amplification Trends in Metro Optical Networks: Amplets 204
4.5 Optical Transmitters 205
4.5.1 Lasers 205
4.5.2 Vertical Cavity Surface Emitting Lasers 211
4.5.3 Modulation Technology 212
 Contents ix

4.6 Optical Receivers in Intensity-Modulated Direct-Detection

Systems 217
4.6.1 Photodetectors 217
4.6.2 Front-End Amplifier: Signal-to-Noise Ratio 219
4.6.3 Digital Signal Detection: Noise, Interference, and Bit
Error Rate 221
4.6.4 Analog Systems: Carrier-to-Noise Ratio 227
4.7 The End-to-End Transmission Channel 228
4.7.1 Modulation Formats 229
4.7.2 Forward Error Correction 231
4.7.3 Equalization 233
4.8 Coherent Optical Systems 234
4.9 Performance Impairments in a Network Environment 235
4.9.1 Cross-Talk 235
4.9.2 Signal Power Divergence 239
4.9.3 Chirp-Induced Penalty 240
4.9.4 Optical Filter Concatenation: Distortion-Induced Penalty 240
4.9.5 Polarization Mode Dispersion Impact on
System Performance 241
4.10 Optical and Photonic Device Technology 241
4.10.1 Couplers and Switches 242
4.10.2 Reciprocity 255
4.10.3 Nonreciprocal Devices 257
4.10.4 Optical Filtering Technology 257
4.10.5 Multiwavelength Switch Technology 266
4.11 Wavelength Conversion and Signal Regeneration 274
4.11.1 All-Optical Wavelength Conversion 275
4.11.2 Opaque Wavelength Conversion and Signal Regeneration 278
4.12 Optical Switch Architectures 281
4.12.1 Space Switches 281
4.12.2 Wavelength-Selective Switches 288
4.13 Performance Evaluation: Methodology and Case Studies 297
4.13.1 Physical-Layer Simulation: Three-Step Approach 298
4.13.2 WDM Network Simulation Case Studies 301
4.14 Problems 311

5 Static Multipoint Networks 324

5.1 Shared Media: The Broadcast Star 324
5.2 Representative Multiplexing and Multiple-Access Schemes 327
5.2.1 Time-Wavelength-Division Multiplexing/Multiple
Access 328
5.2.2 Subcarriers 336
5.2.3 Code Division Multiple Access 352
x Contents

5.3 Traffic Constraints in Shared-Channel Networks 367

5.3.1 Balanced Traffic 370
5.3.2 Unbalanced Traffic 370
5.4 Capacity Allocation for Dedicated Connections 371
5.4.1 Fixed-Frame Scheduling for Stream Traffic 371
5.4.2 Fixed-Frame Scheduling for Packet Traffic 383
5.5 Demand-Assigned Connections 389
5.5.1 Blocking Calculations in WDMA Networks 390
5.5.2 Blocking in Combined Time-Wavelength-Division
Networks 395
5.6 Packet Switching in the Optical Layer 399
5.6.1 Uncontrolled Scheduling: Random Access 401
5.6.2 Scheduling with Loss 403
5.6.3 Lossless Scheduling: Reservations 405
5.6.4 Perfect Scheduling 407
5.6.5 Dynamic versus Fixed Capacity Allocation 408
5.7 The Passive Optical Network 409
5.7.1 ATM and Fixed-Frame PONs 412
5.7.2 Ethernet-Based PONs 414
5.7.3 WDM PONs 416
5.7.4 Optical-Wireless Access 420
5.7.5 Recent Trends 422
5.8 Summary 424
5.9 Problems 425

6 Wavelength/Waveband-Routed Networks 432

6.1 Introduction 432
6.2 Physical Topologies 434
6.3 Wavelength-Routed Networks: Static Routing
and Channel Assignment 442
6.3.1 Flow Bounds: Matching the Physical
and Logical Topologies 444
6.3.2 Nonblocking Stations 448
6.3.3 RCA as a Graph Coloring Problem 449
6.3.4 Rings 452
6.3.5 Ring Decomposition of General Mesh Networks 458
6.3.6 Multistar Wavelength-Routed Networks 462
6.3.7 RCA as an Optimization Problem 464
6.3.8 Heuristics for Static RCA 474
6.4 Wavelength-Routed Networks: Dynamic Routing
and Channel Assignment 484
6.4.1 Some Basic Routing and Channel Assignment Algorithms 484
 Contents xi

6.4.2 Case Study: Bidirectional Rings 491

6.4.3 Performance of Dynamic Routing Rules on Meshes 494
6.4.4 Case Study: An Interconnected Ring 495
6.4.5 Routing Multicast Connections in WRNs 497
6.5 Linear Lightwave Networks: Static Routing Rules 507
6.5.1 Routing of Optical Paths 509
6.5.2 Optical Connections: λ-Channel Assignment 516
6.5.3 Significance of Nonblocking Access Stations in LLNs 518
6.5.4 Local Access to LLNs 519
6.5.5 Routing Waveband and Channel Assignment on the
Petersen Network 521
6.5.6 Channel Assignment 528
6.5.7 Multistar Linear Lightwave Networks 540
6.6 Linear Lightwave Networks: Dynamic Routing Rules 544
6.6.1 Point-to-Point Connections 544
6.6.2 Routing Multicast Connections in LLNs 558
6.7 Problems 568

7 Logically-Routed Networks 576

7.1 Introduction: Why Logically-Routed Networks? 576
7.1.1 Multitier Networks: Grooming 581
7.2 Point-to-Point Logical Topologies: Multihop Networks 585
7.2.1 ShuffleNets 587
7.2.2 Families of Dense Logical Topologies 589
7.3 Multihop Network Design 591
7.3.1 Logical-Layer Design 591
7.3.2 Physical-Layer Design 594
7.3.3 Traffic Grooming in Point-to-Point
Logical Topologies 597
7.4 Multipoint Logical Topologies: Hypernets 607
7.4.1 Capacity of a Multipoint Subnet 611
7.4.2 Families of Dense Hypernets 613
7.4.3 Kautz Hypernets 615
7.4.4 Hypernet versus Multihop 628
7.4.5 Multicast Virtual Connections 631
7.5 Hypernet Design 632
7.5.1 Logical-Layer Design 632
7.5.2 Physical-Layer Design 634
7.5.3 Traffic Grooming in Multipoint Logical
Topologies 637
7.5.4 Multistar Realizations 639
7.6 Summary 641
7.7 Problems 642
xii Contents

8 Survivability: Protection and Restoration 647

8.1 Objectives of Protection and Restoration 648
8.2 Current Fault Protection and Restoration Techniques in
the Logical Layer 650
8.2.1 Point-to-Point Systems 650
8.2.2 SONET Self-Healing Rings 654
8.2.3 SONET Self-Healing Ring Interconnection Techniques 657
8.2.4 Architectures with Arbitrary Mesh Topologies 663
8.3 Optical-Layer Protection: Point-to-Point and Ring Architectures 669
8.3.1 Point-to-Point Systems 669
8.3.2 Self-Healing Optical Rings 672
8.4 Optical-Layer Protection: Mesh Architectures 677
8.4.1 Shared Optical Layer Line-Based Protection 679
8.4.2 Optical Path-Based Protection 692
8.4.3 Segment Protection 700
8.4.4 Survivability Techniques for Multicast Connections 702
8.5 Summary 703
8.6 Problems 706

9 Optical Control Plane 714

9.1 Introduction to the Optical Control Plane 716
9.1.1 Control-Plane Architecture 719
9.1.2 Control-Plane Interfaces 719
9.1.3 Control-Plane Functions 721
9.2 Overview of Multiprotocol Label Switching 722
9.2.1 Packet Transport through an MPLS Network 722
9.2.2 MPLS Protocol Stack 727
9.2.3 MPLS Applications 728
9.3 Overview of Generalized Multiprotocol Label Switching 729
9.3.1 Link Management in GMPLS 731
9.3.2 Routing in GMPLS 734
9.3.3 Signaling in GMPLS 742
9.4 Conclusions 751

10 Optical Packet-Switched Networks 756

10.1 Optical Packet-Switched Network Architectures 758
10.1.1 Unbuffered Networks 759
10.1.2 Deflection Routing 764
10.1.3 Performance Analysis of Deflection Routing 766
10.1.4 Buffering: Time Domain Contention Resolution 770
10.1.5 Buffering and Wavelength Conversion: Time/Wavelength
Domain Contention Resolution 778
 Contents xiii

10.1.6 Comparison of Contention Resolution Techniques for

Asynchronous OPS Networks 782
10.1.7 Hybrid Electronic and Optical Buffering 784
10.2 OPS Enabling Technologies 787
10.2.1 Packet Synchronization 788
10.2.2 All-Optical 2R or 3R Regeneration 788
10.2.3 Optical Switching 788
10.2.4 Wavelength Conversion 789
10.2.5 Optical Header Processing 789
10.2.6 Optical Buffering 789
10.3 OPS Network Testbed Implementations 791
10.3.1 CORD Testbed 791
10.3.2 KEOPS Testbed 793
10.3.3 WASPNET Testbed 796
10.4 Optical Burst Switching 798
10.4.1 Just Enough Time Protocol 801
10.4.2 Just In Time Protocol 803
10.4.3 Contention Resolution in OBS Networks 806
10.5 Optical Label Switching 808
10.5.1 All-Optical Label Swapping 809
10.5.2 Contention Resolution Techniques 811
10.5.3 OLS Network Implementations 811
10.6 Conclusions 820
10.7 Problems 822

11 Current Trends in Multiwavelength Optical Networking 828

11.1 Business Drivers and Economics 828
11.1.1 Cost Issues for WDM Point-to-Point Systems 831
11.1.2 Cost Issues for WDM Rings 832
11.1.3 Cost Issues for WDM Cross-Connect Networks 833
11.1.4 Open versus Closed WDM Installations 835
11.2 Multiwavelength Optical Network Testbeds 838
11.2.1 Optical Networks Technology Consortium 838
11.2.2 All-Optical Network Consortium 839
11.2.3 European Multiwavelength Optical Network Trials 839
11.2.4 Multiwavelength Optical Network 840
11.2.5 National Transparent Optical Networks Consortium 840
11.2.6 The Importance of the Testbeds in Driving the
Telecommunications Infrastructure 840
11.3 Metropolitan Area Networks 841
11.3.1 Metro Network Unique Characteristics 841
11.3.2 Defining the Metropolitan Networking Domain 842
xiv Contents

11.3.3 Metro Network Evolution 844

11.3.4 Metro Networking State of the Art 847
11.4 Long-Haul and Ultra Long-Haul Networks 854
11.4.1 Current Considerations in Wide Area
Network Architectures 854
11.4.2 Some Recent Commercial Network Deployments 856
11.5 New Applications and Services 858
11.5.1 Wavelength-on-Demand 858
11.5.2 Virtual Private Optical Networks 858
11.5.3 Bandwidth Trading 859
11.6 Conclusions 861

A Graph Theory 869

A.1 Graphs 869
A.1.1 Cycle Double Covers 872
A.1.2 Eulerian Graphs 872
A.1.3 Planar Graphs 873
A.1.4 Matchings in Graphs 873
A.1.5 Graph Coloring 874
A.1.6 Digraphs 875
A.1.7 Moore Bounds 875
A.1.8 Max Flow–Min Cut 876
A.2 Hypergraphs 877
A.2.1 Undirected Hypergraphs 877
A.2.2 Directed Hypergraphs 878

B Fixed Scheduling Algorithm 879

B.1 Column/Row–Expansion Algorithm 880
B.2 Decomposition into Permutation Matrices 883
B.3 Column/Row–Compression Algorithm 883

C Markov Chains and Queues 884

C.1 Random Processes 884
C.2 Markov Processes 885
C.3 Queues 887
C.3.1 The M |M |1 Queue 888
C.3.2 The M |G |1 Queue 888
C.3.3 Little’s Formula 889
 Contents xv

D A Limiting-Cut Heuristic 890

D.1 The Multicommodity Flow Problem and Limiting Cuts 890
D.2 A Heuristic 891
D.2.1 Swap (X, Y ) 891
D.2.2 Limcut 892

E An Algorithm for Minimum-Interference Routing

in Linear Lightwave Networks 893
E.1 The Image Network 893
E.2 The Min-Int Algorithm 894
E.3 Minimum Interference 895

F Synopsis of the SONET Standard 896

G A Looping Algorithm 900

Acronyms 903

Index 915
Figures

1.1 Multilayered network. page 6

1.2 Physical picture of the network. 8
1.3 Layered view of an optical network. 13
1.4 Alternative network approaches. 15
1.5 Hierarchical network. 17
1.6 Road map. 24
2.1 Layered view of optical network connections. 29
2.2 A typical connection. 32
2.3 Client server associations in an optical transport network. 33
2.4 Fiber resources. 34
2.5 Wavelength and waveband partitioning of the optical spectrum. 35
2.6 Network picture based on spectrum partitioning. 38
2.7 Tree physical topologies. 41
2.8 Directional coupler. 42
2.9 A 16 × 16 star coupler. 43
2.10 Static routing node. 44
2.11 Directed star. 45
2.12 Space switch connection matrices. 46
2.13 Unidirectional ring. 47
2.14 Crossbar switch. 48
2.15 Clos switch. 49
2.16 Recursion for Benes switch. 49
2.17 An 8 × 8 Benes switch. 50
2.18 Generalized optical switch. 53
2.19 δ–σ linear divider-combiner. 54
2.20 A node without loopback connections. 55
2.21 Three-stage realization of a waveband-space switch. 56
2.22 Multiwaveband directional coupler. 57
2.23 WADM–NAS combination. 58
2.24 Wavelength conversion as a linear operation. 64
2.25 Wavelength interchanger. 65
2.26 Wavelength-interchanging switch. 65
2.27 WIXC implementation. 66
2.28 Wavelength-routed network. 67
xviii Figures

2.29 Network access station. 68

2.30 Example of a logical connection between two NASs. 69
2.31 Optical transmitter. 70
2.32 Optical receivers. 72
2.33 Heterodyne receiver and spectra. 73
2.34 Overlay processor. 75
2.35 Logically routed network. 78
2.36 SONET DCS. 81
2.37 ATM cell format. 82
2.38 ATM switch connections. 83
2.39 TCP/IP and OSI. 84
2.40 Taxonomy of multiwavelength networks. 86
3.1 End systems: full connectivity. 92
3.2 Star physical topology. 92
3.3 Bidirectional ring physical topology. 93
3.4 The control plane in an optical network. 98
3.5 Connection management system. 99
3.6 Star coupler example. 105
3.7 Time-shared medium. 106
3.8 A TDM/TDMA schedule. 108
3.9 TDM/T-WDMA. 110
3.10 CSMA/CD. 114
3.11 CSMA/CD collision. 115
3.12 NAS equipped for packet switching. 116
3.13 Packet switching in the optical layer. 118
3.14 MAC protocol in the layered architecture. 120
3.15 Wavelength-routed star network. 123
3.16 Channel assignment example. 125
3.17 Nonblocking access link. 129
3.18 Bidirectional ring: single access fiber pair. 130
3.19 Bidirectional ring: two access fiber pairs. 132
3.20 A mesh network. 133
3.21 Inseparability. 136
3.22 Two violations of DSC. 137
3.23 Inadvertent violation of DSC. 137
3.24 Avoidance of DSC violations. 138
3.25 Color clash. 139
3.26 Seven stations on a mesh. 141
3.27 Tree embedded in mesh. 142
3.28 Multistar network. 144
3.29 Embedded star on a bidirectional ring. 146
3.30 Seven-node hypernet. 148
3.31 Assumed channel spacings. 150
3.32 A logical switching node in an optical network. 152
 Figures xix

3.33 Eight-node ShuffleNet. 154

3.34 ShuffleNet embedding. 155
3.35 Details of ShuffleNet node. 157
3.36 Twenty-two node hypernet. 159
3.37 Hypernet embedding. 160
4.1 A point-to-point optical connection. 167
4.2 Refractive index profiles for fibers. 169
4.3 Snell’s law. 169
4.4 Ray propagation in a step-index fiber. 170
4.5 Ray propagation in a graded-index fiber. 171
4.6 Cylindrical coordinates. 173
4.7 Commercial fiber cables. 175
4.8 Attenuation as a function of wavelength. 176
4.9 Broadening of pulses due to dispersion. 178
4.10 Dispersion coefficients as a function of frequency. 179
4.11 Limitations due to nonlinear effects in multiwavelength systems. 186
4.12 Soliton. 188
4.13 Three types of microstructured fibers. 189
4.14 Basic erbium-doped fiber amplifier structures. 192
4.15 Energy levels in EDFA. 192
4.16 EDFA gain profile. 194
4.17 Illustration of noise figure. 196
4.18 Raman gain coefficient in bulk silica as a function of frequency shift. 198
4.19 Hybrid distributed-discrete amplification. 200
4.20 Signal and pump power in hybrid system. 200
4.21 Fabry–Perot laser. 206
4.22 Single-frequency lasers. 208
4.23 Laser array. 210
4.24 Typical VCSEL structure. 211
4.25 Pulse and accompanying chirp. 213
4.26 Mach–Zehnder interferometer. 215
4.27 Typical structure of an EA-DFB transmitter. 216
4.28 Absorption and chirp (linewidth enhancement factor) parameters
versus reverse bias voltage for a typical EA-DFB transmitter. 216
4.29 Photodiode. 217
4.30 Transimpedance amplifier. 219
4.31 Binary receiver. 221
4.32 Typical waveforms in an IM/DD system. 222
4.33 Eye diagram. 223
4.34 Ideal detection. 225
4.35 BER as a function of Q. 226
4.36 Transmission channel processing operations. 228
4.37 Modulation formats. 229
4.38 FEC encoding/decoding functions. 232
xx Figures

4.39 Transversal decision-directed equalizer. 233

4.40 Heterodyne receiver. 234
4.41 Types of cross-talk. 237
4.42 Power penalty with homodyne cross-talk. 238
4.43 Controllable directional coupler. 245
4.44 Mach–Zehnder switch. 246
4.45 Two-stage Mach–Zehnder switch. 247
4.46 Y-branch switch. 248
4.47 Gate array switch. 249
4.48 Laser-activated bubble switch element. 250
4.49 2D mechanical switch using micromachined mirrors. 251
4.50 3D MEMS switch. 252
4.51 3D gimbaled mirror. 253
4.52 Liquid crystal holographic switch. 254
4.53 Two hologram N × N liquid crystal holographic switch. 255
4.54 Illustration of reciprocity. 256
4.55 Optical isolator. 258
4.56 Fabry–Perot filter and its spectral response. 260
4.57 MI filter. 262
4.58 MI filter array. 263
4.59 FBG used as a drop filter. 264
4.60 A Mach–Zehnder WADM. 264
4.61 Arrayed waveguide grating. 266
4.62 Acousto-optic tunable filter. 267
4.63 Liquid crystal MWS. 269
4.64 A MEMS-based WADM. 270
4.65 An MI filter-based WADM. 272
4.66 Wavelength-dilated switch. 273
4.67 Optoelectronic wavelength converter. 275
4.68 Performance of a difference frequency converter. 277
4.69 Opaque conversion and regeneration. 278
4.70 SA-based regenerator. 279
4.71 Nonlinear Mach–Zehnder regenerator. 279
4.72 Nonlinear optical loop mirror regenerator. 280
4.73 Optical crossbar switch. 282
4.74 Path-independent loss crossbar switch. 282
4.75 Circuit layout for 8 × 8 optical crossbar switch. 284
4.76 Router/selector. 285
4.77 Benes switch. 285
4.78 Orders of cross-talk. 286
4.79 Enhanced performance switch. 287
4.80 Space dilation. 288
4.81 OADMs in a network. 289
 Figures xxi

4.82 Parallel and serial OADM architectures with capability for m

wavelength add/drops. 290
4.83 Functional diagram of an OADM based on wavebands and
wavelengths. 290
4.84 Typical B&S OADM architecture. Assumes (1+1) protection. 292
4.85 Typical 4 × 4 ROADM based on 4 × 1 wavelength selective switch
and B&S-type of architecture. 292
4.86 Transparent OXC. 293
4.87 Opaque O-E-O OXC. 294
4.88 Opaque O-O-O OXC. 294
4.89 Hybrid waveband/wavelength switch. 297
4.90 Wavelength-domain simulation. 299
4.91 WADM chain. 302
4.92 WADM structure and simulation model. 302
4.93 Simulation results for the WADM chain. 303
4.94 Ring interconnect network architecture. Worst-case paths between
A and B are indicated. 304
4.95 Histogram of all cross-talk terms accumulated at receiver B for the
worst-case path of Figure 4.94. 306
4.96 Cross-talk-induced Q penalty in dB versus dominant cross-talk term
power level. 307
4.97 Q-channel performance for the worst-case path of Figure 4.94
assuming OC-192 bit rate and EA-modulated transmitters. 308
4.98 A DWDM metro network deployment scenario. All rings represent
typical SONET OC-12/48/192 designs. DWDM is deployed only
between the superhub nodes (dark squares) in ring (solid) or possible
mesh (dotted) configurations. 309
4.99 DWDM metro network case study based on the network deployment
scenario presented in Figure 4.98. Nodes represent only superhub
stations with typical distances (not shown to scale). 309
4.100 Simulation results for path A-F-D in Figure 4.99 comparing
Q-channel performance with and without EDC. 311
5.1 Star networks. 325
5.2 A 3 × 3 example. 330
5.3 TDM/T-WDMA channel allocation schedules. 331
5.4 Illustrating channel reuse in an FT-TR system. 332
5.5 Illustrating optical spectral efficiency. 334
5.6 SCMA example. 337
5.7 Transmitting and receiving stations equipped for SCMA. 337
5.8 Subcarrier spectra. 340
5.9 Effect of OBI. 342
5.10 TDM/T-SCMA. 346
5.11 SCM/SCMA. 349
xxii Figures

5.12 SCM/WDMA/SCMA. 350

5.13 SCM/WDMA/SCMA example. 351
5.14 Block diagram of a direct-detection CDMA system. 354
5.15 Waveforms for a direct-detection CDMA system. 356
5.16 Orthogonal optical codes. 357
5.17 Parallel CDMA transceiver structure. 359
5.18 CDMA with all-optical processing. 359
5.19 Multidimensional codes. 361
5.20 FBG encoder for FFH-CDMA. 361
5.21 A 3D CDMA system. 363
5.22 Realization of coherent optical CDMA. 364
5.23 Shared-channel broadcast medium. 367
5.24 Normalized traffic matrices. 370
5.25 CASs for systems with a full complement of channels. 376
5.26 CASs for Examples 4, 5, 6, and 7. 378
5.27 Heterogeneous traffic scheduling. 379
5.28 Logical multicast CAS. 382
5.29 Single-server queue. 385
5.30 Throughput versus traffic intensity. 387
5.31 Markov chain model for demand-assigned traffic. 391
5.32 Comparison of Engset and Erlang models. 393
5.33 Normalized throughput versus traffic intensity. 394
5.34 Normalized throughput versus traffic intensity. 395
5.35 Matching time slots. 396
5.36 Framed system blocking probabilities. 397
5.37 Illustrating rearrangeability. 398
5.38 Slotted ALOHA. 403
5.39 Tell-and-go protocol. 404
5.40 Lossless scheduling. 406
5.41 Queues for perfect scheduling. 407
5.42 Passive optical network. 410
5.43 BPON frame. 412
5.44 Transmission scenario in a BPON system. 413
5.45 PON equipped for decentralized control. 416
5.46 LARNet. 417
5.47 RITE-Net. 418
5.48 WDM PON. 419
5.49 Integrated system for dual services. 421
5.50 Dual services testbed. 421
5.51 DWDM/TDM PON. 423
6.1 Number of vertices in known maximal graphs. 435
6.2 Thirty-eight-vertex graph. 435
6.3 Tessellations of the plane. 436
6.4 Undirected deBruijn and Kautz graphs. 437
 Figures xxiii

d
6.5 Construction for min . 438
d
6.6 Plot of min as a function of N . 438
6.7 Internodal distances in random networks. 439
6.8 Recursive grid. 440
6.9 Hierarchical Petersen graph. 441
6.10 Limiting cuts for four networks. 447
6.11 Three-node network. 449
6.12 Illustrating RCA in a wavelength-routed network. 451
6.13 A four-fiber SPRING. 453
6.14 A two-fiber SPRING. 454
6.15 Bidirectional ring. 455
6.16 Five-node WDM ring. 458
6.17 Ring decomposition. 459
6.18 Bridged ring overlay. 462
6.19 A multistar network. 462
6.20 Layered view of RCA. 465
6.21 External traffic in flow conservation equations. 467
6.22 Wavelength savings by increasing fibers. 476
6.23 Mean values of N λ versus α. 477
6.24 Minimum values of N λ versus α. 478
6.25 Flow chart of the Monte Carlo algorithm. 481
6.26 Time trace of Monte Carlo algorithm. 483
6.27 An example of SPD routing. 488
6.28 Blocking on an 11-node WDM ring. 490
6.29 Gain in blocking; 11-node WDM ring, simulation. 491
6.30 Fairness ratio; 11-node WDM ring, simulation. 492
6.31 Fairness ratio improvement versus interchanger density; 11-node
WDM ring with 32 wavelengths. 493
6.32 Simulation and asymptotic analysis; 195-node interconnected WDM
rings. 496
6.33 Blocking improvement with wavelength interchange; 195-node
interconnected WDM rings. 497
6.34 Fairness ratio improvement with wavelength interchange; 195-node
interconnected WDM rings. 498
6.35 Fairness ratio improvement versus interchanger density; 195-node
interconnected WDM ring, 32 wavelengths. 499
6.36 Multicast connection in a transparent network. 500
6.37 A P×P split-and-deliver switch. 501
6.38 A P×P multicast-capable optical cross-connect based on a
split-and-deliver switch. 502
6.39 A P×P multicast-capable optical cross-connect based on splitter
sharing. 503
6.40 Multicasting in a network with sparse splitting capabilities. 506
xxiv Figures

6.41 Petersen network. 509

6.42 Structure of a nonblocking access station for an LLN. 510
6.43 Optical paths. 512
6.44 Optical connection hypergraph. 518
6.45 Local access subnets on the Petersen network. 520
6.46 Embedded star on tree T A . 523
6.47 Waveband assignments: W = 5. 526
6.48 Connection interference graph. 529
6.49 Connection interference graph for Equation (6.60). 530
6.50 Optical connection hypergraph. 534
6.51 Fixed-frame scheduling for four LCs. 534
6.52 Directed hypernet GKH (2, 8, 4, 4). 542
6.53 Color clash. 547
6.54 Illustrating inseparability. 548
6.55 Illustrating Min-Int. 550
6.56 Random network. 551
6.57 Max Reuse versus Min Reuse channel allocation. 552
6.58 k-SP routing. 553
6.59 k-SP versus Min-Int routing. 554
6.60 Blocking in networks with multifiber links. 555
6.61 Blocking in networks with multiple wavebands. 556
6.62 Example of a multicast connection. 559
6.63 Example of a tree decomposition using MBFS-1. 562
6.64 Example of a tree decomposition using MBFS-4. 563
6.65 Illustrating routing on a tree. 564
6.66 Blocking probability for multicast connections. 568
7.1 Why logically-routed networks? 577
7.2 A schematic of a point-to-point LRN. 579
7.3 Two-tier architecture. 582
7.4 The architecture of a grooming node with optical bypass. 584
7.5 ShuffleNet: δ = 3, k = 2, N = 18. 587
7.6 Maximum throughput per node for ShuffleNet. 588
7.7 deBruijn and Kautz digraphs. 590
7.8 A traffic matrix and matched LCG. 593
7.9 ShuffleNet on Atlantis. 595
7.10 Benefit of traffic grooming. 598
7.11 A node in a SONET over WDM ring. 599
7.12 Advantage of grooming static traffic in SONET over WDM rings. 601
7.13 Construction of an auxiliary graph for grooming. 605
7.14 Layered view of a hypernet. 608
7.15 Hypernets. 609
7.16 Illustrating fan-out in hypergraphs. 613
7.17 Shuffle hypernet. 614
7.18 Orders of K H (2, D, r ). 616
 Figures xxv

7.19 Duality construction. 618

7.20 Directed hypergraph construction via duality. 619
7.21 Directed hypergraph construction via edge grouping. 620
7.22 Tripartite representation of G K H (2, 42, 3, 28). 622
7.23 Routing header. 624
7.24 Comparison of hypernets and multihop networks. 628
7.25 Multicast tree in D K H (1, 4, 2). 631
7.26 Multicast-capable logical-grooming switch. 638
8.1 Path versus line protection. 651
8.2 (1 + 1) SONET protection. 652
8.3 (1:1) SONET protection. 653
8.4 (1:N) SONET protection. 654
8.5 Single- and dual-access ring interconnection configurations. 658
8.6 Two-fiber UPSR-to-UPSR ring interconnection. 660
8.7 BLSR-to-BLSR ring interconnection. 661
8.8 BLSR-to-UPSR ring interconnection. 662
8.9 (1 + 1) Protection in the optical layer. 670
8.10 (1:1) Protection in the optical layer. 671
8.11 (1:N) Protection in the optical layer. 671
8.12 (1 + 1) Optical protection and (1:N) electronic protection for
a WDM system. 672
8.13 Four-fiber WDM SPRING architecture. 673
8.14 Four-fiber WDM SPRING surviving a link failure. 674
8.15 Four-fiber WDM SPRING surviving a node failure. 675
8.16 Two-fiber WDM SPRING architecture. 676
8.17 A taxonomy of survivability schemes. 677
8.18 Rerouting around a failed link. 682
8.19 Directed cycles in a planar graph. 683
8.20 Directed cycles in a nonplanar graph: K 5 . 683
8.21 Face traversal for a planar national network. 684
8.22 Orientable CDC of the ARPANet. 686
8.23 Seven-node planar network with default protection switch settings. 687
8.24 Seven-node planar network after a link failure. 687
8.25 Failure recovery using the p-cycle approach. 688
8.26 Generalized loopback example. 691
8.27 (1+1) dedicated protection architecture. 692
8.28 (1:3) shared protection in a mesh network. 693
8.29 Spanning trees used in optical path protection. 697
8.30 Shared risk groups. 698
8.31 SRG classification. 699
8.32 SLSP protection scheme. 701
8.33 Examples of different types of islands centered on node 22. 702
8.34 Examples of segment and path-pair protection of multicast sessions. 703
8.35 Example illustrating the arc-disjoint and MC-CR algorithms. 704
xxvi Figures

9.1 A mesh optical network. 715

9.2 Example of an optical node architecture. 716
9.3 Provisioning a connection between two routers through
an optical network. 717
9.4 Control plane architecture. 720
9.5 Control plane interfaces. 721
9.6 MPLS header format and MPLS packet format. 724
9.7 Two LSPs in an MPLS packet-switched network. 725
9.8 Label stacking. 727
9.9 MPLS protocol stack. 728
9.10 Link bundling illustration. 740
9.11 LSP hierarchy in GMPLS. 741
9.12 User-Network Interface. 743
9.13 Provisioning in GMPLS. 744
9.14 Path and Resv message flows in RSVP for
resource reservation. 746
9.15 RSVP message format. 747
9.16 Protection signaling using GMPLS RSVP-TE. 751
10.1 Optical packet-switching node. 760
10.2 A generic OPS node architecture for an unslotted network. 761
10.3 A generic OPS node architecture for a slotted network. 761
10.4 An FDL-based synchronizer. 762
10.5 A generic packet format for a slotted network. 762
10.6 Optical packet contention. 764
10.7 Petersen network graph. 766
10.8 Paths from A to D. 766
10.9 Input buffered optical packet switch with WSXC. 771
10.10 Input buffered optical packet switch using multiple space
switch planes. 772
10.11 FDL input buffer. 772
10.12 Example of head-of-the-line (HOL) blocking. 773
10.13 Feed-forward delay line architecture. 773
10.14 Feedback delay line architecture. 774
10.15 Dump-and-insert buffer architecture. 775
10.16 Typical packet sequence in DI buffers for a 4 × 4 optical switch. 776
10.17 Generic node architecture with TOWCs at the input lines. 779
10.18 Details of output buffers for TOWC switch. 780
10.19 Packet-loss probability versus number of FDLs with and without
wavelength conversion. 780
10.20 Generic node architecture with TOWCs that are shared among
input lines. 781
10.21 Node architectures for different contention resolution schemes:
single-wavelength delay line, multiwavelength delay line, wavelength
 Figures xxvii

conversion, and wavelength conversion with multiwavelength

buffering. 783
10.22 Switch architecture with electronic buffering
and wavelength conversion. 784
10.23 All-optical buffering and switching architecture. 790
10.24 Physical implementation of the CRO device. 792
10.25 CORD testbed. 792
10.26 Packet format for the KEOPS project. 794
10.27 Proposed unicast node architecture for the KEOPS project. 794
10.28 Proposed multicast/broadcast node architecture
for the KEOPS project. 795
10.29 SLOB architecture. Each stage is a photonic switch element (PSE). 796
10.30 WASPNET optical packet switch. 797
10.31 Optical packet switching and Optical burst switching. 799
10.32 OBS architecture concept. 800
10.33 Just enough time (JET) protocol. 802
10.34 Just in time (JIT) protocol. 804
10.35 Segmentation of a burst. 807
10.36 Packet-loss probability versus load for different contention resolution
policies in OBS. 808
10.37 OLS network. 809
10.38 All-optical processor for OLS. 810
10.39 OLS subcarrier transmission system. 812
10.40 OLS network node. 813
10.41 Network node architecture for an OLS testbed demonstration. 814
10.42 FSK/IM orthogonal labeling scheme used in the STOLAS project. 814
10.43 Optical router and optical label swapper used in the STOLAS project. 815
10.44 Architecture of the edge-router in OPSnet. 816
10.45 Architecture of the core-router in OPSnet. 817
10.46 Core node configuration for label swapping and packet switching. 819
10.47 OCSS technique. 819
10.48 Experimental setup for multihop packet transmission
and multirate payload. 820
10.49 K 3,3 network. 822
11.1 Six Central Offices, including two hubs, with capacity exhaust. 831
11.2 Application of WDM point-to-point systems to alleviate
capacity exhaust. 832
11.3 Six Central Offices, including two hubs, with capacity exhaust. 833
11.4 Economic case for WDM rings. 834
11.5 Node in Central Office: Electronic cross-connect. 835
11.6 Economic case for WDM optical cross-connect. 836
11.7 Open WDM network architecture: Opaque network. 837
11.8 Integrated closed WDM network architecture. 837
11.9 Current legacy SONET/SDH design in U.S. metropolitan regions. 842
xxviii Figures

11.10 Metro WDM interconnected-ring simulation case study. 844

11.11 Typical ILEC metro network in the 2004 time frame. 845
11.12 Typical view of a superhub in the ILEC metro network of
Figure 11.11 (2004 time frame). 846
11.13 Typical view of a current superhub in an ILEC metro network. 847
11.14 Typical deployment of 10 GbE technology in the metro environment. 848
11.15 Typical metro network ring architectures. 850
11.16 Typical vendor generic WADM node architecture for the metro
network application space of Figure 11.15. 851
11.17 Metro network case study. 853
11.18 Wavelength-brokering operational model. 860
11.19 Infrastructure swapping. 860
11.20 Multicarrier recovery. 861
A.1 A maximal independent set. 870
A.2 The complete graph K 5 . 870
A.3 The complete bipartite graph K 3,3 . 871
A.4 Orientable cycle double cover for K 3,3 . 872
A.5 Multigraphs: Non-Eulerian and Eulerian. 873
A.6 Maximum matching of a bipartite multigraph. 874
A.7 A diclique. 876
A.8 A cut. 877
B.1 Example of decomposition of an NQDS matrix. 881
B.2 Example of fixed-frame scheduling. 882
C.1 Two-state chain. 886
C.2 Birth–death process. 887
C.3 A queue. 888
E.1 Image network. 894
F.1 SONET STS-1 frame and overhead channels. 897
F.2 Creating an OC-N signal. 898
F.3 Structure of a concatenated SONET frame. 898
G.1 Looping, first step. 901
G.2 Final switch settings. 901
Tables

3.1 Tree routing on mesh 149

3.2 Seven-station comparisons 150
3.3 Wavelength assignments for ShuffleNet on a ring 156
5.1 SCM/WDMA/SCMA example 351
5.2 Multicast connections 382
6.1 Orders of some graphs 437
6.2 Comparative performance of three RCA heuristics 480
6.3 Routing table 517
6.4 Waveband routing in the Petersen graph: W = 3 524
6.5 Waveband routing in the Petersen graph: W = 5 527
6.6 Comparison of performance of various configurations of the
Petersen network 538
6.7 Sizes of trees generated with the MBFS-d algorithm 567
7.1 ShuffleNet routing on Atlantis 596
7.2 Relative costs for different ring networks 602
7.3 Kautz hypergraphs 615
7.4 Performance of K H (2, D ∗ , r ) 627
7.5 Performance comparison of D K H (d, D, s) to multihop networks 630
7.6 Hypernet trees on Atlantis 636
Preface to the Second Edition

The first edition of this book was published when optical networks were just emerging
from the laboratory, mostly in the form of government-sponsored testbeds. Since then
there have been fundamental changes in many aspects of optical networking, driven
by the move from the laboratory to commercial deployment and by the twists and
turns of the world economy. The investment climate in which optical networks have
developed has had two major swings as of this writing. During the technology bubble that
began at the end of the 20th century, investment in research, product development, and
network deployment increased enormously. The activities during this time of euphoria
produced advances in the technology base that would not have been possible without
the extraordinary momentum of that period. At the same time, commercial network
deployment provided a reality check. Some ideas that were pursued in the late 1990s
dropped by the wayside because they did not meet the test of commercial viability, and
new ones came along to take their place. When the bubble burst after less than a decade
of “irrational exuberance,” the pendulum swung the other way. Investors and executives
who a short time earlier thought the sky was the limit now wondered if demand would
ever materialize for all of the fiber capacity in the ground. At this writing a more reasoned
approach has taken hold; that seemingly elusive demand has materialized and, hopefully,
a more rational and sustainable growth period will ensue.
This is the context for the second edition. It is designed to build on the foundations
laid out in the first edition while reflecting the new developments of the past 9 years:
a maturing underlying technology, new tools for network control, and a recognition of
the latest directions of optical network deployment and research. These new directions
include cost-effective metropolitan area network architectures tailored to the strengths of
current optical transmission and switching equipment, passive optical networks to bring
high-speed access to the end user, hybrid optical/electronic architectures supporting the
merging of multiwavelength and Internet technologies, and networks of the future based
on all-optical packet switching.
As in the first edition, the emphasis of this book is on concepts and methodologies that
will stand the test of time. The first three chapters provide a qualitative foundation for
what follows, presenting an overview of optical networking (Chapter 1), the multiwave-
length network architecture and its supporting components (Chapter 2), and a high-level
view of the different network structures considered throughout the book (Chapter 3). A
more detailed picture is provided in the remaining chapters, with a survey of enabling
technology (Chapter 4) and in-depth studies of the three basic network structures: static
xxxii Preface to the Second Edition

multipoint networks (Chapter 5), wavelength/waveband routed networks (Chapter 6),

and optical/electronic (logically routed) networks (Chapter 7). The remaining chapters
complete the networking picture: survivability (Chapter 8), network control (Chapter 9),
optical packet switching (Chapter 10), and current trends (Chapter 11).
The first three chapters are suitable for the reader who wishes to gain an understanding
of multiwavelength networks without delving deeply into the analytical tools for net-
work design and the physical underpinnings of the optical technology. These beginning
chapters, together with Chapter 11, would be suitable for a short undergraduate course
for electrical engineers and computer science majors.
The first seven chapters provide a largely generic framework for understanding net-
work architectures, performance, and design in an abstract setting. An exception is
Chapter 4, which surveys enabling technology from theory to practice, thereby provid-
ing the necessary background concerning the physical limitations and possibilities of
the network technology. The material through Chapter 7, together with selected material
from the remaining chapters (depending on the reader’s orientation), can form the basis
of a comprehensive graduate course, introducing the student to the latest developments
in the field and suggesting a host of different research directions.
The networking developments since the publication of the first edition have served to
reorient and expand our treatment in significant ways.
r Recognizing the importance of current activity in the “last mile” (fiber to the
home/premises), and in metropolitan area networks, we have added a new section
on passive optical networks (PONs) in Chapter 5,1 and we have included new material
in Chapters 4 and 11 to connect our generic networking approach to recent metro
network developments.
r Chapter 4 was substantially expanded and updated to provide a glimpse of the im-
pressive new trends in photonic and electro-optic technology. Some of the new and/or
expanded topics are photonic crystal fibers, Raman amplification, supercontinuum
generation, amplification trends in metro networks, and forward error correction and
equalization to improve transmission performance. There are also new and expanded
sections on wavelength conversion and signal regeneration with emphasis on all-
optical techniques, and a new section on microelectromechanical system (MEMS)
devices. The treatment of optical switch architectures has been significantly enlarged
with a focus on cost-effective architectures and opacity versus transparency. More
emphasis is placed on the effects of signal impairments, including a new section on
performance impairments in a network environment. Also, new case studies are in-
cluded that illustrate methodologies for evaluating the performance of metropolitan
area networks.
r Chapter 8, on survivability, protection, and restoration, was extensively updated, con-
sistent with the growing importance of optical layer fault management in current

1
It is interesting to note that the PON, epitomized by the broadcast star, was the first structure that demonstrated
the possibilities of optical networking in the 1980s. However, it was largely ignored for large-scale network
deployment until recently, when it has again come into its own as the vehicle of choice for extending optical
networks to the end user.
 Preface to the Second Edition xxxiii

networks. It contains recent work on the subject, including shared line-based protection
in mesh networks, path-based protection, ring-based protection, segment protection,
the treatment of shared risk groups, and recovery of multicast connections.
r Chapter 9, describing the control plane, was added to present the latest developments in
optical network control. It describes the control plane architecture as it has developed
through the recent activities of several standards organizations. The chapter offers a
detailed discussion of Multiprotocol Label-Switching (MPLS) and Generalized MPLS
(GMPLS) as it applies to optical networks.
r Chapter 10, on optical packet-switched networks, was added to provide an introduction
to this emergent field.2 It provides a window on a cutting-edge research area that has
the potential to offer the next breakthroughs in optical networking.
r Chapter 11, on current networking trends (replacing the original Chapter 9), is a
completely updated description of the current networking environment. This includes
a historical perspective describing the pioneering network testbeds, business drivers,
and current trends in metro, long-haul, and ultra long-haul netwoks. Included also
are some new applications that have emerged on the commercial scene, such as
wavelengths on demand, virtual private optical networks, and bandwidth trading.
r This edition places increased emphasis on the practical aspects of hybrid (i.e., elec-
tronic/optical and wavelength/waveband) architectures. This includes the importance
of grooming, which is required to pack electronically multiplexed channels efficiently
into an optical wavelength channel, and to pack wavelength channels into wavebands.
Also, the existence of transparent (purely optical) and opaque (electronic/optical) alter-
natives to network design is stressed throughout. These practical aspects of networking
have become important as optical networks have found their place in the real world.
Exercises are provided for most of the chapters, and many of them suggest avenues for
future study. The book is meant to offer several different alternatives for study depending
on the interest of the reader, be it understanding the current state of the field; acquiring
the analytical tools for network performance evaluation, optimization, and design; or
performing research on next-generation networks.
2
Although the idea of using packet switching in optical networks is not new, it has attracted renewed interest
as the technology for purely optical packet processing has developed over the past few years, and the
advantages of merging Internet and WDM technology have become apparent.
Acknowledgments

The first edition of this book had its origins in 1990, when we organized a small group
within the Center for Telecommunications Research (CTR) at Columbia University, to
investigate lightwave networks. Among the many colleagues, students, and friends who
contributed in various ways to the first edition, there are several who have continued
to interact with us in the preparation of the second edition. We specifically express our
thanks to Eric Bouillet, Aklilu Hailemariam, Gang Liu, and G.-K. Chang. Special thanks
go to Ioannis Roudas for useful discussions and comments. Mischa Schwartz, who was
singled out as our guiding spirit in the first edition, is still an indefatigable contributor
to communication networking and a continuing inspiration to us.
We are especially indebted to Neophytos Antoniades for coauthoring Chapters 4 and
11 of the second edition. His understanding of the role of physical layer simulation and
the evolution of optical networks in the metropolitan area domain provided invaluable
additions to this edition.
We also express our thanks to Phil Meyler at Cambridge University Press for his
support and encouragement, and to Anna Littlewood at Cambridge and Barbara Walthall
at Aptara for their help in putting everything together.
Finally, Thomas Stern expresses his profound gratitude to his wife, Monique, for her
everlasting support; Georgios Ellinas is deeply grateful to his mother, Mary, and sister,
Dorita, for their unyielding support and understanding during this endeavor; and Krishna
Bala is greatly indebted to his wife, Simrat, and children, Tegh and Amrita, for their
patience and support.
Multiwavelength Optical Networks, Second Edition
1 The Big Picture

Since the beginning of the 21st century there has been a burgeoning demand for com-
munications services. From the ubiquitous mobile phone, providing voice, images, mes-
saging, and more, to the Internet and the World Wide Web, offering bandwidth-hungry
applications such as interactive games, music, and video file sharing, the public’s ap-
petite for information continues to grow at an ever-increasing pace. Underneath all of
this, essentially unseen by the users, is the optical fiber-based global communications
infrastructure – the foundation of the information superhighway. That infrastructure
contains the multiwavelength optical networks that are the theme of this book.
Our purpose is to present a general framework for understanding, analyzing, and
designing these networks. It is applicable to current network architectures as they have
evolved since the mid-1990s, but more importantly it is a planning and design tool for
the future. Our approach is to use a generic methodology that will retain its relevance as
networks, applications, and technology continue to evolve.

1.1 Why Optical Networks?

Since the fabrication of the first low-loss optical fiber by Corning Glass in 1970, a
vision of a ubiquitous and universal all-optical communication network has intrigued
researchers, service providers, and the general public. Beginning in the last decades
of the 20th century enormous quantities of optical fiber were deployed throughout the
world. Initially, fiber was used in point-to-point transmission links as a direct substitute
for copper, with the fibers terminating on electronic equipment. Glass fiber was and
is the ideal medium because of its many superior properties: extraordinary bandwidth,
low loss, low cost, light weight and compactness, strength and flexibility, immunity to
noise and electromagnetic interference, security and privacy (it is difficult to tap them),
and corrosion resistance. Although all of these qualities make the fiber a technological
marvel, fibers do not become networks until they are interconnected in a properly
structured architecture. For our purposes, an optical network is a telecommunications
network with transmission links that are optical fibers, and with an architecture that is
designed to exploit the unique features of fibers. (Most of the communication systems
in use today, including many specialized networks such as cable TV and mobile phone
systems, have optical fiber in them somewhere; however, this does not make them
optical networks.) As we shall see, suitable architectures for high-performance lightwave
2 Multiwavelength Optical Networks

networks involve complex combinations of both optical and electronic devices. Thus,
as used here, the term optical or lightwave network does not necessarily imply a purely
optical network, but it does imply something more than a set of fibers interconnecting
electronic switches.
As optical and photonic technology has advanced, applications to point-to-point trans-
mission have preceded advances in networking. For example, it was clear in the early
years of optical fiber transmission that by introducing wavelength division multiplexing
(WDM) on existing fibers the capacity of a fiber link could be increased manyfold at
minimum cost. However, it was only since the early 2000s that the optical switching
technology necessary to convert isolated fiber transmission links to optical networks
matured sufficiently to permit the commercial deployment of these networks. In the
mid-1990s, the optical network (as opposed to optical fiber transmission alone) was
still a “blue sky” concept. New optical and photonic devices were being developed and
incorporated into experimental networks. But full-fledged multiwavelength networks
integrating optical transmission, switching, and user access were still in the research and
development stage. At that time the technology push for networking was out in front,
but demand for the seemingly unlimited capacity of these networks was essentially
nonexistent. As this is being written, the promise of optical networking is finally being
fulfilled. The demand pull for these networks has materialized. As low-cost broadband
services are made available to the general public, demand for Internet-based applications
continues to increase. Equipment manufacturers, communications carriers, and service
providers have joined in moving optical networking from feasibility studies to commer-
cial viability in both cost and performance. The focus in the networking community has
now shifted to organization, control, manageability, survivability, standardization, and
cost-effectiveness, a trend that reflects the maturing of the optical technology as well as
the recognition that the optical network is the only way of supporting current and future
demand. These networks have played a critical role in reducing communications costs,
promoting competition among carriers and service providers, and thereby increasing the
demand for new services.
In addition to the technology push and demand pull, a number of other recent devel-
opments are contributing to the expansion and effectiveness of optical networks. One is
the accelerating removal of the bottleneck in the “last mile” – the distribution network
that is the bridge between the high-speed fiber core network and the end users. Until the
last decade of the 20th century this distribution network – composed of twisted pairs
of copper wires connecting each residential subscriber to the local telephone Central
Office – was specifically engineered to a limited bandwidth of 3000 Hz. As a result
the user bit rates were restricted to a tiny trickle. This low-speed access link separated
the various high-speed communications and computing devices located on the premises
of the end users (e.g., PCs, TV displays, and music/image/video storage equipment in
the home) from the high-capacity network serving this equipment. Considering that
the processors in today’s PCs operate at speeds six orders of magnitude faster than a
low-speed access link, and the optical fibers in the network have bandwidths nine orders
of magnitude wider than the bandwidth of the access link, it is obvious that access was –
and is – a severe problem. As long as the last-mile bottleneck is present, the information
 The Big Picture 3

superhighway is still a dirt road; more accurately, it is a set of isolated multilane high-
ways with cow paths for entrance and exit ramps. The introduction of broadband access
to residential customers by the telephone carriers and the cable operators is a step toward
eliminating those cow paths. However, digital subscriber line (DSL) and cable modems
are half-measures at best. Direct access to the fiber network by the end user [i.e., fiber
to the home (FTTH) or business user] is the ultimate way of removing the bottleneck
so that the network remains effective as demand for bandwidth grows. Although FTTH
was deployed many years ago in a few demonstration projects, it did not take hold for
several reasons, including cost and the absence of services of interest to the customers.
Today that has changed because of the proliferation of broadband Internet services. De-
ployment of glass is now moving from the network core through fiber access networks
to the end users. This will undoubtedly stimulate interest in new broadband services that
take advantage of high-speed access and in turn produce demand for more bandwidth.
At this writing, most of the world’s installed fiber capacity is underutilized – arguably
due to the last-mile bottleneck. That should change rapidly as progress in the removal of
the bottleneck results in a quantum jump in network traffic, making high-performance
optical networks indispensable.
Higher level issues such as deregulation, new ideas for improving the economics
of networking, and standardization of control and management techniques in multi-
vendor networks are also contributing to the growing effectiveness of optical networks.
Deregulation, which began in the United States in 1984 with the dismantling of AT&T,
has brought with it a new level of competition, with long-haul carriers, local carriers,
Internet service providers (ISPs), and cable operators poaching on each other’s domains
and using optical fiber capacity to do so. Bandwidth trading has been introduced as
a way of improving the utilization of fibers and thereby optimizing profits. A carrier
with idle capacity sells it to another carrier with excess demand. This type of exchange
requires sophisticated control and management tools for network reconfiguration. More
generally, any large network requires complex control and management systems and
intelligent network elements for performance monitoring, network reconfiguration, and
fault recovery. The systems, protocols, and equipment for performing these functions in
traditional telephone and data networks were built over many years by the public carri-
ers and equipment manufacturers. The new optical networks require similar tools, and
this is especially important in multivendor environments. These are now making their
appearance in the form of a proposed control plane for optical networks and protocols
for systems management in these networks. As more sophisticated control and manage-
ment functions are incorporated into optical networks the network operators are in a
better position to offer high-quality service to their customers, improving the operator’s
revenue stream and customer loyalty.
Above all, the lessons of the past show us that tomorrow’s networks must be flexible and
versatile enough to adapt to a continuing barrage of new and as-yet-unknown services.
It is interesting to note that when optical networks were still in an embryonic form,
the typical uses envisaged for them were high-tech applications such as high-resolution
medical image archiving and remote supercomputer visualization – basically usages
generated by a minuscule, elite segment of the population. Today these applications
4 Multiwavelength Optical Networks

represent but a tiny part of the global network traffic, submerged in a torrent generated
by the common man, who has only recently gained access to the enormous opportunities
our worldwide communication system has to offer. The networks we conceive today
must be “futureproof ” so as to be ready for the next unforeseen developments.

1.2 Objectives of an Optical Network Architecture

Today’s and tomorrow’s optical networks must provide the capacity, connectivity, and
intelligence necessary to link together a global community of information providers and
consumers. A well-designed network performs this function efficiently and reliably. To
facilitate a systematic study of networks that achieve this goal, it is useful to formulate
a generic model in the form of a multiwavelength network architecture (MWNA). As
background for the MWNA we briefly review the current network structures and the
services they support.
Until the end of the second millennium, the world of networking consisted of
two separate spheres: the traditional telephone networks mainly devoted to providing
voice services (operated in a circuit-switched mode) and data networks (operated in a
packet-switched mode) for communication between computers. Each type of network
was specially engineered and optimized for its own type of service. Circuit switching was
the preferred approach to voice transmission, because the voice signal was transmitted as
a continuous stream, whereas packet switching was invented to carry data traffic because
data signals were bursty in nature, making the circuit-switching approach very ineffi-
cient. Because the voice networks operated by the public carriers contained virtually
all of the world’s installed communication capacity, the early data networks were con-
structed as overlays on these networks, running on lines leased from the public carriers –
mainly AT&T in the United States and the government administrations in Europe.
The traffic flow in the early data networks was minute compared to voice traffic –
essentially confined to businesses, universities, and research laboratories. For this rea-
son the main players in data networking were originally government, research, business,
and educational organizations and data-processing equipment manufacturers.
As optical fiber became the dominant transmission medium, various standards for ex-
ploitation of fiber were developed, including the synchronous optical network (SONET)
standard in the United States and a similar synchronous digital hierarchy (SDH) stan-
dard in Europe. The SONET/SDH transmission, multiplexing and switching equipment,
adapted primarily to circuit-switched applications, was soon augmented by asynchronous
transfer mode (ATM) switches and Internet Protocol (IP) routers (cell-switched and
packet-switched, respectively) to handle a wide variety of data and multimedia services.
By the late 1990s the traditional separation of voice and data networks changed signifi-
cantly. In a very short time we moved from a voice-centric world to a data-centric world,
and, more importantly, the techniques of carrying data (packet switching) were extended
to an infinite variety of services having no resemblance to those in the traditional com-
puter world. Internet/Web services, running the gamut from interactive computer games
through telemedicine to peer-to-peer file sharing, now use IP for transmitting anything
 The Big Picture 5

from computer data to video (Internet Protocol TV; IPTV) to old-fashioned voice (Voice
over IP; VoIP).1
This brings us to the characteristics and requirements of the services supported by the
optical networks discussed here. These are extremely diverse in terms of connectivity,
bandwidth, performance, survivability, cost, and a host of other features.
Consider the common Internet services offered to the general public (e.g., e-mail
and search engines). They serve a vast globally distributed user community. In terms
of connectivity, these types of services push networking to its ultimate limits; any
end user wants rapid connectivity to anyone or anything in the network. However, in
terms of performance, they are undemanding – they can tolerate errors, delays, and
occasional downtimes due to congestion, programming bugs, and equipment failure.
Total costs may be high, but they are spread over an enormous user base resulting in a
very low cost per user.
In contrast, consider a different type of application, the virtual private network (VPN).
This is a subnet carved out of a larger network by a telecommunications carrier and
put at the disposal of a single enterprise, which typically controls and manages it.
Consequently it has a much smaller user group with more intense utilization per user,
far fewer active connections, and tighter control of network performance, including
security and reliability. Customer costs per user will be higher, but this is offset by
higher performance and more responsiveness to the needs of the customer.
Another example is telemedicine, which requires high-quality communication (e.g.,
high fidelity medical image encoding and transmission, and rapid response) and where
cost is secondary. Different requirements apply to public safety services (e.g., police,
fire, and disaster relief), which depend on a high degree of survivability, fault recovery,
and availability2 in the face of equipment and line failures, natural disasters, or malicious
attacks. Transmission quality is secondary. Similar requirements hold to a lesser degree
for financial services (e.g., banks and brokerage houses). In public safety and financial
service applications, cost is not the primordial issue.
To ensure satisfactory service, large users of network services (e.g., enterprises op-
erating VPNs) enter into service-level agreements (SLAs) with the service providers.
For example, the SLA might specify a level of availability, network delay, packet loss,
and other features. These represent promises from the provider to the user, and as such
they must be backed up by suitable controls within the underlying network to achieve
the performance stated in the SLA. These controls are enforced within a large network
by identifying differentiated services, that is, traffic flows that are singled out to be pro-
vided with a predictable quality of service (QoS) (e.g., limits on packet loss and delay).
Traffic routed through a large network can be tagged to recognize its class of service
(CoS), thereby facilitating the satisfaction of service requirements through mechanisms
such as priority packet queueing, bandwidth allocation, and service recovery priority.

1
The increased interest in VoIP, because of its low cost and growing ubiquity is, to paraphrase Shakespeare,
the most unkindest cut of all from the computer community to the traditional telephone carriers.
2
Availability is the percentage of time that a network is operational. For example, “five 9s” (99.999%)
availability, which is a goal for public carriers, implies 5.25 minutes of downtime per year.
6 Multiwavelength Optical Networks

Music/video
Internet access POTS VPN VoIP Telemedicine file sharing

Services
Layer

IP IP
Logical
Layers ATM

SONET

Wavelength
on demand

Physical
Layer

Figure 1.1 Multilayered network.

The various functions executed by the network operator, such as load balancing and
QoS-based traffic handling, are known as traffic engineering.
Considering the wide diversity of these service requirements (and we have only
mentioned a small sample), it is a challenge to support all of them on a single network.
Yet this is not only possible, but it is generally the most efficient way of doing the job. As
we will show, the building blocks are now available to assemble multiwavelength optical
networks that can sustain large user populations with diverse service requirements of the
type just described. This means that, ideally, the MWNA must be structured to offer a
special set of features adapted to each service it supports. To see how this is achieved it
is convenient to think of the network in terms of its constituent layers, with client–server
relations between the neighboring layers. An illustration is the multilayered view shown
in Figure 1.1. The architecture is composed of an underlying optical infrastructure –
the physical layer – which provides basic communication services to a number of
independent logical networks (LNs) residing in the logical layer. Each LN organizes
the raw capacity offered by the physical layer, adapting it to the needs of the clients
it serves, shown in the services layer of the figure. For example, the SONET network
shown in Figure 1.1 uses optical wavelength channels provided by the physical layer,
transmits optical signals on them, and carries multiplexed communication channels on
those signals. The SONET channels can be tailored to support a wide variety of services;
two services shown in the figure are plain old telephone service (POTS) and a VPN.
In our example, the SONET layer also supports an ATM layer that in turn supports
a client IP layer providing Internet access services to end users. Another independent
IP network shown in Figure 1.1 is supported directly by the physical layer, providing
a telemedicine service, VoIP, and a music/video file-sharing service. In addition, the
 The Big Picture 7

physical layer provides purely optical connections directly to end users via demand-
assigned wavelengths (also known as clear channels), thereby bypassing the logical
layer altogether.
Thus, the logical layer shown in the figure contains several LNs; some are stacked in a
client–server relationship, and others are independent of each other, offering specialized
features to the service layer. Stacked logical layers; e.g., IP over ATM over SONET
over WDM, have both advantages and disadvantages. For example, different services
(e.g., POTS and VPNs) require channels running at different bit rates. The SONET layer
supports these different speeds and in addition provides a grooming function, packing the
diverse channels onto a common optical wavelength, using time division multiplexing.
This “fills up” the wavelength channel for efficient utilization. However, stacked layers
mean additional equipment, which is costly, introduces delays and potential points of
failure, and is difficult to manage. Therefore, it is desirable to reduce superfluous layers
wherever possible. For example, the IP equipment manufacturers propose to provide
IP over WDM, short-circuiting the commonly used configurations involving stacked
intermediate layers.3
Another view of an optical network is the physical picture of Figure 1.2, showing the
network elements in the layers of Figure 1.1. Here the physical layer is portrayed for
simplicity as a transparent purely optical core.4
The “glue” in the physical layer that holds the transparent optical network together
fits roughly into two basic classes: the optical network nodes (ONNs), which connect
the fibers within the network, and the network access stations (NASs), which interface
end-user systems and other nonoptical equipment to the network. Shown as rectangles in
Figure 1.2, the NASs (or stations for short) provide the terminating points (sources and
destinations) for the optical signal paths within the physical layer. The communication
paths continue in electrical form outside the purely optical part of the network, either
terminating at end systems (for example, PCs, telephones, and servers) or traversing
electronic multiplexing and switching equipment (e.g., ATM switches, IP routers, or
SONET digital cross-connect systems [DCSs]), shown as hexagons in Figure 1.2. The
ONNs (or nodes for short), shown as circles in Figure 1.2, provide the switching and
routing functions that control the optical signal paths (also called lightpaths), configuring
them to create desired source-destination connections. The stations and nodes contain
the optoelectronic and photonic components of the network: lasers, detectors, couplers,
filters, optical switches, amplifiers, and so on. These components work together with
the fibers to produce the required optical signal connectivity. Although the underlying
optoelectronic and photonic technologies have matured considerably since the mid-
1990s, they are not as well developed as their electronic counterparts. Thus, electronics
(in the logical layer) is currently an equal partner with photonics (in the physical layer).

3
In many cases, superfluous LN stacking results from a reluctance of carriers to write off a large investment
in legacy systems, which would require a complete revision of existing control and management structures.
4
As will be seen, the physical layer in optical networks often includes electronic components in the form
of signal regenerators or electronic switch fabrics, so it is not always purely optical nor is it completely
transparent. We have more to say about purely optical signal paths and the meaning of transparency in
Sections 1.3 and 1.4.
8 Multiwavelength Optical Networks

Internet
access
Tele- File
VoIP
medicine sharing
IP
POTS VPN
A

S IP

Super-
e computer
o

NAS

ONN

IP IP router Optical Ether

A ATM switch

S SONET DCS

Figure 1.2 Physical picture of the network.

The line between the optical and electronic parts of the network has become fuzzy as
technology has advanced, but at this point in our discussion we retain the simplified view
that the physical layer is transparent and optical, whereas the logical layers are electronic
and “opaque.”
The electronic switching node plays the same role in the logical layer as ONNs play
in the physical layer. Our generic term for an electronic switching node (each hexagon
in Figure 1.2) is a logical switching node (LSN). The LSNs sort, multiplex, switch, and
route signals in the various LNs. In this way they create virtual connections among the
entities they serve. These entities may be service provider equipment or user equipment
(end systems), as in the case where the LSNs are IP routers connecting ISP servers to
customer PCs, or they may be higher layer switching nodes in a client network, as in the
case of SONET DCSs serving ATM switches serving IP routers.
Although the focus of this book is the physical layer and its optical components, the
logical layers are an integral part of the overall network architectures we discuss here.
Therefore, our MWNA includes the logical layers and their electronic components. An
understanding of the design and operation of multiwavelength optical networks requires
an awareness of the close coupling between the physical layer and the logical layers it
serves.
The networks we examine will generally be designed to serve large, heterogeneous,
geographically dispersed user populations. Given this fact, and the various service
requirements discussed, we can infer a list of general design and operating objectives:
 The Big Picture 9

r Connectivity
– Support a very large number of end systems
– Support a very large number of concurrent connections, including multiple connec-
tions per station and per end system
– Support multicast connections efficiently
r Performance
– High aggregate network throughput (hundreds of terabits per second)
– High fiber transmission capacity (terabits per second)
– High user bit rate (gigabits per second)
– Small end-to-end delay
– Low error rate (digital)/high signal-to-noise ratio (analog)
– Low processing load in nodes and stations
– Adaptability to changing and unbalanced loads
r Technology: cost-effective design and utilization
– Access stations: limited number of optical transceivers per station, limited complex-
ity of optical transceivers, simple tuning techniques
– Optical network nodes: high throughput, minimal signal impairment, and low
complexity
– Logical switching nodes: efficient channel grooming, simple packet-routing proce-
dures, and controlled traffic load
– Network: Limited number and length of cables and fibers, efficient use (and reuse)
of the optical spectrum, controlled signal impairment in the physical layer, mini-
mization of logical layer complexity.
r Structural features
– Scalability
– Modularity
– Survivability (fault tolerance)
r Control and management
– Efficient, rapid, automated connection provisioning, and reconfiguration
– Built-in intelligence in the network elements for monitoring and control
– Efficient and rapid automatic fault identification and recovery
– An integrated network management system to monitor and coordinate all network
layers

As we look at existing and proposed network architectures, it is important to keep

these goals in mind.

1.3 Optics versus Electronics: The Case for Transparent

Multiwavelength Networks

There are certain functions that come naturally to each technology. Referring to the
somewhat idealized view of a network in Figure 1.2 based on the assumption of a purely
optical physical layer, there is a clean separation between optical/photonic technology,
10 Multiwavelength Optical Networks

on the one hand, and electronic technology, on the other. The NASs represent the
optoelectronic interface (denoted by the boundary labeled e/o) between the electronic
domain (the equipment outside the purely optical portion of the network) and the optical
domain, sometimes called the optical ether. This interface is the point of demarcation
between the physical layer and the logical layers. In a typical purely optical physical
layer the optical signal paths are as transparent as a piece of glass.
The stations provide the basic functions of getting the light into the fibers (with lasers)
and getting it out (with photodetectors). When the signals are in optical form, photonic
technology is well suited to certain simple signal-routing and switching functions within
the nodes. With static photonic devices, it is fairly easy to perform functions such as op-
tical power combining, splitting, filtering, and wavelength multiplexing, demultiplexing,
and routing. By adding suitable control, the static devices can be controlled dynamically
(switched) at slow to fast speeds (milliseconds in the case of mechanical or thermal
control and microseconds or nanoseconds in the case of electronic control).
The enormous usable bandwidth of a single fiber (tens of terahertz) is at the same
time a great asset and a great challenge. It is technologically impossible to exploit all
of that bandwidth using a single high-capacity channel. Thus, to make efficient use of
the fiber it is essential to channelize its bandwidth. This is most easily accomplished
by superimposing many concurrent signals on a single fiber, each on a different wave-
length; that is, by using WDM. Thus this book focuses on multiwavelength or WDM
network architectures. The relative ease of signal manipulation in the wavelength (or
optical frequency) domain, as opposed to the time domain, suggests that current optical
technology is particularly suited to multiwavelength techniques. In WDM networks each
optical transmitter (receiver) is tuned to transmit (receive) on a specific wavelength, and
many signals operating on distinct wavelengths share each fiber – possibly more than
100 in dense WDM (DWDM) transmission systems.
It should be observed that all photonic routing and switching functions within the opti-
cal domain in these networks are linear operations. Thus, at the optical level the network
typically consists of only linear devices, either fixed or controllable. It is the property
of linearity that makes multiwavelength networking simple and cost-effective. To dis-
tinguish these linear networks from other types of optical networks, we refer to them
frequently as transparent optical networks. Typical nonlinear operations performed in
networks include signal detection, regeneration, reading, and modifying the information
in the signal, buffering, and logic functions (e.g., packet routing based on header infor-
mation). Although many nonlinear functions can be performed in the optical domain
with present-day technology the current state of the art for these nonlinear devices is
not nearly as advanced as it is for linear components. For these reasons, we frequently
use the terms transparent optical network and purely optical network interchangeably
in this book.5 Nonlinearities make the signal path opaque rather than transparent. Some
of the advantages of keeping nonlinear operations out of the signal path are (1) the

5
As the state of the art progresses photonic technology is becoming a viable alternative for many nonlinear
signal processing operations, so that the linkage between “transparent” (i.e., linear) and “purely optical” is
becoming tenuous.
 The Big Picture 11

end-to-end optical path behaves as a literally transparent6 “clear channel” so that there
is nothing in the signal path to limit the throughput of the fibers (a transparent channel
behaves very much like an ideal communication channel with almost no noise and a
very large bandwidth), (2) the architecture of the optical network nodes can be very
simple because they have essentially no signal processing to do (optical node simplicity
also means simplicity of network control and management), and (3) system upgrades
involving changes in speed, format, and protocol are easily implemented.
There are also downsides to transparency. First, problems caused by equipment failures
tend to propagate throughout the network, making fault management a more complex
issue than in nontransparent networks. Similarly, impairments such as switch cross-talk,
noise, fiber dispersion, and nonlinear effects accumulate over long paths, limiting the
geographic “reach” of an optical connection. Second, by definition, in-band information
(e.g., control information carried in packet headers, such as source and destination
address, sequence number, channel number, and parity check bits) cannot be used while
the signal is in optical form. Because of this, a transparent physical layer cannot perform
the various processing functions required in packet switching.
It is important to note that the in-band control information carried with the data
in IP (packet-based) or ATM (cell-based) networks is the key to achieving a high
degree of virtual connectivity in these networks. Typically, many virtual connections
are multiplexed on each network link and sorted (switched) on a packet/cell basis at
each IP router or ATM switch using information contained in the packet or cell headers.
Maintaining transparency in the physical layer eliminates the intelligence necessary
to process this information and therefore tends to produce an optical “connectivity
bottleneck” in transparent networks. For all of the aforementioned reasons, there is a
case to be made for opaque optical networks [Bala+95].
The properties of electronics are complementary to those of optics. Electronic pro-
cessing is ideal for complex nonlinear operations, but the limited speed of electronic
and optoelectronic devices (e.g., electronic switches, memory devices, and processing
units), and the high processing load imposed on electronics in broadband networks,
causes a well-known “electronic bottleneck” in optical transmission systems. Putting an
electronic termination on an optical fiber reduces the potential multiterabit-per-second
throughput of the fiber to a multigigabit-per-second trickle: the maximum speed that can
be expected of an electronic signal. This is the origin of the highway/cow path analogy
we used in Section 1.1. More succinctly, optics is fast but dumb, whereas electronics is
slow but smart.
A final caveat: whenever we speak of enabling technology it must be understood that
it is a fast-moving target. Thus, the state of the art is rapidly evolving in the direction of
smarter optics and (somewhat) faster electronics. One consequence of smarter optics is
that nonlinear operations can be introduced into a purely optical network using optical
processing. For example, optical packet switching can be realized either through purely
6
Transparency implies that signals with any type of modulation schemes (analog or digital), any bit rate, any
type of format, and using any kind of protocol can be superimposed and transmitted without interfering
with one another and without their information being modified within the network. Opaque networks do not
have these properties.
12 Multiwavelength Optical Networks

optical processing or a combination of electronics for header processing and optics for
switching, resulting in an opaque optical network capable of very high speed packet-
switched operation (see Chapter 10). Another example of smarter optics is the use of
optical processing for signal regeneration and wavelength conversion within an otherwise
linear signal path (see Chapter 4). Faster electronics is a more questionable issue, because
as we push the electronic speed limits, costs rise rapidly.

1.4 Optics and Electronics: The Case for Multilayered Networks

Because of the size and complexity of the networks we are considering, and because
of the fact that lightwave technology alone cannot satisfy our networking objectives,
we now return to the multilayered model of Figure 1.1 with a more detailed look at
the logical layers. Elaborating on the discussion of Section 1.2, we continue with the
assumption of a clear separation between optics (in the physical layer) and electronics (in
the logical layers); that is, the network has as its physical foundation a multiwavelength
purely optical network. Superimposed on the physical layer are one or more LNs, each
of which is designed to serve some subset of user requirements and is implemented as
an electronic overlay superimposed on the physical layer. Just as a transparent optical
network has a physical topology composed of ONNs and fiber links, a logical network
has a logical topology composed of LSNs and logical links. A logical link is an electronic
transmission channel joining two LSNs. It is carried on an optical path provided by the
underlying physical layer.
Each LN organizes the connections it offers to its clients in a specific way, with its own
layered architecture. Different LNs may be managed independently or in coordination
with others. Typically, an LN will provide services to end systems in the form of virtual
connections traversing paths in the logical topology. In MWNAs involving stacked LNs
a logical layer acting as a server for a client logical layer (e.g., the ATM layer serving
the IP layer in Figure 1.1) will offer its services to the client layer in the form of virtual
connections, which can be used as logical links in the client layer. The physical layer
makes a large pool of bandwidth available to the LNs in the form of transparent end-
to-end connections. These high-bandwidth optical channels may be used to provide a
dedicated communications backbone for an LN as in IP over WDM, or the channels
may be demand assigned for temporary activities such as response to changing traffic
distributions or recovery of faults.
In descending the layers in the optical network architecture, several connections in
each client layer are typically multiplexed on a single connection in the corresponding
service layer, resulting in fewer but “thicker” connections carried on the optical paths in
the underlying physical layer. This results in a connection granularity ranging from fine
to coarse as we move down the stack. In each LN the electronic switching equipment
acts as a “middleman,” taking high-bandwidth channels offered to it by the layer below
it and organizing them into lower bandwidth channels with a format acceptable and
cost-effective for the end users and/or client layer it serves. For example, the layer in
Figure 1.1 providing IP over WDM makes use of wavelength channels offered to it by
 The Big Picture 13

Logical
LL1 LL2 LL3
Layer

Physical Optical Layer

Layer Fiber Layer

Figure 1.3 Layered view of an optical network.

the physical layer and “packages” the bandwidth so as to support the flow of IP packets
among its end users. Sophisticated network users requiring high bandwidth and the
flexibility of a clear channel can dispense with the services of an LN to obtain direct
access to demand-assigned wavelength channels as shown in the figure, without the
intervention of an electronic middleman.
Returning to the physical picture, Figure 1.2 illustrates in more detail how the end
users interface to the network through various layers of logical (electronic) switching
equipment, and then through the NASs to the physical layer. For example, a user of
Internet access services is shown accessing the network via an IP router on top of
an ATM switch on top of a SONET DCS, which uses the services of the transparent
physical layer. At the other extreme, a supercomputer is shown accessing the physical
layer directly through an NAS.
In Figure 1.3, the architecture of the overall network is shown in a more formal
view, reduced to its constituent layers, each one providing support for the layer above
it, and using the services of the layer below it – a natural extension of the idea of
layered architectures to optical networks. The layering formalism is introduced here to
partition a complex set of interactions among network components into a small number
of more manageable pieces. The characteristics of the logical layers (LLs) depend on the
architectures of the various LN overlays. The physical layer, representing a purely optical
network, is now shown divided into an optical layer and a fiber layer. The former contains
the optical connections supported by the fibers, and the latter embodies the layout of the
physical infrastructure itself: the fibers, switches, and optical transceivers. In Chapter 2,
the layered view of optical networks is expanded into a complete multiwavelength
network architecture.
Why is a hybrid approach required for optical networks? Although some early struc-
tures proposed for local area networks (LANs) and metropolitan area networks (MANs)
were purely optical, the current state of the art suggests that neither optics nor electronics
alone can provide all the desired features listed in Section 1.2. Using current technology,
purely optical wavelength-selective switches are capable of interconnecting as many as
100 fibers operating in a DWDM mode, switching and routing each wavelength inde-
pendently and yielding aggregate throughputs in the range of many terabits per second
while supporting hundreds to thousands of optical connections running at speeds of the
order of 10 Gbps each. On the other hand, electronic packet/cell switches are currently
limited in throughput to approximately 100 Gbps.7 However, they can support a much
larger number of relatively low-bit-rate virtual connections.
7
A similar, but not as severe, throughput limitation applies to “opaque” optical switches that are often used
in the physical layer to replace purely optical ONNs. An opaque switch converts the optical signals to
14 Multiwavelength Optical Networks

The difference in the two approaches is the granularity we spoke of earlier. Optical
switches are operated most easily in a circuit-switched mode, in which the information-
bearing units (optical wavelength channels) being switched are few but large (in band-
width) and the holding times for a given switch configuration are long (seconds or
more). Circuit-switched operation of the optical nodes is perfectly suitable for the phys-
ical layer shown in Figures 1.1 and 1.3. Dedicated connections supporting the various
LNs are normally held in static configurations for durations of hours, days, or more.
Demand-assigned connections are held typically for minutes or hours. Thus, the num-
ber of circuits being set up and taken down per unit time is relatively small. This type
of operation requires little processing and provides a high aggregate throughput. Con-
versely, electronics is employed in situations in which there are many information units
(e.g., individual packets or cells) being switched per unit time. Because the units are
typically small (in number of bits) and because each unit is processed individually,
this leads to a heavy processing load, with a relatively low throughput limited by the
processing power of the switch. Wide area networks (WANs) must handle both large
and small information units: hence the need for marrying both electronic and optical
switching technologies. The hybrid approach exploits the unique capabilities of each
while circumventing their limitations.
To illustrate the importance of combining optical and electronic technology let us
explore the demands placed on large networks. Broadband WANs must be capable of
supporting high connectivity, high throughput, and heterogeneous traffic mixes, and they
must be flexible in meeting changing demands and unforeseen circumstances such as
equipment and link failures. Consider a WAN serving as a backbone that interconnects
a large number of users. To reduce access costs, the traffic from small users should be
aggregated before it enters the WAN, through the intermediary of an access network –
a high-speed LAN, a fiber access network, or an electronic switch (e.g., an IP router
owned by an ISP). Each of these represents a gateway to the WAN, and there might be
10,000 of these in a network of global scale, each one serving hundreds of active users.
The aggregate traffic in and out of each gateway might require connectivity to perhaps
1000 other gateways on the network at any one time, and the total traffic injected into the
network might be of the order of tens of terabits per second. (This example corresponds
in orders of magnitude to present-day numbers on the Internet.)
Let us examine the optical, electronic, and hybrid alternatives for supporting these
users. Consider first a network supporting gateway interconnection, using purely optical
switching. It would resemble a modified version of Figure 1.2, with the access gate-
ways attached directly to NASs without the intervening (electronic) LSNs, as shown in
Figure 1.4(a). Costs dictate that there will be considerably fewer optical nodes than
NASs (i.e., each node may serve many stations). In the purely optical case, each NAS
interfaces one or more gateways to the backbone, so that the NAS must connect its gate-
ways to the other gateways on the network through individual optical connections. This
could require millions of relatively low-bit-rate connections (as many as 10,000,000 for
electronic form for purposes of switching, and in the process it performs signal regeneration, wavelength
conversion, and signal monitoring as well. Although this makes an opaque switch much more versatile than
a transparent optical ONN, it also produces an electronic bottleneck within the node.
 The Big Picture 15

Gateway Gateway Gateway

Gateway
Gateway Gateway
Gateway Gateway Gateway
Gateway
Gateway Gateway Gateway

Gateway

e
Gateway
o
Gateway

IP
IP
IP
IP
NAS

ONN IP
Gateway IP
Optical Ether
IP Gateway
IP IP router
IP IP
IP

Gateway Gateway
Gateway Gateway
Gateway Gateway Gateway Gateway

(a) Purely optical switching (b) Electronic switching

Gateway Gateway Gateway

Gateway
Gateway Gateway

Gateway IP IP

e
IP Gateway
o

NAS

ONN
Gateway
IP IP router Optical Ether

IP Gateway
IP Gateway
Gateway
Gateway

(c) Hybrid network

Figure 1.4 Alternative network approaches.

the connectivity postulated here). It is the high-connectivity requirement that makes a

purely optical approach very difficult. To implement this connectivity, each station must
maintain connections, to 1000 other NAS/gateways (assuming 1 gateway per NAS). This
means either equipping each station with that many optical transceivers or providing ex-
tremely rapid optical connection switching. No matter how it is realized, this degree of
connectivity is well beyond the reach of current optical technology.
A purely electronic version of Figure 1.2 has its own problems. This would be im-
plemented with electronic rather than optical switches at the network nodes, as shown
in Figure 1.4(b), where the ONNs have been replaced by IP routers. This reduces the
physical layer to a set of isolated point-to-point transmission links terminating on the
electronic routers. Electronics can easily support the required connectivity via virtual
connections. However, the electronic processing bottleneck at the switches makes it
difficult and expensive to sustain the required multiterabit throughput on the backbone.
Because optical switching is still in the early stages of penetrating large networks,
current architectures are tilted more to the electronic side (with multiple stacked logi-
cal layers) than the optical side. However, pressures of increasing demand, performance,
cost-effectiveness, and fault tolerance are moving networks toward hybrid configurations
of the type shown in Figure 1.2.
16 Multiwavelength Optical Networks

In a hybrid architecture both high connectivity and high throughput are achieved
easily and efficiently. For example, an LN composed of IP routers provides the necessary
connectivity through sorting and routing packets at each LSN (see Figure 1.4[c]). The
physical layer supports the required throughput over dedicated, high-bandwidth (optical)
backbone connections carrying aggregated traffic on logical links between the IP router
ports. This example illustrates the fact that the test of a good lightwave network is
whether it can achieve both high throughput and high connectivity at a reasonable cost.
Another advantage of the hybrid approach is the versatility achieved by combining
optical and electronic technology. This is especially important from the point of view of
reconfigurability. For example, frequent changes in fine granularity traffic routing due to
changes in customer demand are most easily accommodated by switching in the logical
layers; i.e., by using the intelligence built into the IP routers. However, rerouting of
masses of coarse granularity traffic due to equipment or link faults is more quickly and
efficiently handled in the physical layer by resetting the optical paths through the ONNs
using automatic fault recovery mechanisms. Without a reconfigurable layer underneath
them, the logical links between the IP routers are fixed. We have an IP network whose
logical topology is written in stone.8 Conversely, without an intelligent logical layer
above it, the physical layer cannot manipulate the fine-granularity traffic flows among
end users. In other words, the logical layers need a reconfigurable physical layer just as
the physical layer needs reconfigurable logical layers.

1.5 Network Hierarchies

Just as there is a case to be made for a multilayered network architecture, there is also
a rationale for a hierarchically structured network. Traditional carriers have all found
hierarchies to be useful. In telephone networks, individual subscribers are connected to
a nearby Central Office, and Central Offices in the same area are interconnected by a
MAN, typically spanning a region of a few hundred kilometers at most. These are usually
in the form of rings. Finally, the MANs are interconnected by long-haul networks with
mesh topologies. These include ultra long-haul transmission links such as transoceanic
cables. In going from local subscriber access through MANs to long-haul networks,
efficiency and manageability demand that the granularity of the connections increases,
just as it does in going down the layered architecture of Figure 1.1.
The new optical network architectures have similar hierarchies as shown in the exam-
ple of Figure 1.5. The mesh network in the center is the core long-haul network, which
joins the MANs in the form of rings. End users connect to the MAN through access
networks joined to the MAN at gateways indicated by the shaded circles in the figure.
These access networks might be in optical form (e.g., p⁀ assive optical networks [PONs])
or electrical form (e.g., traditional LANs or electronic switches). Their purpose is to
aggregate traffic from individual users for more efficient and cost-effective transmission
8
One of the reasons for the use of stacked LNs such as ATM and SONET underneath an IP logical network
is that the lower logical layers can be used for offering a reconfigurable logical topology to the IP network,
partially replacing the functions of a reconfigurable physical layer.
 The Big Picture 17

Access point

ONN

Figure 1.5 Hierarchical network.

on the network. A characteristic of the hierarchical structure is that as one moves closer
to the end user, the number of entities attached to the network grows exponentially.
Referring to Figure 1.5, there are many MANs attached to the core, many access net-
works attached to each MAN, and many end users attached to each access network.
Conversely, the amount of traffic originating from each entity declines exponentially as
one moves closer to the end user. Because there are so many end users with modest
individual traffic demands, the equipment located close to the end users in the access
networks must be inexpensive and simple but does not need to be high capacity or high
performance. Conversely, the long-haul network contains relatively few optical links
and ONNs (the clear circles in Figure 1.5), but they must support high throughput with
carefully controlled signal quality and high reliability. For example, transmission on
long-haul links is cost-effective when a large amount of capacity is packed into each
fiber. This means using large numbers of closely spaced wavelengths; i.e., DWDM with
subnanometer wavelength separations and high bit rates (10 Gbps or higher) on each
wavelength. Achieving these numbers over distances of thousands of kilometers is not a
trivial feat and requires costly transmitting, amplifying, and receiving equipment.
The MAN must meet criteria somewhere in between these two extremes. For example,
a standard called coarse wavelength division multiplexing (CWDM) is adapted to MANs
and uses spacings of the order of 20 nm. This wide spacing combined with fairly short
transmission distances means that relatively inexpensive transmission equipment can be
used in these applications.
In addition to the cost–performance issues in the various levels of a hierarchy, there are
important control and management issues. The complexity of control and management in
a large network grows rapidly with the number of entities being managed. For example,
in real-time decision making for fault recovery, speed is of the essence, but speed
is compromised if the fault recovery system must deal “microscopically” with every
18 Multiwavelength Optical Networks

fine-grain active connection in the network. If all traffic among end users in a large
network was transported and controlled in the form of low-speed individual connections
(e.g., kilobits per second for a voice call) the number of connections to keep track of
in the high-capacity long-haul core would be overwhelming, making the core network
unmanageable. Instead, as we move from the end user, through the MANs and into the
core, various levels of multiplexing are used, so that the fine-grain end-user connections
are bundled into a much smaller set of coarse granularity connections on the long-
haul links. Manageability is maintained because there is a small number of high-speed
connections (and a small amount of hardware) to control rather than an enormous number
of low-speed connections. On the other hand, control and management in a lower level
of the hierarchy – a single MAN or access network – can be handled more or less
autonomously, dealing only with the relatively small number of “local” entities without
an overall view of the complete network. Another way of seeing this is in terms of the
multilayered view in Figure 1.1. In a high-speed core of an optical network, control
activity is largely confined to high-speed highly multiplexed connections in the physical
layer. In moving out to the end users the control functions shift to the higher logical
layers, but because the view is local, the total number of entities being controlled is
still relatively small.

1.6 A Little History

The idea of a high-speed optical transmission system (in free space) was considered as
early as 1958, when the laser was conceived [Schawlow+58], and guided wave optical
transmission was exhibited in the laboratory in the mid-1960s [Kao+66]. However,
practical optical transmission systems did not become possible until the production of
the first low-loss fibers and the invention of the semiconductor laser diode, both around
1970. By refining the optical transceivers and reducing fiber loss, the effectiveness of
unamplified optical transmission systems (measured in bit-rate-distance product) grew
roughly at an exponential rate from the early 1970s to the late 1980s, with bit rates as high
as 8 Gbps over distances of 100 km achieved in the mid-1980s [Miller+88]. The first
optical-fiber transatlantic cable (using electronic repeaters) was laid in 1988. The distance
limitations due to fiber attenuation disappeared in the late 1980s, almost overnight, with
the emergence of the erbium-doped fiber amplifier (EDFA) [Desurvire+87, Mears+86,
Mears+87]. Over the ensuing years interest in long-distance optical transmission using
EDFAs grew rapidly [Saito+90]. In laboratory experiments, in which long distances are
simulated using closed loops with amplification, and in which fiber dispersion effects are
eliminated using solitons, transmission distances have been extended essentially without
limit. For example, [Nakazawa+93] reported a 10-Gbps soliton system operating over
a total distance of 106 km.
During the late 1970s to the middle 1990s, fiber transmission capacity on a sin-
gle wavelength roughly doubled each year. Afterward, the focus was on multiwave-
length transmission, resulting in a significant jump in aggregate transmission bit rates
to the terabits-per-second range in the late 1990s using WDM. A recent example of
 The Big Picture 19

high-capacity long-distance transmission is reported in [Charlet+04]. It involved a

DWDM system using 149 channels at 50-GHz spacing running at 42.7 Gbps each,
for a total capacity of 6 Tbps over 6120 km.
While the transmission limitations in both capacity and distance were being sur-
mounted, important developments were taking place at other levels as well. The SONET
and SDH standards were developed in the late 1980s [Ballart+89, Boehm90]. Both of
them pertain to optical transmission links carrying synchronous bit streams terminated
by electronic switches.
Soon after SONET and SDH came on the scene, the concept of a broadband integrated
services digital network (B-ISDN) was introduced as a means of supporting all sorts
of multimedia services on a common network [CCITT92]. In the 1990s, much activity
was devoted to developing ATM as the preferred transport service for B-ISDN. The
cell-based transport technique in ATM (essentially a fixed-length, fast packet-switching
system) lent itself at the time to a wide variety of multimedia applications and at the
same time was well adapted to high-speed switching techniques [de Prycker91].
All of the developments discussed so far were carried out within the traditional voice
communications carrier community. However, the early data networks also influenced the
structure of today’s optical networks. The first data networks were developed in the 1970s
mainly for business users, utilizing packet switches designed by computer equipment
manufacturers to work together in a closed network environment employing proprietary
protocols. Examples were IBM’s SNA (Systems Network Architecture) and Digital
Equipment’s DEC-NET. Typical applications were airline reservation systems and time-
shared computing. Governmental organizations joined with ARPANet in the United
States, Datapac in Canada, and Cyclades and Transpac in France. Although Datapac
and Transpac were public data networks, ARPANet and Cyclades were experimental
and contained the precursors of today’s dominant global packet-switched network – the
Internet. Today’s Internet routing equipment and IP, the Internet Protocol, evolved from
the hardware and software developed in the 1970s for ARPANet, which was a pioneering
communication network sponsored by the Advanced Research Projects Agency (ARPA)
of the U.S. government. ARPANet was developed as an experiment in computer resource
sharing, and its driving force was the computer community.
In a competing development, the International Standards Organization (ISO) Open
Systems Interconnection (OSI) Reference Model was created in the 1980s as an attempt
by international organizations to produce a common standard that would move data
communications out of the proprietary networking world, making multivendor networks
feasible. However, at the same time that the OSI model was being promoted worldwide
by international standards bodies, the Internet was quietly and spontaneously spreading
throughout the world. Proprietary networking protocols, which had previously become
de facto standards were displaced by IP, the language of the Internet. ISPs became key
players in the networking community, and the World Wide Web, invented in 1989, pro-
vided the impetus for a flood of new services requiring a high-capacity, high-connectivity
infrastructure, and high-speed user access. The new multimedia applications necessi-
tated broadband access, and, as mentioned in Section 1.1, the telephone and cable carriers
responded, respectively, with DSL and cable access. Going beyond these copper-based
20 Multiwavelength Optical Networks

systems there has been considerable activity by the local exchange carriers (LECs) in
fiber-based access: fiber to the curb or node (FTTC or FTTN), building (FTTB), cabinet
(FTTCab), and home (FTTH).9 Although most of the developments discussed so far
occurred in the logical layers of the network, they were and are continuing to produce
challenges throughout the layers of the network architecture in terms of increased net-
work loads and ever escalating performance requirements. It is ultimately the task of the
underlying optical infrastructure to respond to these challenges.
Spurred by the early developments in optical and photonic technology, interest in
purely optical networks began in the mid-1980s [Henry89]. However, the technological
barriers to the deployment of large-scale networks remained formidable until the advent
of the fiber amplifier. Systems efforts during the pre-EDFA era were focused on a
simple architecture appropriate for LANs or MANs: the broadcast-and-select network
[Mukherjee92a]. In a typical network of this type, each access station is equipped with
a single laser transmitter capable of generating light at a fixed wavelength and contains
a single optical receiver capable of being tuned to the wavelength of any transmitter.
Signals from all transmitting stations are combined in a centrally located optical star
coupler, a passive device that broadcasts an attenuated version of the combined signals
back to each receiver. By selecting the appropriate wavelength, each receiver can accept
the signal injected by the corresponding transmitter, thereby creating a transparent
connection from the transmitter to that receiver.
Probably the earliest prototype of a broadcast-and-select network was LAMBDANET
[Goodman+86, Goodman+87]. Broadcast-and-select networks do not scale well to large
sizes primarily because they rely on rapid tuning of optical transceivers over a wide
range of wavelengths, they waste optical power, and, most important, they make poor
use of the optical spectrum. More general mesh topologies were soon proposed to
eliminate the constraints of broadcast-and-select networks. In these networks, alternate
paths together with wavelength routing produce possibilities for reuse of the optical
spectrum [Hill88] as well as recovery from failures.
At about the same time that wavelength routing was proposed, the multihop concept
was suggested to obtain high connectivity without requiring expensive, tunable optical
transceivers. Multihop networks were early examples of the hybrid approach described
in Section 1.4 [Acampora87, Mukherjee92b], relieving the connectivity bottleneck at
the optical level by adding packet or cell switches in an electronic LN overlay. From the
late 1980s to the present, activity intensified in optoelectronic and photonic technology
in the demonstration and deployment of new network architectures and, most recently,
in optical network standardization, control, and management.
Some of the important recent technological advances occurred in fiber technology
and amplification, extending the usable optical fiber spectrum over a contiguous win-
dow from 1200 to 1600 nm. Other advances occurred in microelectromechanical systems
(MEMS). These devices were the basis of new high port-count optical switches, tun-
able filters, and related devices. A wide variety of high-performance devices for WDM
9
Early demonstrations of fiber access date back to 1981 with FTTH by Northern Telecom and 1989 with
passive optical access networks by British Telecom.
 The Big Picture 21

based on guided-wave technology were also developed, including tunable lasers, fil-
ters, integrated switch fabrics, and optoelectronic subsystems such as integrated optical
receivers.
As the fundamental multiwavelength technology for transparent networks was ma-
turing, experimental work was (and is) continuing in more speculative areas entailing
nonlinear optical devices for opaque networks. Units such as all-optical switches, optical
logic devices, and optical storage elements are of interest to move the nonlinear opera-
tions now executed in electronics down to the optical level. (Among the enablers of these
devices are photonic crystal fibers [PCFs] whose structures can be tailored to produce
a variety of highly nonlinear effects.) Applications are in optical packet switching and
optical computing.
In the 1990s, ambitious optical network testbeds were deployed in the United States,
Europe, and Japan, involving the maturing multiwavelength technology as well as incor-
porating the management and control equipment necessary for making these networks
operational and reliable [IEEE93, IEEE96, IEEE98]. These testbeds showed for the first
time that optical technology could be taken out of the laboratory to produce cost-effective
operational networks. Many of the technologies and concepts developed in the testbeds
led to commercial products and network deployments that are basic components of our
current network infrastructure. Undoubtedly the rapid advances in the enabling technol-
ogy for optical networks, and its accelerated commercialization at the end of the 1990s,
can be largely attributed to the massive infusion of capital to the various players during
the “technology bubble” at that time. This produced a host of new start-ups and spin-offs
as well as expansions of ongoing activities in the large equipment manufacturers.
Although many of the companies that originated the new products have disappeared, the
technological progress remains, and will serve as a foundation for the networks of the
future.
The period of economic consolidation following the bursting of the bubble brought
with it a more down-to-earth view of networking, essentially focusing on the areas of
prime concern to the network operator and the customer: cost-effectiveness, reliability,
fault recovery, and manageability. Focus on costs has led to a trend toward eliminating
superfluous electronic and electro-optic components in networks (e.g., the use of IP
over WDM without intervening logical layers) and improving operating efficiencies
through enhanced traffic grooming and fault recovery techniques. Focus on cost has
also produced a concentration on bread-and-butter issues such as economic designs in
MANs, in contrast to the “hero experiments” in the 1980s, where the objective was to
break records for long-haul transmission – at any cost.
Focus on manageability has resulted in solid advances in control and management,
including taking control techniques designed for the logical layers of the network and
adapting them to the physical layer. As an example, the control protocol known as
Generalized Multiprotocol Label Switching (GMPLS) has been proposed for application
to control functions in the physical layer. As its name implies, GMPLS is a generalization
of Multiprotocol Label Switching (MPLS), which was designed as an improvement over
the packet-forwarding techniques used in IP networks.
22 Multiwavelength Optical Networks

This recent activity provides the context for this book: an advancing and maturing
technology base, a steady increase in demand for network capacity and performance,
and an industrial base that has positioned itself to meet these demands.

1.7 Overview and Road Map

Lightwave networks can be characterized broadly in terms of three basic features:

(1) physical (fiber) topology, (2) functionality in the links, the network nodes, and
the access stations, and (3) control algorithms for assigning, routing, and multiplexing
connections.
In keeping with the focus of this book on the physical layer, we elaborate on these
issues here, emphasizing how they apply to the optical infrastructure. A rich physical
topology provides many alternate paths among access stations, increasing the aggregate
capacity of the network as well as its potential survivability and adaptability to changing
load patterns. But the properties of the physical topology cannot be exploited without
sufficient functionality in the links, the network nodes, and the access stations. Also,
logical network overlays are generally required to adapt the bandwidth offered by the
physical layer to the needs of the end users.
By link functionality we mean good transmission properties (large bandwidth-
distance product). Useful functional properties of nodes and stations include controllable
switching and multiplexing features. Without controllability in the nodes and stations,
optical channel assignments and signal paths must remain fixed at all times so that
connections are frozen, and the network has no flexibility in responding to changing
conditions. Conversely, a high degree of node and station controllability under the
supervision of a network control and management system improves the efficiency of
resource utilization, allows the network to maintain satisfactory performance in the face
of fluctuating demand, and enables it to reconfigure itself in case of component failures.
Of course, controllability implies the existence of suitable control algorithms to coor-
dinate the functions of the various network entities. Three basic features – topology,
functionality, and control – interact closely to influence overall network performance.
As might be expected, there are many opportunities for cost–performance trade-offs.
Thus, high functionality in the nodes and access stations improves performance, but this
comes at a price. The same can be said of the richness of the physical topology. Also,
to compensate for the limitations of the network resources, one can attempt to optimize
performance through sophisticated control algorithms. However, optimality generally
comes at the price of controller complexity. Thus, the cost–performance trade-offs
involve complex interactions among all the basic features.
In a field of engineering where the technology is mature (e.g., in digital electronics),
systems can be understood, analyzed, and designed with only a limited understanding of
the physical principles involved. (A designer of a personal computer is not particularly
concerned with electromagnetic theory.) However, when the enabling technology is
rapidly evolving, as in the case of optical networks, a more thorough understanding of
technology and its relation to system performance is required. Thus, a good grasp of
 The Big Picture 23

optical networks requires an understanding of the interrelations between two bodies of

knowledge that traditionally have been treated separately: the physics of the underlying
devices and the mathematical methodology required to analyze, design, and control
systems incorporating these devices.
The emphasis of this book is on methodology rather than devices. However, we
weave the physical and mathematical sides of networking into an integrated whole by
linking physical constraints with performance analysis and design concepts whenever
possible. In addition, we attempt to integrate current practice, generic models, and
futuristic concepts. By emphasizing linkages across traditional lines, our intention is
to break down the compartmentalization that tends to hinder progress. We recognize
that integration across a broad range of material presents a challenge to the reader. The
interconnections make it difficult to isolate sets of topics matched to the background and
interests of each reader, and some readers may feel caught in a tangled web. To help the
disoriented traveler, we provide a simple road map in Figure 1.6 as a guide through the
labyrinth.
As shown in the figure, the first three chapters are required for an understanding
of the rest of the subject matter. These are accessible to readers with only a limited
background in networking and physical principles. After the broad view presented in
this chapter, Chapter 2 introduces the multiwavelength network architecture, describing
the layers of connectivity in a wavelength division multiplexed network. The chapter
focuses on the functionality of optical network elements – links, nodes, and access
stations – and their relation to network performance. Chapter 3 gives an overview of
the various layers of network connections. Purely optical networks are discussed first,
starting from the simplest (static) networks and then considering the two controllable
classes: wavelength-routed networks and linear lightwave networks.
The former class supports point-to-point connections, whereas the latter supports
multipoint connections, representing a more general view of transparent optical networks
and their functionality. Because the physical layer alone is generally not sufficient to
serve the needs of network users, the chapter concludes with a discussion of logically
routed networks – the multilayered networks of Figure 1.1, consisting of electronic
overlays on an optical infrastructure. Chapters 2 and 3 are largely qualitative and serve
as introductions and “pointers” to material explored quantitatively in later chapters.
The rest of the book may be read more selectively. It is linked to the earlier material,
as shown in Figure 1.6. For those with only a limited understanding of the physical
side of networking, Chapter 4 provides a concise treatment of physical principles and
device technology. It is not intended to be all-encompassing, because other works are
solely devoted to these topics; for example, see references [Agrawal02, Saleh+07].
The material in the first part of the chapter (through Section 4.9) focuses mainly on
fundamentals and generic concepts. The latter part of the chapter is devoted to current
technology with a focus on physical and technological constraints that limit network
performance. These include limits on WDM channel-packing density, optical receiver
performance, geographic reach of optical connections, and technological limitations
imposed on optical switches. The chapter concludes with methodologies for performance
evaluation.
24 Multiwavelength Optical Networks

Chapter 1
The Big
Picture

Chapter 2
The Layered
Architecture
and Its
Components

Chapter 3
Network
Connections

Chapter 6 Chapter 7
Chapter 4 Chapter 5
Wavelength/ Logically
Enabling Static Multipoint
Waveband Routed
Technology Networks
Routed Networks Networks

Chapter 8 Chapter 10 Chapter 11

Chapter 9
Survivability: Optical Packet- Current Trends
Optical
Protection Switched in Multiwavelength
Control Plane
and Restoration Networks Optical Networking

Figure 1.6 Road map.

The remaining chapters are more specialized. Chapters 5 through 7 present a thorough
treatment of the generic multiwavelength optical network.10 Static networks based on
shared optical media are covered in Chapter 5. The chapter discusses multiplexing and
multiple-access techniques, traffic flow constraints, capacity assignment, packet switch-
ing in the optical layer, and access network applications of passive optical networks.
Wavelength-routed and linear lightwave networks are examined in Chapter 6, with dis-
cussions of routing and channel assignment, as well as the relationship between optical
switch functionality and network performance. Chapter 7 deals with logically routed
networks. It is here that the overall multilayered network design problem appears for the
first time. We present methodologies for designing a logically routed network to satisfy
a prescribed traffic requirement while observing the constraints imposed by a given fiber
topology, the limitations of the network components, and the limited capacity of the
available optical spectrum.

10
The necessary background in graph theory, Markov chains, and queueing theory is included in Appendices A
and C, and some algorithms for special aspects of network switching, provisioning, and control appear in
Appendices B, D, E, and G.
 The Big Picture 25

Chapter 8 considers the very important issue of optical network survivability and
fault recovery, specifically ring and mesh topologies and line and path-based recovery
techniques. Chapter 9 examines the issue of network control, focusing on the most recent
proposals for the optical network control plane, based on GMPLS. Chapter 10 is the
most forward-looking part of the book, dealing with recent progress in optical packet,
burst, and label switching. We conclude in Chapter 11, tying the generic concepts of
earlier chapters to recent trends in network deployment.
An infinite variety of additional engineering issues arise when operating networks
in the real world, and these are often missed in an abstract view of things. Parts of
Chapters 4 and 7, the opening part of Chapter 8, and all of Chapters 9 and 11 provide
examples of contemporary technology, network design, and network operation, as well
as the trade-offs between optical and nonoptical networking solutions. These will be of
particular interest to those involved in near-term network deployment. However, this is
the most “perishable” material in the book. For example, Chapter 9 is important for an
understanding of current standardization efforts in network control. However, these are
continuing to evolve as this is being written, so that techniques of optical network control
and management can be expected to change and progress in future network deployments.
Some of the more advanced and speculative sections in the book may be skipped
initially by readers learning about optical networks for the first time. In Chapter 5,
Section 5.2.3, Code Division Multiple Access, and parts of Sections 5.5 and 5.6, which
deal with demand-assigned connections and packet switching in the optical layer, can be
bypassed by those new to the field. The same is true for parts of Chapter 6, including the
material on ring decomposition in Section 6.3.5, optimization in Section 6.3.7, and some
of the more specialized topics on LLNs in the latter part of the chapter. In Chapter 7,
most of the material on point-to-point logically routed networks will be of interest to all
readers, whereas Sections 7.4 and 7.5 on hypernets are more futuristic in nature than the
rest of the chapter. The latter sections of Chapter 8, on optical layer protection in mesh
topologies, are currently important open research areas, and the same can be said of all
of Chapter 10.
Readers involved in research on next-generation networks would normally focus on
the complement of the subject matter of interest to the novice. Thus, for example, rings,
which are “old hat” to the researcher, could be skimmed in favor of more advanced
topics; for example, optimization and LLNs in Chapter 6. Researchers might also focus
on hypernets in Chapter 7, general optical layer protection in Chapter 8, and the full
range of optical packet-switching issues in Chapter 10. In summary, the reader is invited
to customize an itinerary suited to his or her interests in exploring the fascinating world
of optical networking. We hope you enjoy the journey.

Bibliography

[Acampora87] A. S. Acampora. A multichannel multihop local lightwave network. In Proceedings

of the IEEE Global Telecommunications Conference (Globecom), pp. 1459–1467, Tokyo, Japan,
November 1987.
26 Multiwavelength Optical Networks

[Agrawal02] G. P. Agrawal. Fiber-Optic Communication Systems, Third Edition (Wiley Series in

Microwave and Optical Engineering). New York: Wiley 2002.
[Bala+95] K. Bala, R. R. Cordell, and E. L. Goldstein. The case for opaque multiwavelength
optical networks. In Proceedings of the IEEE/LEOS Summer Topical Meeting, Keystone, CO,
August 1995.
[Ballart+89] R. Ballart and Y.-C. Ching. SONET: Now it’s the standard optical network. IEEE
Communications Mag., 27(3):8–15, 1989.
[Boehm90] R. J. Boehm. Progress in standardization of SONET. IEEE Lightwave Communication
Systems, 1(2):8–16, 1990.
[CCITT92] CCITT. Recommendation i.121, Broadband Aspects of ISDN. Geneva: CCITT,
1992.
[Charlet+04] G. Charlet, E. Corel, J. Lazaro, A. Klecamp, R. Dischler, P. Tran, W. Idler,
H. Mardoyan, A. Konczykowska, F. Jorge, and S. Bige. WDM transmission at 6 Tbps ca-
pacity over transatlantic distance, using 42.7 Gbps Differential Phase-Shift Keying without
pulse carver. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., PDP 36, Los
Angeles, CA, February 2004.
[de Prycker91] M. de Prycker. Asynchronous Transfer Mode: Solution for Broadband ISDN.
Chichester, UK: Ellis Horwood, 1991.
[Desurvire+87] E. Desurvire, J. R. Simpson, and P. C. Becker. High-gain erbium-doped traveling-
wave fibre amplifier. Optics Lett., 12(11):888, 1987.
[Goodman+86] M. S. Goodman, H. Kobrinski, and K. W. Lo. Application of wavelength division
multiplexing to communication network architectures. In Proceedings of the IEEE Int’l Conf.
Commun. (ICC), pp. 931–934, Toronto, Canada, 1986.
[Goodman+87] M. S. Goodman, C. A. Brackett, R. M. Bulley, C. N. Lo, H. Kobrinski, and M. P.
Vecchi. Design and demonstration of the LAMBDANET system: A multi-wavelength optical
network. In Proceedings of the IEEE Global Telecommunications Conference (Globecom),
pp. 1455–1458, Tokyo, Japan, November 1987.
[Henry89] P. S. Henry. High-capacity lightwave local area networks. IEEE Communications
Mag., 27(10):20–26, 1989.
[Hill88] G. R. Hill. A wavelength routing approach to optical communications networks. In
Proceedings of the IEEE Conference on Computer Communications (Infocom), pp. 354–362,
New Orleans, LA, March 1988.
[IEEE93] IEEE. Special issue on broad-band optical networks. IEEE/OSA J. Lightwave Technol.,
11(5/6), 1993.
[IEEE96] IEEE. Special issue on multiwavelength optical technology and networks. IEEE/OSA
J. Lightwave Technol., 14(6), 1996.
[IEEE98] IEEE. High-capacity optical transport networks. IEEE J. Select. Areas Commun., 16(7),
1998.
[Kao+66] C. K. Kao and G. A. Hockham. Dielectric-fiber surface waveguides for optical fre-
quencies. Proc. IEEE, 113(7):1151–1158, 1966.
[Mears+86] R. J. Mears, L. Reekie, S. B. Poole, and D. N. Payne. Low-threshold, tunable CW
and Q-switched fibre laser operating at 1.55 µm. Electron. Lett., 22(3):159, 1986.
[Mears+87] R. J. Mears, L. Reekie, I. M. Jauncey, and D. N. Payne. Low-noise erbium-doped
fibre amplifier operating at 1.54 µm. Electron. Lett., 23(19):1026, 1987.
[Miller+88] S. E. Miller and I. P. Kaminow (Editors). Optical Fiber Telecommunications II. New
York: Academic Press, 1988.
 The Big Picture 27

[Mukherjee92a] B. Mukherjee. WDM-based local lightwave networks—Part I: Single-hop sys-

tems. IEEE Network, 6(3):12–27, 1992.
[Mukherjee92b] B. Mukherjee. WDM-based local lightwave networks—Part II: Multihop sys-
tems. IEEE Network, 6(4):20–32, 1992.
[Nakazawa+93] M. Nakazawa, K. Suzuki, E. Yamada, H. Kubota, Y. Kimura, and M. Takaya.
Experimental demonstration of soliton data transmission over unlimited distances with soliton
control in time and frequency domains. Electron. Lett., 29(9):729–730, 1993.
[Saito+90] S. Saito, T. Imai, T. Sugie, N. Ohkawa, Y. Ichi-hashi, and T. Ito. An over 2,200 km
coherent transmission experiment at 2.5 Gbps using erbium-doped fiber amplifiers. In Proceed-
ings of the IEEE/OSA Optical Fiber Commun. Conf., Paper PD2, San Francisco, CA, January
1990.
[Saleh+07] B. E. A. Saleh and M. C. Teich. Fundamentals of Photonics, Second Edition, (Wiley
Series in Pure and Applied Optics). New York: Wiley, 2007.
[Schawlow+58] A. L. Schawlow and C. H. Townes. Infrared and optical lasers. Phys. Rev.,
112(6):1940, 1958.
2 The Layered Architecture and
Its Resources

Ultimately, the performance of a network is limited by the quantity and functionality of

its physical resources. In this chapter we examine the various functions performed in a
multiwavelength network, emphasizing the role of the optical resources (located in the
physical layer of Figure 1.3) in providing connectivity and throughput. For the most part
we use the terms transparent optical, purely optical, and just optical interchangeably
to refer to entities in the physical layer. The implication is that there is a clean break
between the underlying technology and functionality in the physical layer and that
in the logical layer. The physical layer contains optical components executing linear
(transparent) operations on optical signals, whereas the logical layers contain electronic
components executing nonlinear operations on electrical signals. In reality, as mentioned
in Chapter 1, the picture in real networks is more nuanced. For example, some simple
signal processing (either electronic or optical) may be present in the physical layers
of today’s networks, making them “opaque” to a greater or lesser degree. Conversely,
as optical technology for signal processing matures, it is beginning to make its way
into the logical layers. Nevertheless, the somewhat simplified view of a transparent
(linear) optical layer underlying an electronic (nonlinear) logical layer is very helpful in
providing a generic model for most multiwavelength networks. It will be used throughout
this book, with exceptions duly noted as they appear. To provide a proper framework
for the discussion that follows, we start in Section 2.1 with a description of layers and
sublayers of the multiwavelength network architecture.
The functional characteristics of the optical resources – the network links, the optical
network nodes (ONNs), and the network access stations (NASs) – are discussed in
Sections 2.2, 2.3, and 2.4, respectively. They provide the basic functionality for setting
up and routing optical connections through the network. We refer to a purely optical
network as a wavelength- or waveband-routed network (WRN). However, as pointed out
in Section 1.4, the additional functionality provided by an electronic overlay is often
required to satisfy network requirements. Thus, in Sections 2.5 and 2.6 we consider
hybrid networks that contain a purely optical network infrastructure complemented by
an electronic overlay. Section 2.5 deals with special-purpose overlay processors (OLPs)
for enhancing the performance of the optical network while maintaining many of the
desirable features of transparency.
In Section 2.6 we add logical switching nodes (LSNs) to create logically routed
networks (LRNs). The LRN relies on the resources of the physical layer to provide high
 Layered Architecture 29

throughput, and the processing power and routing capabilities in the LSN to provide
high connectivity.

2.1 Layers and Sublayers

In Figure 2.1(a) the layered view of the optical network, first introduced in Figure 1.3, is
expanded into the sublayers of the multiwavelength network architecture. To fix ideas,
think of the higher layers as representing a packet/cell-switched client logical network
(LN), including a virtual connection layer, which might correspond to the IP or ATM
logical network, shown in Figure 1.1. The characteristics of each logical layer and its
sublayers are determined by the specific structure of that client network (e.g., SONET,
ATM, and IP). The physical layer acts as a server for one or more LNs. For example,
in the case of Figure 1.1 the optical layer has two LN clients: the IP network and the
stacked IP over ATM over SONET network.
The multiwavelength network architecture to be described is representative of a
WAN operating with dedicated and/or demand-assigned optical connections providing a

Virtual Connection
Logical Logical Path
Layer Logical Connection
Transmission Channel
Optical Optical Connection
Physical Layer λ -Channel
Layer Optical/Waveband Path
Fiber Fiber Link
Layer Fiber Section

Sublayers
(a)

NAS Access NAS

eo Link Network Link oe

TP OT OA OR RP
ONN ONN
RP OR OT TP
Fiber Section
Fiber Link
Optical/Waveband Path
λ-Channel
Optical Connection
Transmission Channel
Logical Connection
(b)

Figure 2.1 Layered view of optical network connections.

30 Multiwavelength Optical Networks

common infrastructure for its client LNs. In smaller or simpler networks (e.g., optical
access networks) some of the functions described may be nonexistent.
In describing the architecture we start in the optical layer of Figure 2.1(a) and work
down from the interface between the logical and physical layers to the fiber layer. In terms
of hardware, the logical–physical layer interface is located at the external (electronic)
ports of the NASs. A detailed look at a typical point-to-point connection is presented
in Figure 2.1(b). As indicated, the access stations terminate the optical connections
and serve to interface electronic end systems (typically user equipment) or electronic
switches (typically communications carrier equipment) to the optical network.1 (As
mentioned, the access station acts as the boundary between the electronic and optical
domain, moving information between its external ports on one side of the electro-optic
interface and optical ports on the other.) On its optical side, each NAS connects to
an ONN through an access link, which consists of one or more pairs of access fibers.2
Optical signals are exchanged between the station and the network by optical transceivers
in the station. Each node is connected to neighboring nodes by pairs of fibers, which
constitute internodal network links. We call the graph that describes the interconnections
of the network links and network nodes the physical topology of the network. A long
link may contain one or more optical amplifiers (OAs) to compensate for attenuation in
the fiber, so that the link becomes a series connection of several fiber sections between
amplifiers. For simplicity we usually assume that fibers are used in unidirectional pairs
to support bidirectional transmission. (In practice, single fibers are sometimes used for
bidirectional transmission to reduce cost.)
The physical layer of our multiwavelength network architecture is designed to capture
the essential generic features of optical networks for purposes of design, modeling,
performance evaluation, and control. It is divided into an optical layer, encompassing
the various aspects of the optical connection: signal generation and transmission, routing,
survivability, and so on, and a fiber layer, comprising the aspects of the supporting fiber
infrastructure.
The optical connection provides an end-to-end optical communication channel ter-
minating on an optical transmitter/receiver pair. The characteristics of the optical layer
in Figure 2.1 are a function of the way in which the optical spectrum is partitioned.
The smallest entities in the partition are λ-channels, each of which is assigned a dis-
tinct wavelength (or optical frequency). These are the basic information carriers in the
physical layer. To keep things simple at this point, the λ-channels are assumed to be
routed (switched) independently by the ONNs. (This is the way most multiwavelength
networks are configured at present.) A more efficient two-tiered partitioning involving
wavebands is introduced in Section 2.2.
Each point-to-point optical connection (OC) is carried on a λ-channel and is created
by (1) assigning a λ-channel to the source transmitter and destination receiver on a

1
In cases when the switch technology is electronic but it has little more signal processing functionality than
a purely optical switch, we sometimes make an exception and lump the switch into the optical layer (with
an appropriate explanation).
2
For survivability, a station might connect to two (or more) network nodes, a configuration called dual homing
in current telecommunication networks. In dual homing a facility, such as a telephone Central Office, is
connected to two hubs (nodes) of a backbone network.
 Layered Architecture 31

selected wavelength and (2) by establishing an optical path (OP) through a sequence of
network nodes to carry that wavelength from source to destination. These actions are
normally under the control of a network manager, the details of which we leave for later.
We call a unidirectional connection between external ports on a pair of source and
destination NASs a logical connection (LC). This connection carries a logical signal
in some agreed-on electronic format (e.g., ATM cells, IP packets, SONET digital bit
streams, or analog video). Each logical signal format is tailored to the needs of a particular
client LN or end system. All LCs draw on the resources of the physical layer, as shown
in Figure 1.1. Each LC is carried on an optical connection through the intermediary
of a transmission channel. The transmission channel performs an adaptation function,
converting the logical signal to a transmission signal. This conversion (as well as certain
multiplexing functions) takes place in the transmission processor (TP), with the reverse
operations implemented in the reception processor (RP). Just as the logical signal
format must be adapted to the LN that it serves, the transmission signal format must be
matched to the requirements of the optical equipment that carries it. For example, the
transmission signal must be chosen such that it, and the optical signal that carries it,
satisfies the bandwidth constraints of the optical transmitter (OT) and the optical receiver
(OR), as well as the permissible spectrum occupancy of the optical channel carrying it.
The optical connections are established through the coordinated actions of the nodes
and stations. It is the function of the terminating NASs to tune their transceivers to
the assigned wavelength of the λ-channel, thereby establishing the optical connection.
Similarly, it is the role of the nodes to create an optical path on that wavelength, thereby
establishing the support for a potential connection. Activation of a connection on the
OP only occurs when the transceivers are properly tuned and the path is “lit.”
The main consideration in breaking down the optical layer into sublayers is to account
for multiplexing, multiple access (at several layers), and switching. Using multiplexing,
several logical channels may be combined on a single λ-channel originating from one
station; using multiple access, λ-channels originating from several stations and operating
on the same wavelength may carry multiple LCs to the same destination; and through
switching, many distinct optical paths may be created on different fibers in the network,
using (and reusing) λ-channels on the same wavelength.3
In the fiber layer, the optical path is carried by a succession of fiber links that connect
the network nodes. The fiber layer is subdivided further into the fiber link sublayer
supported by the fiber section sublayer. Amplification may occur within the fiber layer
either in lumped or distributed form (see Section 4.4). Each of the internodal connections
along an optical path constitutes an optical hop.
A typical connection illustrating these concepts is shown in Figure 2.2. A virtual conn-
ection (VC) between a pair of end systems is carried on a logical path (LP) that consists
of two LCs (two logical hops) terminated by LSNs. The first and second LCs are carried
on two- and three-hop OCs, respectively. (The access links are not counted as hops.)
Although this discussion has been limited to point-to-point connections to simplify
the exposition, the terminology also applies to multipoint connections. As a matter of

3
Optical spectrum reuse is a recurring theme throughout the book because it is the key to scalability of optical
networks to large sizes.
32 Multiwavelength Optical Networks

VC

LP

LC LC

OC OC

VC Virtual Connection Network Access Station

LP Logical Path Optical Network Node
Optical Amplifier End System
Logical Switching Node

Figure 2.2 A typical connection.

fact, optical multipoint connections play a unique and important role in the network
architectures to be discussed later.
At this point, a brief comparison with another layered view of the optical network is in
order. Figure 2.3 is a schematic portrayal of ITU-T recommendation G.872 [ITU-T01]
for optical transport networks. It shows three layers: the optical channel (OCh) layer, the
optical multiplex section (OMS) layer, and the optical transmission section (OTS) layer.
(Another layer, the physical media layer [PM] dealing with the properties of fibers, not
shown in the figure, is the server for the OTS layer.) As in all ITU-T network standards,
the focus is on establishing agreed-on requirements for operational networks. This is in
contrast to the objectives of our layered architecture, which is designed not as a standard
but rather as a way of thinking about networks. Because the G.872 standard is directed
toward facilitating interworking of operational networks, it pays special attention to
means for monitoring and ensuring integrity of the connections in each layer as well as
supervisory functions for network management.
To decode the figure, some explanation of the nomenclature used in ITU networking
standards is needed [ITU-T00]. In each layer, an adaption processing function (e.g.,
OCh A Source) accepts information from the client layer above it, and processes it
to allow transfer over a trail in its layer network, in this case the OCh layer network.
(The OCh A Sink executes the reverse processing function.) The trail is a bidirec-
tional “transport entity,” an abstraction that represents transfer of information between
its termination points in the indicated (OCh) layer. The access points (AP) are the
reference points marking the borderline between a trail and its associated adaptation
functions. The next processing function encountered in descending the OCh layer is the
 Layered Architecture 33

OCh_A OCh_A
source sink

OCh Trail
OCh AP OCh AP
OCh_TT
OCh_TT
OCh layer source
sink
network
OCh_NC
OCh TCP OCh TCP

OMS/OCh_A OMS/OCh_A
source sink

OMS Trail
OMS AP OMS AP
OMS layer OMS_TT OMS_TT
source sink
network

OMS_NC
OMS TCP OMS TCP
OTS/OMS_A OTS/OMS_A
source sink

OTS Trail
OTS AP OTS AP
OMS_TT
OMS_TT
source
OTS layer sink
network
OTS_NC
OTS TCP OTS TCP

Figure 2.3 Client server associations in an optical transport network (From [ITU-T01, Figure 1].
Copyright
c 2001 ITU. Reproduced with the kind permission of The International Telecommu-
nications Union.)

OCh trail termination (TT) processing function, which completes the processing neces-
sary to present information (including added overhead) to the layer below, in this case
the OMS layer network. The OCh network connection (NC) represents a series of links
joining the two ends of the OCh trail, and the OCh terminal connection points (TCP)
bind the TT source and sink to the bidirectional connection.
The OCh, OMS, and OTS layers combined correspond roughly to the optical layer
in Figure 2.1(a). (The PM layer corresponds to the fiber layer.) They provide for an
end-to-end optical connection, including routing (OCh layer), λ-channel multiplexing
(OMS layer), transmission parameters (OTS), and the various overhead and supervisory
functions required for network operations and management. Except for the position of
the routing and transmission functions, there is an approximate correspondence between
these layers and individual layers in Figure 2.1(a): OTS/OCh, OMS, and PM correspond
to transmission channel/optical connection/λ-channel, optical/waveband path, and fiber
link, respectively. The reason for the different positioning of the routing and transmission
functions in the two architectures is that the G.872 standard is focused on an opaque
network, where optical connections are terminated at the ends of each transmission
link, meaning that routing is a higher layer function and transmission is a lower layer
function than in Figure 2.1(a). In a transparent network represented by the physical layer
of Figure 2.1(a), those positions are reversed.
34 Multiwavelength Optical Networks

A layered architecture is important because it helps to deal with the complexity of

a network by breaking it down into separate parts. However, it is nothing more than
a framework, which often has to be bent into shape and changed to adapt to specific
features that continue to evolve in optical networks. Now let us examine each of the key
components of the network in more detail.

2.2 Network Links: Spectrum Partitioning

A large number of concurrent connections can be supported on each network link

through successive levels of multiplexing as discussed earlier. A schematic view of the
fiber resources illustrating this concept appears in Figure 2.4.
A typical network link consists of a cable containing several (sometimes more than
100) fibers, which are used as bidirectional pairs. (Given the cost of the right-of-way and
labor in laying a cable, it would be very inefficient to deploy a cable that consists of only
a single fiber pair.) This bundling of fibers into one cable is an example of space division
multiplexing in the fiber layer. The usable optical spectrum on a single fiber is wide
enough to carry many high-speed optical connections through several additional levels of
multiplexing. In our discussions, wavelength division is the basic multiplexing technique,
with each fiber carrying connections on many distinct wavelengths (λ-channels). This is
indicated by a coarse and fine division of the optical fiber spectrum within the wavelength
dimension in Figure 2.4. (An explanation of the coarse/fine spectrum division is given
later.) First, the spectrum is divided into wavebands, and these are divided further into
λ-channels. The assigned wavelengths of the λ-channels must be spaced sufficiently far
apart to keep neighboring signal spectra from overlapping. (Overlapping signal spectra
cause interchannel cross-talk; in other words, interference among the signals at the

Network Link

Space Wavelength Time

{ Transmission Channels Out

λ-Channels Out
Cable Fibers
Wavebands

λ-Channels In

{ Transmission Channels In

Figure 2.4 Fiber resources.

 Layered Architecture 35

Usable spectrum
λ1 λ2 λm

f λ
10 GHz/
GHz/nm
.08 nm
(a) λ-channel Spacing for Separability at Receivers

λ1 λ2 λm

f λ
100 GHz/0.8 nm
(b) λ-channel Spacing for Separability at Network Nodes

10 GHz

λ1,1 λ2,1 λ10,1

100 GHz/0.8 nm

100 GHz/
w1 0.8 nm w2 wm
f λ
100 GHz/
.08 nm
(c) Wavebands

Figure 2.5 Wavelength and waveband partitioning of the optical spectrum.

optical receiver.) Figure 2.5(a) shows an example of a WDM wavelength assignment, in

which the λ-channels are spaced 10 GHz apart in optical frequency.
Note that an optical wavelength scale is also shown, because wavelength is often used
in specifying optical components. Because wavelength λ and frequency f are related by
f λ = c, where c is the velocity of light in the medium, we have the relation
(cλ)
f ≈ − 2 (2.1)
λ
between small changes in frequency  f and wavelength λ. This gives a correspon-
dence between wavelength and frequency intervals of 100 GHz ≈ 0.8 nm in the range of
1550 nm, where most modern lightwave networks operate.4 One of the facilitating forces
behind commercial development of WDM has been standardization by the ITU of an op-
tical wavelength grid and a set of transmission bands covering a range of 1260–1675 nm
(see Section 4.3). A grid laying out alternatives for close wavelength/frequency spacing,
dubbed dense wavelength division multiplexing was established in 2002 [ITU-T02a].
The DWDM grid is centered at 193.1 THz (1552.52 nm) and specifies a variety of reg-
ular spacings from 100 to 12.5 GHz. Recognizing the need for much larger wavelength
spacing to accommodate wavelength drift when inexpensive components are used, the
coarse wavelength division multiplexing standard was adopted for metropolitan area
4
This is the region of minimum attenuation in an optical fiber.
36 Multiwavelength Optical Networks

networks. It specifies 18 wavelengths with a 20-nm spacing over the 1270–1610 nm

wavelength range [ITU-T02b].
Returning to our example of wavelength assignment, the total bandwidth of each
channel (i.e., the optical bandwidth occupied by the modulated laser signal) is shown
(see Figure 2.5, shaded areas) to be somewhat less than the frequency spacing to leave
sufficient “guard bands” to allow for imprecision and drift in laser transmitter tuning
and to make it possible to separate adjacent signals at the receivers with reasonably
simple receiving equipment (see Chapter 4). The 10-GHz channel spacing shown in
this example is sufficient to easily accommodate λ-channels carrying aggregate digital
bit rates on the order of 1 Gbps or occupying analog bandwidths on the order of
1 GHz.5 Efficient use of channels with this much capacity frequently requires sharing
one λ-channel among several lower rate transmission channels. A common way of doing
this is by time division multiplexing/multiple access techniques, indicated by the time
dimension in Figure 2.4 (see Section 3.2.1). As mentioned in Chapter 1, the grooming
operation of multiplexing several lower rate transmission channels onto a higher capacity
λ-channel for achieving efficient use of the channel plays an important role in network
management (see Section 7.1). The 10-GHz channel spacing shown in Figure 2.5(a) is
very dense by current standards, but it would be feasible for accommodating many high-
capacity λ-channels on a single network link, given the characteristics of typical optical
transceivers. However, we must also anticipate the limitations of the ONNs because
they have the job of routing each optical connection on a specified path through the
network. Recall that in transparent networks containing optical switching nodes, several
λ-channels may be multiplexed on each fiber, with each routed selectively on a different
path based on its wavelength. To accomplish this the nodes must have the capability
to “recognize” each λ-channel on an incoming fiber independently and to direct it to
a specified output fiber. This is referred to as wavelength-selective routing or simply
wavelength routing.
Now the characteristics of current optical node technology (see Chapter 4) suggest
that the relatively close channel spacings that can be resolved by optical receivers (e.g.,
10 GHz) are too close to permit independent wavelength routing of each channel at
the nodes of a transparent optical network.6 In fact, in today’s switched multiwavelength
networks the λ-channel spacings are typically more than 100 GHz. Thus, there is an order
of magnitude mismatch between the spacings required in the λ-channel layer and the
optical path layer. A wider spacing of the λ-channels to 100 GHz, permitting individual
wavelengths to be routed independently, is shown in Figure 2.5(b). Note that this reduces
the channel density and hence the fiber throughput by a factor of 10.
Such a loss of throughput is certainly to be avoided if possible. A brute force (and
costly) way of solving this problem would be to refine the switching node technology
to produce an order of magnitude improvement in its resolution in the transparent case,

5
In the digital case, this represents a modulation efficiency of 0.1 bps/Hz, which is poor for electronic systems,
but typical for optical systems.
6
One reason for this is that many nodes may be traversed on a long optical path. At each node the imperfections
in signal resolution (optical filtering) for closely spaced channels result in signal attenuation, distortion, and
switch cross-talk, which accumulate along the path in a transparent network.
 Layered Architecture 37

or to drop transparency altogether, using an “opaque” switch (see Section 2.3.2.4). A

less costly approach, and one that partially circumvents the problem, is shown in the
two-tiered spectrum partition in Figure 2.5(c). Here we define a waveband as the smallest
segment of the spectrum that is optically resolvable in the node (i.e., in the optical path),
in contrast to the λ-channel, which is the smallest unit resolvable in the access station
by a tunable receiver.7 Below the plot of λ-channels in Figure 2.5(c) is another plot
showing a coarser partition of the spectrum into wavebands. In this example, we show
wavebands with a width of 100 GHz placed at intervals of 200 GHz (i.e., separated
by guard bands of 100 GHz). Wavebands with this width and separation are easily
resolvable by waveband-selective optical nodes without pushing the current state of
the art; that is, each waveband can be recognized and switched (routed) independently
at each network node without introducing excessive signal impairment. Each of these
wavebands has enough capacity to support many individual λ-channels. (Ten are shown
in the figure.) This second partition of the spectrum requires positioning of wavebands
in the optical layer as shown in Figure 2.1(a). Now the waveband carries an optical
path and the optical/waveband path may carry several λ-channels. Several wavebands
may now be multiplexed on each fiber (waveband multiplexing) and several λ-channels
may be multiplexed on each waveband (wavelength multiplexing). The resultant space
waveband and wavelength multiplexing is indicated schematically in Figure 2.4.
In subsequent chapters, we discuss both the one-tiered and two-tiered approach to
spectrum partitioning. Clearly, the partitions of Figure 2.5(a) and 2.5(b) can be consid-
ered special cases of Figure 2.5(c). If the sparse spacing of Figure 2.5(b) is used to allow
each λ-channel to be switched individually, this corresponds to an m-waveband system
with one λ-channel per waveband. If the dense spacing of Figure 2.5(a) is used to achieve
higher throughput, with channels spaced so close that they cannot be switched individ-
ually, then we have a single-waveband network. (Of course, the distinction between
the two cases depends on the characteristics of the network nodes. If the nodes have
no waveband selectivity then any spectrum spacing is equivalent to a single-waveband
partition.) To simplify the terminology, from now on we refer to networks using the
spacing of Figure 2.5(b) as wavelength-routed networks, and those using the spacing
of Figure 2.5(c) as waveband-routed networks. A more graphic depiction of the three
alternatives in spectrum partitioning and their influence on network operation is shown
in Figure 2.6.
A multiwavelength network with λ-channels on m wavelengths, λ1 , λ2 , . . . , λm , spaced
far enough apart to be switched independently (a wavelength-routed network), can be
envisioned as m copies of one network, each with the same physical topology, as shown
in Figure 2.6(b). An optical connection between a pair of stations runs on an optical
path laid out on the copy of the network corresponding to the connection’s assigned
λ-channel. The layers of the figure illustrate nicely the concept of wavelength continuity:
A signal generated at a given wavelength in a transparent network must remain on that

7
We assume a simple, direct detection receiver; in other words, a photodetector possibly preceded by a
tunable optical filter, which can select and detect the instantaneous power in any desired λ-channel (see
Section 2.4.2).
38 Multiwavelength Optical Networks

1, 2, m

(a) Single Waveband

1,1 10,1

w1 w1

2 w2 w2
1,2 10,2

m 1,m 10,m

wm wm

(b) Wavelength-routed (c) Waveband-routed

Figure 2.6 Network picture based on spectrum partitioning.

wavelength (i.e., on the corresponding copy of the network) from source to destination.8
If the wavelength spacing is so close that wavelength-selective switching is impossible
(the single-waveband case), all λ-channels are forced to share a single copy of the
network; see Figure 2.6(a). Finally, if a two-tiered partitioning is used, then we have a
waveband-routed network, in which there are as many copies of the network as there are
wavebands, and several (in this case 10) λ-channels share each copy of the network, as
in Figure 2.6(c).
How do these approaches compare? Using the numbers in the example of Figure 2.5,
we find that the wavelength-routed case has a fiber throughput reduced by a factor
of 10 compared with the single-waveband case. This is the price paid for being able
to route each λ-channel independently. The waveband-routed case is a compromise
between the two extremes of complete wavelength routing and none at all. The exam-
ple of Figure 2.5(c) uses 10 λ-channels in each waveband for an aggregate capacity
five times that of the wavelength-routed case and one half of the single-waveband case.
(The same effect would be achieved if the 10 λ-channels were replaced by a single
very-high-speed λ-channel operating at 10 times the speed of the original channels.)
As shown in Section 3.4, waveband routing imposes some special routing constraints
not present in wavelength-routed networks. The constraints are due to the fact that all
λ-channels sharing a common waveband are forced to “stick together” on the same
optical path.9 These partially offset the throughput gains produced by waveband routing.

8
An exception to this occurs in networks containing wavelength interchangers (see Section 2.3.3).
9
The constraints on waveband routing of groups of λ-channels within the same waveband are identical to
those affecting wavelength routing of groups of time division-multiplexed channels groomed onto the same
wavelength.
 Layered Architecture 39

The coarse/fine (waveband/λ-channel) partition has the following advantages:

1. Regardless of the wavelength density on the fiber, it is less costly to switch/route the
optical signals as a small number of aggregated groups of channels, with each group
contained in a continguous waveband, rather than as a large number of individual
λ-channels. (The cost of a wavelength-selective switch typically increases at least
linearly with the number of segments of the spectrum it switches independently.)
2. In many cases (e.g., multipoint optical connections), it is actually desired to route
a “bundle” of λ-channels on a common path through the network – something
that is especially easy to do if these channels lie within a common waveband (see
Section 3.4.2).
3. Network management at the level of the optical nodes is simplified if there are a
small number wavebands to keep track of instead of a large number of λ-channels.

Given these considerations, a fundamental issue in network design is how a net-

work deals with the question of waveband versus wavelength routing. As pointed out
in Section 1.6, the first optical networks were broadcast-and-select, so that routing
was wavelength independent; i.e., the case of Figure 2.6(a). In the mid-1990s, a num-
ber of multiwavelength network testbeds were developed, based on wavelength routing
(Figure 2.6[b]), which typically used λ-channel spacings of several nanometers. How-
ever, these testbeds were driven more by proof-of-concept considerations rather than
economic ones. As commercial deployment began at the turn of the millennium, eco-
nomic considerations pointed toward waveband routing (Figure 2.6[c]), and at this writ-
ing commercial products exist that implement waveband routing, wavelength routing,
and hybrid waveband/wavelength routing. It should be noted that whenever λ-channels
are bundled together for the purpose of waveband routing, it is important from the
point of view of efficiency to “fill up” each waveband with the maximum number of
constituent λ-channels; this constitutes grooming at the wavelength/waveband level.

2.3 Optical Network Nodes: Routing, Switching, and Wavelength Conversion

The functions of the optical path layer (establishing transparent paths for the optical sig-
nals) are implemented in the ONNs. Therefore, the more functionality these nodes have,
the more flexible is the network in reacting to fluctuating user demand, changing loads,
and equipment problems, and the better it will perform under all types of conditions.
One way of viewing functionality is in terms of the way various degrees of freedom (or
dimensions) are exploited. Most networks make use of the space dimension by assigning
different signals to different fibers. Through multiplexing, multiwavelength networks
make use of the wavelength dimension and perhaps the time dimension if time divi-
sion multiplexing is used. However, the benefit obtained from the various dimensions
depends on what happens to the multiplexed signals as they pass through the network
nodes. In this section, we examine the basic types of node functionality in transparent
networks, classified in increasing order of complexity as follows:
40 Multiwavelength Optical Networks

r Static nodes
– Directional couplers (including fixed signal combiners, splitters, and star couplers)
– Static routers (including waveband multiplexers and demultiplexers)
r Dynamic (switching) nodes
– Permutation switches
– Generalized switches
– Linear divider-combiners (LDCs)
– Waveband-space switches
– Wavelength-interchanging switches

Given the multiplicity of options and terminology in currently deployed networks, we

need to simplify our language in discussing generic network concepts. Therefore, we
will usually use the term waveband when discussing components operating on either a
waveband or wavelength basis (because a wavelength is a special case of a waveband,
which carries one λ-channel). However, there are times when our generic terminology
may conflict with terms widely accepted in the commercial world, in which case we
shall revert to the more commonly accepted term; for example, our waveband-space
switch is commonly known as a wavelength-selective cross-connect, whether it switches
wavelengths or wavebands or both.
It should be recalled that ONNs, as defined here, perform only transparent (linear)
operations on the optical signals. In networks possessing some degree of opacity, ad-
ditional nontransparent nodes may be present with nonlinearities (usually electronic)
in the signal path. These include network elements where optical signals are con-
verted to the electrical domain to facilitate switching, wavelength conversion, or signal
regeneration. These devices are also treated here and in Sections 2.5 and 2.6. The
network nodes as well as other network elements discussed in this chapter are pre-
sented in terms of idealized generic mathematical models. This is the approach that is
most useful in network analysis and design. Detailed discussions of practical realiza-
tions, physical characteristics, and performance impairments in these devices appear in
Chapter 4.

2.3.1 Static Nodes

Figure 2.7 illustrates three simple optical network topologies with nodes (shown as
circles) that are comprised solely of static devices. (These structures are suitable mainly
for LANs and MANs.) In each, the physical topology is in the form of a tree. This means
that no alternate paths are available for optical path routing, so that the role of the nodes
is relatively limited. In each case we assume that the node serves to combine and/or
divide (split) signal power without any wavelength selectivity, so that signals propagate
through the network in the direction of the arrows. (In the figure, each unidirectional
fiber is shown individually with an arrow that indicates its direction of propagation.)
Thus, in the broadcast star network of Figure 2.7(a), n access stations are joined by
access links to a single node, which is an n × n star coupler, combining all n inbound
signals and broadcasting them on each outbound access fiber.
 Layered Architecture 41

n 2

Star
4 Coupler 3

(a) Broadcast Star Network

Head
Bus 1 2 3 n End

(b) Folded Bus

Head End
A B

T R T R
1 2

R T R T R T R T
3 4 5 6
Trees (Forest)
(c) Tree Topology

Figure 2.7 Tree physical topologies.

In the folded bus shown in Figure 2.7(b), optical power from each station is coupled
onto the upper bus through a 2 × 1 combiner. The transmitted signals from the n access
stations propagate to the right, so that a combination of all signals appears at the head
end and is then distributed from the lower end of the bus, through 1 × 2 dividers (optical
taps) to the receivers in the stations. The network of Figure 2.7(c) contains six stations
joined in a tree topology. Signals from the set {1, 2} are combined and then multicast
to the set {3, 4, 5, 6} and those from the set {3, 4, 5, 6} are combined and multicast to
the set {1, 2}. Because there are two fiber-disjoint trees here carrying signals in opposite
directions, it is possible to run signals in both directions reusing the same wavelengths
without causing any interference between them. This is a rudimentary illustration of
how the fiber topology can help or hinder the reuse of the optical spectrum. (Spectrum
reuse is not possible in the star and bus topologies of Figures 2.7[a] and 2.7[b].) The
star and bus will be of special interest to us later. Note that if there is no wavelength
42 Multiwavelength Optical Networks

selectivity in the nodes, the two networks are functionally the same. Each is a broadcast
network that can potentially support n(n − 1) unidirectional point-to-point connections
among the n stations (not counting connections from a station back to itself). (However,
the star is more efficient than the bus in conserving signal power – see Problem 4.)

2.3.1.1 Directional Couplers

Each of the nodes in the previous examples can be built using one or more 2 × 2
directional couplers. A 2 × 2 directional coupler, is an optical four-port device, which
we represent as shown in Figure 2.8, with ports 1 and 2 designated as input ports and
1′ and 2′ designated as output ports.10 Optical power enters the coupler through fibers
attached to the input ports, is combined and divided linearly, and leaves via fibers attached
to the two output ports. Provided that the signals entering each input port originate at
distinct optical sources,11 the action of the coupler can be expressed in terms of input
signal powers P1 and P2 and output powers P1′ and P2′ . The power relations are given
by
P1′ = a11 P1 + a12 P2
(2.2)
P2′ = a21 P1 + a22 P2 .
For ideal symmetric couplers, the power transfer matrix A = [ai j ] is of the form


(1 − α) α
A= , (2.3)
α (1 − α)
where the parameter α may take on any value between 0 and 1. As defined in Equa-
tion (2.3), this is a passive and lossless device, meaning that power is conserved for the
signals passing through it. If the parameter α is fixed, we call it a static device. If α can be
varied through some external control (for example, electronic, thermal, or mechanical),

1 1′

2 2′

–α<
0< –1

Figure 2.8 Directional coupler.

10
By designating certain ports as inputs and others as outputs, we imply that the power propagates in the
indicated directions only, a condition that depends on the internal characteristics of the coupler as well
as the external network to which it is connected. (See Section 4.10.2 for a more complete description of
multiport devices in terms of incident and reflected waves.)
11
This ensures that the optical fields of signals interacting with each other in the coupler are not coherently
related, so their relative phases and polarizations can be ignored. In this case, the operation of the device
can be expressed as an input/output power relation, which does not depend on the specific relations among
the fields.
 Layered Architecture 43

1 1
2 2
3 3
4 4
5 5
6 6
7 7

Outputs
Inputs

8 8
9 9
10 10
11 11
12 12
13 13
14 14
15 15
16 16

Figure 2.9 A 16 × 16 star coupler.

then the device is dynamic or controllable. Now suppose, for example, that α = 1/2
(fixed). If signals are present at both inputs, the device acts as a 2 × 2 star coupler – see
the n = 2 case of Figure 2.7(a).
If we set P2 = 0, it acts as a power divider – as used on the lower bus in Figure 2.7(b),
and if we use only output port 1′ , terminating port 2′ with an (absorbing) dummy load,
it acts as a combiner – as used on the upper bus in Figure 2.7(b). In the latter case it is
important to observe that even if the coupler is lossless, there is an inevitable combining
loss (dissipated in the dummy load).
In reality, all physical devices are lossy. In a lossy 2 × 2 directional coupler, for
example, a11 + a21 < 1 and a12 + a22 < 1, indicating the presence of excess losses due
to device imperfections. Excess losses are to be distinguished from combining losses,
which are inevitable, even in ideal devices.
Other more elaborate types of static nodes can be constructed by interconnecting
2 × 2 couplers. For example, Figure 2.9 shows a 16 × 16 star coupler built using an
array of 32 2 × 2 star couplers arranged in a Banyan structure [Hui90, p. 99]. A setting
of α = 1/2 is used for each coupler. (If each 2 × 2 is lossless, then the complete 16 × 16
system will be lossless.)
In the dynamic case, an ideal directional coupler would be fully controllable through
the range α = 0 (the bar state) to α = 1 (the cross state). Controllable directional cou-
plers are used commonly as building blocks for optical switches, using only the bar and
cross states and thus acting as binary switching devices [Hinton93] (see Section 2.3.2.1).
If both input/output ports are used, they act to permute the two inputs between the two
output ports. If only a single input/output port is used, the device acts as a simple on/off
switch. But the directional coupler is also useful as a building block for more complex
network nodes, in which α is used as a continuously variable parameter. For example,
using only ports 1 and 1′ , the device becomes a variable attenuator.
The power transfer relations introduced in Equation (2.2) generalize naturally to larger
numbers of ports and to waveband selective configurations. For an r × n multiport device
44 Multiwavelength Optical Networks

operating on m wavebands, they can be written in the form

P ′ (wk ) = A(wk )P(wk ), k = 1, 2, . . . , m, (2.4)

where A is an n × r power transfer matrix, P(wk ) is a column vector of r input powers,

and P ′ (wk ) is a column vector of n output powers. The waveband argument indicates that
power is specified as a function of waveband. The element, ai j (wk ), of A(wk ) represents
the power transfer ratio from input port j to output port i valid for waveband wk .12 As
in the 2 × 2 case, the elements of a general power transfer matrix may be either static or
controllable. In the latter case, the matrix represents a waveband-selective switch (see
Section 2.3.2).

2.3.1.2 Static Routers

A serious defect of the broadcast star in Figure 2.7(a) is that without waveband or
wavelength selectivity, all signals entering the star coupler are combined on all outbound
fibers. This limits the total throughput to the usable capacity of a single fiber, thereby
reducing the capacity available to any one station by a factor of 1/n. One way of
improving the throughput of this network is by replacing the star coupler with a static
routing node.
Figure 2.10 shows a static router with n input and n output fibers, each carrying up
to n distinct λ-channels, with wavelengths chosen from a set  = {λ0 , λ1 , . . . , λn−1 }.
(For simplicity we discuss this device in the context of wavelength routing, but the
discussion can be transposed to the waveband routing domain by replacing wavelength
with waveband throughout. In spectral partitions with one λ-channel per waveband, the
two terms are equivalent.) Each input fiber is connected to a 1 × n wavelength demul-
tiplexer (WDMUX), which spatially separates the wavelengths on the fiber. Similarly,
each output fiber is connected to an n × 1 wavelength multiplexer (WMUX), which is
identical to the demultiplexer but is used in the opposite direction of signal flow, that
combines the different wavelengths onto that fiber. We assume that a fiber has a capacity
of n λ-channels. The wavelengths are assumed to be individually recognizable by the

WDMUX WMUX
λ0 λ0
λ1 λn 1
1 1
λn 1 λ1

2 2

n n

Figure 2.10 Static routing node.

12
By using a discrete set of wavebands, it is implied that the multiport operates uniformly on all signal
frequencies within each waveband.
 Layered Architecture 45

demultiplexer, so that this is a wavelength-routed network. (We consider the whole

“fabric” of WDMUXs, WMUXs, and their interconnecting fibers as contained in a
single “black box”; that is, they act as a single node.)
There are n 2 fibers between the input and output stages, connected in a way that
prevents identical wavelengths from different input ports from being combined on the
same output port, thus avoiding interference among the different channels. The physical
path that a signal takes through the node is determined uniquely by its wavelength and
port number, with the routing rule as follows: A signal on input port j, carried on
wavelength λk , is routed to output port i, where k = (i − j) mod n. For example, in
a 4 × 4 network, a signal entering on port 2, to be routed to output port 1, would use
λ3 .13 Using the matrix notation of Equation (2.4), the power transfer ratios for the n × n
system using the preceding “arithmetic” routing rule are ai j (wk ) = 1 when k = (i − j)
mod n and ai j (wk ) = 0 otherwise. Note that if input and output ports with the same
label correspond to fiber pairs carrying signals in opposite directions in the same link,
then the wavelength λ0 produces loopback connections. Thus, if loopback connections
are excluded, only n − 1 wavelengths are required for routing.
Multiports with similar but more general wavelength permutation connectivity have
been termed Latin routers [Barry+93], after the Latin square, an n × n array, where
each entry contains one of n numbers with no number appearing in a row or column
more than once. The Latin square is equivalent to a power transfer matrix, A , for a
Latin router, where all entries for some number, k, correspond to the 1s in ai j (wk ). The
reference [Barry+93] describes some multistage fabrics for general Latin routers.
If it is used in conjunction with tunable transceivers, the static wavelength router
functions as a permutation switch (see Section 2.3.2.1). For example, consider the router
of Figure 2.10 placed at the center of the directed star network of Figure 2.11, replacing
the star coupler. (The directed star is similar to the undirected case in Figure 2.7[a]
except that the transmitters and receivers have been separated to reflect the fact that
they may belong to different access stations.) A connection is established between any
transmitter/receiver pair by tuning both transmitter and receiver to the unique wavelength
that is routed on a path between the desired input/output ports.14

Transmitting Receiving
Stations Stations
1 1

2 2

n n

Figure 2.11 Directed star.

13
Although Figure 2.10 shows a node constructed with bulk components, integrated optic static routers have
also been fabricated in the form of arrayed waveguide gratings (see Section 4.10.4).
14
Another way of making the router function as a permutation switch is by placing wavelength interchangers
(see Section 2.3.3) in the input/output ports, as indicated by the rectangles in Figure 2.10.
46 Multiwavelength Optical Networks

More generally, up to n signals, each at a distinct wavelength in the set , can be

routed from each input fiber to each output fiber in this manner. Thus, suppose each
transmitting station is equipped with an array of n transmitters each tuned to one of
the n wavelengths in , with a similar arrangement of tuned receiver arrays in the
receiving stations. Using all transmitters and receivers simultaneously would create
full connectivity among all transmitting and receiving stations at maximum throughput
(n 2 fully utilized channels). In this arrangement, the use of the more complex (but static)
node architecture together with the transceiver structures that can exploit it allows for
an n-fold reuse of the optical spectrum. Compare this with the broadcast star, in which
the throughput is limited to n fully utilized channels – the available capacity of a single
fiber.

2.3.2 Dynamic Nodes

The simplest dynamic optical node is a space switch with no waveband selectivity. It
routes the signals on each input port to one or more selected output ports. Space switches
come in two flavors: permutation and generalized. In an n × n permutation space switch,
connections between input and output ports are point-to-point; neither one-to-many nor
many-to-one connections are allowed. Thus, the admissible connection patterns are in
the form of permutations, as illustrated by the connection matrix in Figure 2.12(a).15 (A
permutation switch is square, and its connection matrix has a single 1 in each row and
each column.) The specific connection pattern is selected under external control (i.e.,
switches are active devices in the sense that an external control agent is present).16 An
n × n nonblocking permutation switch can create any one-to-one connection pattern, so
that it has n! possible input/output patterns or connection states. This seems like a lot,
but relatively speaking it is not.

Output Port Output Port

1 2 3 4 1 2 3 4 5

1 1 1 1 1 1
Input Port

Input Port

2 1 2 1 1 1

3 1 3 1 1

4 1 4 1 1

(a) Permutation (b) Generalized

Figure 2.12 Space switch connection matrices.

15
Note that the connection matrices in the figure are transposes of the corresponding power transfer matrices.
16
The terms active and passive are tricky. Unless amplification is present in the signal path, an optical node,
or any other device in the network for that matter, is “seen” by the optical signals flowing through it as a
passive device because no energy is supplied to the signal as it traverses the device. We generally use the
term passive in that sense. Nevertheless, energy is required for control, so a power source must be present
for controllable devices. Hence, controllable (dynamic) devices are often termed active in the literature,
and static devices are called passive.
 Layered Architecture 47

A generalized space switch can create any input/output pattern, including one-to-many
and many-to-one connections. A connection pattern for a generalized 4 × 5 switch is
illustrated in Figure 2.12(b). (Connections are shown by 1s.) Counting all the possibilities
in the case of an r × n generalized switch, we find 2nr connection states. Because they
are capable of creating multipoint as well as point-to-point connections, two additional
control possibilities exist in generalized switches. For each one-to-many connection, it
is possible to specify how the input power splits among the various outputs, and for
each many-to-one connection, the combining ratios among the various input powers
can be specified. We call generalized switches that control these power-dividing and
-combining ratios linear divider-combiners. Because combining and dividing ratios can
take on a continuum of values, LDCs have an infinite number of connection states.
It is important to note that generalized switches have the ability to terminate a connec-
tion, thus “dumping” its signal internally (by setting all elements of a row of the power
transfer matrix equal to zero), whereas permutation switches do not. Similarly, they have
the ability to set the power to zero on any output port (by setting all elements of a column
equal to zero), whereas permutation switches do not. The connection termination prop-
erty has important practical implications, because in a network built around permutation
switches, unwanted closed paths may be created inadvertently, causing degradation in
network operation.
To illustrate, consider the network in Figure 2.13, which consists of n access stations
each containing a pair of OTs and ORs connected to a unidirectional ring via a network
node composed of a 2 × 2 coupler operating as an add/drop switch. With its coupler in
the cross state (the add/drop state), a station is coupled into the ring so that it transmits
to the next station on its right, which is in the add/drop state, and it receives from the
station to its left, which is also in the add/drop state. With its coupler in the bar state, the
station is bypassed. Figure 2.13 shows a condition in which station 1 transmits to station
n and n transmits to 1. A problem arises if all couplers are in the bar state. In this case a
closed loop exists.
At first glance, this would not appear to create a problem. However, in a network with
long fiber spans, there would normally be amplifiers in the links. Assuming that the
gains are adjusted to compensate for the losses over the links and nodes, we have a ring

n 2
OR
OT

1
OT
OR

OT OR

Figure 2.13 Unidirectional ring.

48 Multiwavelength Optical Networks

with a loop gain of unity. Because each amplifier generates some spontaneous emission
noise, the gain in the ring will cause the noise power to grow without bound (or at least
until the amplifiers saturate). Problems of this type have been observed in WDM ring
networks, in which closed paths may exist for certain wavelengths, with the amplified
noise robbing power from and interfering with signals at other wavelengths. Permutation
switches therefore must be interconnected with care!

2.3.2.1 Permutation Switches: Characterization and Complexity

There is a vast amount of literature on switches dating from the early days of telephony.17
Although technology has progressed from electromechanical to electronic to photonic
(optical), the basic structure of the switch “fabric” has remained largely unchanged. One
of the simplest structures that is adapted to both permutation and generalized switch
operation is the crossbar. As shown in Figure 2.14, an r × n crossbar switch consists of
r input lines, n output lines, and r n cross-points located at the intersections of the lines.
The implementation of the cross-points has progressed from electromechanical relays
through electronic gates to controllable optical couplers and micromirrors.
In an n × n crossbar switch, a permutation connection is made by closing one cross-
point in each row and each column. (Multicast connections can be made by closing more
than one cross-point in a row.) The main problem with the crossbar, used as an n × n
permutation switch, is that the number of cross-points grows as n 2 , which is far more
than necessary to create all possible permutation connections. Because optical switching
devices are generally costly, switch fabric realizations with a minimum of these devices
are desirable.
A common method of constructing large switch fabrics that are more economical
regarding cross-points is through a multistage fabric. A popular configuration is a
three-stage arrangement, which can be configured to produce many different types
of switches. As we shall see, multistage configurations usually can be realized with
far fewer cross-points than the crossbar, especially for large-size switches. Figure 2.15
shows an n × n symmetric three-stage Clos network (see [Hinton93, p. 93]) composed

1 2 n

Figure 2.14 Crossbar switch.

17
For a good view of developments from classical theory to early developments in photonic switching, see
[Benes65, Hui90, Hinton93].
 Layered Architecture 49

k k

p m m p

1 1

n n

k k

Figure 2.15 Clos switch.

of k p × m switches in the first stage, mk × k switches in the middle stage, and km × p

switches in the third stage, where n = kp. One possible way of realizing each of the
smaller switches is as a crossbar.
For n × n permutation switches with n a power of 2, the general Clos structure of
Figure 2.15 can be “factored” recursively to produce a Benes switch fabric, which is
one of the most economical permutation switches in terms of cross-points. The idea is
shown in Figure 2.16. In the first factorization, we take p = m = 2 so that there are
two (n/2) × (n/2) switches in the middle stage, and all first and third stage switches are
2 × 2. Now, each middle-stage switch can be factored in the same way, and the process
repeated until the middle-stage switches are 2 × 2. The result is an array consisting only
of 2 × 2 switches. The Benes switch for the case n = 8 is shown in Figure 2.17. It is

n × n
2 2

n × n
2 2

Figure 2.16 Recursion for Benes switch.

50 Multiwavelength Optical Networks

1 1
2 2

8 8

Figure 2.17 An 8 × 8 Benes switch.

left as an exercise for the reader to show that a general n × n Benes switch can be built
using (n log2 n − n/2) 2 × 2 elements.
Multistage switch fabrics such as the Benes, composed of 2 × 2 elements, are particu-
larly attractive in the optical domain. The reason is that each element (a 2 × 2 permutation
switch) is operated in only two states: the bar and cross states defined in Section 2.3.1.1.
Therefore, it can be implemented optically using a controllable directional coupler, with
a power transfer matrix of the form of Equation (2.3). Systems composed of arrays of
2 × 2 directional couplers, operated in various modes, appear frequently in this book.
We have already seen an example in the Banyan network of Figure 2.9, in which each
element was operated as a (static) star coupler.
The minimum number of binary switching elements required for a permutation switch
can be deduced by equating connection states to switch states. A switch having S connec-
tion states requires at least log2 S binary switching elements (giving S switch states). Be-
cause a permutation switch has S = n! connection states, we find that the minimum num-
ber of switching elements for any realization using any type of binary devices is log2 (n!),
which (using Sterling’s formula) is approximately (n log2 n − 1.44n) for large n.
Comparing this number with the number of elements required in the Benes switch, we
find that for large n the Benes fabric is very close to optimal in its use of hardware.

Characterization of Blocking
Permutation switches are classified in terms of their blocking characteristics. In dis-
cussing blocking performance, it is assumed that we are considering sequences of
demand-assigned connections between input and output ports on the switch, with re-
quests for connection establishment and termination occurring at random points in time.
An n × n switch is “rearrangeable” or rearrangeably nonblocking if there exists a set
of paths through the switch fabric (i.e., a set of switch states) that realizes each of the
n! connection states. The term rearrangeable comes from the fact that it may be neces-
sary to rearrange currently active connections to support a request for a new connection
between a pair of idle input and output ports. Rearrangeable switches have two problems:

1. For any given set of desired connections (any given permutation of inputs to outputs),
the determination of the required device settings to route the connections generally
requires complex computation.
 Layered Architecture 51

2. Connections in progress may have to be interrupted momentarily while rerouting is

taking place to accommodate new connections.

The good news about rearrangeable switches (of which the Benes is one) is that they
can be realized with a minimum amount of device hardware. The price of hardware
simplicity is control complexity. A control algorithm for the Benes switch, known as
the looping algorithm, is described in Appendix G. It has a complexity of the order of
(nlogn)2 for an n × n switch.
Wide-sense nonblocking networks are those that can realize any connection pattern
without rearranging active connections provided that the correct rule is used for routing
each new connection through the switch fabric. These require more hardware than re-
arrangeable networks, and still need intelligent routing algorithms, but have the ad-
vantage that active connections need not be interrupted. Algorithms for wide-sense
nonblocking networks are not necessarily complex. As an example, it is left as an ex-
ercise for the reader to devise a connection algorithm for the path-independent loss
crossbar in Figure 4.74, which is wide-sense nonblocking.
Strict-sense nonblocking networks, sometimes called simply nonblocking, require no
rearrangement and no complex routing algorithm. New connection requests are allowed
to use any free path in the switch. As might be expected, these networks require still
more hardware than the wide-sense nonblocking fabrics – this cost is traded for the lack
of connection disruption and simplicity of routing.
The Clos network of Figure 2.15 can be constructed to be either wide sense or
strictly nonblocking. A necessary condition on the number of middle-stage switches
to make the network wide-sense nonblocking is m = ⌊2 p − p/k⌋, where ⌊x⌋ denotes
the largest integer equal to or less than x [Smith+76]. If k = 2, then a necessary and
sufficient condition to make it wide-sense nonblocking is m = ⌊3n/2⌋ [Benes65]. A
necessary and sufficient condition to make the Clos network strict-sense nonblocking
is m = 2 p − 1. For our purposes, strict-sense nonblocking switches are the simplest
and most practical but not always cost-effective. If the sizes of the switches in its three
stages are optimized, it can be√shown that a strict-sense nonblocking Clos switch can be
realized with approximately 4 2n 3/2 cross-points.
Another way of realizing a strictly nonblocking switch is through space dilation of
the original Benes fabric. An m-fold space dilation of an n × n Benes switch is created
in the general three-stage form of Figure 2.15. The dilated switch is built using n 1 × m
switches in the input stage, n m × 1 switches at the output, and m Benes networks in the
middle stage. With m = log2 n, it turns out that this system, called a Cantor network,
is strictly nonblocking because a free path through the fabric can always be found
for any pair of idle input/output ports via one of the m middle switches. This switch
has O(n[log2 n]2 ) cross-points, which is an improvement over the strictly nonblocking
version of the Clos switch for sufficiently large n.
In terms of the number of cross-points required for strictly nonblocking switches,
multistage fabrics do not start to become worthwhile until the switch size is fairly large
(n > 32). However, for rearrangeable or wide-sense nonblocking switches, multiple
stages are definitely more economical.
52 Multiwavelength Optical Networks

At this point it is worth noting the contrasting features of the traditional and opti-
cal network switching requirements. In traditional networks, switches often have very
large numbers of ports, and cross-points (realized electronically) are very inexpensive.
Conversely, optical switches are orders of magnitude smaller and their cross-points are
orders of magnitude more expensive. Now, in large switches it is possible to take ad-
vantage of laws of large numbers to reduce dramatically the hardware complexity of
the switch in return for accepting some small blocking probability (meaning that not all
possible connection patterns can be realized). Typical telephone switches with n on the
order of 105 operate with blocking probabilities on the order of 10−3 or less. This is not
particularly relevant to optical switches, however, because their size is more likely to be
on the order of 10.
In traditional networks, rearrangeable and wide-sense nonblocking switches are rarely
used because the computational complexity involved in finding and executing an ap-
propriate routing algorithm is overwhelming for large-size switches and is not worth
the savings in switch hardware. In the optical domain, however, switches are small, so
routing complexity for rearrangeable and wide-sense nonblocking switches is reduced.
Furthermore, in transparent WDM networks the routing algorithm is used only in the
space dimension, so algorithmic complexity is that required for a single waveband only,
whereas the hardware savings are roughly proportional to the number of wavebands
being switched. Also, computational speed can be increased by doing computations for
all wavebands in parallel. These considerations suggest that it is worth taking a second
look at the rearrangeable and wide-sense nonblocking category for optical applications.
Another issue that is important in optical switches is cross-talk. In electromechanical
or electronic switches, the cross-points are close to ideal; a switch is either definitely
open or closed. In optical switches, the binary elements (controllable couplers) typically
used as cross-points may have as much as a 5% cross-state “leakage” when they are
set to the bar state and vice versa. This necessitates using switch fabrics with special
designs involving extra binary elements to reduce cross-talk (see Section 4.12.1). Thus,
a minimum cross-point realization is not always the best.

2.3.2.2 Generalized Switches

Because the number of connection states of a generalized r × n switch is 2nr, it is clear
that we need at least nr binary elements to realize it. This is the number of cross-points
used in the traditional crossbar switch, and one way of realizing generalized optical
switches is through an optical version of the crossbar.
Figure 2.18 shows one possible realization [Sharony94]. It consists of three stages, the
first and third of which are static and the middle stage is controllable. Stage 1 consists
of an array of r 1 × n signal dividers (splitters), and stage 3 is a similar array of n r × 1
signal combiners (which could be realized with the same type of hardware, used in the
opposite direction of signal flow). The middle stage consists of r n binary on/off switches,
which could be realized, for example, using the 2 × 2 coupler of Figure 2.8 or the
Y-branch switch of Figure 4.46. Any generalized connection pattern of the type shown
in Figure 2.12(b) can be created by turning on a middle-stage switch for each 1 entry
in the connection matrix. Assuming that the splitting (combining) stages produce equal
 Layered Architecture 53

r 1
Combiner
1

11 12 1r

21 22 2r
1 n
Splitter

n1 n2 nr
1 2 r

Figure 2.18 Generalized optical switch.

splitting (combining) ratios for each signal, this switch will produce an input/output
power relation in the form

P ′ = A P, (2.5)

where P and P ′ are vectors of input and output powers, respectively, and the elements
of the n × r power transfer matrix A are

1/nr if switch i j is on
ai j = (2.6)
0 otherwise.

2.3.2.3 Linear Divider-Combiners

Although the crossbar is an extremely simple configuration, it has the disadvantage that
the two passive stages produce a combined signal attenuation (splitting loss followed by
combining loss) of 1/r n. A different approach, which has less inherent loss and at the
same time offers complete generality in setting combining and dividing ratios, is shown
in Figure 2.19. We call this a δ–σ LDC. It consists of a power-dividing stage followed
by a power-combining stage, both of which are controllable.
A 4 × 4 example is shown in Figure 2.19(a). It has a two-stage power-dividing network
fed by four input fibers followed by a combining network (the reverse of the dividing
network) connected to four output fibers. Each box in the figure represents a 2 × 2 con-
trollable coupler operated as a continuously adjustable power divider or combiner. This
structure generalizes in a straightforward way to an r × n LDC – as shown schematically
54 Multiwavelength Optical Networks

11 11

21 12
1 . . 1′
. .
n1 1r

12 21

22 22
2 . . 2′
. .

. n2 2r .
. .
. .
1r n1
2r n2
r . . n′
. .
nr nr
(a) (b)

Figure 2.19 δ–σ linear divider-combiner.

in Figure 2.19(b) – that requires a total of [r (n − 1) + n(r − 1)] couplers. If we denote

by δi j the fraction of power from input port j directed by its divider to the combiner for
output port i ′ and denote by σi j the fraction of power received from the divider serving
input port j and combined onto output port i ′ , then the elements of the power transfer
matrix for this LDC are given by

ai j = δi j σi j . (2.7)

The dividing and combining ratios are varied by setting the coupler states appro-
priately. In the ideal case of lossless couplers, the ratios are subject to the physical
constraints

δi j = 1 (2.8)
i

and

σi j = 1. (2.9)
i

Within these constraints, arbitrary combinations of inbound signals at the node can
be directed to each outbound fiber. For example, if the coupling state for each 2 × 2
coupler in Figure 2.19(a) is set at 0.5, the LDC operates as a 4 × 4 star coupler, with
ai j = 1/16 for all i and j. If each 2 × 2 coupler is used only in a binary (bar/cross) state,
the system operates as a permutation switch. Note that any input signal can be “dumped”
(i.e., absorbed in the switch) by directing it to an output port with a combining ratio that
is set to zero. For example, a signal at input port i can be terminated by setting δik = 1
and σik = 0. Similarly, the power at output port j ′ can be set to zero by letting δk j = 0
and σk j = 1.
Note that the switch of Figure 2.18 can be converted to an LDC by making the
dividing and combining ratios in Equation (2.6) more general. If the on/off switches are
replaced by continuously variable attenuators, the elements of the power transfer matrix
 Layered Architecture 55

Figure 2.20 A node without loopback connections.

become ai j = αi j /nr , where αi j is the transmission constant for the i j th attenuator in

the switch. The only functional difference between this and the δ–σ configuration is that
the constraints on ai j given in Equations (2.7), (2.8), and (2.9) are replaced by

0 ≤ ai j ≤ 1/nr. (2.10)

These switch fabrics are representative of a very large class of configurations. Many
variants are possible for adapting to special situations or for reducing cost. For example,
less hardware is needed if full connectivity is not required. Figure 2.20 illustrates a
case of a switching node of degree three (connected to three bidirectional links), where
loopback connections along the same link are not required. A δ–σ LDC for this node
can be realized with 6 (instead of 12) 2 × 2 couplers as shown.
As another example, the large controllable component count in the LDC of Figure 2.19
can be reduced, at the cost of additional splitting (combining) loss, by replacing the
controllable dividing (combining) stage with a fixed dividing (combining) stage, as is
used in the switch of Figure 2.18. The elements of the power transfer matrix of an r × n
LDC are then given by

ai j = δi j /r (2.11)

for a fixed combining stage and

ai j = σi j /n (2.12)

for a fixed dividing stage.

Different switch architectures and switch configurations produce different effects on
signal power levels at the output ports. For example, considering square switches built
from ideal lossless couplers, the Benes switch and the δ–σ LDC are lossless when used as
permutation switches, whereas the optical crossbar incurs a loss of 1/n 2 for any switch
configuration. When configured as a broadcast star, however, both the crossbar and
the δ–σ LDC produce a 1/n 2 combining/splitting loss, whereas the Banyan realization
56 Multiwavelength Optical Networks

(Figure 2.9) produces the minimum possible loss of 1/n. (The Banyan structure does
not, however, contain enough couplers to be used as a nonblocking switch.)

2.3.2.4 Waveband-Space Switches

By adding waveband and/or wavelength selectivity in various ways, the basic switch
fabrics described earlier can be converted to waveband/wavelength-selective switches
(WSSs). In this section we begin the discussion with transparent switches without wave-
length interchange. In keeping with the idea that any spectral selectivity in a transparent
optical switch is assumed to be limited to wavebands – the “coarse” subdivisions of
the optical spectrum defined in Section 2.2, we focus on the waveband as the elemen-
tary spectral unit being switched. Later in the section some opaque architectures are
introduced.
In an m-waveband WSS, signals carried on m wavebands in the set W =
{w1 , w2 , . . . , wm } are multiplexed on each input fiber. An r × n WSS directs these
signals in a waveband-selective manner to the n output ports. Thus, the switch operates
on mr inputs. (Although there may be several λ-channels grouped on each waveband,
as far as the switch is concerned the waveband is the smallest recognizable entity.) The
switching pattern is independently controllable for each waveband.
Figure 2.21 shows one way of realizing an n × n WSS using a three-stage configu-
ration, with the middle stage composed of m switching layers, each one operating on
signals in one waveband. The m wavebands on each input fiber are separated spatially
into m layers by WDMUXs in the first stage. Each layer is a waveband-independent
n × n space switch, which operates on all signals in one waveband. The space switching
can be of any type, from permutation to LDC. The switched signals are recombined on
the output fibers by WMUXs in the third stage.
This realization of a WSS used a combination of fixed (passive) waveband-selective
devices (WMUXs and WDMUXs) and dynamic but waveband-insensitive switching
devices. A more compact way of realizing the system would be to use multiwaveband

WDMUX WMUX
n n
w1 wm w1 wm
1 1 1

n n
2 2
2
. .
.
. .
.
. .
.
n n

n n

Figure 2.21 Three-stage realization of a waveband-space switch.

 Layered Architecture 57

P1(w1), … P1(wm ) P1(w1), … P2(wi ), P2(wj ), … P1(wm )

P2(w1), … P2(wm ) P2(w1), … P1(wi ), P1(wj ), … P2(wm )

Switch Control

Figure 2.22 Multiwaveband directional coupler.

switches (MWSs) where switching and waveband- or wavelength-selective functionality

are integrated into the same device. Waveband-selective directional couplers (i.e., 2 × 2
MWSs) have been constructed using a variety of technologies, with a capability of
independently and simultaneously switching several wavebands under external control
(see Section 4.10.5). The power transfer matrix for a lossless symmetric 2 × 2 MWS is
of the form of Equation (2.3), where the parameter α is now a function of waveband w.
Figure 2.22 shows a 2 × 2 MWS operating on m wavebands. The switch is set in
the cross state for wi and w j and in the bar state for all other wavebands. Each of
the space-switching fabrics of Figures 2.17, 2.18, and 2.19 can be converted to an
m-waveband WSS simply by replacing each switching device by an MWS operating on
m wavebands. Of course, the distinction between a WSS, which we show as combining
many individual components performing separate functions, and an MWS, shown as an
integrated device, is largely arbitrary. If we draw a black box around the collection of
elements in Figure 2.21 and do not look inside, it might be considered to be an MWS.
There are, however, many optical switching devices that are integrated to the point
that they cannot be subdivided even conceptually into components performing separate
functions; these are unquestionably MWSs (see Section 4.10.5).
The power relations in an WSS can be described using the power transfer matrix of
Equation (2.4). In a permutation switch, A(wk ) is a permutation matrix for each wk in
the set W, whereas in a waveband-selective LDC, each element of A may take on a
continuum of values. The elements of A will, of course, change in response to external
control. Thus, each switch configuration is defined by m independently controllable
matrices A(w1 ), A(w2 ), . . . , A(wm ) – one for each waveband. For example, in the
realization of Figure 2.21, the setting of the kth layer of the switch would be identified
with A(wk ).
The WSS is one of a family of switching devices commonly called optical cross-
connects (OXCs). The most commonly used WSSs operate as wavelength-selective or
waveband-selective permutation switches, which are also called wavelength-selective
cross-connects (WSXCs). For n × n m-waveband WSSs, there are (n!)m connection
2
states for the permutation case and 2n m connection states in the generalized case.
Because mn entities are being switched when the fibers carry m wavebands, one
might first think that the permutation or generalized switch should have (nm)! and
2
2(nm) connection states, respectively. This is not so because of the wavelength con-
tinuity constraint: Wavelengths of signals do not change in passing through a WSS,
58 Multiwavelength Optical Networks

so less switching alternatives exist in the wavelength dimension than in the space
dimension.
To avoid a plethora of details and special cases, we have restricted the discussion here
to fairly general models of switches. Of course, in practice there are many variants of
the basic switch, and many different realizations adapted to particular sets of needs. We
summarize them briefly now, directing the reader to Section 4.12 for more details on
switch architectures and Section 4.10 for the enabling switch technologies.
First, there is the issue of hardware realization, and the related question of transparency
versus opacity. Transparent switches require purely optical realization using photonic
components of the type described in Section 4.10. An example of a transparent OXC,
sometimes called a photonic cross-connect (PXC), is shown in Figure 4.86. It is based
on the three-stage architecture of Figure 2.21, with signals remaining in optical form
throughout the switch. This arrangement is termed an O-O-O switch because signals
enter and leave the switch in optical form, and the fabric is also optical. Local users can
access the switch on add/drop ports via NASs whose access links are connected directly
to switch ports. This corresponds to our model of a transparent ONN with an associated
NAS. The optical connections terminating on the OXC must use the same wavelengths
as those carried on the internodal links connected to the switch ports.
Figure 2.23 shows a special case of the architecture of Figure 4.86 with a single
input and output port and multiple access fibers linking it to a colocated NAS. Known
variously as a wavelength add/drop multiplexer (WADM), an optical add/drop multi-
plexer (OADM) or a reconfigurable optical add/drop multiplexer (ROADM) it can be
realized in many different forms, and is the workhorse of current metropolitan area
networks (see Section 4.12.2). We have already seen an add/drop switch, in the form
of a 2 × 2 coupler, in Figure 2.13. The WADM is a direct generalization of that to the
multiwavelength case, where it independently adds and drops each wavelength on a pair
of inbound and outbound unidirectional internodal links. In the realization shown in

W A DM

W D MUX WM U X
λ1

λ2

λm

λ1 λ2 λm
OT OR OT OR ... OT OR
NAS
TP/R P

...
Figure 2.23 WADM–NAS combination.
 Layered Architecture 59

Figure 2.23, a controllable directional coupler is associated with each wavelength to be

added/dropped at the local NAS. Thus, a separate pair of access fibers is required for
each wavelength. With the coupler in the bar state the wavelength is passed through the
WADM, and in the cross state the wavelength is dropped to the local receiver, and a
signal on the same wavelength can be added to the signal set exiting on the output port.18
This is an especially useful and cost effective network element for accessing a ring or
an intermediate point on a long link in a mesh network (see Figure 4.81).
Although transparency has many advantages, most currently deployed networks
use opaque switching nodes. Figure 4.87 shows an opaque OXC using a three-stage
O-E-O arrangement, wherein the switch fabric is electronic, requiring the demultiplexed
wavelengths to be converted to electrical form before entering the switch and then back
to optical form at the output ports of the switch. For various practical reasons, the conver-
sion between λ-channels and electrical signals is implemented as a two-stage process,
first shifting the wavelengths of the λ-channels typically to the 1.3-µm band using a
transponder and then converting the resultant 1.3-µm optical signals to electrical form.
(A transponder accepts a signal in optical form, converts it to electrical form, provides
some degree of regeneration, and then converts it back to optical form on a different
wavelength.) The devices for converting the 1.3-µm signal to and from electrical form
at the electrical switch fabric ports are called transceivers; they are composed of an
optical transmitter and receiver pair operating in the 1.3-µm band. They perform the
electrical-to-optical and optical-to-electrical signal conversion, respectively. Note that
this architecture offers the possibility of connecting electrical signals from client logical
networks (e.g., SONET, ATM, or IP) directly to add/drop ports on the electrical switch
fabric. Electronics creates a bottleneck, limiting the bit rates of the signals passing
through the fabric. Furthermore, because electronics is generally neither bit-rate nor
format independent, changes in the signal bit rate and format require a complete change
of the switch fabric.
Finally, Figure 4.88 shows an opaque O-O-O architecture similar to the O-E-O case,
except that the switch fabric is optical, eliminating the need for transceivers. The opaque
O-O-O differs from the transparent case in Figure 4.86 because electronics is present in
the form of transponders translating the optical signals to the 1.3 µm band as previously
noted. This destroys the transparency of the connections. (Signals from client networks
again access the switch fabric directly, but in this case they are in optical form at
1.3 µm.) In the opaque O-O-O and O-E-O architectures, the transponders are costly and,
again, create an electronic bottleneck. However, on the positive side they clean up the
signals, avoiding the propagation of signal impairments, and as a by-product they offer
the possibility of wavelength conversion in the switching node. However, conversion
is possible only if the switch fabric is configured as a nonblocking permutation switch
interconnecting all ports. Thus, an m wavelength n × n OXC requires a single mn × mn
switch fabric rather than m n × n fabrics (as in Figure 2.21) if it is to perform wavelength

18
In cases where the add/drop connections remain fixed for long periods of time, a more economical static
version of the WADM can be constructed by replacing the controllable couplers by fixed connections.
60 Multiwavelength Optical Networks

conversion. The former is more costly than the latter, especially when m is large. Finally,
an advantage of the two opaque configurations is the fact that the signal is converted
to electrical form at some point, meaning that there is easy access to control and
management overhead carried in the bit stream. This allows for realization of various
signal monitoring, control, and recovery functions.
In any of the three switch architectures described above there remains the issue of
whether they should be operated on a waveband or wavelength basis. The opaque archi-
tectures can switch wavebands and wavelengths equally well once they are demultiplexed,
and can perform wavelength conversion. In either case they must operate on each wave-
length individually, so that they cannot take advantage of the cost advantages of bundling
λ-channels into wavebands: demultiplexing is difficult and costly for many closely spaced
channels as compared to fewer widely spaced bands. Furthermore, the complexity of
the switch fabric grows rapidly with the number channels being switched. As DWDM
systems move to large numbers of wavelengths on a fiber this becomes a serious issue.
The transparent OXC offers simplicity and generally lower cost than the opaque
case but it is better suited to waveband than wavelength operation and cannot perform
wavelength conversion. Recall that a disadvantage of waveband-switching mentioned
earlier is that all λ-channels in a common waveband stick together in the optical path layer.
But in a purely waveband-oriented network architecture not all of the optical connections
grouped into a common waveband will terminate at the same source and destination
ONNs. When at least one connection terminates, its waveband must be multiplexed
(demultiplexed); i.e., its optical path must terminate. Optical path termination involves
additional processing and equipment at the nodes and therefore should be reduced to
a minimum. This problem can be eliminated or at least reduced if all λ-channels for
connections terminating at a common node are grouped into a single waveband that
terminates at that node. The remaining through traffic is carried on wavebands routed
through the node. The hybrid switch of Figure 4.89, which has a waveband-switching
layer (WSXC), and a wavelength-switching layer (λSXC), is adapted to this approach.
Through traffic is carried on wavebands that are routed through the switch without
being passed to the λSXC layer. Waveband w3 , carrying terminating traffic is passed
to the λSXC layer for demultiplexing and further processing in an NAS. To achieve
the efficiencies that are attainable with a hybrid switch, traffic terminating at common
nodes must be grouped into common wavebands wherever possible, a grooming process
mentioned earlier in Section 2.2.

2.3.2.5 Wavelength-Space-Time Switches: Optical Packet Switching

When used as a node in a large circuit-switched network, the WSS described earlier
would normally be reconfigured on a relatively slow timescale. For example, dedicated
connections used to provision one of the specialized LNs shown in Figure 1.1 might be
held for very long periods of time, much like connections through digital cross-connects
in traditional networks. For demand-assigned connections, the timescale would be faster
– on the order of minutes or hours. For efficient operation in a circuit-switched mode,
the time required to establish and terminate each connection (including routing com-
putations, signaling, and switch reconfiguration) should be small compared with the
 Layered Architecture 61

holding time (for example, on the order of seconds for demand-assigned connections).
This is no problem using the WSS structures discussed earlier. However, there are other
cases when much faster switch reconfiguration timescales are required. For example,
consider a wavelength-routed network (one λ-channel per waveband), in which each
λ-channel entering a switch is carrying several logical channels time division-
multiplexed into slots (of microsecond duration) in a fixed periodic time frame (see
Section 3.2.1). In this case each time slot is equivalent to a subchannel of the λ-channel.
Now suppose we want to switch the subchannels independently among different space
ports on a time slot-by-time slot basis. This requires rapid switching (on a microsecond
time scale) to route each time slot independently to a different output port. With proper
time slot synchronization among all the entering signals, the switch could be reconfigured
to create a different space connection pattern for each wavelength in each time slot,
producing switching in three dimensions: a wavelength-space-time switch.
Assuming a frame with l time slots, the total number of connection states for an
m-wavelength n × n (permutation) switch would be (n!)lm. In a wavelength-space-time
switch, the time dimension behaves much like the wavelength dimension: The number
of connection states does not grow as rapidly with l (or m) as it does with n. This
reflects time slot continuity through the switch; in other words, information entering
in a particular time slot leaves in the same time slot. As the central node in a star
network, this kind of switch is an interesting alternative to the passive star. Using time
and wavelength, a set of lm logical connections (one for each time division subchannel
and each λ-channel) can occupy each input port, resulting in a total of lmn independently
switched connections using m wavelengths.
Extending this idea to WANs is much more difficult. Because of time slot continuity,
the optical wavelength-space-time switch is not as versatile as its electronic counterpart.
Traditional electronic space–time switches in wide area networks use time slot inter-
change (TSI) in addition to space switching. With TSI, the order of the time slots in
the frame is permuted as they pass through the switch. This produces more connection
states but requires buffering, which is much more difficult in the optical domain than
the electrical domain.
With or without buffering, time domain switching is not a transparent process because
of the necessity of aligning the slot boundaries with the switching times. Once trans-
parency is abandoned, the temptation is to continue in this direction to fully optical packet
switching (OPS). In a WDM network, OPS is characterized by two features: (1) time
division multiplexing of information packets on each wavelength and (2) routing pack-
ets through ONNs based on control information in the packet headers. As mentioned in
Section 1.3 the use of this in-band control information for routing provides a high degree
of virtual connectivity in a packet-switched network. However, implementation in the
optical domain requires optical buffering, because (1) packets are typically of variable
length (as in IP networks) or are separated by idle periods, so that they are multiplexed
asynchronously, (2) packets arriving at a node from different sources are not synchro-
nized in time, and (3) packets contending for transmission on a given output line at a
node may require buffering to resolve contention. As optical technology progresses
the feasibility of optical buffering and signal processing improves. Thus, although
62 Multiwavelength Optical Networks

speculative today, OPS may be a viable option in the near future and the
wavelength-space-time switch will come into its own (see Chapter 10).

2.3.2.6 Cross-Talk
So far we have avoided consideration of device imperfections and their effect on the
performance of switching devices; these are largely relegated to Chapter 4. However,
these play an important role in the evaluation of cost–performance trade-offs in each of
the switch architectures described earlier. One important performance consideration in
a switch is cross-talk, which results when some of the power from an input signal leaks
through to an unintended output port. (This was one of the transmission impairments
listed in Section 1.3 as a downside to transparency.) For example, the directional coupler
defined in Equation (2.3) is in the bar state when its control parameter α = 0. However,
for any real device α can be made close to but not exactly equal to zero. If α = ǫ > 0
when the switch is meant to be in the bar state, we have on port 1′
P1′ = (1 − ǫ)P1 + ǫ P2 , (2.13)
where the presence of P2 (at the same nominal wavelength as P1 ) represents undesired
leakage of the signal entering on input port 2 to output port 1′ . This leakage constitutes
cochannel heterodyne cross-talk between the signals entering on the two input ports
(see Section 4.9). When a switch is composed of a multistage fabric, as in Figure 2.17,
there are many sources of cross-talk on a signal path within one ONN, compounding
this problem. Furthermore, when an optical path consists of several hops, the cross-talk
accumulates from each optical node along the path. The accumulated cross-talk causes
interference at the optical receiver, with a consequent deterioration of the bit error rate
(BER) or SNR.
In addition to cochannel heterodyne cross-talk, leakage paths can also cause
cochannel multipath cross-talk, which occurs when a portion of the signal power leaks
through an unintended path within a switch fabric and recombines with the original sig-
nal. The recombination may be “downstream” on a parallel path or “upstream” on a feed-
back path. Multipath cross-talk is more troublesome than heterodyne cross-talk because
the recombining optical fields originate from the same source so that phase and polariza-
tion relations influence the magnitude of the resultant interference. In waveband-space
switches interchannel cross-talk may also be present, either due to imperfect waveband
demultiplexing or imperfect operation of multiwaveband switches. These phenomena
and means of combatting them are explored in more detail in Chapter 4.
The main point to be observed here is that cross-talk effects increase with the size of a
switch, the number and packing density of the wavebands, and the imperfections in the
basic building blocks. As shown in Chapter 4, it is possible to offset some of the cross-
talk effects due to imperfect components by using more complex switch architectures.
However, these require a higher component count and therefore an increased cost. Herein
lies a basic three-way cost–performance trade-off: Large-size switches that handle many
densely packed wavebands are desirable for increasing optical network connectivity and
throughput but require high-quality (and hence expensive) component technology and/or
a high-component count for satisfactory performance.
 Layered Architecture 63

Another cost–performance issue stems from the relation between component count
and switch functionality (closely related to switch states). The component count is
O(n log n) for n × n rearrangeable switches and increases as we go to wide-sense non-
blocking, strictly nonblocking, and generalized switches, and LDCs. The more switch
states, the larger the required component count and/or the more complex the compo-
nents (i.e., the higher the cost). But, as mentioned at the beginning of this section,
switches with a large number of states offer better network performance in terms of
throughput, connectivity, flexibility, and survivability. These trade-offs between network
performance and switch functionality are discussed in more detail in later chapters.

2.3.3 Wavelength Converters

The basic optical transmission channel considered here has been a λ-channel, which
was generally assumed to remain on a fixed wavelength from end-to-end in transparent
networks – a condition called wavelength continuity. Here for the first time we intro-
duce the operation of wavelength conversion, which violates the wavelength continuity
condition.
Wavelength conversion can be implemented electronically, electro-optically, or opti-
cally and it can be executed as a (nearly) transparent or opaque operation. We have already
seen instances of wavelength conversion, performed as a by-product of operations in the
electronic domain in opaque OXCs. Transponders operating on single demultiplexed
optical signals at input ports translate the input signal frequencies to a common band,
the resultant signals are switched either optically or electronically and then converted
to optical output signals at a different wavelength. As a result of this wavelength con-
version capability the opaque O-O-O and O-E-O OXCs take on the functionality of
a wavelength-interchanging switch, also known as a wavelength-interchanging cross-
connect (WIXC). A signal entering the switch at any port and any wavelength on a
designated grid can be routed to any output port on any other wavelength on the grid –
with the constraint that signals routed to the same output port are carried on different
wavelengths to avoid interference. Of course, a transponder is itself an opaque wave-
length converter, executing the conversion operation by using an optical-to-electrical
followed by an electrical-to-optical signal conversion. The transponder is an example
of a converter based on “strongly” nonlinear devices, in this case the optical detector
and transmitter. Other opaque wavelength converters based on strong nonlinearities are
discussed in Section 4.11.2.
We turn our attention now to a class of wavelength converters using wavelength
mixing, called coherent converters. These are based on “weak” nonlinearities, and for
all practical purposes they are transparent. An ideal (transparent) wavelength converter is
a single input/output device that converts the wavelength of a λ-channel appearing on its
input port to a different value at its output port but otherwise leaves the optical signal(s)
unchanged. Converters based on strong nonlinearities typically operate on a single input
signal and are not transparent to bit rate and modulation format, whereas wavelength-
mixing converters operate transparently on multiple signals within a broad band of
wavelengths. Within this category, devices based on four-wave mixing and difference
64 Multiwavelength Optical Networks

Signal Wavelength Output

s Converter c p s

Pump
p
(a )

0 p
c s p
2
(b)

Signal Low-Pass Output

s Filter c p s

cos2 pt

(c )

Figure 2.24 Wavelength conversion as a linear operation.

frequency generation (also called three-wave mixing) have been demonstrated. Our
discussion here is confined to difference frequency generation. Section 4.11.2 discusses
both types in more detail.
An illustration of the operation of a difference frequency-generating parametric
wavelength converter or difference frequency converter (DFC) [Yoo+95] is shown in
Figure 2.24(a). A signal at optical frequency νs is applied to the input port and mixed
with a pump signal at frequency ν p . The DFC contains an optical medium that possesses
a second-order (square-law) nonlinearity, so that various sum and difference frequen-
cies may be present at the output. If the proper phase-matching conditions are satisfied
within the converter, only the difference frequency νc = ν p − νs is present. Expressing
this in terms of wavelengths, we have 1/λc = 1/λ p − 1/λs . Typical spectral relations
are shown in Figure 2.24(b). Here, three input signals are shown in a band at nominal
frequency νs , and the pump frequency is chosen at slightly less than double the signal
frequency. For these frequency relations, the output spectrum is an attenuated mirror
image of the input spectrum, reflected about the “mirror frequency” ν p /2.19
Although the DFC depends on nonlinearity to produce frequency conversion, it can
actually be modeled mathematically as a (time-varying) linear device, as shown in
Figure 2.24(c). In the model, the input signal is multiplied (modulated) by a sinusoid
at the pump frequency, and the product is low-pass filtered to retain only the difference
19
This means that the spectrum of each input signal is inverted. The inversion is removed by an even number of
passes through DFCs. Spectrum inversion can be used to an advantage in compensating for fiber dispersion
(see Section 4.3.2.3).
 Layered Architecture 65

p 3 2 1 p 1 2 3
s2 s1
2 2
(a) (b)

Figure 2.25 Wavelength interchanger.

frequency terms. Thus, for all practical purposes, it produces wavelength conversion
while exhibiting the transparency we expect of any linear device. The conversion is
controllable by varying the pump frequency. If only two input signals are present, and it
is required to interchange their optical frequencies, this can be done by placing the mirror
frequency midway between the signal frequencies, producing a wavelength interchanger
as shown in Figure 2.25(a).
Because the DFC acts as a linear device, it obeys the same conversion rules for
superimposed signals as for individual ones; that is, it is a “bulk” conversion device.
For example, suppose a set of signals at equally spaced optical frequencies is present at
the input of a DFC. Then, by placing the mirror frequency midway between two of the
signal frequencies, as shown in Figure 2.25(b), the DFC interchanges signal frequencies
in pairs: ν1 → ν−1 , ν−1 → ν1 , ν2 → ν−2 , ν−2 → ν2 , and so on. It is also possible to
use two pump frequencies to produce a more elaborate interchange pattern. (More than
two pump frequencies make the device impractical, introducing unwanted converted
frequencies that cannot be filtered out.)

Wavelength-Interchanging Switches
An example of a 2 × 2 WIXC is shown in Figure 2.26. Three optical signals are active
on the input ports. S1 (λ1 ) and S2 (λ2 ) appear on port 1 at the indicated wavelengths,
and S3 (λ1 ) enters on port 2. The switch is shown in the bar state for S2 and S3 and in
the cross state for S1 . In addition, the wavelength of S1 has been changed to prevent
a wavelength conflict at output port 2′ . This example generalizes in a natural way to
switches of higher dimension. It can be realized as an opaque switch using the OXC
architectures of Figures 4.87 and 4.88.
Another approach using transparent components is shown in Figure 2.27. It uses
a combination of 2 × 2 WSXCs and difference frequency wavelength interchangers
to realize a general WIXC [Antoniades+96]. The architecture is based conceptually

S1( λ 1), S2(λ2) S2(λ2)

1 1

WIX C
S3(λ 1) S1(λ 2), S3(λ 1)
2 2

Figure 2.26 Wavelength-interchanging switch.

66 Multiwavelength Optical Networks

S2(λ2)
S1(λ1) S3(λ2) S3(λ1) S2(λ2)
WSXC WSXC
WSXC WSXC
S3(λ1) S1(λ2), S3(λ1)
S1(λ1) S1(λ2)
DFC DFC

λ1 X = X
λ2 = = =

Figure 2.27 WIXC implementation.

on a modified Benes structure called a twisted Benes switch fabric, with DFCs placed
between the stages. The figure shows a realization of the 2 × 2 two-wavelength WIXC in
the example of Figure 2.26, implemented using this approach. It requires three WSXCs
and two DFCs used as single-wavelength interchangers. The settings of the WSXCs (bar
or cross for each wavelength) to produce the signal configuration shown in Figure 2.26
are indicated in Figure 2.27.
The realization procedure yields a general structure for any n × n m-wavelength
WIXC, provided that n and m are powers of 2. For example, a 2 × 2 four-wavelength
WIXC can be built with five 2 × 2 wavelength selective switches and four four-
wavelength interchangers. The general structure, based on the twisted Benes fabric,
requires 2 log2 (nm) − 1 multiwavelength switching stages and 2 log2 (nm) − 2 wave-
length interchanger stages. Because of their relation to the Benes switch, these structures
are rearrangeably nonblocking. To add another connection on an idle pair of ports and
wavelengths, it may be necessary to rearrange active connections.
One application of WIXCs is to circumvent connection blocking problems associated
with the wavelength continuity constraint. Recall that all optical signals multiplexed on
the same fiber in a multiwavelength network must be assigned distinct wavelengths so
that they can be distinguished at the receivers without interchannel interference. But the
wavelength continuity condition sometimes makes it impossible to satisfy the distinct
wavelength requirement, resulting in blocked connections. In these cases a WIXC can
be used to resolve the conflict.
To illustrate, consider the wavelength-routed network of Figure 2.28(a), operating with
two independently routed wavelengths λ1 and λ2 . Two optical connections are active,
with optical signal S2 (λ2 ) carrying a connection from station 2 to 4 on wavelength λ2 ,
and signal S3 (λ1 ) connecting station 3 to 6 on wavelength λ1 . Suppose a new connection
is required from 1 to 5. If a wavelength-continuous λ-channel is used for the new
connection, neither of the two wavelengths can be assigned to it without causing a
conflict. A solution to the problem is shown in Figure 2.28(b). Here, the new connection
is carried on signal S1 , with its wavelength changed from λ1 on the first part of its path
to λ2 on the last part. The central node of the network is the 2 × 2 WIXC, which was
shown previously in isolation in Figure 2.26. The port labels in that figure match the
corresponding points in Figure 2.28(b). The applications of WIXCs in more general
network settings are discussed in Chapter 6.
 Layered Architecture 67

3
WIXC S3(λ 1)

1 W 5
S2(λ 2)

2 4 6
( a)

3
2 S3( λ1)
S1( λ 1) S1(λ 2)
1 2
1 W 5
S2(λ 2)
1 WIXC

2 4 6
( b)

Figure 2.28 Wavelength-routed network.

2.4 Network Access Stations

The functions in the logical connection, transmission channel, and λ-channel layers
of our network architecture, shown in Figure 2.1, are implemented in the NASs. The
NAS uses the services of the optical path layer to provide LC services to end systems
or electronic switching equipment attached to its external ports. These devices will
generally belong to a client network. Thus, the access station is involved in two functions.
First, it interfaces the external LC ports to the optical transceivers. Second, it implements
(in the transceivers) the functions necessary to move signals between the electronic
and optical domains. These functions can become quite complex, especially when the
connections are multipoint. In this section we discuss the structure of the NAS, focusing
on the functionality required for supporting point-to-point LCs and optical layer signals.
Section 4.7 goes into more detail concerning the signal transmission and reception
functions of the NAS, and Section 3.2.1 deals with the functions specific to multipoint
connections: multiplexing, multicast, multiple access, and demultiplexing.
A typical access station is presented in Figure 2.29. The transmitting side consists of a
transmission processor (TP) with a number of LC input ports and transmission channel
output ports. The output ports of the TP are connected to optical transmitters (OTs), and
each transmitter is in turn connected to an outbound access fiber (using a signal combiner
or WMUX if there are several transmitters multiplexed on one fiber). The function of
the TP is to convert each logical signal to a transmission signal in a form suitable for
modulating a laser in the OT. In some cases, for example, on-off keying (OOK), where
a binary digital signal turns the laser on and off, there is little or no processing. In other
cases, the conversion may itself involve a modulation process at a subcarrier level –
for example, subcarrier phase-shift keying (PSK) and frequency-shift keying (FSK) for
digital signals and AM or FM for analog signals (see Section 5.2.2). These represent
68 Multiwavelength Optical Networks

a b
OT WMUX

TP eo
a b
OT

Logical c
Connection ONN
Ports
e d Access
OR WDMUX Fiber pair

RP eo
e d
OR

Figure 2.29 Network access station.

a second level of modulation preceding the laser modulation operation. Transmission

signals are present at points a in Figure 2.29, and optical signals are generated at points b,
with the multiplexed optical signals appearing at c. The TP may also perform coding
and multiplexing functions. For example, redundancy might be added for forward error
correction (FEC) on poor-quality optical connections. Also, the data streams from several
incoming LCs may be multiplexed onto a common λ-channel for purposes of grooming
or for distribution to several destinations via an optical multicast connection.
Assuming that it is tunable, each OT transmits an optical signal on a wavelength
assigned by the network manager. The OTs may be fabricated in the form of an integrated
array of fixed tuned lasers multiplexed onto the access fiber. If only one laser in the array
is active at a time, this combination acts as a single transmitter, tunable over the set
of wavelengths generated by the array. Alternatively, assuming that all wavelengths are
distinct, the system might transmit simultaneously on several fixed wavelengths, each
carrying its own LC.
On the receiving side, the optical signal arriving on an inbound access fiber is split
or wavelength demultiplexed and the resultant signals are passed on to the optical
receivers (ORs). If an OR is not preceded by a WDMUX, it will generally need its
own optical filter, which may be tunable to select an assigned λ-channel. The ORs
convert the optical power to electrical transmission signals, which are versions of the
original transmitted signals corrupted to some degree by noise and other transmission
impairments. Additional electronic operations are required in the reception processor
(RP) to convert the corrupted transmission signal to one or more logical signals. For
example, digital signals may be regenerated to produce a “clean” logical signal (possibly
with some bit errors), and analog signals may be filtered to remove noise, distortion, and
interference accumulated in the transmission process. Also, the inverse of the various
operations executed in the source TP must be performed in the destination RP, including
error correction, demultiplexing, electrical demodulation, and so on. In addition, when
several incoming signals from different sources are multiplexed on the same λ-channel on
 Layered Architecture 69

Logical Connection
NAS [A, B] NAS
A B

TP RP

a Transmission Channel e

Electrical
OT Optical Connection OR
Optical (A, B) 1

b 1 m -channel m 1 d

WMUX WDMUX

Optical Path
c A, B w 1

ONN ONN

w1
w2
ONN

Figure 2.30 Example of a logical connection between two NASs.

the same access fiber (as in many-to-one connections), the RPs and ORs may implement
the receiving side of a multiple access protocol (see Sections 3.2.1, 5.2, and 5.7). Finally,
signals received and processed in the RP may be reinserted into the network via a TP
in the same NAS instead of exiting into the logical layer. In this case the RP/TP pair
form a unit that we call an overlay processor. These devices are useful for signal
regeneration, wavelength conversion, grooming, and other functions in the optical layer
(see Section 2.5).
An example of the functions performed by a pair of NASs, A and B, in creating a
point-to-point LC is shown in Figure 2.30. An LC, denoted [ A, B] in the figure, originates
at one of the external input ports of station A. It may be multiplexed in the TP with other
LCs entering at other input ports. The LC is carried on a transmission channel (point a),
which modulates a laser in the OT, producing an optical signal carried on a λ-channel
at wavelength λ1 , appearing at point b. The OT is the point of origin of the optical
connection, denoted ( A, B)λ1 , which carries LC [ A, B] to station B. If there are other
OTs in the station, their signals may be multiplexed together, with the combined signals
appearing on the outbound access fiber at point c. This point is the origin of an optical
path, denoted A, B w1 to indicate that the path is carried on waveband w1 . As shown
70 Multiwavelength Optical Networks

in Figure 2.30, the optical path traverses three ONNs and terminates on the inbound
access fiber of station B. (Another optical path on waveband w2 , with a different source
and destination, is shown traversing the same network nodes.) At the entrance to the
station, optical connection (A, B)λ1 is demultiplexed from any other optical connections
that might have entered the station on the same access fiber. The resultant demultiplexed
signal (point d) is converted to a transmission channel in the OR and finally undergoes
any required additional electronic operations (e.g., regeneration and demultiplexing) in
the RP. The logical connection terminates at one of the output ports of station B.
In the previous example, a single bidirectional pair of access fibers served to connect
the access station to a port on an ONN. (Only the transmitting side of station A and the
receiving side of station B were shown.) Other configurations might use multiple pairs
of access fibers, each one connecting a transmitter–receiver pair to a separate port on
a network node or to separate network nodes. Many variants are possible. An example
is the WADM of Figure 2.23 attached to an NAS through several fiber pairs. (This is a
natural arrangement when the NAS and WADM are colocated.) One advantage of this
configuration is that there is no need for receiver tuning or wavelength demultiplexing
because only one wavelength reaches each receiver. Other access arrangements adapted
to special network applications appear in Section 3.3.2.
In the next two subsections, we discuss the basic functions of the OT and OR. More
detailed discussions appear in Sections 4.5, 4.6, and 4.8.

2.4.1 Transmitting Side

Figure 2.31 shows the details of the modulation process. (The points a and b corre-
spond to similarly labeled points in Figure 2.29.) Two possible OT configurations are
shown. In a directly modulated laser diode (see Figure 2.31[a]), the laser drive current is
modulated by the transmission signal s(t) appearing at point a, whereas in an externally
modulated transmitter, an electro-optic modulator is driven by the modulating waveform
s(t) to vary the optical signal emitted by the laser diode. (The mechanisms, advantages,
and disadvantages of direct and external modulation are discussed in Section 4.5.3.)

Laser Diode
a b Ss(f )
s(t)

(a) Direct Modulation

f
0 Bs
Se( )
Laser Diode
a External b
s(t)
Modulation
0 s
Be
(b) External Modulation (c) Spectrum

Figure 2.31 Optical transmitter.

 Layered Architecture 71

The optical field E(t) of a laser tuned to an optical frequency νs can be represented as
the real part of a complex signal:
E(t) = Re E(t)e j2πνs t .
 
(2.14)
The complex envelope E(t) can be written in the form

E(t) = 2I (t)e jϕ(t) , (2.15)
where I (t) = |E(t)|2 /2 is the instantaneous optical signal intensity (proportional to
power P) and ϕ(t) is the instantaneous phase. If the laser is intensity modulated by s(t),
we have
I (t) = I0 [1 + ms(t)] 0 < m ≤ 1, |s| ≤ 1, (2.16)
where I0 is the intensity of the unmodulated signal and m is the modulation index. In
analog or subcarrier modulation, s(t) is a continuously varying signal, and in digital
modulation (e.g., OOK), s(t) is in the form of a sequence of pulses. The phase ϕ
contains components due to the complexities of the modulation process (pure intensity
modulation is not achievable in practice) as well as random phase fluctuations due to the
laser itself.20 If the modulating signal s(t) is restricted to a bandwidth Bs , then s and E
have power spectral densities Ss ( f ) and Se (ν), respectively, as shown in Figure 2.31(c).
Note that only the positive frequencies of Se (ν) are shown.
Recall that in amplitude modulation (as opposed to intensity modulation), the band-
width of the modulated signal is just twice that of the baseband modulating signal.
However, in intensity modulation of a laser, the bandwidth Be of the optical field is
considerably more than this because of the square root in Equation (2.15) as well as the
extraneous phase and frequency modulation represented by ϕ(t).

2.4.2 Receiving Side

There are several common OR structures. The simplest is the direct detection receiver
shown in Figure 2.32. In the tunable version shown in Figure 2.32(a), the optical signal
is first passed through an optical filter and is then detected by a photodetector (PD) to
produce a photocurrent i(t) at point e. (Points d and e in Figure 2.32 correspond to the
same points in Figure 2.29.) The spectrum Se (ν) of the input optical field E(t) and the
spectrum Si ( f ) of the output photocurrent are shown in Figure 2.32(b). Note that two
λ-channels are shown on the access fiber, at optical frequencies ν1 and ν2 .21
The receiver is tuned to select the former by appropriately positioning its optical filter
transfer function HOF . (Of course, this selection is possible only if the two optical signal
spectra do not overlap.)
20
The randomness in the lasing process also produces fluctuations in intensity, called relative intensity noise
(RIN; see Section 4.6.4). The effect of RIN is omitted here.
21
The set of optical frequencies that reaches the input of a particular receiver depends on the selectivity
(if any) of the splitting/WDMUX device in front of it, as well as the waveband selectivity and setting of
the network node. Recall that a waveband-selective switching node directs all signal power in a selected
waveband to each desired output port. The waveband may contain a set of several closely spaced λ-channels,
from which the desired channels must be selected by the receiver.
72 Multiwavelength Optical Networks

PD

Signal OF i(t)
d e

(a) Tunable Direct Detection Receiver

Se( )

HOF

0 1 2

Si(f )

f
(b) Spectra

PD1

OF1
e

Signal i(t) Switch

d e
OFm
e

PDm
(c) Arrayed Receiver

Figure 2.32 Optical receivers.

Ideally the photodetector, acting as a photon counter, produces a photocurrent that is

an exact replica of the instantaneous optical intensity impinging on it; in other words, it
acts as a square-law detector of the optical field. More precisely,
R
i(t) = R I (t) = |E(t)|2 , (2.17)
2
where R is the responsivity of the photodetector. Thus, if intensity modulation is used
to generate the optical signal E(t), as in Equation (2.16), this receiver will recover
the transmission signal s(t). It is important to note, however, that a direct detection
receiver recovers no phase information, so it cannot be used to recover optical phase- or
frequency-modulated signals.22

22
By appropriately shaping the optical filter transfer function HOF , it is possible to introduce phase and
frequency discrimination in front of the photodetector, thereby making it possible to detect optically phase-
and frequency-modulated signals; e.g., optical frequency-shift keying and differential phase-shift keying
(see Section 4.7.1).
 Layered Architecture 73

PD
(t)
i(t) x(t)
d Coupler BPF ED
LO e

LO

(a)

Si (f)

HI

f
0 fI 2 LO
(b)
Sx(f)

(c)

Figure 2.33 Heterodyne receiver and spectra.

Precision tunable filters (especially widely and rapidly tunable filters) are difficult
to fabricate and hence are costly. Another way of producing the equivalent of receiver
tunability is to use an array of m photodetectors, each preceded by a fixed optical filter
tuned to a different frequency, as in Figure 2.32(c). Using an m × 1 (electronic) switch,
an equivalent receiver tunable over the m filter frequencies is obtained (point e′ ). Without
the switch, the resultant system receives m simultaneous signals (point e).
As indicated, the direct detection receiver normally cannot recover frequency and
phase information. Furthermore, it has less-than-ideal noise discrimination properties.
These problems are essentially absent in the coherent heterodyne receiver (see Section 4.8
for noise performance). In principle, heterodyne detection gives the best possible recep-
tion and is the most versatile. Modeled after the ubiquitous heterodyne radio receiver,
the heterodyne approach is shown in Figure 2.33. The incoming optical signal E(t) at
frequency νs is first combined in a passive coupler with a signal at optical frequency
νLO , generated by a local oscillator, LO (a tunable laser). The combined signal is de-
tected by the photodetector. Because the photodetector behaves as a square-law device,
the photocurrent i(t) contains a component at intermediate frequency f I = νs − νLO ,
where the local oscillator frequency νLO is chosen very close to the signal frequency,
giving an f I at a point in the radio frequency spectrum that is chosen for convenience
in subsequent electronic signal processing. The signal i(t) is passed through a bandpass
filter (BPF) with transfer function HI centered at f I , and the resultant filtered signal is
detected again by an electrical detector to produce the desired output signal x(t).
74 Multiwavelength Optical Networks

Next we show that the heterodyne receiver eliminates the need for a tunable optical
filter. Consider the case when the λ-channel at optical frequency ν1 is to be selected
from the pair of incoming optical signals at neighboring frequencies ν1 and ν2 . Assuming
that the polarizations of the two input signals and the local oscillator signal are all aligned,
the combined optical field at the input to the photodetector will be

E(t) = Re E 1 e j2πν1 t + E 2 e j2πν2 t + E L e j2πνL O t ,

 
(2.18)

where

E i (t) = |E i (t)|e jϕi (t) (2.19)

and

Ii (t) = |E i (t)|2/2. (2.20)

Now, assuming that the local oscillator is tuned to frequency νLO = ν1 − f I and its
power is much greater than the incoming signal power, the photocurrent at the output of
the detector is approximately
 
i(t) ∼ I L + I1 I L cos(2π f I t + ϕ1 − ϕ L ) + I2 I L cos(2π [ν2 − ν L O ]t + ϕ2 − ϕ L ),
(2.21)
where low-power terms have been neglected.23
The positive frequency spectrum Si ( f ) of the photocurrent is shown in Figure 2.33(b).
Note that the phase of the desired signal ϕ1 is present in the photocurrent of
Equation (2.21), so that phase- and frequency-modulated information can be recov-
ered at the receiver.
The bandpass filter selects the signal centered at frequency f I and rejects the other
signal. The final electrical direct detection stage recovers the desired signal x(t) ∼ E 1 (t)
at point e, with the spectrum Sx ( f ), as shown in Figure 2.33(c).
The heterodyne receiver has three advantages over the direct detection receiver:

1. It has inherently better noise reduction properties, producing better SNRs or, equiv-
alently, BERs.
2. It can demodulate all types of optical signals, including those using frequency and
phase modulation.
3. It is tunable without requiring a tunable optical filter.

The primary (and significant) disadvantage of this receiver is that it is complex and
expensive to build.

2.5 Overlay Processors

As indicated in previous sections, an optical WAN eventually “runs out of steam”

if all connections are required to be optically transparent end to end. One reason is

23
Equation (2.21) is an idealized expression that ignores the effects of inexact polarization alignment, local
oscillator phase noise, and other problems in physical realization.
 Layered Architecture 75

that the “reach” of an optical connection is limited by transmission impairments such

as noise, dispersion, nonlinear distortion, and optical node cross-talk, all of which
accumulate along a transparent path. Even when reach is not the issue, there are many
cases when some additional functionality in the optical layer serves to enhance network
performance. An example is the use of an opaque switch in an otherwise transparent
network to achieve the combined benefits of signal regeneration, wavelength conversion,
and switching, all in one network element. In all of these situations, some additional
(nonlinear) functionality must be inserted at the upper edge of the physical layer (in the
transmission channel sublayer of Figure 2.1).
We use the term overlay processor (OLP) as a catchall for devices that perform any of
these functions. Typically, the OLP interfaces to the optical transceivers but not always
to the logical layers. Overlay processing is generally format dependent, so transparency
is lost when signals traverse OLPs. A typical network access station containing an OLP
is shown as station U in Figure 2.34. Note that the OLP replaces the TPs and RPs, and
in the case shown it has no interface with the outside world. (For simplicity only one OT
and OR are shown, without the usual WMUX and WDMUX.) The purpose of the OLP
is merely to process signals in electronic form at the outputs of the optical receivers and
reinsert the processed signals into the network via the optical transmitters. There are
many possible functions for such devices. We provide two examples.

λj

OT

OLP

OR

λi

Figure 2.34 Overlay processor.

76 Multiwavelength Optical Networks

2.5.1 Regeneration
Consider establishing a logical connection between stations X and Y in Figure 2.34, in
which the stations are beyond each others’ reach in the sense that the quality of the signal
would be unacceptable if it were transmitted purely optically from X to Y. A solution
to the problem is to relay the signal through station U, which contains an OLP whose
function is to regenerate the signal so that a clean version of it is reinserted into the
network and retransmitted from U to Y. The resultant LC from X to Y is thus carried on
two concatenated optical connections.
The OLP accepts the signal in electrical form from the OR, detects (reshapes) the bit
stream, retimes it, and passes the regenerated version of the signal on to the OT. These
are the functions typically performed by a regenerative repeater (see Section 4.11). Of
course, this type of regeneration is a highly nonlinear operation, and it is generally
confined to signals operating with a well-defined format, say a 1-Gbps digital bit stream
carried on a binary intensity-modulated transmission channel. As we have seen, network
elements of this type, which combine an optical receiver, some degree of regeneration,
and an optical transmitter, are called transponders. If a transponder transmits on a
different wavelength than that of the received signal, it is also performing a wavelength
interchange operation.
Although there may be some finite BER accompanying this process, the degradation
due to bit errors is far less than what would be experienced by trying to extend the purely
optical path beyond its reach. More elaborate variants of the regenerator OLP might
include processors that support format conversion, error correction, and other signal
transformations at the transmission channel level.

2.5.2 Wavelength Interchange

In routing optical connections through a transparent optical layer containing no wave-
length interchangers, wavelength continuity is required end to end. Thus, even if two
stations are within reach of each other optically, it may be impossible to make a con-
nection between them if no wavelength-continuous optical path is free to support it
without interfering with other connections. (An example of this problem was shown
in Figure 2.28, for which optical wavelength interchange was presented as a possible
solution.) The equivalent result may be obtained by using a wavelength-interchanging
OLP.
Referring again to Figure 2.34, suppose a connection is required from station X to
Y, but no wavelength-continuous path exists. If paths can be found from X to U and
from U to Y, say on wavelengths λi and λ j , respectively, then an end-to-end path can
be relayed through U using the appropriate wavelengths. First, an optical connection is
established from X to U using λi and from U to Y using λ j . This carries an optical
signal from X to U, which is converted to electronic form in the receiver of station U
and is passed through the OLP to the OT in the station. The transmitter, tuned to λ j ,
places the resultant signal on the optical path to station Y. The function being performed
in the station is equivalent to wavelength conversion, but it is accomplished in the
 Layered Architecture 77

electronic domain. Although the operation of the OLP in this case is minimal24 – simply
connecting the receiver and transmitter back to back – the effect on network operation
may be substantial.
As is shown in Section 6.4, significant enhancement of network performance through
reduced connection blocking is sometimes achieved by using wavelength interchange
in an otherwise transparent optical network. As in the case of regenerative repeaters,
wavelength-interchanging overlay processors are format dependent, but in return for
this dependence they can provide signal regeneration functions as a by-product of the
conversion function. (We have already seen that signal regenerators can perform a
wavelength interchange function as a by-product of regeneration.)
In large, geographically dispersed transparent networks, enhancement of optical layer
performance can be realized by placing a number of NASs equipped with versatile OLPs
at strategic points throughout the network. If these stations contain multiple transceivers,
they can act as optical layer “servers” available to provide regeneration, wavelength con-
version, and possibly other optical layer enhancement functions to several simultaneous
connections. The term translucent has been coined to describe networks of this type.
These functions are especially important in cases when two or more isolated and inde-
pendently managed optical networks are to be concatenated by extending connections
from one network to the other. In this case the OLP performing network concatenation
would have separate access links to the different networks and would be controlled
jointly by them both.

2.6 Logical Network Overlays

Moving up to the logical layer in the multiwavelength network architecture, there are
many possible logical network overlays that can be used for implementing various user
services over an optical network. These overlays are typically client networks that will
themselves have layered architectures. Without entering into the details of the logical
network structure, a generic arrangement for a logically routed network is shown in
Figure 2.35. (The rationale for LRNs was explained in Section 1.4.) The hexagons in the
figure are logical switching nodes, which communicate with each other through logical
links, forming a logical topology superimposed on the physical layer. By showing the
logical topology as a collection of nodes and links in the logical layer, we simplify the
representation of the logical network, concealing unnecessary detail in the supporting
optical network. The end-to-end connections in an LN constitute the virtual connections
in the top layer of the architecture of Figure 2.1.
A blown-up view of a typical LSN appears in Figure 2.35, which shows a logi-
cal switch (LS) interfaced to an ONN in the optical network through an NAS. The
LS may also be interfaced to end systems through ports external to the network.
An LS without interfaces to end systems performs a transport (essentially routing)
function only, whereas one with external interfaces provides user access as well. For
24
The OLP may also perform some linear filtering in the case of analog signals or regenerating in the case of
digital signals.
78 Multiwavelength Optical Networks

LS
Logical Switching
Node
NAS

ONN

LSN

Logical Layer

Physical Layer

LSN Including
NAS

Figure 2.35 Logically routed network.

example, the LN overlay may be an IP network whose LSs (IP routers) are a mix of core
routers (without an external interface) or edge routers (with an external interface).
In Section 2.3.2.4, we described several versions of optical cross-connects (illustrated
in Figures 4.86 through 4.88). Each of these can be used as an implementation of
the ONN/NAS combination supporting the LSN of Figure 2.35. Each performs “pass-
through” functions, routing optical channels through it, and add/drop functions shunting
“local” traffic to and from the attached LS. However, there are differences among the
three switches in terms of the granularity of the switching operations and the functions
performed on the transiting traffic. The transparent switch performs the routing function
at a coarse level of granularity: either wavebands or λ-channels. The O-O-O switch
operates at a λ-channel level, and the O-E-O switch may operate at a sub-λ-channel
level if it demultiplexes time division-multiplexed transmission channels within a
λ-channel before switching them, a common feature of this type of switch. (In this case
the O-E-O switch is capable of rearranging these subchannels as part of the routing func-
tion for purposes of grooming.) As noted, the opaque switches also perform regeneration
and wavelength translation functions. Thus, as far as the transiting traffic is concerned
the switches operate as either transparent ONNs or overlay processors, as described in
Section 2.5. With respect to the add/drop traffic, each can operate as a full-fledged LSN
 Layered Architecture 79

belonging to a SONET, ATM , IP, or other LN, as suggested in Figures 4.87 and 4.88. At
the logical layer, the LS may perform “edge” functions, terminating virtual connections
to end systems, transit functions such as routing virtual connections over logical paths,
or other functions such as logical layer grooming, fault recovery, traffic engineering,
and so on.
By showing the LS interfaced to the physical layer via an NAS we have implicitly
assumed that it is electronic. In future networks the LS may be an optical packet or
burst switch, in which case it would access the physical layer directly through the ONN
(see Chapter 10). Effectively, this moves the boundary between optics and electronics
upward to the top of the logical layer of the architecture in Figure 2.1.
Each logical link in the LN overlay is realized as a unidirectional or bidirectional
logical connection (see Figure 2.2) between a pair of source and destination access
stations, and each LC is in turn carried on an optical connection, routed through the
fibers on an optical path. (Typically, the optical path may involve several optical hops.)
The logical and physical topologies are independent of each other, with each logical link
being embedded in a dedicated path in the physical topology. A logical path supporting
a virtual connection is reconfigurable by changing the LSN settings in the LN. Similarly,
the logical topology is reconfigurable by changing the ONN settings in the physical
layer. Thus, if network reconfiguration is required because of changing traffic patterns
or equipment failures, it can take place in the logical or physical layer or both.
The generic LRN shown in Figure 2.35 may take many specific forms, depending on
the structure of the LS. Consistent with current and evolving standards, we now briefly
describe four important types of LRNs in use today: SONET networks, ATM networks,
IP networks, and MPLS networks.

2.6.1 SONET Networks

SONET networks are based on transmission links carrying digital synchronous trans-
port signals (STS) at standard rates that range from STS-1 (51.48 Mbps) to STS-192
(9.95324 Gbps) and higher.25 These signals are supported by corresponding optical
carriers, OC-1 to OC-192, that are constructed by scrambling the STS-n signal and
converting it to optical form. The basic STS-1 is carried in a 125-µs frame containing
transport overhead plus the information payload. Higher rate STS-n signals are formed
by byte-interleaving STS-1s. The frame structure is fairly complex, incorporating over-
head for both communications and maintenance functions.26 In terms of our generic
layered architecture the STS-n (electrical) signal is associated with a logical connection
and the OC-n (optical signal) is carried on an optical connection.
The logical switch in a SONET network is a SONET digital cross-connect system
(DCS) or add/drop multiplexer (ADM). The DCS and ADM play the same roles in the

25
In Europe and Asia the North American SONET standard is replaced by a more or less equivalent SDH
(synchronous digital hierarchy) standard.
26
One of the most important features of the SONET standard is a comprehensive set of maintenance and
protection switching functions implemented using the transport overhead; see Chapter 8, Appendix F, and
[Bernstein+04].
80 Multiwavelength Optical Networks

logical layer that a WSXC and WADM play in the physical layer. In current SONET
networks, the DCSs are interconnected through internodal point-to-point fiber links
carrying OC-n signals and are connected to local equipment through electronic add/drop
ports. The function of a DCS is to demultiplex, switch (route), and remultiplex the signals
with which it interfaces. An ADM performs operations similar to a DCS except that the
ADM has only two internodal ports, like the WADM of Figure 2.23.
The inter-DCS connections (corresponding to the logical links in Figure 2.35) are
operated as dedicated connections, and the signals carried on the internodal links range
from STS-1 to STS-48 and higher. At its external local equipment interface, the DCS
or ADM typically exchanges digital signals with end systems at DS-1 (1.544 Mbps)
or DS-3 (44.736 Mbps) rates. The end systems using these signals may be located in
telecommunications carrier Central Offices or on the premises of large users. A DCS
or ADM demultiplexes signals on its inbound internodal links down to a low rate (e.g.,
STS-1 or STS-3), sorts and cross-connects these signals (dropping some signals to end
systems), and remultiplexes the cross-connected signals (including signals added from
end systems) onto the outbound internodal links. Thus each internodal logical link oper-
ating at a high bit rate typically carries many time division multiplexed lower rate signals.
The DCS performs a grooming function on these signals as well as a routing function.
An end-to-end connection terminating at externally attached end systems will typically
traverse several SONET DCSs, resulting in a multiple-hop source-destination path. The
SONET network typically provisions and holds these connections over long time periods.
In currently operational telecommunication networks the SONET network uses point-
to-point optical fiber transmission links for its inter-DCS connections, without the in-
tervention of a reconfigurable physical layer. This is shown in Figure 2.36(a), in which
the fibers are interfaced directly to the DCS through its own optical transceivers. In the
first implementations these transmission links were single wavelength, carrying only
one bidirectional logical link. They were therefore grossly underutilized. However, with
the trend toward WDM transmission, the transmission links are now being operated
on multiple wavelengths, so they can carry several parallel logical connections either
from the same or different electronic equipment. In this case, the link terminating equip-
ment, which interfaces the DCS (or similar equipment) to the fiber transmission link,
is called a WDM terminal, wavelength terminal multiplexer, or WDM transport system.
Typically, the interconnections between the link terminating equipment and the elec-
tronic switching equipment are optical. For example, a SONET DCS might transmit
and receive OC-n optical signals in the 1300-nm band on its input/output ports. The
output signals from the DCS are connected to WDM terminals via a short-reach fiber.
Each terminal then converts these signals to electrical form and remodulates them onto
another optical carrier at a wavelength (normally in the 1550-nm range) suitable for long-
distance WDM transmission. The reverse operations are carried out at the receiving end
of the link transmission link. This arrangement is similar to the O-E-O cross-connect in
Figure 4.87. The intermediate connection between the DCS and the short-reach fiber is
made via a 1300-nm transceiver, and the connection from the short-reach fiber to the
long-haul transmission link requires a transponder of the form previously described in
Section 2.3.
 Layered Architecture 81

DCS

DCS NAS

(a) (b)

WDM Terminal
DCS SONET DCS

Figure 2.36 SONET DCS.

One difficulty with this arrangement is that there is no possibility for optically re-
configuring the SONET network, either for load redistribution or for fault recovery.
Furthermore, if the equipment interfaced to the WDM terminal does not generate a
throughput comparable with the capacity of the transmission link, much of the fiber ca-
pacity will be wasted. Finally, each WDM terminal can only connect the DCS to one other
switching node. A more flexible and efficient arrangement is shown in Figure 2.36(b).
An NAS, together with an ONN (e.g., a wavelength-selective cross-connect), replaces
the WDM terminals on the three links in Figure 2.36(a). Now, the logical connections
from the DCS can be routed to different destinations under the control of the optical
node. By superimposing the DCS on a purely optical network, additional flexibility is
introduced, with optical multiplexing and logical network reconfiguration made possible
by control in the optical layer.
It is interesting to view the functions of the SONET switch in terms of granularity. The
signals it exchanges with the optical network are highly aggregated and well adapted
to the capacity of the optical transmission channels. Those exchanged with the end
systems are typically of a much finer granularity, adapted to the needs of the attached
telecommunications equipment. In the DCS, information channels of relatively fine
granularity are sorted individually and then reaggregrated for transport on the logical
links. This illustrates the use of electronics (in the DCS) to provide the fine-granularity
grooming needed for high connectivity and transport efficiency while using optics to
carry coarse-granularity traffic at high throughputs.

2.6.2 ATM Networks

An ATM logical network is very similar to a SONET network except that the basic
entities processed by the LSNs are cells rather than synchronous bit streams. The logical
switch in these networks is an ATM (cell) switch. ATM was developed to support
broadband services that require a wide range of bandwidths and quality of service.
82 Multiwavelength Optical Networks

VPI/VCI, Header (5 bytes)

Data Payload (48 bytes)

Figure 2.37 ATM cell format.

It uses the ATM cell as the basic information unit: a fixed-length structure that contains
53 bytes, of which 5 constitute the header and the rest are data (Figure 2.37). The header
contains a virtual channel identifier (VCI) and virtual path identifier (VPI), assigned as
part of a call setup process, as well as header error control, flow control, and priority
and other information.
Typically an ATM network is configured in the form of a virtual topology that con-
sists of dedicated virtual paths (VPs), acting as virtual internodal links joining pairs of
ATM switches. (Each VP may traverse several ATM switches between its end points.)
End users request demand-assigned virtual channels (VCs) to support each applica-
tion, with the bandwidth and other characteristics of the VC designed to provide the
required quality of service for that application. The VPs are sized to carry many asyn-
chronously multiplexed VCs, which are the entities carrying cells end to end in an ATM
network. Because the cells are multiplexed asynchronously on the internodal links, the
VC and VP information is necessary for sorting and routing individual cells within
the ATM switches. Each cell header is read by the switch, and the cell is directed
to the output port specified by the VCI, the VPI, and the routing table stored in the
switch.
As shown in Figure 2.38, the ATM switches are joined together through internodal
logical links and are interfaced to end systems through external ports. An end system
could be a workstation, a gateway to an access network, a supercomputer, a switch in a
client network carried over ATM (e.g., an IP router), or any other device equipped with
an ATM interface. As in the SONET case, a typical connection between end systems
(carried on a VC) will traverse several ATM switches between source and destination,
resulting in a multihop logical path. The internodal links connecting the ATM switches
can be realized in various ways. Figure 2.38 shows three possibilities.
In Figure 2.38(a), the internodal ports on the switch are connected directly to point-to-
point digital transmission links, which might be fiber links terminated with the necessary
optical transceivers. These links form a logical topology that cannot be reconfigured
without physically changing the connections. (This is analogous to the SONET DCS
network connection of Figure 2.36[a].)
An alternative is to use a hybrid ATM over SONET structure as shown in
Figure 2.38(b). Now the ATM switch acts as an end system attached to the external
ports of a SONET DCS, so the ATM network is a client network of the SONET network
and each ATM internodal logical connection is carried on a SONET STS routed through
the DCS to another ATM switch. (The network of ATM switches may represent just one
 Layered Architecture 83

ATM

ATM NAS
ATM

(a) (b) (c)

ATM or ATM ATM switch

S SONET DCS
including optical
transceivers

Figure 2.38 ATM switch connections.

of many client networks served by the SONET infrastructure, and the SONET network
may in turn be one client LN supported by a larger optical network infrastructure.) In
the latter case reconfiguration is possible at the SONET level by changing the routing
within the DCS, and in the physical layer of the optical network by reconfiguring the
OXCs.
Figure 2.38(c) shows a third alternative, in which the ATM switch accesses an optical
network directly through its own NAS. This is analogous to the SONET DCS configu-
ration shown in Figure 2.36(b). As in the SONET case, a logical topology for the ATM
network is constructed by creating the desired logical connections among the NASs, and
reconfiguration is possible through control in the optical layer.

2.6.3 IP Networks
An IP logical network is a packet-switched LN where the information units are IP
packets and the LSNs are IP routers (packet switches). IP packets are of variable length
and generally much longer than an ATM cell. The function of an IP router is to examine
incoming packets, determine destination addresses, compare them with the contents of a
routing table stored in the router, and then forward the packets to the appropriate output
interface.
As mentioned, logical networks typically have their own layered architectures. The
Internet protocol stack, called TCP/IP, is shown in Figure 2.39,27 for comparison with
the Open Systems Interconnection (OSI) Reference Model. Layers 2 and 3 in these
architectures concern us the most. Layer 2, called the network interface layer in TCP/IP
27
The Transmission Control Protocol (TCP) is at level 4 in the figure and is used for reliable transport over
an end-to-end path that may consist of multiple links, and the Internet Protocol (IP) comprises layers 2
and 3.
84 Multiwavelength Optical Networks

7 Application

6 Presentation Application

5 Session

4 Transport 4 Transport

3 Network 3 Internet

2 Data link control 2 Network interface

1 Physical

OSI TCP/IP

Figure 2.39 TCP/IP and OSI.

and the data link control (DLC) layer in OSI, has the function of ensuring reliable and
ordered delivery of data frames over a point-to-point link. It uses the services of the
physical layer, which provides for transmission of the underlying bit stream on a point-
to-point physical medium such as a wire or fiber. Layer 3, the Internet layer in TCP/IP
and the network layer in OSI, is concerned with routing – in our case, routing in the
logical layer of Figure 2.35. As we consider aspects of network control in later chapters,
these lower layers of the protocol architectures will reappear in the discussions.
The Internet is a highly dynamic structure composed of many individual intercon-
nected IP networks with frequently changing topologies. Thus, one of the fundamental
capabilities of IP routers is the discovery of the current network topology and the creation
of routing tables that adjust themselves automatically to changes in network topologies
caused by link and equipment modifications and failures and by addition or deletion of
nodes, fibers, or wavelengths between routers.
Networks of IP routers can be configured over optical infrastructures using the various
options suggested previously for SONET and ATM networks, with the additional option
of running the IP network as a client on top of ATM switches and/or SONET DCSs.

2.6.4 MPLS and Its Extensions

As the number of users, the traffic volume, and the variety of high-speed real-time
services increase on IP networks, efforts are expanding to make the IP protocols and
other packet- and cell-switched systems more effective. In IP networks in particular,
the processing requirements together with the growing demand for service differenti-
ation present a challenge. A development that attacks these problems is Multiprotocol
Label Switching (MPLS). It is a method of streamlining the processing of packet- and
 Layered Architecture 85

cell-switched protocols within the logical network, enforcing specialized treatment of

differentiated services by implementing CoS and QoS requirements and providing a tool
for optimizing network performance. MPLS manages flows by appending simple fixed
length labels to packets, which are used in label-switching routers (LSRs) for expediting
packet forwarding. The labels are assigned during a call setup process much like that
used for ATM. A label on a packet inbound to an LSR is compared to a routing table
that makes a forwarding decision and at the same time swaps the inbound label for
an outbound label. Thus, the sequence of labels and the LSR routing tables define vir-
tual paths for the packets. This label-switching procedure, reminiscent of those used in
the earliest packet-switched data networks of the 1970s [Tymes81], facilitates efficient
packet processing and generalizes to many other applications. One of these is GM-
PLS, which adapts the control functions in MPLS to other types of networks, including
MWNs (see Chapter 9 for more details). Label switching is also being tried in optical
packet-switching networks (see Chapter 10).
We explore logically-routed network (LRN) structures in considerable detail in
Chapter 7. As shown there, LRNs can be constructed with logical connectivity that
is more general than a collection of point-to-point logical connections. Through the use
of optical multicast connections in the underlying physical layer, logical hypernets can
be built with logical topologies that consist of multipoint logical connections. Hypernets
are ideal structures for supporting multicast virtual connections, which are becoming
increasingly important in many Internet and multimedia applications.

2.7 Summary

Having presented the essential resources of the networks being explored in this book,
we can now summarize their relationship to network functionality. Figure 2.40 shows a
taxonomy of multiwavelength networks, classified according to
r Physical topology
r Optical connectivity
r Optical node functionality
r Station and overlay functionality
r Network architecture
The physical topologies break down into trees (no alternate paths exist) or a general
topology (providing alternate paths and hence better response to changing loads and
faults). The optical connectivity may be static or controllable, depending on the func-
tionality of the optical network nodes. In the simplest case, the nodes are static coupling
devices with no spectral selectivity. Better performance, including wavelength reuse, is
achieved with static wavelength- and waveband-selective nodes, and the most versatile
networks contain controllable nodes, ranging from space switches to waveband-space-
time switches, wavelength-interchanging cross-connects, and optical packet switches.
The space and waveband switches may operate in a simple permutation mode or may
provide generalized switching functions to support multipoint optical connectivity.
86 Multiwavelength Optical Networks

Physical Topology Tree General

Optical Connectivity Static Controllable

Optical Node Wavelength w-s, w-s-t

Static Coupler Space Switch WIXC OPS
Functionality Router Switch

Permutation Generalized LDC

Transceiver TP/RP Overlay Processor

Station and Overlay Fixed/Tunable/ MUX/MA Regeneration, Logical Switch
Functionality Arrayed MAC/PS Wavelength
Conversion

Purely Logically
Network Architecture Optical Routed

Broadcast/ Wavelength Linear

Multihop Hypernet
Select Routed Lightwave

Optically Packet,
Burst, Label-Switched

Figure 2.40 Taxonomy of multiwavelength networks.

The station and overlay level in the taxonomy of Figure 2.40 includes equipment
ranging from optoelectronic to purely electronic. The functionality of this equipment –
the NASs, overlay processors, and logical switches – complements the capabilities of
the resources in the optical network nodes. Optoelectronic station equipment ranges
from a single fixed-tuned optical transceiver to multiwavelength transmitter and receiver
arrays, with the latter configuration providing multiple simultaneous optical connec-
tions. In addition, the electronic TPs and RPs in a station may be equipped to execute
a wide variety of functions that increase the communication efficiency, fidelity, and
connectivity of the optical connection. Electronic multiplexing/demultiplexing in the
TP/RPs can be used for grooming low-granularity channels onto higher rate transmis-
sion channels. FEC can be implemented in the TP/RPs to improve the bit error rate
on a logical channel. In a network with a controllable optical path layer, a multipoint
connection can be constructed through the combined efforts of the TP/RPs residing
in the NASs and generalized switches residing in the ONNs. First, a multipoint opti-
cal path is set up by using splitting/combining functions in the generalized switches;
then a multipoint optical connection is created on the optical path through appropriate
multiplexing/multiaccess functions in the TP/RPs. A particularly flexible means of op-
erating multipoint connections is through packet switching in the optical layer, imple-
mented through a media access control (MAC) protocol executed within the TPs and RPs.
Network access stations might also contain overlay processors executing signal
regeneration to extend optical reach and wavelength interchange to improve network
 Layered Architecture 87

performance. They might also contain some simple opaque switching equipment to
achieve regeneration, wavelength interchange, and grooming in a single integrated sub-
system. The OLP may be viewed as a “shim” slid in between the physical layer and the
logical layer of Figure 2.1(a). It has neither the transparency of the lower optical layers nor
the functionality of the upper logical layers. Full logical layer functionality is achieved by
the (electronic) logical switch (e.g., DCS, ATM switch, or IP router) included at this
level of the network equipment. (Opaque optical switches, which have not been
explicitly shown in the taxonomy, would lie somewhere between the optical node level
and the overlay level.)
Network architectures built on these components may be purely optical and transpar-
ent, containing only the transparent physical layer in Figure 2.1(a), they may be opaque,
containing one or more opaque OXCs, or they may be hybrid, containing one or more log-
ically routed network overlays. The transparent networks range from the simple broadcast
star, with a tree topology and containing no control in the ONNs, to wavelength-routed
and linear lightwave networks, with general topologies and substantial controllability.
In wavelength-routed networks, each optical connection is point-to-point and is routed
independently through the network on its own λ-channel. In the linear lightwave net-
work, multicast optical connections that provide the necessary optical infrastructure for
hypernet LRNs are supported. In optical networks containing opaque OXCs signals are
regenerated in the ONNs, which means that wavelength translation is possible, and the
network is no longer transparent to speed and formatting of optical signals. Furthermore,
overhead information in the data streams can be used for control and management.
Logical switching in hybrid networks equipped with electronic overlays provides
additional intelligence, greatly increasing the potential connectivity of the network.
These networks may be based on point-to-point logical connection topologies, in which
case they are called multihop networks, or they may have hypernet structures, which
achieve greater connectivity by combining logical switching in an electronic overlay
with multipoint optical connections in the physical layer. As discussed in Section 2.6,
it is also possible to implement logical switching in a purely optical but nontransparent
architecture using optical packet, burst, or label-switching techniques. In this way, the
connectivity benefits of logical switching are combined with the throughput advantage
of optical processing.
Subsequent chapters explore each of these major classes of networks in detail.

2.8 Problems

1 List and discuss the factors that limit channel spacing in wavelength-routed networks
and waveband-routed networks.
2 What are the factors limiting throughput in the following types of networks:
(a) A star coupler-based LAN
(b) A purely optical wavelength-routed WAN
(c) A WAN consisting of electronic switches joined by point-to-point WDM links
88 Multiwavelength Optical Networks

3 For the 8 × 8 Benes switch:

(a) If all 2 × 2 switching elements are in the bar state, determine the connection state
of the switch.
(b) Find a switch setting (a set of device states) to produce the following connection
state:

Input Output
1 7
2 5
3 4
4 8
5 1
6 2
7 6
8 3

4 Compare the power distribution rule for the star coupler with that of the folded bus.
(Assume in both cases that there are no excess losses in the devices or attenuation in the
fibers.)
(a) Show that in the 16 × 16 star coupler of Figure 2.9 each output port receives the
sum of the powers entering at all input ports attenuated by a factor of 16.
(b) Show that in a folded bus with 16 stations, configured to act as a star coupler, the
best one can achieve is an attenuation factor of 256. (Hint: See Section 4.10.2.)
(c) Generalize this to the n × n case.
5 Prove that an n × n Benes switch uses n log2 n − n/2 binary elements.
6 Compare the number of cross-points (binary switching elements) for the n × n
crossbar and Benes switches with log2 (n!) for n = 4, 8, 16.
7 Compare the number of cross-points in a 16 × 16 strictly nonblocking Clos switch
to those in the same size Benes switch. Assume that the parameters in the Clos switch
are p = k = 4 and that each smaller switch is realized as a crossbar.
8 Consider the 4 × 4 P-I crossbar switch in Figure 4.73. The switch fabric consists of
waveguides placed at 45◦ angles to the horizontal. Each 2 × 2 switching element can
be in either the cross state, where no connection is made between the two waveguides
crossing it, or the bar state where the waveguide connection is made whose direction is
indicated by the arrow on the switching element.
Show by example that this is a wide-sense nonblocking switch; i.e., find a set of switch
states that will block the switch. Then, give an algorithm for routing connections through
the switch in a nonblocking manner. Generalize this to the n × n case.
9 This problem concerns cost trade-offs in WSS realization using either the three-
stage architecture of Figure 2.21 or using MWSs as the elementary switching devices.
Assume that the device costs are Ccc for a 2 × 2 controllable coupler (or on/off device),
 Layered Architecture 89

mCmux for an m-wavelength MUX or DMUX, nC p for an n-fold passive splitter or

combiner, and mCmws for an m-wavelength 2 × 2 MWS.
(a) For a four-waveband 8 × 8 permutation switch using a Benes fabric, find the range
of values of Cmws over which the MWS realization is more economical than the
three-stage realization. Express your answer in terms of the other cost parameters.
(b) Repeat the previous part for a generalized switch fabric of the type shown in
Figure 2.18.
(c) Generalize the previous results to m-waveband n × n switches.
10 Consider a 48 × 48 permutation switch connecting three cables, each containing
16 bidirectional fiber pairs.
(a) Indicate how many connection states are required, propose a possible strictly non-
blocking switch design based on binary switching elements, and indicate how many
binary elements it uses.
(b) Now, suppose that the switch connections are required to be made only on a fiber-to-
cable basis without loopback connections. That is, each inbound fiber in one cable
must be connected to an outbound fiber in a prescribed (different) cable, but it does
not matter which outbound fiber is used. Again, indicate how many connection states
are required, propose a possible strictly nonblocking switch design, and indicate how
many binary elements it uses (hopefully fewer than in the previous case). In counting
connection states, do not distinguish between different fibers in the output cables.
11 Show how the physical constraints of the various switch fabrics lead to
Equations (2.8), (2.9), (2.10), (2.11), and (2.12).
12 In the WIXC of Figure 2.27, assume that the inputs are as shown in the figure. Find
the device states (i.e., the settings of the three WSXCs) to send S 2 (λ1 ) and S 3 (λ2 ) to the
upper output port and S 1 (λ1 ) to the lower output port.
13 Show a design for a 2 × 2 four-wavelength WIXC using five 2 × 2 WSSs and four
wavelength interchangers.
14 List as many functions as you can for OLPs, and suggest realizations.
15 Invent some new forms of LRN overlays on optical networks.
16 Discuss some of the problems of managing and controlling a network that is
independently reconfigurable at several different layers: virtual, logical, and physical.
For example, how should fault recovery be managed? How should reconfiguration be
managed in the face of congestion?

Bibliography

[Antoniades+96] N. Antoniades, K. Bala, S. J. B. Yoo, and G. Ellinas. A parametric wavelength

interchanging cross-connect (WIXC) architecture. IEEE Photon. Technol. Lett., 8(10):1382–
1384, 1996.
90 Multiwavelength Optical Networks

[Barry+93] R. A. Barry and P. A. Humblet. Latin routers: Design and implementation. IEEE/OSA
J. Lightwave Technol., 11(5/6):891–899, 1993.
[Benes65] V. E. Benes. Mathematical Theory of Connecting Networks and Telephone Traffic.
New York: Academic Press 1965.
[Bernstein+04] G. Bernstein, B. Rajagopalan, and D. Saha. Optical Network Control: Architec-
ture, Protocols, and Standards. Boston, MA: Addison-Wesley 2004.
[Hinton93] H. S. Hinton. An Introduction to Photonic Switching Fabrics. New York: Plenum
Press 1993.
[Hui90] J. Y. Hui. Switching and Traffic Theory for Integrated Broadband Networks. Norwell,
MA: Kluwer Academic Publishers 1990.
[ITU-T00] ITU-T. Generic Functional Architecture of Transport Networks. ITU-T Recommen-
dation G.805, 2000.
[ITU-T01] ITU-T. Architecture of Optical Transport Networks. ITU-T Recommendation G.872,
2001.
[ITU-T02a] ITU-T. Spectral Grids for WDM Applications: DWDM Frequency Grid. ITU-T
Recommendation G.694.1, 2002.
[ITU-T02b] ITU-T. Spectral Grids for WDM Applications: CWDM Frequency Grid. ITU-T
Recommendation G.694.2, 2002.
[Sharony94] J. Sharony. Architectures of Dynamically Reconfigurable Wavelength Routing/
Switching Networks. Ph.D. thesis. Department of Electrical Engineering, Columbia Univer-
sity, New York, 1994.
[Smith+76] D. G. Smith and M. M. Rahmnekhan. Wide-sense nonblocking networks, and some
packing algorithms. In Int. Teletraffic Cong., pp. 542-1–542-4, Melbourne, Australia, 1976.
[Tymes81] L. R. Tymes. Routing and flow control in TYMNET. IEEE Trans. Commun.,
29(4):392–398, 1981.
[Wu92] T.-H. Wu. Fiber Network Service Survivability. Norwood, MA: Artech House 1992.
[Yoo+95] S. J. B. Yoo, C. Caneau, R. Bhat, and M. A. Koza. Wavelength conversion by quasi-
phase-matched difference frequency generation in AlGaAs waveguides. In Proceedings of the
IEEE/OSA Optical Fiber Commun. Conf., Paper PD14-2, San Diego, CA, February 1995.
3 Network Connections

The multiwavelength network architecture described in Section 2.1 contains several lay-
ers of connections. By exploiting the various alternatives in each layer, it is possible to
produce a rich set of transport network configurations. This chapter explores how a de-
sired connectivity pattern can be established using the combined functionality contained
in the various layers. The approach is to examine the properties of different classes of
networks through a sequence of simple illustrative examples. The design objective in
each example is to provide a prescribed connectivity to a set of end systems. Each of the
network classes illustrated in this chapter is discussed in more detail in later chapters, as
is the issue of optical network control.
Our first example is shown in Figure 3.1. Five geographically dispersed end systems
are to be fully interconnected by a transport network, which is to be specified. The end
systems might correspond to physical devices such as supercomputers that interact with
each other, or they may be gateways (interfaces) to local access subnets (LASs) serving
industrial sites, university campuses, or residential neighborhoods.1
In all of these cases, a dedicated set of connections is desired (shown as dashed
lines in the figure), providing full connectivity among all the sites. Figure 3.2(a) shows
one possible transport network, whose physical topology (PT) is a star, in which the
central node is a star coupler of the type shown in Figure 2.7(a). Each end system is
connected to the star through its own network access station. Full connectivity requires
20 unidirectional logical connections. The logical topology (LT) for the network is shown
in the logical connection graph (LCG) in Figure 3.2(a), where each link represents a pair
of unidirectional logical connections carrying signals in opposite directions. A bipartite
representation of the LCG is also shown, in which transmitting and receiving stations
are separated, so an arc from vertex i to j represents an LC originating at station i (the
left vertex) and terminating at station j (the right vertex).
Each network access station interfaces four pairs of logical connections to its attached
end system to produce the fully connected logical topology shown in the figure. The LT
shows the connectivity realized in the transport network as “seen” by the end systems but
suppresses the details of the underlying physical layer. When a set of LCs is realized over
a given transparent optical infrastructure, the logical topology is said to be embedded in
the supporting PT. In the case at hand, the LT is a set of links that connects directly all pairs
of end systems. The burden of supporting full logical connectivity is borne completely
1
A common form of LAS is the passive optical network (PON), whose gateway to the larger network (the
end system in Figure 3.1) is known as an optical line terminal (OLT). See Section 5.7 for details of PONs.
92 Multiwavelength Optical Networks

Figure 3.1 End systems: full connectivity.

1
2

5 2 3

4 3 5

(a)

1 1 1

1
2 2

5 2 3 3
5 2

4 4

4 3 4 3 5 5

(b)

PS
5 2

4 3
(c)

Figure 3.2 Star physical topology.

 Network Connections 93

5 2

4 3 LCG

Figure 3.3 Bidirectional ring physical topology.

by the optical network. This requires multiplexing of many optical connections on each
fiber, together with replication of optical and/or electronic equipment within the NASs
to support multiple connections. If WDM is used on the star, each LC requires a distinct
wavelength and its own transmitter/receiver pair. (No spectrum reuse is possible; see
Section 3.2.) Thus a fully connected LT requires four optical transceivers in each station,
with each transmitter operating on a different wavelength, for a total of 20 wavelengths.
Another possible realization, in the form of a transparent optical bidirectional ring
containing wavelength-routing optical network nodes, is shown in Figure 3.3. Each end
system is connected through its own NAS to an ONN. (The bidirectional NAS-ONN
combination can be implemented in the form of a pair of unidirectional WADM-NAS
structures of the form shown in Figure 2.23.) This time if we use WDM it turns out
that only three or four wavelengths are needed to support the 20 logical connections,2
indicating that a substantial degree of spectrum reuse is possible if wavelength routing
is employed. As a third option, returning to the simpler star network, we can reduce
the connectivity burden on the physical layer by inserting a logical switching node
(LSN) between each end system and its NAS as shown in Figure 3.2(b). This provides
a switching function at the electronic level, which will share the task of providing
full connectivity. (Another alternative for reducing the connectivity burden uses packet
switching in the optical layer, as illustrated in Figure 3.2(c). It will be discussed in
Section 3.2.2.)
To fix ideas, let us assume that the LSNs are IP edge routers as described in Section 2.6.
Their function is to route data in the form of IP packets among the attached end systems.
I/O ports on the routers are connected to NASs on the network side and to the end
systems on the user side. A possible LCG and its bipartite representation for this case
is shown in Figure 3.2(b). Only five LCs are present now, requiring a total of five
wavelengths and one optical transceiver per station and producing a logical topology in
the form of a unidirectional ring connecting the five LSNs. This is a simple example of
an IP over WDM network. It is now the LSNs and their interconnecting logical links
that are seen by the end systems as their transport network, in contrast to the cases of
Figure 3.2(a) or Figure 3.3, in which the transport network was purely optical. So far we
2
The exact number of required wavelengths depends on the stations’ access connections (see Section 3.3).
94 Multiwavelength Optical Networks

do not have complete connectivity among the end systems. This is provided by moving
up to the virtual connection layer of Figure 2.1(a) and superimposing a set of 20 virtual
connections on the logical topology of Figure 3.2(b) to form a fully connected virtual
topology that corresponds to the dashed lines in Figure 3.1.
Note that the virtual connections must be supported by logical paths consisting, in
most cases, of more than one logical hop. The network of LSNs realizes the virtual
connections by sorting and forwarding the cells according to their destinations. If the
LSNs are IP routers, these operations are executed by reading the packet headers and
directing the packets to the outbound port on the router specified in its routing table.
This is a simple example of a transparent optical network with a logically routed or
multihop electronic overlay. In this case the physical layer is required to support only
five logical links, with the remaining connectivity supplied electronically in the overlay.
In comparison with the purely optical star, the wavelength requirement has been reduced
by a factor of 4. However, many virtual connections must be now multiplexed on each
logical link. This logical layer multiplexing (and switching) is possible because of the
intelligence in the LSNs. Each logical connection is now shared by 10 multiplexed
virtual connections (why?), so that the total capacity of the network is reduced to 10%
of the capacity available in the case of the bidirectional ring (see Section 3.5).
It is important to note that in each of these examples, the logical topology is inde-
pendent of the physical topology into which it is embedded. Furthermore, it is generally
reconfigurable and therefore adaptable to changing conditions. For comparison, consider
traditional (opaque) networks made up of point-to-point transmission links (optical or
otherwise) that interconnect LSNs. The logical topologies of these networks are identical
with their physical topologies and therefore are frozen once the transmission links are
in place. A commonly used physical configuration is a bidirectional ring, which super-
ficially resembles the transparent optical ring of Figure 3.3.3 The nodes in the traditional
ring are logical switches, so the logical topology resembles that of Figure 3.2(b) (except
that the logical ring is generally bidirectional). In contrast, the transparent optical ring of
Figure 3.3 is configured to produce a fully connected LT, a topology that could be recon-
figured by changing settings in the ONNs and NASs without reconnecting fiber links.
It should now be clear that there are many possibilities for realizing connectivity
in multiwavelength optical networks. The resultant logical and virtual topologies are
generally reconfigurable and offer many opportunities for trading cost and performance.
The remainder of this chapter illustrates how connectivity is created in each layer,
focusing on relations among physical constraints, device functionality, connectivity, and
throughput. To avoid unnecessary complications, the exposition proceeds from simple to
complex classes of networks based on the level of controllability in the network nodes:
r Static (broadcast-and-select) networks
r Wavelength-routed networks
r Linear lightwave networks (LLNs)
r Logically-routed (hybrid) networks (LRNs)
3
Bidirectional rings based on point-to-point optical fiber links are used commonly in many networking
applications, the most common being SONET rings.
 Network Connections 95

These four categories correspond to the network classes shown from left to right in
the taxonomy of Figure 2.40. The static category comprises transparent optical networks
based on static network nodes with no wavelength selectivity; i.e., single-waveband net-
works. (We are arbitrarily excluding wavelength selectivity of the type present in static
wavelength routers to simplify our network categorizations.) Any signal introduced into
such a network propagates along all possible paths, tending to “flood” the network,
and eliminates possibilities for reusing the optical spectrum on different fibers. For all
practical purposes, this limits the PT to a tree (or a collection of separate trees) and
eliminates the function of the optical path sublayer in Figure 2.1(a). Hence the size,
throughput, and flexibility of static networks are limited. (They are typically used as
LANs.) The only way to support many simultaneous connections in these networks is
through optical multiplexing, multicast, and multiaccess techniques. The basic features
of this approach are discussed in Section 3.2. By adding either static or dynamic wave-
length selectivity to the network nodes, we obtain wavelength-routed networks (WRNs),
wherein spectrum reuse and hence improved performance becomes possible through the
use of appropriate connection management. These networks normally have more general
physical topologies.
Wavelength-routed networks are discussed in Section 3.3. Consistent with the way
in which typical WRNs are operated, the discussion there is limited to point-to-point
optical connections, which can be realized in networks equipped with wavelength-
selective cross-connects. In Section 3.4 the discussion is extended to LLNs, which are a
generalization of wavelength-routed networks, characterized by the fact that their optical
nodes are waveband-selective linear divider-combiners (LDCs) rather than wavelength-
selective permutation switches. The properties of LLNs are more general than those of
wavelength-routed networks, in two respects:

1. They are waveband rather than wavelength routed. In other words, their spectrum
partitioning can be visualized as shown in Figure 2.5(c) rather than Figure 2.5(b).
2. They support multipoint optical connections in addition to point-to-point connections.

The multipoint optical switching capability of an LLN offers a wealth of connection

alternatives. Among them is the possibility of tying subsets of NASs together into fully
connected multipoint subnets (MPSs). This is a particularly effective way of creating
high connectivity in large networks, as well as providing a transparent optical support
adapted naturally to logical multicast connections.
Section 3.5 introduces LRNs, constructed using electronically switched overlays sup-
ported by a transparent optical substructure. Networks with two types of logical topolo-
gies are described: those in which the basic logical link is (fixed capacity) point-to-
point and those with multipoint logical links. We call the latter category hypernets.
Although point-to-point logical topologies (multihop networks) can be implemented
over wavelength-routed optical networks, hypernets require an LLN physical layer to
support their multipoint links. The shared-capacity multicast connections in the logical
layer are supported by MPSs in the physical layer. The chapter concludes with a summary
and comparison of the different approaches to realizing network connectivity.
96 Multiwavelength Optical Networks

To set the stage for the discussion of these different network structures, we must
have at least a rudimentary idea of how connections can be managed. Management and
control issues in large networks encompass an enormous range of critical functions,
some of the most important being:
r Fault management (including fault detection and service recovery)
r Performance management (monitoring the performance of the various network ele-
ments and connections)
r Configuration management (provisioning and reconfiguration of network connections)

Associated with these are signaling and many other operations, administration, and
maintenance functions. These topics are still in their embryonic form for optical net-
works and must evolve as the networks themselves evolve (see [Bernstein+04, Maeda98,
Wei+98]). The international standards bodies break down the generic architectural is-
sues in traditional networks into three parts by defining a transport plane, a control
plane, and a management plane, and this approach is now being applied to optical net-
works. The transport plane encompasses equipment associated with data transfer, the
control plane is concerned with the operations required for connection provisioning
and reconfiguration, and the management plane deals with the myriad other operations
administration and maintenance functions required to ensure satisfactory network per-
formance, including performance monitoring, fault diagnosis, statistics gathering, and
so on. In the next section, we briefly address the issue of connection management and
control; that is, we focus on the functions of the control plane. The objective is to provide
some understanding of what operations must take place to provision connections in an
optical network, and how these operations might be implemented.

3.1 Connection Management and Control

In any large network, connections do not remain static. In the case of demand-assigned
connections, users become active, request specific connections, and then disconnect.
In the case of dedicated connections, even though the intention is to retain a fixed
connection configuration, changes in traffic conditions or network faults may dictate
changes in routing, wavelength assignment, and bit rates while the connection is in
progress. Our repeated use of the term connection implies that information flow is
normally connection oriented in the networks we are discussing.
Any connection-oriented network requires a system for connection management. Our
layered view of the network allows us to look at optical, logical, and virtual connection
management as three (almost) separate issues. At each layer, the protocols for connection
management consist of three phases: connection establishment, information transfer, and
connection release. In the most elaborate case, exemplified by the virtual, logical, and
physical layers in the network of Figure 3.2(b), these phases must occur within each layer.
Typically, the higher the layer, the more frequent are the connection changes. Thus, the
optical connections would normally be provisioned on a dedicated basis, with changes
 Network Connections 97

made only for fault recovery. The logical connections might also be fixed perhaps with
some modifications to support changing load distributions. On the other hand, most of
the virtual layer connections would be demand assigned.
Looking at the connectivity from the top down, suppose Figure 3.2(b) represents
an ATM layer that serves as a logical network acting as a transport network for an IP
network with its own virtual topology. (This is a case of two stacked logical layers over
the physical layer: IP over ATM over WDM.) The logical nodes in Figure 3.2(b) are ATM
switches, and the end systems are IP routers. The ATM logical topology is configured
to act as a server for the IP client network, whose virtual topology is carried on ATM
virtual connections. A connection between IP routers is realized on a virtual channel
carried on a specified virtual path in the ATM network.
To support the expected traffic demand, the ATM network in this example is configured
on a dedicated LT (in this case a unidirectional ring) capable of carrying the anticipated
load. This is set up when the ATM network is initialized, at which time the appropriate
LCs are created using the available resources in the physical layer. (In larger networks,
additional NASs providing services to other client layers might access the same optical
infrastructure.)
If traditional point-to-point transmission links were used between the ATM switches,
the logical and physical topologies of the ATM network would be identical, so initial-
ization of the logical network would be straightforward. In the point-to-point case, once
the logical topology has been chosen, ports on the switches (LSNs) are interconnected
accordingly using point-to-point links that are either owned by the network operator or
leased from another carrier. However, creation of an arbitrary target LT over a transparent
optical network is a more complex operation – an inevitable consequence of the many
connection alternatives available in the physical layer.4
Now let us focus on connections within the physical layer of a WRN. Optical connec-
tion establishment consists of setting up a path through a sequence of optical network
nodes for an assigned wavelength and then tuning source transmitter and destination
receiver(s) to the chosen wavelength.5 These are the operations occuring in the optical
path and λ-channel sublayers, respectively. Thus, connectivity at this level is realized
through the combined action of the optical transceivers and the network nodes. For
demand-assigned connections, connection establishment and release involves signaling
between the party initiating the action and a network manager, followed by issuance of
routing and wavelength assignment commands from the network manager to the net-
work elements (stations and network nodes) involved in the connection. (In the case of

4
When transmission links are leased from another carrier, our optical network becomes a client network and
the carrier’s network is the server network. Now, there are similar connection alternatives in the physical
layer, but they are handled by the carrier providing the links. The operator of the logical network is normally
shielded from this part of the problem, in effect isolating management of the logical layer from that of the
physical layer.
5
At this point, it is assumed that transceiver tuning remains fixed for the duration of a connection. In later
sections, we consider multipoint connections involving rapid retuning of transceivers during the information
transfer phase of the connection.
98 Multiwavelength Optical Networks

Control Plane

Control domains
Control agent

Client network node (LSN)

Optical network node (ONN)

Figure 3.4 The control plane in an optical network.

dedicated connections, the various connection provisioning operations would usually be

executed manually off-line.)
As networks get more complex, with multivendor equipment and with multiple optical
network providers interacting with each other, a high-level view of the network as viewed
from the control plane has become essential. Figure 3.4 shows an abstract view of the
control plane in an optical network consisting of two control domains. The two domains
might represent networks owned by different providers or subnets owned by the same
provider but controlled separately.
The optical network, consisting of ONNs and optical links, is a server for a higher
layer logical network, represented by the three client nodes (e.g., IP routers and/or ATM
switches), which we represent using our generic LSN symbols. The control plane is
shown in the form of interconnected control plane agents, which are abstractions of the
control functions that take place in both the client (logical) and server (optical) networks.
The realization of these agents may take many forms, from controllers physically asso-
ciated with each network element (a distributed control approach) to a single controller
remotely linked to several network elements (centralized control approach). The asso-
ciations between the control plane agents and the networks and network elements they
control are indicated by dashed lines in Figure 3.4, and the various interactions between
control agents are represented by solid lines.
Network control requires message exchange among control plane agents through a
dedicated Data Communications Network (DCN) using some standard signaling proto-
col. The DCN may be supported by the optical network being controlled using separate
 Network Connections 99

Control
Processor

Controller ONC
Access Switch Control/Monitoring
Station

Signaling Switch Control Interface

Optical
Switch

1 A B 3

ONN

C ONM

Network Network
Management Management
Access Station Processor

Figure 3.5 Connection management system.

control channels or by an independent infrastructure, and it may be used to support

communication with other network management systems as well.
We relegate a more comprehensive treatment of general network control to Chapter 9
and continue here with a discussion of the specific control functions that must be
performed in the optical network elements.
The exact form of the connection management system (e.g., either centralized or
distributed) is not important at this point, nor are the details of the messaging system
(e.g., either supported by the information-bearing fibers or realized as a separate data
network). To fix ideas, Figure 3.5 shows one way the connection management system
might operate. The connection management function is divided into local ONN control
functions and an overall network connection management function. Each optical network
node contains an optical switch (e.g., a WSXC) together with a control agent: an optical
node controller (ONC) that executes the local control functions. An optical network
manager (ONM) disseminates commands for switch reconfiguration. In this example,
the ONM is attached through an access link to node C of the optical network. It consists
of a network management processor accessing the optical network through an NAS.
The ONM might be part of a larger network management system (NMS) encompassing
logical and virtual layers as well as the optical layer, in which case it would interact
100 Multiwavelength Optical Networks

with the NASs, the ONNs, and possibly higher layer NMSs to receive optical connection
requests and supervise their execution.
In this example, we assume that the ONM, the various ONCs, and all NASs intercom-
municate via a DCN multiplexed onto the information-bearing fibers.6 For simplicity we
assume here that a single dedicated wavelength shared among all network elements is
used for supporting control signals.
As shown in Figure 3.5, the ONC is composed of an access station connected optically
to a port on its switch through an access fiber pair and a control processor connected
electrically to the switch control interface. Commands received from the ONM are
interpreted in the control processor of the ONC, which issues electrical commands to
configure the switch as required. (The ONC might also send information back to the
ONM regarding performance monitoring. Parameters to be monitored might include
switch status, power levels of signals transiting the switch, faults on the links, and so
on.)

3.1.1 Optical Connections

Suppose station 1 in Figure 3.5 wishes to set up an optical connection to station 3.
Using the signaling network it sends a connection establishment request to the ONM,
which determines an optical path, say the one-hop path from node A to B, and selects
an available wavelength. The ONM then signals nodes A and B to set up the correct
path for the assigned wavelength and signals stations 1 and 3 to tune to that wavelength.
The information transfer phase can then begin. Connection release would normally be
initiated by a signal from one of the communicating stations to the ONM, which would
then signal all participating network entities to release the connection. Note that if the
network nodes contain generalized switches, then optical multicast paths can be set up
in the form of trees, wherein the optical signal power is split, with a portion reaching
each destination station. For example, station 1 can multicast to 2 and 3 through a tree
rooted at the source node ( A in this case). Furthermore, many-to-one optical paths such
as a path from 1 and 2 to 3 can be set up in a similar fashion using a tree rooted
at node B, which combines the signals from the transmitting nodes. The total time
required for establishing and releasing an optical connection includes signaling time
(including propagation delays), routing and wavelength assignment computation time,
optical node switching time, and transmitter/receiver tuning time. Because this total can
be substantial (possibly on the order of 100 ms for propagation times alone), optical
connections should generally be held for much longer periods of time (seconds or more)
for efficient operation.

3.1.2 Logical Connections

Once an optical connection exists, one or more logical connections can be carried on
it through the assignment of suitable transmission channels, which serve to adapt each
6
If there are optical line amplifiers along the fiber links, they would normally include their own control agents
for surveillance and reconfiguration purposes.
 Network Connections 101

LC to its supporting optical connection. Referring to Figure 2.29, a point-to-point LC

originates at an external port on the transmission processor (TP) of the source station,
which converts the logical signal to a suitable transmission signal, appearing at point a
in the figure. This is in turn converted to an optical signal, appearing at point b, which
may be multiplexed with other optical signals before it leaves the NAS. The process
is reversed at points d and e in the receiving side of the destination station. Because
all operations concerning LCs and their associated transmission channels take place
within NASs, logical connection management does not require the intervention of the
ONCs once the supporting optical paths are in place. As a result, connection recon-
figuration within a logical network can often take place through distributed connection
management protocols executed cooperatively among the participating NASs without
calling on the functions of lower layer network entities (optical path layer and below).
(Examples of this appear in subsequent sections.) This can greatly simplify the various
network management and control functions and provides further justification for viewing
the multiwavelength network in terms of independent logical and physical layers and
sublayers.

3.1.2.1 Stream and Packet Traffic

During the information transfer phase, it is useful to distinguish between two types of
data flow on a logical connection: stream type and asynchronous (packet or cell) traffic.
Stream traffic, exemplified by SONET STSs described in Section 2.6, typically supports
voice, fixed bit-rate video, and similar applications. A synchronous optical network in-
frastructure such as SONET or SDH can also be used as an underlying server for a packet-
or cell-switched client network, as in ATM or IP over SONET. In the case of stream
traffic, the information entering and exiting a logical port on an NAS is in the form of
a continuous, synchronous bit stream. To support this type of traffic, the transmission
channel carrying the LC must act as a dedicated, fixed-capacity “pipe.” In the asyn-
chronous case, the information flow is in the form of random bursts carried in data
packets or cells. In this case, each packet or cell carries explicit addressing information
in a header, processed by various network entities (e.g., IP routers, ATM switches, or
MPLS LSRs) involved in routing the data along a path to its destination.
In the case where a synchronous network serves as the infrastructure for a packet-
or cell-switched network the distinction between stream type and asynchronous traffic
is not always obvious. Thus, in packet-based systems, a random sequence of packets is
typically embedded into a synchronous bit stream by encapsulating the packets into data
frames separated by “idle” characters. Similarly, when a sequence of ATM cells is carried
on a SONET connection, the data cells are carried within an STS bit stream interspersed
with idle cells. In this case, the useful information is “bursty” or asynchronous, but the
bit stream carried by the optical network is steady or synchronous. For our purposes, the
traffic type is characterized by the way in which the NASs operate on the bit stream. If
the bits are processed transparently, without distinguishing the individual information-
bearing packets/cells from the idle bits, it is considered to be stream type. If the individual
packets/cells are extracted from the bit stream when it enters the transmission processor,
and headers are read for the purposes of multiplexing, scheduling, routing, and so forth,
102 Multiwavelength Optical Networks

the traffic is considered to be asynchronous. The latter mode of operation corresponds to

packet- or cell-switching within the optical layer. Typically, this requires that transmission
and reception processors execute a media access control (MAC) protocol designed for
sharing one or more λ-channels dynamically among many logical connections. Although
synchronous traffic is always connection oriented, asynchronous traffic may be either
connection oriented or connectionless. For example, traffic in IP networks is usually
connectionless: Packets are launched into the network as individual entities without
establishing a virtual connection.

3.1.2.2 Multipoint Logical Connections

The efficient realization of one-to-many and many-to-one logical connectivity among
stations is an important issue in optical networks. A station maintains one-to-many
logical connectivity if it is the source of more than one simultaneous logical connection.
Similarly it maintains many-to-one logical connectivity if it is the destination for more
than one LC. For a station accessing a network node through a single pair of fibers, one-
to-many logical connectivity requires multiplexing several transmission channels on the
outbound access fiber, and many-to-one connectivity requires that transmission channels
from several source stations share the inbound access fiber using some multiple access
procedure. These multipoint connections are crucial for achieving high connectivity in
optical networks.
When multiplexing and multiple access are used for multipoint connections, the issue
of addressing arises. Addressing methods are linked to the particular multiplexing tech-
niques being used. For example, with synchronous time-division multiplexing (TDM),
addressing is implicit in the position of a data segment in a periodic time frame, and
thus very little processing is required for the multiplexing/multiaccess operations in the
NASs (see Section 3.2). However, in the asynchronous case, each packet/cell header
must be processed individually in the station’s transmission processor (reception pro-
cessor) to execute the multiplexing (demultiplexing) function on an outbound (inbound)
transmission channel. Thus, packet switching in the optical layer typically carries with
it considerable computational overhead within the NASs.

3.2 Static Networks

As we have seen, in static networks the physical layer is stripped down to its bare
essentials: passive splitting/combining nodes interconnected by fibers to provide static
fiber connectivity among some or all pairs of optical transmitters and receivers (OTs
and ORs). This was illustrated in the examples of Figure 2.7, which show full potential
connectivity from each transmitting to each receiving station in the case of a star or
bus, and partial connectivity in the case of a tree. Because the optical paths cannot
be modified, connection control in static networks is confined to the λ-channel and
transmission channel sublayers. This means that if the NASs have a means of signaling
among themselves, connection management can be executed in the NASs without the
intervention of a separate optical network manager. Consider the broadcast star example
 Network Connections 103

of Figure 3.6, which consists of a star coupler connecting transceivers in three stations.
This is an example of a static shared medium network.
As described in Section 2.3.1, the star coupler combines signals from all transmitting
stations and broadcasts them to all receiving stations.7 Thus a receiver can make a logical
connection with any transmitting station by selecting the information destined for it and
discarding the rest. This is the essence of the broadcast-and-select method of operating
shared-medium optical networks. Its most attractive feature is that the broadcast function
permits complete optical connectivity among all pairs of transmitting and receiving
stations because every receiver “sees” every transmitter. This has a downside, however.
Two transmitters emitting optical signals at the same time must transmit on distinct
wavelengths. If they use the same wavelength, they will interfere with each other at all
receivers, destroying any useful information flow.
This distinct channel assignment (DCA) constraint prevents any reuse of the optical
spectrum in broadcast networks. It should be noted that broadcast-and-select is not
limited to networks based on star couplers. For example, any of the tree networks in
Figure 2.7 can be operated in a broadcast-and-select mode. Each of them combines
signals from a set of transmitters and broadcasts them to a set of receivers. (In the case
of Figure 2.7[b], all transmitted signals are combined at the head end of the bus and
broadcast to all receivers, and in the case of Figure 2.7[c], each of the two unidirectional
trees provides signal combining and broadcasting for the subsets of transmitters and
receivers which it interconnects.) We now examine the operation of static networks
using the broadcast star as an illustrative example.

3.2.1 Point-to-Point and Multipoint Connections

Although the discussion thus far has focused on point-to-point connections, the broadcast
star is a natural medium for creating multipoint connections for enhanced connectivity,
and these play an important role in the discussion that follows. To keep track of what is
happening, we use different notation for connections in each layer, as introduced first in
Section 2.4:
r A unidirectional point-to-point optical path from station a to station b is denoted as
a, b , and a multicast path from station a to the receiving set {b, c, . . .} is denoted as
a, {b, c, . . .} .
r A unidirectional point-to-point optical connection from station a to station b is denoted
as (a, b). To show the wavelength of a λ-channel carrying the connection, say λk , we
use the notation (a, b)k . A multicast optical connection from station a to the receiving
set {b, c, . . .} is denoted as (a, {b, c, . . .}).8
r A unidirectional point-to-point logical connection from an external port on station a
to one on station b is denoted as [a, b], and a multicast LC from a to a set {b, c, . . .}
7
If transmitter–receiver pairs are shown with corresponding numbers, as in this case, they are assumed to
belong to the same station, and the network is called an undirected star; otherwise, it is referred to as a
directed star.
8
An optical path from a station a to a receiving set B can carry any optical connection from a to a receiving
set C ⊆ B.
104 Multiwavelength Optical Networks

is denoted as [a, {b, c, . . .}]. In a multicast LC, the transmitting station sends the
same information to all receiving stations. (If station a transmits different information
streams to a set of destination stations, the connections would be designated [a, b],
[a, c], . . . .).

As we shall see, there is a significant difference between the realization of a set of

one-to-many LCs composed of several point-to-point connections (say, [a, b] and [a, c])
and the related multicast connection [a, {b, c}]. The former requires multiplexing of
two distinct information streams at the transmitting station, whereas the latter does not.
This means, for example, that a multicast LC can be supported by creating a multicast
optical path in the form of a directed tree that delivers the same signal to several
destinations.9
Summing up the possibilities at the logical level, there are three possible types of
logical connectivity on the transmitting side:

1. One-to-one: A source station maintains a point-to-point logical connection with a

single destination.
2. Multicast: A source station maintains a multicast logical connection with several
destinations.
3. One-to-many: A source station maintains several point-to-point logical connections
to different destinations.

Note that an n-fold multicast connection can always be realized as a set of n point-to-
point connections, but it is generally wasteful of communication resources to do so. On
the receiving side the possibilities are one-to-one, in which a destination station maintains
a logical connection with a single source, and many-to-one, in which a destination station
maintains several point-to-point logical connections from different sources. The term
multipoint is used generically to designate an arrangement including one-to-many, many-
to-one, and/or multicast logical connections.
Now let us illustrate the layers of connectivity using the example of Figure 3.6. The star
coupler provides a permanent broadcast optical path of the form i, {1, 2, 3} from each
transmitting station to the set of all receiving stations (including loopback paths from
each station to itself) independent of the wavelength being used. Thus, suppose station 1
transmits on wavelength λ 1 . Then a point-to-point optical connection (1, 2) 1 is set up by
tuning a receiver in station 2 to λ 1 , and a multicast optical connection, say (1, {2, 3}) 1 ,
is created by tuning receivers 2 and 3 to λ 1 . Because the signal power originating from
station 1 is split among all receiving stations, exactly the same information reaches
both stations 2 and 3. It is important to note that the optical path produced by the
broadcast star offers the potential for various types of point-to-point and multicast
optical connections, but the actual connections are realized through the actions of the
access stations (transmitter/receiver tuning). Furthermore, the optical power from all
transmitters is always delivered to all receivers whether they use it or not.

9
Multicast connections can also be created in the logical path layer of Figure 2.1(a) by forming a directed
tree in the electrical domain.
 Network Connections 105

NAS 1

[1, 2] 1a 1b 1d 1e
TP1 OT1 OR1 RP1
[1, 3]

2a 2b 2d 2e
[2, 3] TP2 OT2 OR2 RP2 [1, 2]

3a 3b 3d 3e [1, 3]
TP3 OT3 OR3 RP3
[2, 3]

(a)

RP3

[1, 3]
3d 3e
OR3
(1, {2, 3})
(2, {2, 3}) [2, 3]
1 1

Synchronized
2 2 Gate
or
Packet Filter
3 3 and Buffer
(b) (c)

Figure 3.6 Star coupler example.

Moving up to the logical layer, suppose the point-to-point LCs [1, 2], [1, 3], and [2, 3]
are to be realized, as shown in the bipartite LCG in Figure 3.6(b). Both one-to-many and
many-to-one LCs are present. Depending on the throughput requirements and available
network resources, these connections can be realized in various ways. The simplest
approach is by “brute force,” using WDM and wavelength division multiple access
(WDMA). If three wavelengths – λ 1 , λ 2 , and λ 3 – are available, then three optical
connections – (1, 2) 1 , (1, 3) 2 , and (2, 3) 3 – are set up, which in turn carry the cor-
responding LCs. Note that this requires two optical transmitters in station 1 and two
receivers in station 3 and provides the full capacity of a λ-channel to each LC. (The
NASs in Figure 3.6[a] are assumed to contain the necessary number of transceivers, even
though they are not shown explicitly.) One-to-two logical connectivity from station 1
is realized by multiplexing the two wavelengths λ 1 and λ 2 onto its outbound fiber
(WDM), and two-to-one connectivity at station 3 is realized by accessing the station
with λ-channels on wavelengths λ 2 and λ 3 superimposed on its inbound fiber (WDMA).
106 Multiwavelength Optical Networks

Now suppose the two point-to-point connections [1, 2] and [1, 3] are carrying the
same information. They can be replaced by a multicast LC [1, {2, 3}]. The multicast
connection could be supported by the same optical connections as used for the point-
to-point case simply by modulating the two optical transmitters in station 1 by the
same (electrical) transmission signal. However, a more economical way of achieving
the same result is to use a single multicast optical connection (1, {2, 3}) 1 , eliminating
one transmitter in station 1 and using only two wavelengths instead of three while
achieving the same throughput. Because the multicast connection [1, {2, 3}] is still just
one connection, it can be realized without multiplexing. This simple example shows that
there are often several alternatives for realizing connections at the logical and optical
levels, so judicious choices are important to obtain cost-effective results.
For larger networks, a high degree of multipoint connectivity is often required, as
suggested in the example of Section 1.5. However, the WDM/WDMA approach of
dedicating a λ-channel to each LC becomes prohibitively expensive and quickly exhausts
the available optical spectrum as the number of connections increases. Thus, other
methods of achieving high logical connectivity while using less network resources
(wavelengths and transceivers) must be employed. These typically involve combinations
of wavelength division and other multiplexing and multiple access techniques. Two
additional cases: time division and time-wavelength division will suffice for illustration.

3.2.1.1 Time-Division Techniques

TDM and time division multiple access (TDMA) are methods of realizing multipoint
connectivity that are particularly cost-effective in terms of optical resources. At any
one time, a large network will be carrying many optical connections, some dedicated
and some demand assigned. All of these are sharing the network resources subject to
certain admission control rules, executed by the ONM to ensure satisfactory operation
of the network. The dedicated connections can be considered to be fixed for the pur-
poses of the current discussion. With demand-assigned (circuit-switched) traffic, the
connection pattern evolves with time as active connections are released and new ones
are established. These events generally occur on long time scales (seconds or more)
for reasons explained at the beginning of this chapter. In addition to this “slow” time
sharing of network resources, when TDM/TDMA is used sharing also occurs on a much
faster (microsecond/nanosecond) time scale by interleaving concurrent connections in
time.
To illustrate the different time scales, Figure 3.7 shows the signals appearing at one of
the receivers when the LCs in the example of Figure 3.6 are realized with one transmitter
or receiver per station and only one wavelength λ 1 by time sharing the optical medium. Is
this a sequence of circuit-switched connections or a set of simultaneous time-multiplexed
multipoint connections? It all depends! If the time scale is on the order of seconds or
more, then the figure shows circuit-switched connections, with the idle times between the

[1, 2] [1, 3] [2, 3]

t

Figure 3.7 Time-shared medium.

 Network Connections 107

information flows used for connection establishment and release. However, if the time
scale is microseconds, there is no time for connection changes between information
flows, and the picture must represent TDM/TDMA. In this case, the two LCs [1, 2]
and [1, 3] are time division multiplexed onto a transmission channel in station 1, and
the optical transmissions from stations 1 and 2 each access the same λ-channel during
different time intervals – an example of TDMA. The time interleaving might be on a
packet, cell, byte, or bit basis.
The essential differences between the circuit-switched and multiplexed cases concern
how the connections are managed. In the circuit-switched case, each connection is
released before the next is established; in the multiplexed case, both connections are
established before the information transfer phase and are maintained simultaneously in
time. (Because the optical path is static in this example, all connection management
can occur via signaling among the access stations.) Note that in either case the total
amount of time required to transmit a given number of bits (the throughput in bits
per second, neglecting idle periods) is the same and is limited by the capacity of the
shared λ-channel. (In the WDM/WDMA case, in which three λ-channels were used, the
maximum throughput was three times as much as in TDM/TDMA, illustrating the fact
that there is no free lunch. If less optical resources are used, less transmission capacity
is available.)
Now let us examine the multipoint case in more detail. Observe that because all
information is transmitted on wavelength λ 1 , all transmitters and receivers remain tuned
to λ 1 at all times, creating the multicast optical connections (1, {2, 3}) 1 and (2, {2, 3}) 1 .10
Using TDM at the transmission channel level, the information streams for connections
[1, 2] and [1, 3] arriving at the two input ports on station 1 in Figure 3.6(a) are multiplexed
onto a single transmission signal (point 1a) that modulates the transmitter OT 1 . The
TP contains buffers and synchronization equipment for interleaving the two information
streams. The transmission channel supporting these multiplexed LCs is carried on the
optical connection (1, {2, 3}). The LC [2, 3] is carried on optical connection (2, {2, 3})
without any multiplexing.
Because both optical connections are carried on the same wavelength, an optical
multiple access scheme is necessary to keep the two signals from interfering with each
other at the receivers and to distinguish the different information streams. In this example,
we use TDMA, in which the two optical signals are transmitted in nonoverlapping
“bursts” by switching the transmitters on and off intermittently. The overall connection
arrangement is therefore a combination of TDM at the logical (transmission channel)
level and TDMA at the optical (λ-channel) level. At each receiver the optical signal is first
converted to electronic form (at points 2e and 3e in Figure 3.6[a]) and then demultiplexed
to extract the information destined for it. The rest is discarded. Demultiplexing requires
synchronized gates or “packet filters” in the RP, as shown in Figure 3.6(c) for station 3.
The packet filters are required if the data stream interleaving is done on a packet/cell
basis (see Section 3.2.2).

10
Note that the loopback connection from 2 to itself is not needed but is an unavoidable result of the broadcast
topology.
108 Multiwavelength Optical Networks

[1, 2]1 [1, 2]2 [1, 3]1 [1, 2]3 [1, 2]4 [1, 3]2
OT1

[2, 3]1 [2, 3]2

OT2

From 1 From 2 From 1 From 2

Coupler
F1 F2
From 1 From 2 From 1 From 2
OR2
F1 F2
From 1 From 2 From 1 From 2
OR3
F1 F2

Figure 3.8 A TDM/TDMA schedule.

A typical schedule of transmissions for this example is shown in Figure 3.8. In the
figure, transmissions occur in synchronized time slots of equal length, where the notation
[a, b]n in a slot means that it is used for the LC [a, b], with n indicating that this is the
nth segment of the data stream transmitted on that connection. The first two time traces
in the figure show the information streams as they appear at the transmitters OT 1 and
OT 2 , and the third shows the streams interleaved as they appear at the coupler. Note
that the different propagation delays between the stations and the coupler indicate that
station 1 is farther away from the coupler than station 2. The last two traces represent
the streams as they appear at receivers OR 2 and OR 3 , showing that station 3 is closer
to the coupler than station 2. Multiple access synchronization is required at the trans-
mitters to produce nonoverlapping optical signals at the coupler (and hence at each re-
ceiver), taking into account different propagation delays from each transmitter to the star
coupler.
Although the multiplexed stream from one transmitter can be contiguous in time,
small guard times are required between signals from different transmitters to provide
tolerance for small errors in synchronization. The schedule shown has a fixed periodic
frame pattern, with F 1 and F 2 being two successive frames containing four slots per
frame. In this case, the destination address of each data segment is implicit in its position
in the frame, so no explicit addressing need be carried with the information itself to
indicate to which logical channel it belongs. This is typical of a fixed (as opposed to a
dynamic) capacity allocation. (The fraction of the frame time allocated to each logical
channel is proportional to its allocated capacity.)
Several operations that require exchange of control information are involved in initial-
izing the connections. A frame schedule must be determined by a connection manager
and must be made known to all participating stations. Also, each station must acquire
frame, slot, and bit synchronization. These functions take place during the connection
establishment phase. The only control information needed during the information trans-
fer phase is that required for maintaining timing synchronization – a process known
 Network Connections 109

as ranging. This is generally an operation that requires very little overhead.11 To keep
latency times and buffer sizes small, the slot and frame sizes should be small. However,
for efficient operation the slot size should be much larger than the guard time. These two
considerations dictate schedules with a small number of slots per frame, with slot and
frame sizes typically on the order of microseconds for transmission channel bit rates on
the order of gigabits per second. These aspects of connection control and synchroniza-
tion are typical of all fixed frame (synchronous) TDM/TDMA systems regardless of the
underlying technology.
The previous TDM/TDMA example illustrated how time division techniques can be
used to create multipoint logical connectivity with a minimum of optical transmitting
and receiving hardware, and minimal use of the optical spectrum. In realizing multipoint
connections, four successive operations were involved:
1. Logical channel multiplexing in the transmitting stations
2. Optical multicast, enabling each transmitter to reach several receivers
3. Optical multiple access, enabling several transmitters to access each receiver
4. Logical channel demultiplexing in the receiving stations
All multipoint connections in optical networks generally rely on the same fundamental
operations, as shown in subsequent chapters.

3.2.1.2 Time-Wavelength-Division Techniques

The previous illustrations showed how either TDM/TDMA or WDM/WDMA could be
used to achieve multipoint connectivity. The former was economical in the usage of
optical hardware and the optical spectrum but had limited throughput due to the fact
that transceivers were idle part of the time and only one λ-channel was used. The latter
had high throughput but required more optical transceivers and more spectrum. Is there
a way of obtaining the best features of both? Figure 3.9 shows another example using
the broadcast star network of Figure 3.6. This time, suppose full logical connectivity
is required (without loopback connections) as shown in the LCG of Figure 3.9(a).
The logical channels are now created using multiple wavelengths as well as multiple
time slots. In contrast to the WDM/WDMA case, each station contains only a single
transceiver. However, to achieve the necessary multipoint connectivity, some tunability
of the transceivers during the information transfer phase of the connections is needed.
Because optical connectivity between a given transmitter/receiver pair is created by
tuning both to a common λ-channel, a number of tuning options are available: Use fixed
transmitters with tunable receivers (FT-TR), tunable transmitters with fixed receivers
(TT-FR), or tunable transmitters with tunable receivers (TT-TR). In this simple example,
we use the FT-TR option, in which case each transmitter is tuned permanently to a distinct
11
In systems based on undirected stars or their equivalent, as in the current example, all stations see all
transmissions, including their own. This makes it relatively easy to execute mutual synchronization. The
RP in each station can observe the arrival times of all signals, and the station’s TP can use this information
to adjust the transmitted bursts to fall correctly into their allocated time slots. In cases in which stations do
not see their own transmissions (for example, directed stars, or stars without loopback connections), some
feedback for synchronization purposes must be provided from the receiving stations to the transmitting
stations via a separate control channel or a return data channel. See Section 5.7.
110 Multiwavelength Optical Networks

T R
1 1

2 2

3 3
(a) LCG

[1, 2] [1, 3] [1, 2]

Channel 1

[2, 3] [2, 1] [2, 3]

Channel 2

[3, 1] [3, 2] [3, 1]

Channel 3 t
Slot 1 2 3
F1
(b) Time-wavelength schedule

Figure 3.9 TDM/T-WDMA.

wavelength (that is, a unique λ-channel is “owned” by each transmitter). Each transmitted
signal fans out through the star coupler, reaching all three receivers (including its own).
If only one-to-one logical connectivity were desired, then all connections could be
realized by equipping each destination station with a single receiver that remains tuned
to the desired transmitter’s wavelength for the duration of a connection. To rearrange
the connections on a circuit-switched basis, the receivers would be retuned, but because
receiver tuning occurs as part of connection establishment, low tuning speeds (milli-
seconds) would be acceptable. Here we are seeking multipoint logical connectivity,
which can be achieved by using a single, rapidly tunable OR at each station and us-
ing TDM in the transmitters. This arrangement potentially provides full, simultaneous
logical connectivity among all sources and destinations. Any receiver wishing to select
information from a given transmitter tunes to that transmitter’s wavelength whenever the
transmission signal contains information for it, resulting in a time-wavelength-division
multiple access (T-WDMA) scheme.
To create a prescribed set of source-destination LCs, appropriate TDM schedules
must be set up for each transmitter so that the desired information can be extracted by
each receiver from the various transmitted information streams using a suitable tuning
rule. This creates a TDM/T-WDMA system. In contrast to the TDM/TDMA case, in
which transmitters operated in nonoverlapping bursts to avoid interference, the TDM/
T-WDMA system can operate with all transceivers active at all times, providing more
efficient usage of the optical equipment. A possible time-wavelength schedule using a
periodic frame with two equal size time slots is shown in Figure 3.9(b), which displays
the interleaved transmissions on each channel (wavelength) as seen at the star coupler.
Each transmitter multiplexes the logical channels destined to the other two receivers
into a fixed TDM frame, with all stations transmitting simultaneously on different
 Network Connections 111

wavelengths. The transmitters must be synchronized with each other at the star coupler, as
shown in Figure 3.9(b), so that the TDM frames for each wavelength are aligned to allow
conflict-free receiver tuning schedules.12 A receiver conflict occurs if information from
two different transmitters, destined for the same receiver, is transmitted during overlap-
ping time intervals, making it impossible for the receiver to pick up both transmissions.
In a conflict-free schedule, no destination station may appear more than once in each time
slot because this would require the station’s receiver to tune to two different wavelengths
at the same time. Note that the conflict-free condition is satisfied in Figure 3.9(b). In the
first slot, stations 1, 2, and 3 tune to wavelengths 3, 1, and 2, respectively, and in the sec-
ond slot, they tune to 2, 3, and 1, respectively. Because all channels and all transceivers
are active at all times, the network resources are utilized to maximum efficiency.
Note that the four ingredients of multipoint connectivity are present in the TDM/
T-WDMA case:
1. Logical channel multiplexing at the transmitters using TDM
2. Optical multicast via the star coupler
3. Optical multiple access using T-WDMA
4. Logical channel demultiplexing by tuning the receivers to the right wavelength at the
right time
Multipoint connections can be realized using many types and combinations of multi-
plexing and multiple access. These are discussed in more depth in Chapter 5.

3.2.2 Packet Switching in the Optical Layer: The MAC Sublayer

The fixed-capacity allocation produced by the periodic frame structure in the previous
examples is well adapted to stream-type traffic. However, in the case of asynchronous
packet traffic13 this approach may produce very poor performance. To see why, consider
two basic characteristics of applications typically carried using packet switching:
1. Burstiness: The “bursty” nature of the information flow tends to produce poor trans-
mission efficiency if fixed-capacity allocation is used.
2. Need for high logical connectivity: In typical packet-switched applications, it is
common for a station to require simultaneous logical connections with a large number
of other stations. (This is one of the primary advantages of packet-switched transport.)
Although the aggregate traffic may be large, each logical connection between a pair
of stations typically carries only a very small fraction of the total.
These considerations suggest that dedicating a fixed capacity to each LC is difficult,
inefficient, and offers poor quality of service. For example, imagine extending the
12
As in the TDM/TDMA case of Figure 3.8, this may require accounting for unequal propagation delays to
the coupler. In certain extreme cases involving long distances and/or large fiber dispersion, synchronization
at the coupler is insufficient to avoid conflicts in TDM/T-WDMA. These occur when the differences in
propagation times for different wavelengths may be great enough so that different λ-channels are aligned
at the coupler but are misaligned by the time they reach the receivers.
13
Henceforth, we usually employ the term packet generically to refer either to fixed-length cell-based or
variable-length packet-based data flow.
112 Multiwavelength Optical Networks

broadcast star example of Figure 3.6 to 100 stations, each supporting an end system
operating as a gateway to an LAS and requiring an LC to each of the 99 other gateways.
Full connectivity using TDM/TDMA would require 9900 time slots (one for each LC),
which would lead to poor optical transceiver utilization and large delays.
The poor transceiver utilization results from the fact that each transmitter or receiver
is idle 99% of the time. The delay is due to the fact that when a user in one LAS has traffic
ready to be transmitted to another LAS (say, a file to be transferred), it needs a large-
capacity allocation immediately. Instead, it has only one slot available out of 9900 in a
long frame. For example, with a 1-Gbps channel bit rate, the effective capacity allocated
to one connection is only 101 Kbps. The inferior performance is a consequence of the
fact that we are treating packet traffic exactly as if it were stream traffic and assigning a
fixed-capacity pipe to each LC.
However, by implementing a packet-switching function that reaches down into the
optical layer it is possible to maintain a very large number of LCs simultaneously
using dynamic capacity allocation. This is accomplished by exploiting the essential
advantage of the packet approach: packet-by-packet addressing and control. In packet
traffic the identity of the LC to which a packet belongs (that is, addressing information),
together with other control information, is carried in each packet’s header. Traditional
packet-switched transport networks (e.g., IP networks) process the packet headers in
their logical switching nodes (packet switches), where they perform packet sorting and
routing and other control functions. They typically have a sparsely connected logical
topology.
For example, if the network of Figure 3.2(b) is used to support a packet-switched
logical overlay it appears to be an ordinary packet-switched network with a logical
topology in the form of a unidirectional ring. Packets delivered to the ring by the end
systems are forwarded to the next LSN until they reach the destination node, at which
point they exit to the end system. The binary decision forward/exit is made by reading
the packet header. However, if the packets are processed in the transmission/reception
processors of the NASs instead of in the LSNs, the packet-switching functions are
executed in the transmission channel sublayer of the optical layer. This is especially
advantageous in the case of broadcast networks for three reasons:
1. Full logical connectivity for packet streams is supported easily by the fully connected
optical infrastructure already in place.
2. Optical broadcast together with packet processing in the NASs amounts to paral-
lel packet processing using the resources of all stations simultaneously instead of
sequential processing of all packets in each LSN they traverse.
3. An optical broadcast path is the most efficient support for logical multicast connec-
tions.
The reader should note the difference between packet switching in the optical layer
(which is accomplished electronically in the NASs) and optical packet switching, in
which most packet-switching functions, including some header processing, routing, and
possibly packet buffering are executed optically in the ONNs. This is discussed in detail
in Chapter 10.
 Network Connections 113

A schematic representation of a network executing packet switching in the optical

layer is shown in Figure 3.2(c). On the left we show an underlying structure in the form
of our ubiquitous broadcast star, the same physical topology that supported the fully
connected logical topology in Figure 3.2(a) and the logical ring in Figure 3.2(b). Now
the NASs are labeled PS to indicate the presence of an optical layer packet-switching
function replacing the electronic switching executed in the LSNs of Figure 3.2(b). The
logical topology is now shown as a “cloud” containing the five NASs. This is meant to
convey the idea that packet routing is effected using the NASs and the shared broadcast
medium, as opposed to the case of Figure 3.2(a), in which five stations are joined by
20 individual point-to-point LCs.
To achieve full connectivity on the broadcast star without resorting to logical switch-
ing, we use the broadcast-and-select property of the star, in which information in the
form of packets is broadcast to all receivers. Each one selects those destined for it and
discards the others. Although broadcast-and-select with fixed-capacity assignments for
stream traffic was described in Section 3.2.1, we now have the potential for dynamic
capacity assignment. The TPs can schedule packets based on instantaneous demand,
priorities, and quality of service requirements for different traffic classes.
Comparing the logical topologies in Figures 3.3, 3.2(a), and 3.2(c), they all show full
connectivity among the stations over a single logical hop. Their difference lies in how
capacity is assigned. In the case of Figure 3.2(c), the capacity of the broadcast star is
shared dynamically among the five stations, whereas in the other cases, a fixed capacity
is assigned to each logical link. In all cases, the actual physical topology – a ring or star –
is hidden from view and is largely irrelevant.
To delve further into implementation of packet switching in the optical layer, it is
useful to recall its antecedents. Packet switching over a shared medium existed well
before the deployment of optical fibers. Examples of conventional shared media are
terrestrial radio and satellite, in which the medium is free space (the “ether”), and wired
buses and rings, in which the medium is a copper cable. When these were introduced
into data networks, special MAC protocols were developed to adapt the higher layers of
the network architecture to the unique features of the shared medium. The role of the
MAC protocol is to replace the standard data link control protocols designed for a point-
to-point link by procedures for efficient utilization of a medium accessed simultaneously
by many users. Most MAC protocols are designed for LANs and MANs, and the basic
problem they must address is contention resolution; that is, how to control the scheduling
of packet transmissions among many contending stations so that the system performs as
efficiently as possible.
The most commonly used shared optical medium is the broadcast star. Control of
optical media requires a MAC protocol that has many of the same features as the earlier
MAC protocols yet is adapted to the unique properties of optical networks. A good
way to understand how an optical MAC protocol should operate is to examine a LAN
protocol designed for a wired bus: carrier-sense multiple access with collision detection
(CSMA/CD). Employed in Ethernet, this is probably the most widely used MAC protocol
today. To illustrate, consider the system of Figure 3.10(a). Four stations are shown
connected to the bus, which propagates signals in both directions. The station access
114 Multiwavelength Optical Networks

1 2 3 4

(a)

Interface

T/R
tap
(b)

LLC Frame

Data

Preamble ... DA SA ... LLC Frame CRC

Encapsulation
(c)

Figure 3.10 CSMA/CD.

arrangement appears in Figure 3.10(b), indicating that the end system (for example, a
data terminal, personal computer, or workstation) is interfaced to a transceiver coupled
into the cable. (Typically the interface is located in the end system, and the transceiver
is located at the cable tap.) A packet to be transmitted from the end system is passed to
the interface, which encapsulates it into a format suitable for transmission on the cable
and buffers it to await transmission.
A simplified picture of the encapsulated format is shown in Figure 3.10(c). The
information from the end system is first enveloped in a variable-length logical link
control (LLC) frame to which is added a preamble to allow the receiver to synchronize
and recognize the beginning of the frame. A header including destination address (DA)
and source address (SA)14 also precedes the LLC frame, which is followed by a group of
cyclic redundancy check bits (CRC) used for error detection (but not for error correction).
The procedure for transmitting the frames is as follows: On receipt of a data frame
from the end system, the interface sends the encapsulated frame to the transceiver
for transmission on the cable as soon as the transceiver senses the cable to be idle
14
If the LAN is part of a larger network (for example, a subnet of an internet), then the LAN addresses are
local addresses, which differ from global internet addresses. In some of our subsequent discussions of
packet switching in subnets of large optical networks, this local/global distinction is needed.
 Network Connections 115

[1, 2]
T1 t

[4, 3]
T4 t
abort [1, 2]

[1, 2] [4, 3]
R1 t
2

[4, 3] [1, 2]
R4 t
abort [4, 3]

Figure 3.11 CSMA/CD collision.

(the “carrier sense” part of the protocol). All receivers listen to all transmissions and
read their destination addresses, selecting those destined for their station and dropping
the others. This is equivalent to the broadcast-and-select procedure used in a star network.
If, during transmission of a frame, the transceiver detects another frame on the cable,
that is, a collision (the “collision detection” part), it aborts its transmission and schedules
the frame for retransmission at a later time.
A characteristic of this and many other MAC protocols is the possibility that a frame
will be lost, thereby requiring retransmission. The CSMA/CD protocol is a refined ver-
sion of random access, which uses neither carrier sense nor collision detection. However,
the two protocols share the following three properties: (1) there is no coordination among
transmissions from different stations, (2) collisions can occur, and (3) if collisions occur,
retransmissions are used to recover lost packets.
The refinements in CSMA/CD over random access simply reduce the probability of
collisions and mitigate their effects when they occur.
Figure 3.11 shows the events associated with a collision. It is assumed that packet
transmission times are long compared with τ – the end-to-end propagation time on the
bus. A worst case is shown in which stations 1 and 4, at opposite ends of the bus in
Figure 3.10(a), are transmitting to stations 2 and 3, respectively. Time traces as seen at
transmitters T 1 and T4 and receivers R 1 and R4 are shown in Figure 3.11. Observe that
T 1 begins transmission of a frame [1, 2] first, and it takes a time τ to propagate to station
4. Just before it reaches the other station, T4 , thinking the cable is idle, begins to transmit
a frame [4, 3], which propagates to the left, toward station 1.
Both of these transmissions begin to overlap on the bus, garbling the information
received at stations 2 and 3. But shortly after beginning its transmission, R4 detects
the collision with [1, 2], which has now reached station 4. This causes T4 to abort
its transmission as shown. None of this is known to station 1 until approximately
τ seconds later, when the initial portion of [4, 3] reaches the other end of the cable, at
which time R 1 detects the collision and aborts its transmission [1, 2]. The total time
occupied on the cable by this event is slightly more than 2τ . This represents time lost for
useful transmissions and limits the maximum throughput of the system. Furthermore,
116 Multiwavelength Optical Networks

each collision and retransmission causes an additional delay until the frame is finally
transmitted successfully. It should be clear from the figure that the smaller the end-to-end
propagation time τ , the less time is lost in collisions.
In fact, it turns out that the throughput efficiency of CSMA/CD under high loads
is approximately 1/(1 + 5a), where a = τ/m and m is the frame transmission time.
Thus, perfect efficiency (no time wasted in collisions) is approached as propagation
time gets small (i.e., a short cable) and/or frame transmission times get large (i.e, many
bits per packet and/or low transmission bit rates). For example, typical parameters for
Ethernet might be a cable of length 2.5 km, a transmission rate of 10 Mbps, and a frame
length of 1000 bits. Taking the speed of light in the cable as 1.5 × 108 m/s, this gives
an end-to-end propagation time of approximately 10.9 µs, a = 0.109, and a throughput
efficiency of approximately 0.65. The short propagation time means that the overall
delay for successful packet transmission, even after several retries, is quite small.
Unfortunately, all of this changes radically in the context of optical networks. For
MANs and WANs the distances and bit rates make CSMA/CD impossible. This still
leaves local access subnets. In a star-based LAS, a common form of access network, the
star coupler creates a shared medium much like the copper Ethernet bus. However, in this
case, the fibers are typically using unidirectional rather than bidirectional transmission.
Thus each station accesses the star via an OT/OR pair, using a pair of unidirectional
access fibers. In this configuration, the typical distances (of the order of kilometers) and
bit rates on a single λ-channel (of the order of gigabits per second) still yield a ≫ 1
for typical network parameters. Thus, CSMA/CD is impractical except in some special
cases. What are the alternatives?
To suggest some approaches, let us extend the TDM/TDMA and TDM/T-WDMA
examples of Figures 3.8 and 3.9 to packet switching using the star coupler of
Figure 3.6 operating in a broadcast-and-select mode. In these two cases, we created
three and six fixed-capacity LCs, respectively, by using a periodic slot assignment. Now,
one of our objectives is to achieve dynamic capacity allocation; that is, to implement
complete and instantaneous sharing of the broadcast medium on demand. The role of

Data Frames OT
TP

MAC ONN

Data Frames OR
RP

NAS

Figure 3.12 NAS equipped for packet switching.

 Network Connections 117

the MAC protocol is indicated schematically in Figure 3.12, which shows an NAS with
packet-switching functionality. We assume now that the data belonging to LCs [1, 2]
and [1, 3] is delivered from the end system to the NAS over either two separate external
ports or over a single port in the form of randomly interleaved packets. It is then up to
the TP to recognize the individual packets and process them accordingly, encapsulating
them into a properly addressed format (similar to the CSMA/CD format of Figure 3.10).
The encapsulated packets must then be scheduled and transmitted under the supervision
of a suitable MAC protocol (indicated by the MAC block in the station).
How should this protocol operate? It all depends on the objectives to be achieved
and the information and resources available to the participating stations. For simplicity,
let us begin with the assumption of a single, shared channel using only time division
techniques. Suppose the packets are of fixed length, so they can be inserted into slots of
equal length. Now the time traces in Figure 3.8 can be reinterpreted in a packet-switched
context. The time slots are no longer grouped into a periodic frame schedule; instead,
packets are transmitted on a slot-by-slot basis. Ignoring the fixed periodic schedule in
Figure 3.8, the data segments can now be interpreted as packets, where, for example,
the notation [1, 2] 2 represents a packet with source and destination addresses 1 and 2,
respectively, and with sequence number 2.
The reception processors in the stations must now contain suitable packet filters, which
have the functions of stripping the encapsulation, reading local destination addresses,
capturing all packets destined for the station, routing them to the correct external ports,
dropping the unwanted packets, and performing error checking. Thus, in this example
the reception processor in NAS 2 must filter out all packets belonging to LCs [1, 3]
and [2, 3], retaining only those on [1, 2]. Note that dynamic scheduling allows for the
capacity to be allocated based on instantaneous demand: When a station has a packet to
transmit, it schedules it in the next available slot. It is this feature of packet switching
that allows us to treat the broadcast star as a shared-capacity rather than fixed-capacity
medium.
The difficult part of dynamic scheduling is that, in the absence of a periodic frame
structure, some sort of coordination and synchronization among the transmitters is
needed to resolve contention among the stations for the available channel capacity.
It is the function of the MAC protocol to arbitrate contention, attempting to keep
collisions to a minimum and rescheduling transmissions in cases of lost packets. The
simplest protocols (e.g., random access) have no coordination and minimize collisions by
keeping traffic levels low. More complex protocols rely on the exchange of coordination
and scheduling information, which reduces collisions or avoids them altogether. (The
example in Figure 3.8 shows perfect scheduling with no collisions.)
Although the contention problem is the bad news in multiwavelength broadcast net-
works, the good news is that multiple wavelengths are available to use for concurrent
transmissions. To illustrate, consider a packet-switched generalization of the TDM/
T-WDMA case of Figure 3.9, as shown in Figure 3.13. We continue with the assumption
of a slotted system with six active logical channels. Figure 3.13(a) shows four slots in a
possible sequence of packet transmissions using three λ-channels and FT-TR. Note that
the schedule is now assigned dynamically according to the demand of each station, with
118 Multiwavelength Optical Networks

Channel 1 [1, 2]1 [1, 3]1 [1, 3]2 [1, 3]3

Channel 2 [2, 3]1 [2, 1]1 [2, 1]2 [2, 3]2

Channel 3 [3, 1]1 [3, 2]1 [3, 1]2 [3, 2]2

t
Slot 1 2 3 4

(a) FT-TR

[3, 1]2
Channel 1 [3, 1]1 [2, 1]1 [2, 1]2

Channel 2 [1, 2]1 [3, 2]1 [3, 2]2

[1, 3]3
Channel 3 [2, 3]1 [1, 3]1 [1, 3]2
[2, 3]2
t
Slot 1 2 3 4

(b) TT-FR

Channel 1 [1, 2]1 [3, 2]1 [1, 3]1 [1, 3]2 [3, 2]2 [1, 3]3

Channel 2 [2, 3]1 [2, 1]1 [2, 1]2 [3, 1]1 [2, 3]1 [3, 1]2
t
Slot 1 2 3 4 5 6

(c) TT-TR

Conflict
Collision

Figure 3.13 Packet switching in the optical layer.

connection [1, 3] carrying more traffic than the others. Note also that there is a receiver
conflict in the third and fourth slots, with receiver 1 required to tune to two different
channels in slot 3 and receiver 3 required to tune to two different channels in slot 4. In
these cases each receiver must make a choice, picking up one transmission and losing
the other.
Figure 3.13(b) shows a variation of the previous case, using three channels and TT-FR.
The same sequence of packet transmissions appears, but in this case they are carried
on different channels because the channel assignment is determined by the destination
address rather than the source address. As a result, the conflicts that appeared in the
third and fourth time slots in the previous example are now replaced by collisions in
 Network Connections 119

those slots. This is a more serious problem because a collision mutilates both packets,
resulting in double the loss of the FT-TR case.
Figure 3.13(c) shows another version of the same example, using only two channels
and requiring tuning of both transmitters and receivers. Now a properly coordinated
channel–slot schedule is shown, in which all of the packets are transmitted in six time
slots (the minimum possible number with two channels) without losses. Note that the
channel–slot assignment is done on a packet-by-packet basis and results in a different
interleaving of packets than in the previous cases. However, it can be seen from the
sequence numbers that the packets are delivered in the proper order.
Although the previous examples focused on point-to-point LCs, multicast connec-
tions come “free” simply by including the multicast destination addresses in the packet
headers. (Of course, packet scheduling must be arranged to accommodate the multicast
addresses without conflicts.)
In comparing the previous examples, it can be seen that each is characterized by
different constraints on capacity sharing. In TDM/TDMA, the medium (one λ-channel)
is fully shared among all logical channels. In FT-TR, TT-FR, and TT-TR, partial sharing
takes place in which all logical channels originating at a given station may share a
capacity not exceeding one λ-channel, with the same type of sharing occurring for all
channels terminating at a given station.15
The conflicts and collisions (and concomitant packet loss) in the FT-TR and TT-FR
cases resulted from a lack of coordination among the stations. To avoid losses, two
conditions must exist: (1) the transmitting stations must agree among themselves on a
conflict- and collision-free set of channel–slot assignments and (2) the receiving stations
must know these assignments so they can tune properly to pick up the packets destined
for them. Both of these conditions require the following actions: (1) each station must
request permission to transmit to a specified destination, (2) a control algorithm must
be executed (either centralized or distributed among all stations), and (3) a scheduling
decision must be communicated to the concerned parties.
Depending on the network configuration, these actions require exchanges of varying
amounts of control information among the stations. This might be done on one or
more wavelengths or time slots reserved for that purpose, which act as separate control
channels, and in some cases it might require additional optical transceivers. Control
information can also be piggybacked on information packets. In the FT-TR and TT-TR
cases, both transmitting and receiving stations must exchange control information, but
in TT-FR, it is only the transmitters that need to agree on a schedule to avoid collisions
because no conflicts are possible.
In typical packet-switched broadcast-and-select systems, there is usually a compro-
mise between ideal lossless communication (requiring extensive exchange of control
information together with possibly complex scheduling computations) and uncoordi-
nated lossy transmissions. The compromise involves a trade-off between cost (in terms
of NAS complexity and spectrum utilization) and performance (in terms of throughput,
15
All remarks here on conflicts and collisions in shared channel systems are predicated on the assumption of
a single transceiver per station. In cases where stations may have arrays of transmitters and/or receivers,
the rules governing conflicts and collisions are more complex (see Section 5.4).
120 Multiwavelength Optical Networks

loss, and delay). The performance-to-cost ratio for most MAC protocols diminishes as
the number of stations increases.
The procedures governing exchange of control information, conflict resolution, and
recovery from losses are the distinguishing features of MAC protocols. Figure 3.12
indicates as dashed lines the various control paths associated with media access control.
In the most general case, the MAC protocol coordinates many activities executed in the
other components of the NAS, including scheduling, transceiver tuning, and signaling to
and from other stations. Because the specifics of these exchanges vary from one protocol
to another, we defer further remarks to Section 5.6, where a more detailed treatment of
packet switching in the optical layer is presented, and to Section 5.7, where applications
to access networks are discussed.
It is instructive to place the optical layer MAC protocols in the context of the multi-
wavelength network architecture. The relations among the layers are most easily un-
derstood by comparing them with MAC protocols in a conventional layered network
architecture: the OSI Reference Model (see [Schwartz87], p. 15). The second layer of
the OSI model is the DLC layer, whose function is to ensure reliable and ordered delivery
of data frames over a point-to-point link (see Figure 3.14). The DLC layer uses the ser-
vices of the physical layer, which provides for transmission of the underlying bit stream
on a point-to-point physical medium such as a wire or fiber. (Layer 2 in the IP protocol
stack, which corresponds to the DLC layer here, is called the network interface layer.)
As shown in Figure 3.14, in the case of shared media these two layers are replaced
by three: the LLC and MAC sublayers and a new physical layer adapted to multiaccess
on the shared medium. For example, in CSMA/CD, the higher level features specific
to this protocol, including encapsulation and decapsulation, scheduling transmissions
and retransmissions, and error detection, are performed in the MAC sublayer. Lower
level functions associated with the physical medium, including sensing the idle channel

Application
Presentation
Session
Transport

OSI

MWNA
Logical Path Network
LLC Logical Connection DLC LLC
Transmission
MAC MAC
Channel Physical
Physical λ-Channel Physical
Optical Path
Fiber

Figure 3.14 MAC protocol in the layered architecture.

 Network Connections 121

and detecting collisions, as well as bit transmission and reception, are performed in the
physical layer. The remaining high-level functions associated with error-free delivery of
frames in the correct sequence take place in the LLC sublayer. Comparing this layered
view with the hardware implementation of CSMA/CD shown in Figure 3.10(b), the
physical layer would normally be implemented in the transceiver, the MAC sublayer in
the interface, and the LLC sublayer in the interface and/or the end system.
Now let us turn to the optical network case. The position of optical layer packet-
switching functions in the multiwavelength network architecture is shown in Figure 3.14
for comparison with the OSI protocol stack. In this case, similar LLC, MAC, and physical
layers can be defined, which take the place of the logical connection, transmission
channel, and λ-channel sublayers as shown. The location of the MAC and physical
layers at a level corresponding to the transmission channel and λ-channel layers reflects
the fact that medium access generally requires special operations on the bit stream
(frame encapsulation and decapsulation, packet filtering, parity check computation, and
so forth) and may require control of the optical transceivers (rapid tuning for WDMA
operation and/or on/off switching for TDMA operation).
As in the case of conventional packet-switched networks, an LLC frame enveloping
the useful data is prepared in the LLC sublayer and is passed down to the MAC sublayer
together with the local address (i.e., the identity of the NAS) to which it is destined.
The MAC sublayer encapsulates it for transmission on the optical medium in a format
similar to Figure 3.10(c). Included in the encapsulation is the source address (the address
of the transmitting NAS) and the destination address (the destination NAS specified
by the LLC sublayer). The encapsulated frame is then passed to the physical layer,
which does the actual transmission on the optical medium, executing the necessary slot
synchronization and transmitter tuning. Encapsulated frames received by an NAS are
passed up from the physical layer to the MAC sublayer, which performs error detection
and packet filtering (i.e., address checking), strips away the encapsulation, and passes
the successfully received frames up to the LLC sublayer.
In the case of a purely optical transport network (e.g., end systems attached directly to
NASs), as in Figure 3.2(a), the logical link layer would normally reside in the end system,
with the MAC and physical layers in the NAS. However, the transport network may
consist of several packet-switched optical subnets joined together through an overlay that
consists of logical switching nodes (e.g., IP routers). In this case, the subnets function as
multipoint links between the LSNs, and we have two layers of packet switching. Now the
LLC sublayer would be located in the LSNs, where a routing table would determine
the next (optical layer) destination address. In this case the destination address may
be the local subnet address of an intermediate station on a logical multihop path to the
ultimate destination (see the hypernet example in Section 3.5.2).

3.2.3 Additional Comments on Broadcast-and-Select

We have seen that static networks can be operated in a broadcast-and-select mode with
either fixed-capacity or dynamically scheduled (packet-switched) LCs. Fundamental to
broadcast-and-select is the fact that LCs can be created and reconfigured solely by the
actions of the access stations, with logical connectivity ranging from a few point-to-point
122 Multiwavelength Optical Networks

connections to complete multipoint connectivity. The multiplexing and multiple access

procedures necessary for realizing high connectivity rely in important ways on resource
sharing: in the stations and on the fibers. This is a recurring theme throughout the book.
The broadcast-and-select principle based on a star, bus, or tree physical topology
cannot be scaled to large networks for three reasons:
1. Spectrum use: Because all transmissions share the same fibers, there is no possi-
bility of optical spectrum reuse, so the required spectrum typically grows at least
proportionately to the number of transmitting stations.
2. Protocol complexity: Synchronization problems, signaling overhead, time delays,
and processing complexity all increase rapidly with the number of stations and the
number of LCs.
3. Survivability: There are no alternate routes in case of a failure. Furthermore, a failure
at a critical point (either the star coupler or the head end of a folded bus) can bring
the whole network down.
For these reasons, a practical limit on the number of stations in a broadcast star is
approximately 100. Nevertheless, we show in Section 3.4.2 how larger networks can
be constructed using broadcast-and-select subnets as basic building blocks. Using this
technique, it is possible to construct networks that circumvent the limits of geographical
reach imposed by the optical technology, as well as the limits of connectivity and
throughput imposed by economics.

3.3 Wavelength-Routed Networks

In the previous section, we limited consideration to connections over networks with

fixed optical paths and no wavelength selectivity in the nodes. Much more flexibility is
achieved when wavelength selectivity and/or controllability is introduced in the optical
path sublayer producing a wavelength-routed network. These networks have the prop-
erty that each λ-channel can be recognized in the ONNs and routed individually. We
assume throughout that the stations contain transceivers tunable over the full range of
usable wavelengths, and the network nodes are either static (e.g., the static router of
Figure 2.10) or controllable (e.g., wavelength-selective cross-connects). We limit con-
sideration to transparent networks where all optical paths are point-to-point (either fixed
or reconfigurable), so the only way multipoint logical connectivity can be achieved is
by using several point-to-point optical connections (WDM/WDMA). This normally re-
quires one optical transceiver for each connection.16 (Multipoint connections supported
16
In theory, a multipoint connection can be set up in a wavelength-routed network using time and wavelength
division techniques. Thus, if a single transmitting station is connected to several receiving stations, with
each connection realized on a point-to-point optical path using a distinct wavelength, then a one-to-
many connection can be realized by tuning the transmitter rapidly over all wavelengths in the set. A
similar arrangement can be constructed for many-to-one connectivity using receiver tuning. In principle,
many-to-many connectivity can also be achieved by tuning both transmitters and receivers rapidly. However,
this approach becomes exceedingly complicated because each path has a different propagation delay.
Therefore, we do not pursue it further here.
 Network Connections 123

by multipoint optical paths are discussed in the context of linear lightwave networks in
Section 3.4 and in Chapter 6.) To keep things simple, we also assume in this chapter that
there is no wavelength conversion in the network. (Wavelength conversion in WRNs is
considered in Chapter 6.) As indicated in Figure 2.6(b), a wavelength-routed network
with m λ-channels can be viewed as m independently controlled copies of the original
network.
Because all connections are point-to-point, the distinctions between connectivity at
the optical path, optical connection, and LC levels disappear, so it is sufficient to deal
with optical connections only, with the understanding that an optical connection (a, b)i
supports a corresponding LC [a, b] and is supported by a corresponding optical path
a, b . In moving from the static networks of Section 3.2 to the WRNs in this section,
we shift our focus from logical connections created by the actions of the NASs alone
to optical connections created by the coordinated actions of the NASs and the optical
network nodes.
Wavelength selectivity in the ONNs offers considerable advantages in terms of spec-
trum reuse, even without control of the optical paths. Suppose, for example, that we
reconsider the example of Figure 3.2(a), where a star network is used to provide full
interconnection of five end systems. To change this from a broadcast-and-select net-
work, we replace the star coupler with a static wavelength router. Figure 3.15 shows the
resultant network, in which each NAS contains four transceivers tuned to route optical

1 1

2 2

5 5

(a) Wavelength Router

Destination
1 2 3 4 5
1 1 2 3 4

2 4 1 2 3

Source
3 3 4 1 2

4 2 3 4 1

5 1 2 3 4

(b) Wavelength Assignments

Figure 3.15 Wavelength-routed star network.

124 Multiwavelength Optical Networks

connections to the other four NASs and to receive signals from the others. Using the
wavelength assignment rule described in Section 2.3.1.2, these connections can be real-
ized using four wavelengths for a fivefold reuse of the optical spectrum. The wavelength
assignments for the optical connections are shown in Figure 3.15(b).
Comparing the static router with the broadcast star of Figure 3.2(a), we find that five
times as many wavelengths are necessary to support full connectivity in the broadcast
star, reflecting the lack of wavelength reuse in the broadcast network.

3.3.1 Routing and Channel Assignment

In establishing an optical connection, we must deal with both channel assignment and
routing. Channel assignment (executed in the λ-channel sublayer) involves allocating an
available wavelength to the connection and tuning the transmitting and receiving station
to the assigned wavelength. Routing (executed in the optical path sublayer) involves
determining a suitable optical path for the assigned λ-channel and setting the network
nodes to establish that path. Because an optical path in a WRN is associated with a
particular wavelength, a path cannot be set up until a wavelength is allocated.

3.3.1.1 Dedicated and Switched Connections: Rearrangeability

Implementation of routing and channel assignment (RCA) is a quite different problem
depending on whether the optical connections are dedicated or switched. Dedicated
connections are assumed to be held for a relatively long period of time. This would be the
case, for example, for connections provisioned to support a fixed logical network overlay
riding on top of the optical infrastructure. These connections would be chosen at the
design stage of the LN and perhaps modified occasionally in response to changing load
conditions or equipment failure. Switched (demand-assigned) connections, however,
are established and released on demand, with holding times that might be as short as
a minute or less. Thus, a network supporting only dedicated connections can be built
using just static ONNs. For example, a WRN in the form of a ring can be constructed
using static WADMs as network nodes. In the infrequent cases where connections are
modified, the WADMs can be reconfigured manually.
For our purposes the characteristic that distinguishes dedicated from switched con-
nections is that requests for the former occur in the form of a prescribed set, whereas
switched connection requests occur as a random sequence. Thus, the complete set of
dedicated connections for an LN is known and established when that network is ini-
tialized. However, requests for switched connections are signaled to an optical network
manager as they occur, and the ONM has the responsibility of (1) deciding whether to
accept or to block the request and (2) determining a routing and channel assignment. In
practice, the admission control decision may be based on various considerations, includ-
ing the current network load, fairness, priorities, and so forth. However, for simplicity
we assume here that the only factors that govern acceptance of a connection are the
physical constraints on RCA that either permit or block the connection. It should be
 Network Connections 125

clear that the ability of a network to accept a switched connection request depends on its
current state; that is, the pattern of connections that are currently active in the network.
As we shall see, rearrangement of active connections can sometimes create free paths
for new requests that would otherwise be blocked.

3.3.1.2 Channel Assignment Constraints

We illustrate the constraints on channel assignment in WRNs using the example of
Figure 3.16. The figure shows a network containing seven access stations and three
ONNs. Each internodal link consists of a single pair of unidirectional fibers. Because
the physical topology is a tree, there is only one possible optical path between any pair
of stations, so there are no routing decisions to be made. Because there is no wavelength
conversion in this network, each optical connection must obey the wavelength continuity
condition already invoked in Section 2.2. Suppose that the available optical spectrum
contains two λ-channels on wavelengths λ 1 and λ 2 . The unidirectional optical connec-
tions (2, 4), (1, 6), and (5, 7) are to be established. Assume first that they are dedicated
connections. By inspecting the complete set, it is clear that the wavelength assignment
(2, 4) 1 , (1, 6) 2 , and (5, 7) 1 is a feasible choice because connections superimposed on
common fibers are always assigned distinct λ-channels. We encountered this distinct
channel assignment condition earlier in the context of star networks. The DCA condi-
tion is one of the basic optical network routing constraints. It ensures that the optical
signals sharing a fiber do not interfere with each other at the receivers.
Now suppose the same set occurs as part of a sequence of switched connections as
follows. The first request is for connection (2, 4), which is assigned λ 1 . This is followed
by (3, 6) and (5, 7), which are assigned λ 1 and λ 2 , respectively. Then (3, 6) is terminated
and a request for (1, 6) is received. Because λ 1 is now in use on link a and λ 2 on link b,
the current state of the network makes it impossible to add the new connection without
a DCA violation, and it is blocked. Note that blocking was caused by an unfortunate
evolution of the state of activity in the network. However, if it were possible to rearrange
the active wavelength assignments to (2, 4) 1 and (5, 7) 1 , then the new request could be

1 3 6

a b

2 4 5 7

Figure 3.16 Channel assignment example.

126 Multiwavelength Optical Networks

accepted using λ 2 .17 We have encountered an analogous rearrangeability problem before

in the context of permutation switches (see Section 2.3.2.1). In the switch, however, the
entity being rearranged is a path rather than a wavelength assignment. In a larger WRN,
connection rearrangement might involve changing both optical paths (rerouting) and
channel assignments.
This is generally impractical for two reasons:

1. It requires momentary interruption of active connections, which may be unacceptable

to users.
2. Finding a rearrangement that “opens up” the network to the new connection request
typically requires solving an extremely complex combinatorial problem.

We therefore rule out the possibility of rearranging switched connections. Note, how-
ever, that given a prescribed set of dedicated connections, knowledge of all connections
on the list can often be used to accommodate them without conflicts. Thus, we see that
blocking is more likely to occur when assigning a sequence of switched connections
rather than a set of dedicated ones, and it is more likely to occur without wavelength
converters than with them.
As one more variant of this example, suppose we change things by using only a
single wavelength, but putting two fiber pairs in each internodal link; that is, we trade a
twofold reduction in spectrum usage for a twofold increase in fibers. How does network
performance change? In all the cases cited earlier, there were never more than two
connections active on an internodal fiber (and only one connection active on each access
fiber). This implies that all connections can be accommodated on the modified network
while using (and reusing) only one wavelength.
There is an important observation here: Networks with multifiber links perform better
than those with multiple wavelengths. The rule is that a network operating on F wave-
lengths with one fiber pair per link can never perform better (and generally will perform
worse) than the same network with its links increased to F fiber pairs and its spec-
trum decreased to a single wavelength. However, a network operating on F wavelengths
with one fiber pair per link and all of its nodes equipped with wavelength converters is
equivalent to the same network with its links increased to F fiber pairs and its spectrum
decreased to a single wavelength.18
The reason for these relations is that transparent multiwavelength networks without
wavelength converters operate under the constraint of wavelength continuity, but there
is no equivalent constraint in multifiber networks, nor in networks using wavelength
conversion.
To summarize the previous discussion as it applies to wavelength-routed networks
without wavelength converters, we see that routing and channel assignments are subject
to the following two constraints:

17
If wavelength conversion was allowed, the new connection could be accommodated without rearrangement
by assigning it λ 2 on link a and λ 1 on link b.
18
This assumes that each ONN is capable of permuting connections among all fibers on a link.
 Network Connections 127

1. Wavelength continuity: The wavelength of each optical connection remains the same
on all links it traverses from source to destination.
2. Distinct channel assignment: All connections sharing a common fiber must be as-
signed distinct channels. (This applies to access links as well as internodal links.)

Note that the first is a constraint imposed by the laws of physics, whereas the second
is a design constraint required for proper network operation. At this point, it is important
to clarify the meaning of the phrase distinct channel. In the context of this section,
channel means λ-channel, so distinct channel assignment means distinct wavelength
assignment. However, we have already seen a case, in Figure 3.8, in which signals on
the same wavelength were superimposed on the same fibers (the outbound fibers on a
star) without interfering. This was possible because they were nonoverlapping in time.
In that example, the noninterfering channels corresponded to optical signals sharing a
common wavelength but confined to different time slots using TDMA. More generally,
we shall define channels to be distinct if they can be distinguished at a receiver when
superimposed on its access fiber. Distinct channels are normally orthogonal, typically
nonoverlapping in either optical frequency or time. As we consider more complex
multiplexing and multiple access schemes in later chapters, the meaning of distinct
channels is reexamined.
Although the DCA condition is necessary to ensure distinguishability of signals on the
same fiber, it is possible (and generally advantageous) to reuse the same wavelength on
fiber-disjoint paths. This includes paths that may be carried on different fibers contained
in the same (multifiber) link. Channel assignment constraints equivalent to the DCA
condition exist in all types of communication networks. However, the wavelength conti-
nuity condition is unique to transparent optical networks, making routing and wavelength
assignment a more challenging task than the related problem in conventional networks.
The performance of a network depends not only on its physical resources, but also on
how it is controlled. Routing and channel assignment is the fundamental control problem
in large optical networks. Generally, the RCA problem for dedicated connections can be
treated off-line, so computationally intensive optimization techniques are appropriate.
However, routing and channel assignment decisions for switched traffic must be made
rapidly, and hence suboptimal heuristics must normally be used. The RCA problem is
made even more complex when fault recovery is considered. In this case, simultaneous
assignments are needed for both working and backup paths. (See Chapter 8.) The
objective of an RCA algorithm is to achieve the best possible network performance
within the limits of the physical constraints.
The RCA problem can be cast in numerous forms. For example, in the case of dedicated
connections we may wish to accommodate a prescribed set of optical connections using a
minimum number of wavelengths or using the shortest optical paths for each connection
consistent with a limit on the number of available wavelengths. In the case of switched
connections, the objective might be to maximize the offered traffic given a fixed number
of wavelengths and a specified limit on blocking probability. The next section includes
a few simple examples, confined to dedicated connections only. A more comprehensive
look at RCA for both dedicated and switched connections appears in Chapter 6.
128 Multiwavelength Optical Networks

3.3.2 Routing and Channel Assignment Examples

Several examples of RCA are presented in this section to illustrate various possible trade-
offs among network resources and performance. All examples deal with finding optical
connections that provide full logical connectivity among five stations; that is, embedding
a fully connected five-station logical topology on the given physical topology while
observing the constraints of the network. An RCA solution that satisfies the constraints
is given in each case. Although this is presented as a dedicated connection problem, its
solution gives an RCA rule for demand-assigned connections as well. If the five stations
make point-to-point connection establishment/release requests in a random sequence,
the requests can always be accepted without blocking as long as the RCA rule found for
dedicated full connectivity is used.
In these and other examples to follow in later sections, the wavelength assignments
are dependent on the manner in which the NASs are connected to their ONNs. A typical
arrangement, in which a single fiber pair access link connects an NAS to an ONN,
appears in Figure 2.29. In the case at hand, the ONN contains a wavelength-selective
cross-connect, and the two access fibers connect to an input and output port on the
WSXC. In the NAS, several optical transmitters are multiplexed onto the outbound
access fiber, and several optical receivers are fed by the signals demultiplexed from the
inbound access fiber. The transmission/reception processors interface the transceivers to
the external ports and may include electronic multiplexing/demultiplexing and switching
equipment for connecting specified data streams from the transceivers to designated
logical input/output ports. Although all transmitters and receivers may be tunable, the
DCA condition requires that all inbound (outbound) channels (wavelengths) must be
distinct in this case, because they share common fibers.
Figure 3.17 shows a different version of the NAS–ONN connection. Now n, the
number of fiber pairs in the access link, is the same as the number of internodal links
incident on the node (assuming one fiber pair per internodal link). In the most general
case, each access fiber directed outbound from the NAS may multiplex signals from
several fixed or tunable optical transmitters, and each access fiber directed inbound to
the NAS may feed signals to several fixed or tunable optical receivers. Thus, the contents
of each of the n boxes labeled OTs/ORs in Figure 3.17 corresponds to the optical portion
of the NAS in Figure 2.29. Again, the transceivers are interfaced to external input/output
ports through TPs/RPs that may be used for cross-connecting data streams in an arbitrary
fashion between the transceivers and end system equipment. In this case, there are no
limitations on channel assignment due to the access fibers, because the access fibers
can accommodate as many wavelengths as the internodal fibers incident on the network
node.
We shall refer to these multifiber pair access connections as nonblocking access links.
Carrying the nonblocking idea one step further, suppose that the usable optical spectrum
contains  λ-channels and that each OTs/ORs box contains an array of  fixed tuned
transceivers covering the complete set of channels. Now the NAS (and its access fibers)
can support as many optical connections as the internodal links incident on its ONN
can carry. We shall refer to this arrangement as a nonblocking NAS, meaning that no
 Network Connections 129

Internodal Links

1 2 ... n

ONN

...
1 2 n Access
Fiber Pairs

...
OTs/ORs OTs/ORs
... OTs/ORs

NAS

TPs/RPs

...
Input/Output Ports

Figure 3.17 Nonblocking access link.

potential optical connection terminating on that NAS will be blocked due to limitations
in the station or its access link.19

3.3.2.1 The Bidirectional Ring

Our first example is shown in Figure 3.3, wherein a fully connected logical topology
containing five NASs is to be embedded on a ring. Because each LC requires a point-to-
point optical connection in this case, four optical transceivers are needed in each NAS.
The objective now is to specify RCAs using a minimum number of wavelengths. As
mentioned earlier, the feasible wavelength assignments depend on how each NAS is
connected to the network. The simplest approach is the configuration of Figure 3.18(a),
which uses the single fiber pair access connection of Figure 2.29.
In the ring of Figure 3.18(a), the network nodes are WSXCs with three input and
output ports, routing each wavelength around a ring or between a ring and an NAS.
The rings are labeled L(R) for left (right) propagation past the NASs. Each NAS has
four optical transmitters multiplexed onto its outbound access fiber and four receivers
fed from its inbound access fiber. Because this is a two-access-fiber connection, the
optical connections are subject to the DCA condition on the access links; each inbound
(outbound) connection to an NAS must be on a distinct wavelength.

19
The nonblocking condition does not apply to connections from other NASs attached to the same ONN.
130 Multiwavelength Optical Networks

4
3

R
1

NAS

(a) Ring

Destination

1 2 3 4 5

T R 1 1L 2L 3R 4R
1 1

2 1R 3L 4L 2R
Source
3 3R 4R 2L 1L

4 4L 2R 1R 3L

5 2L 3L 4R 1R
5 5
(b) OCG (c) Routing Table

Figure 3.18 Bidirectional ring: single access fiber pair.

 Network Connections 131

A bipartite optical connection graph (OCG) for this example is shown in

Figure 3.18(b). The DCA requirements on the access links translate into a classic
edge-coloring problem on the OCG. Selecting distinct wavelengths is equivalent to col-
oring the edges of the OCG so that all edges incident on the same vertex have different
colors. It is well known that this problem can always be solved in bipartite graphs using
a number of colors equal to the maximum degree of the vertices, which in this case is
four. This is only a necessary condition, however, because DCA must also be observed
on each internodal fiber in the ring.
A solution to the RCA problem, in this case using shortest path routing, is shown in an
RCA table in Figure 3.18(c), which indicates the path and wavelength assigned to each
optical connection. (An entry in the table of the form [nL] means that the corresponding
connection is routed on ring L using wavelength λn .) Shortest path routing means that
each connection is routed on the ring (L or R) yielding the least number of optical hops:
in this case, no more than two. Note that the problem is solved using four wavelengths,
for a spectrum reuse factor of 5.
Another approach uses a nonblocking two fiber pair access link following the inter-
connection structure depicted in Figure 3.17.
In this case, the ONN–NAS combination can be realized as shown in Figure 3.19(a).
Now the NAS consists of two separate pairs of transceivers, with each pair multiplexed
onto a separate access link L or R, connected to one of the rings. Using two access fiber
pairs instead of one removes the DCA constraint on the access links, so fewer wave-
lengths are required. Because the allowable interconnections are limited, the WSXC can
be realized using two 2 × 2 multiwavelength switches, one for each ring, operating as
wavelength add/drop multiplexers. Each WADM is similar to that shown in Figure 2.23
except that each access fiber carries multiplexed add/drop wavelengths instead of a single
wavelength. In effect there are two separate unidirectional ring/WADM–NAS combina-
tions grouped together at each node. The external input/output ports are connected to
the transceivers through the TPs/RPs.
A possible set of wavelength assignments using shortest path routing is shown in
the RCA table in Figure 3.19(b). For example, the source (destination) assignments in
row (column) 2 correspond to the transmitter (receiver) tuning for NAS 2, as shown
in Figure 3.19(a). This RCA requires only three wavelengths instead of four for full
connectivity, because wavelengths can be reused on the same access link (but not on
the same access fiber). It is optimal in the sense that three is the minimum number of
wavelengths required for full connectivity irrespective of how the NAS–ONN connection
is realized. (A more general treatment of RCA on rings appears in Section 6.3.4.2.)

3.3.2.2 A Mesh Network

Now consider the five stations connected to ONNs in the mesh network of Figure 3.20.
Our objective is the same as it was in the case of the ring: Find an RCA using a minimum
number of wavelengths to realize the fully connected OCG shown in Figure 3.18(b).
Again, each NAS must have a total of four transceivers. In this case a lower bound
on the number of required wavelengths is two, no matter what routing is used and no
matter what access connections are used (why?). Now let us first assume single fiber
132 Multiwavelength Optical Networks

ONN 2
L
R

MWS

MWS

L R

1 3 1 3 1 3 1
1 3

NAS 2

TPs/TRs

Input/Output Ports
(a) ONN–NAS Combination

Destination

1 2 3 4 5

1 1L 2L 2R 1R

2 1R 1L 3L 3R
Source
3 2R 1R 2L 1L

4 2L 3R 2R 3L

5 1L 3L 1R 3R

(b) Routing Table

Figure 3.19 Bidirectional ring: two access fiber pairs.

pair access connections, which means that the lower bound is raised to four because four
connections must share each access fiber. It is left as an exercise for the reader to show
that four wavelengths are sufficient using shortest path routing.
Unfortunately, this result provides no improvement in spectrum utilization over the
bidirectional ring, despite the richer physical topology of the mesh. This is because
the access fibers are the bottleneck. To remove this bottleneck, suppose a nonblock-
ing multifiber pair access link is used for each station. (No increase in the number
of transceivers is required.) In this case all connections can be realized using two
 Network Connections 133

4 3

D C

A B

1 2

Figure 3.20 A mesh network.

wavelengths, representing a ten-fold reuse of the spectrum.20 The exact wavelength

assignment is left as an exercise for the reader.
These RCA examples illustrate some key issues in the design and control of
wavelength-routed networks. In particular, they show a trade-off between the extent
of the fiber topology and the optical spectrum requirement: The more fibers there are
and the more “densely” they are interconnected, the fewer wavelengths are needed. In
this trade-off, the functionality of the network node is a limiting parameter because the
WSXC permits only point-to-point connections. The networks described in the next
section broaden our view to include multipoint optical connectivity, thereby expanding
greatly the options in implementing desired connectivity patterns.

3.4 Linear Lightwave Networks: Waveband Routing

The three classes of purely optical networks discussed in this chapter all have the same
basic structure: NASs and optical nodes interconnected by fiber links. Their differences
reside in the functionality of the nodes. The static networks described in Section 3.2
and the wavelength-routed networks of Section 3.3 represent two extremes of ONN
functionality: the former having no wavelength selectivity and no control and the latter
having independent control of each wavelength. In other words, the static case comprises
static single-waveband networks, whereas the wavelength-routed case corresponds to
dynamic single-wavelength-per-waveband networks.
The LLNs described in this section subsume both of the other two classes as special
cases. We use the term LLN to designate any network with nodes that perform strictly lin-
ear operations on optical signals. The most general type of LLN has nodes that function

20
The minimum required number of access fibers varies from station to station, depending on the routing and
wavelength assignments.
134 Multiwavelength Optical Networks

as waveband-selective linear divider-combiners (LDCs) (that is, generalized optical

switches). Recall from Section 2.3.2.3 that an LDC performs controllable optical signal
routing, combining, and splitting. The combining function allows for the multiplexing
of channels from several inbound fibers onto a single outbound fiber, and the splitting
function provides for multicasting of optical power from a single inbound fiber to several
outbound fibers. These functions are required to support multipoint optical connectivity.
As was the case for WRNs, we exclude wavelength conversion from the LLNs and
assume that switching within an optical network node occurs on a slow timescale, with
switches changing their settings only during optical path establishment and release.
Waveband selectivity in the nodes means that the optical path layer routes signals as
bundles that contain all λ-channels within one waveband.
One can conceive of many special cases of LLNs. First, there are the “limiting” cases
mentioned before. The static networks treated in Section 3.2 are examples of networks
having no wavelength selectivity within their ONNs (single-waveband networks), in
which the nodes are fixed but still have combining and splitting functionality. At the
other extreme, the wavelength-routed case of Section 3.3 has no combining and splitting
functions, but spectral resolution is refined to the point where the bundles recognized by
the nodes consist of only one λ-channel.
Another important category is a single-wavelength-per-waveband network possessing
combining and splitting properties (i.e., an “enhanced” WRN, containing LDCs in its
nodes rather than permutation switches). Such a network is capable of establishing optical
multicast connections as well as optical many-to-one connections on a wavelength-by-
wavelength basis. A simple application of optical multicast is a “drop-and-continue”
function at the optical level, providing an efficient way of distributing information on
a common λ-channel to multiple destinations. Optical many-to-one connections are
useful for time division multiplexing of LCs originating at different access stations onto
a common λ-channel. Neither one of these functions can be realized using an ordinary
WRN operating with permutation switches.
Because ordinary wavelength-routed networks are limited to point-to-point optical
paths, there is no need to distinguish among a logical connection, an optical connection,
and an optical path in those networks. However, with the potential for multipoint paths in
LLNs, all layers of connectivity and their interrelations must be examined more carefully.
We already saw an example of this in the broadcast star. For example, in Figure 3.6, using
WDM/WDMA the LCs [1, 2] and [1, 3] are carried on individual optical connections
(1, 2) 1 and (1, 3) 2 , which are in turn carried by the multicast optical path 1, {1, 2, 3} .
However, in the TDM/TDMA case, the same LCs are time division multiplexed
on a single multicast optical connection (1, {2, 3}) carried by the OP 1, {1, 2, 3} . In
each case the optical path reaches all destinations, intended or not, because there is no
controllability or selectivity in the star node. How are the LCs sorted out in the receiving
stations? In the WDM/WDMA case the optical connections (and the LCs they carry)
reach only their intended destinations because the ORs are assumed to tune selectively to
the assigned wavelengths. In the TDM/TDMA case, all optical connections are carried
on the same wavelength, so selecting the proper LC at the receiving stations must be
accomplished at the transmission channel level through time division demultiplexing.
 Network Connections 135

Last, in the TDM/T-WDMA case, LCs must be sorted out by combined actions at the
λ-channel level (receiver tuning) and at the transmission channel level (time division
demultiplexing).
These connectivity alternatives reappear, coupled with the added possibilities of
switching optical paths, as we examine the properties of LLNs.

3.4.1 Routing and Channel Assignment

The two constraints on optical connections applicable to wavelength-routed networks
(wavelength continuity and distinct channel assignment) also apply to LLNs. But when
each waveband contains multiple channels (a condition unique to LLNs), additional
constraints apply within each waveband. These are
r Inseparability: Channels combined on a single fiber and situated within the same
waveband cannot be separated within the network.
r Distinct source combining (DSC): Only signals from distinct sources are allowed to
be combined on the same fiber.
Inseparability is a consequence of the fact that the LDCs operate on the aggregate
power carried within each waveband without distinguishing among signals on different
channels within the band; that is, inseparability is a condition imposed by the architecture
of the optical switch combined with the choice of channel spacing.
Figure 3.21(a) illustrates inseparability. It shows two point-to-point optical connec-
tions (1, 1∗ ) and (2, 2∗ ) assumed to be within the same waveband. Connection (1, 1∗ ) is
routed via the minimum-hop path 1, 1∗ comprising nodes A–B–C–F–G, and connec-
tion (2, 2∗ ) is routed along its own minimum-hop path via nodes A–B–D–E. The label
Si on a link in Figure 3.21 denotes a signal generated at a source station i. Observe that
power from both sources is combined on fiber a. Thus, the DCA condition requires them
to be on distinct channels. For example, they might be carried on different wavelengths
within the same waveband (WDMA) or on the same wavelength but in different time
slots (TDMA). Because the two signals are in the same waveband, they cannot be sep-
arated at node B. Thus, to route them both to their destinations, node B must multicast
both signals to nodes C and D, expanding the intended point-to-point paths to multicast
optical paths 1, {1∗ , 2∗ } and 2, {1∗ , 2∗ } – an unavoidable result of inseparability.
Note the appearance of unintended destinations (underlined). We call these unintended
destinations fortuitous destinations and refer to the unintended paths as fortuitous
paths.
In Figure 3.21, intended paths are shown as solid lines and fortuitous paths are
indicated with dashed lines. These extra paths may potentially cause interference at
the receivers. However, as long as the DCA condition is observed, interference among
superimposed channels can be avoided by tuning in the desired channel and tuning out
the undesired one. For example, in the case of WDMA this would mean tuning the
optical receiver to select the correct wavelength and reject any others, implying that
the selectivity is implemented in the λ-channel layer. However, if TDMA is used to
distinguish the superimposed channels, the receiving station would select the intended
136 Multiwavelength Optical Networks

1 1*
C S1
a S2
A B F G

H D E
2 S1 3*

S2
3 2*
(a)

1 1*
S1 C S1
a
A B F G

S2 H D E
2 3*

S2
3 2*
(b)

Figure 3.21 Inseparability.

channel in its RP by capturing information in the proper time slots. In this case, selectivity
is implemented in the transmission channel layer.21
Inseparability causes connections that share a common fiber on a common waveband
to branch out of their original paths, fanning out onto an optical path in the form of a
directed tree, with new, fortuitous destinations added to the tree as new paths are activated
within the same waveband. The set of destinations (both intended and fortuitous) for each
source consists of all destinations downstream on the tree from that source. Fortuitous
destinations tend to waste fiber resources and power and are therefore to be avoided if
possible. In this case, the fortuitous destinations could have been avoided by rerouting
(2, 2∗ ) on a longer path, via node H , as shown in Figure 3.21(b). Inseparability does
not apply to connections in different wavebands, which are routed independently of one
another. Thus, fortuitous paths associated with inseparability do not exist in wavelength-
routed networks. In WRNs all optical paths are point-to-point so connections from
different source stations on the same wavelength are never combined on the same
fiber.
Now let us examine the distinct source combining constraint.22 The DSC condition
forbids a signal from splitting, taking multiple paths in the network, and then recom-
bining with itself. Figure 3.22 shows two ways a source may combine with itself. In
Figure 3.22(a), the signal power is split into two parallel paths and then later recombined
21
Fortuitous paths occurred in the broadcast star example of Figure 3.2 without referring to them as such.
For example, using WDM/WDMA, all signals reach all three destination stations, but each signal is only
intended for one of them.
22
Without calling it by that name, the DSC requirement was invoked in Section 2.3.1.1 in discussing power
relations in directional couplers.
 Network Connections 137

A B

(a)

A B

(b)

Figure 3.22 Two violations of DSC.

downstream, whereas in Figure 3.22(b), a portion of the signal is fed back onto the
upstream portion of its path. In both cases, the combined signals interfere with each
other, garbling their information.23 Even if routing decisions are made correctly to avoid
DSC violations, some low-level DSC violations may still occur because of imperfec-
tions in hardware. For example, DSC violations in Figure 3.22 could be created because
of small leakage paths through the optical switches in nodes A and B. These effects,
which are unavoidable in any purely optical network (either wavelength or waveband
routed), produce cross-talk, which is tolerable as long as it is kept sufficiently small (see
Section 4.9.1). From now on, when we speak of DSC violations, effects due to nonideal
hardware are ignored.
Figure 3.23 illustrates how a correct but poor routing decision may produce an inad-
vertent violation of the DSC condition. With connections (1, 1∗ ) and (2, 2∗ ) in progress,

1 1*
S1 C S1 + S2

A B F G
f
d S1 + S2 + S3
S2 H S1 + S2 D E
2 3*
S1 + S2 + S3
S3
3 S1 + S2 + S3 2*

Figure 3.23 Inadvertent violation of DSC.

23
Garbling is due to the fact that the signal is combined with a delayed replica of itself, creating interference
at the receiver. Long delays cause interference in the demodulated signal, which translates into noise and
distortion for analog signals and bit errors for digital signals. Short delays (on the order of the optical
signal’s coherence time) produce interference at the level of the optical fields, analogous to multipath
effects in radio transmission.
138 Multiwavelength Optical Networks

1 1*
S1 C S1

A B F G

S2 + S3
S2 H D E
2 3*

S3 S2 + S3
3 2*
(a)

1 S1 + S2 + S3 1*
S1 C
c
a
A B F G
f
b
h d
S2 H D E
2 S1 + S2 + S3 3*
S3
3 S1 + S2 2*
(b)

Figure 3.24 Avoidance of DSC violations.

as shown in Figure 3.21(a), a new connection (3, 3∗ ) is routed via path D–E–F–G. All
three connections are assumed to be in the same waveband but on distinct channels. The
signals as they appear with the new connection in place are shown in Figure 3.23. Due to
inseparability, signal S 3 carries with it (fortuitously) portions of signals S 1 and S 2 after
combining with them on fiber d. This causes S 1 (which split at node B) to recombine
with itself on fiber f , violating the DSC condition. (The same holds true for S 2 .) The
new connection therefore must not be routed along the path shown. It could, however,
use that path if it was assigned to a different waveband, because inseparability would
not apply in that case. Two other solutions to this problem are shown in Figure 3.24.
Figure 3.24(a) solves the problem by rerouting connection (2, 2∗ ) via node H , and in
Figure 3.24(b) the problem is solved by routing (3, 3∗ ) via the path D–B–C–F–G.
Inseparability can also indirectly cause violations of the DCA condition. We call these
color clashes. This may occur when activation of a new connection results in combining
onto the same fiber two or more connections (on the same waveband) that were already
in progress on previously disjoint paths. If these connections use the same channel (for
example, the same wavelength), a color clash occurs. To illustrate, consider the example
in Figure 3.25. In Figure 3.25(a) two connections (1, 1∗ ) and (2, 2∗ ) are shown occupying
disjoint paths. Wavelength λ 1 is assigned to both of them. Now a new connection (3, 3∗ )
is assigned to path D–E–F–G using the same waveband. Because S 3 shares link d with
S 2 , it is assigned a different wavelength, λ 2 . The effect of adding connection (3, 3∗ )
is shown in Figure 3.25(b). Note that, because of inseparability, all three signals are
now carried on fiber f . Because S 1 and S 2 are now on the same fiber, using the same
 Network Connections 139

1 1*
S1 (1, 1*)1 C S1

A B F G

d
H D E
S2 (2, 2*)1
2 3*

3 2*
(a)

1 1*
S1 C S1

A B F G
f
d S2 + S3
S2 H D E
S2
2 3*
(3, 3*)2
S3 2*
3
(b)

Figure 3.25 Color clash.

λ-channel, they produce a color clash. This violation could have been avoided if all three
connections had been assigned distinct wavelengths.
Among the four constraints discussed earlier, wavelength continuity and inseparability
are both consequences of physical laws and technological constraints; the others – DCA
and DSC – are routing requirements imposed to ensure satisfactory operation of the
network. When we discuss routing and channel assignment later, the role of the DCA
and DSC constraints is of paramount importance. The examples used to illustrate the
constraints bring out once again the difficulties associated with demand-assigned as
opposed to dedicated connections. If the state of the network happens to evolve in a way
that tends to “clog up” the fibers as connections are assigned in sequence (dynamically),
then it may be difficult or impossible to assign new connections without violating the
routing and channel assignment constraints. With rearrangeability, however, the original
routing and channel allocations are not irrevocable.
Thus in the example of Figure 3.21, changing the path of (2, 2∗ ) to pass through
node H , as in Figure 3.21(b), would have avoided violation of the DSC condition.
Similarly, changing the wavelength assignment of (1, 1∗ ) to λ 3 in Figure 3.25(b) would
have avoided a color clash. Because we have decided that rearrangeability (resulting
in service disruption) is not a reasonable option, it is especially important that the
algorithms used for dynamic routing and channel assignment in these networks have a
certain amount of foresight built into them. Even though future connection requests may
not be predictable, it is still possible to make optical path and channel allocations in a
way that tends to reduce the probability of blocking for new connections. For example,
the color clash example of Figure 3.25 suggests that λ-channel assignments within a
140 Multiwavelength Optical Networks

common waveband should be made in a way that minimizes the reuse of each wavelength;
an idea that is counterintuitive because wavelength reuse is desirable for conserving the
optical spectrum. We study these questions in more detail in Chapter 6.

3.4.2 Multipoint Subnets in LLNs

In the previous examples, attempts to set up several point-to-point optical connections
within a common waveband led to the unintentional creation of multipoint optical paths,
and consequent extra complications in routing and channel assignment. In view of these
additional complications in waveband-routed networks (LLNs), one might ask what is
to be gained by adding this extra complexity? One answer is that waveband routing
leads to more efficient use of the optical spectrum, given the technological constraints
of the ONNs (see Section 2.2). Another answer is that the multipoint optical path
capability, which showed up accidentally earlier, can be turned to our advantage in
creating intentional multipoint optical connections. Because of this property, LLNs can
deliver a high degree of logical connectivity with minimal optical hardware in the access
stations. This is done by creating controllable multicast and multiaccess optical paths that
support both one-to-many and many-to-one optical connections. This multipoint optical
connectivity property is one of the fundamental advantages of LLNs over point-to-point
wavelength-routed networks. We shall illustrate its power through a few examples in
Section 3.4.3, which compare the WRN and LLN approaches.
Figure 3.26(a) shows a network containing seven stations interconnected by two-fiber
bidirectional links on an LLN with a mesh physical topology. As will be seen in the
examples that follow, it is often useful to create full logical connectivity among spec-
ified “clusters” of stations within a larger network. We call such clusters multipoint
subnets (MPSs). Suppose, for example, that the set of stations {2, 3, 4} is to be in-
terconnected in this way. The LCG defining full connectivity in this subset (without
loopback connections) is shown in Figure 3.26(b), together with the resultant logical
topology (Figure 3.26[c]). Note that the unconnected stations are shown isolated from
each other, and the stations contained in the MPS are shown enclosed in a “cloud,”
indicating that there is full logical connectivity among all of them. Recall that this
type of representation was used earlier in Figure 3.2(c) to symbolize stations sharing
a common broadcast medium. We shall see that these two interpretations are closely
related.
How is full logical connectivity realized on the MPS? One possibility is to create an
optical path on a single waveband in the form of a tree joining the stations, where the path
is “carved out of ” (embedded in) the larger network by setting the LDCs appropriately.
Figure 3.27(a) shows a shortest path tree connecting the three stations. The actual signal
paths on the tree can be chosen in various ways. One convenient approach is to define a
“root node” as the basis for an embedded broadcast star; that is, we route all signals from
the stations to the root (now playing the role of a star coupler), combine them there, and
broadcast the combined signals back to all stations.
Figure 3.27(b) illustrates the equivalence between the tree, rooted arbitrarily at node
C, and a broadcast star. The fibers and nodes along paths taken by the signals to and from
 Network Connections 141

4
4
f 3
D C 3

5
d c

e E g

a b
7
1
1 A B 2
h 2
6
6
(a) Mesh

MPS
T R
2 3
2 2
4
1
3 3
5
6
4 4 7
(b) LCG (c) Logical Topology

Figure 3.26 Seven stations on a mesh.

the root are indicated in Figure 3.27(b), where fibers sharing the same bidirectional link
in the network of Figure 3.26(a) are now shown separately, and nodes traversed twice
are duplicated. (Signal flow is from left to right.) In this and subsequent discussions, we
adopt a standard convention for labeling fibers in two-fiber bidirectional links. Each link
is shown with a reference arrow and a label, where the label designates the fiber carrying
signals in the direction of the reference arrow, and a corresponding primed label denotes
the fiber carrying signals in the opposite direction. For example, in Figure 3.26(a),
fiber a carries signals from node A to E, and a ′ carries signals in the reverse direction.
If a δ-σ LDC is used in node C, then it would be set as shown in Figure 3.27(c). The
solid lines indicate the combining and dividing paths if the node is operated as a star
coupler without loopback paths (a nonreflecting star), and the dashed lines indicate added
paths for loopback operation (a reflecting star). In each case the combining/dividing
ratios are equal for all ports: one-half for the case without loopback and one-third with
loopback, so as to distribute power equally to all stations. (Although the specified LCG
does not include loopback connections in this case, they are useful for control purposes –
for example, synchronization in a TDMA system – and for determining propagation
delays to the root, knowledge of which may be required in a MAC protocol.)
142 Multiwavelength Optical Networks

f
D C 3

Optical Path

c
g

B 2

(a)

T R
2 f
2 B B 2
g g
3
3 C 3 c
f
3
4
4 D D 4

(b) (c) Node C

Figure 3.27 Tree embedded in mesh.

Once an optical path is established to emulate a broadcast star, that is exactly what is
“seen” by the stations. In other words, the desired MPS has been created by embedding
its LCG into the physical topology of a mesh network in the form of a broadcast star. The
embedded star can now be operated exactly like any other shared broadcast medium.
Any multiplexing and multiple-access channel-sharing methods applicable to broadcast
stars can be used, including fixed-capacity allocation, dynamic capacity allocation,
TDM/TDMA, and TDM/T-WDMA. For example, a single λ-channel might be allocated
to the MPS, in which case all stations would share the channel in a TDM/TDMA mode.
If several λ-channels are available within the waveband supporting the MPS, then a
TDM/T-WDMA implementation might be used. In each case, the full capacity of all
allocated λ-channels would be available for sharing among all LCs. It is for this reason
that we represent the MPS as a multipoint link that contains a set of stations sharing a
common capacity.
 Network Connections 143

After the optical path supporting the MPS has been created by fixing the settings of
the participating ONNs, there are no further control functions required within the optical
path layer. Thus, control of the MPS, once it is configured, is completely in the hands
of the NASs. The choice of multiplexing and multiple-access techniques, establishment
and release of logical connections, apportionment of shared capacity, and so forth, can
all be implemented through suitable signaling among the stations in the MPS without
any interaction with the ONNs.
The concept of the MPS is very general. From a top-down point of view, the stations
contained in an MPS (e.g., the stations in the “cloud” in Figure 3.26[c]) appear at the
logical level as a fully connected subset allocated some total capacity, which is shared
among all LCs. Logical connections within an MPS may be multicast as well as point-
to-point. As conditions evolve, the logical aspects of the MPS might be reconfigured by
adding or deleting stations or by modifying capacity allotments. The details of how the
connections are implemented in the physical layer are concealed and are unimportant
from a logical layer point of view.
However, from a bottom-up physical layer view, there are many possible ways of
implementing a given MPS, including the choice of waveband, channel, and optical
path; the use of single or multiple transceivers in the stations; single or multiple fiber
pairs in the access links; different multiplexing and multiple-access schemes; fixed or
dynamic capacity assignment (packet switching); and so forth. Furthermore, it is possible
to reconfigure these physical layer aspects of an MPS in response to changes in load
(e.g., add or subtract channel allocations) or in response to network faults (e.g., route the
supporting optical path around failed links or nodes). Last, many variants of the fully
connected MPS are possible. For example, if we remove the transmitting connections
from stations 3 and 4 in Figure 3.27(b), the result is a multicast optical path from station 2
to stations 3 and 4. Similarly, removing the receiving connections from stations 2 and 3
results in a multiaccess path from transmitting stations 2 and 3 to receiving station 4.
The MPS reappears in the next section as a basic building block of a hypernet.

3.4.3 A Seven-Station Example

Now let us test the LLN concept against the wavelength-routed network approach. The
test case will be the implementation of full logical connectivity among a set of seven
end systems. We assume throughout that each end system is connected to an NAS and
then to a network node through a nonblocking access link. The optical transceivers in
the stations are assumed to be tunable over the full spectrum needed to support the
connections, where each transmitter runs at a bit rate of R 0 bits per second, which
may be taken to be the full capacity of a λ-channel. Implementation on three physical
topologies – (1) the bidirectional ring, (2) mesh, and (3) multistar – is compared.
Implementation on a bidirectional ring is considered a point of reference. We then
compare approaches using the mesh of Figure 3.26 and the multistar configuration of
Figure 3.28(a). The topology of the mesh was chosen arbitrarily and has no relation to the
logical topology being implemented. In contrast to the mesh, the multistar structure was
custom designed to support the seven-station example. The different physical topologies
144 Multiwavelength Optical Networks

T R
1 1

2 2

3 3

4 4

5 5

6 6

7 7

(a)

1 1

1 1
2 2
2 2
3 3

6 6
(b)

Figure 3.28 Multistar network.

were chosen to illustrate their effect on station and optical spectrum resource utilization.
This is the first (but not the last) time we encounter a multistar topology. Until now, the
underlying fiber placements were assumed to be given; that is, the network implementer
was assumed to have no control over the physical topology, which is not necessarily
adapted to the needs of the end systems. When using a multistar arrangement, the
 Network Connections 145

assumption is exactly the opposite. The fiber topology is now put in place expressly to
serve the network being implemented; that is, the design of the physical topology is part
of the overall network design.
At the logical level, three logical topologies are considered: point-to-point (42 uni-
directional LCs), fully shared multipoint, and hypernet (partial sharing). The first is
realized using wavelength routing, and the others are realized using multicast optical
connections in an LLN.

3.4.3.1 Wavelength Routing

In the case of wavelength routing, the nodes in all physical topologies considered are
wavelength-selective permutation switches. Because individual point-to-point optical
connections are required for each LC in this case, six transceivers are needed in each
station. For the bidirectional ring, it can be shown24 that full interconnection of seven
stations requires six wavelengths, which results in a wavelength reuse factor of seven
for the 42 LCs. For the mesh, it is left as an exercise for the reader to show that the
connections can be realized with four wavelengths (a wavelength reuse factor of 10.5),
which is the minimum possible number.
In the multistar topology, each station accesses three stars, using two transceivers on
each star. The two transceivers accessing one star are multiplexed onto a single access
fiber pair. The central node of each star, a WSXC, is set to operate as a permutation
wavelength router of the type shown in Figure 2.10, where the routing rule is as follows:
A signal on input fiber i, carried on wavelength λk , is routed to output fiber j, where k =
j − i mod 3. For example, the star joining stations {1, 2, 6} is shown in Figure 3.28(b).
With the fibers numbered as in the figure, the wavelength assignments for the six optical
connections on the star are (1, 2) 1 , (1, 6) 2 , (2, 1) 2 , (2, 6) 1 , (6, 1) 1 , and (6, 2) 2 . Because
two wavelengths are used on each star, and these can be reused on all stars, the reuse
factor for the 42 connections is 21. For all three physical topologies, the network has a
total capacity of 42R 0 bps, and each LC has a fixed-capacity allotment of R 0 bits per
second.

3.4.3.2 LLN Realization: Fully Shared Logical Topology

In each wavelength-routed network realization, full connectivity was realized necessarily
using point-to-point LCs, with a fixed capacity assigned to each. However, in an LLN
realization, we have the option of replacing point-to-point logical connections with
MPSs; that is, we may group stations together into fully connected subsets, in which the
LCs joining the stations in the subsets share a common capacity.
Extending the idea of the MPS to its ultimate limit in this example we may include
all seven stations in a single subnet. Irrespective of the underlying physical topology,
this can always be accomplished by choosing any convenient tree that links all seven
stations to form an optical path that supports all optical connections among the stations.
Following Section 3.4.2, the tree can be used to emulate a broadcast star. In the case
of the bidirectional ring of Figure 3.29(a), any node can be chosen as the root node for

24
See Section 6.3.4.2.
146 Multiwavelength Optical Networks

4 5

D E 6
3

C F

Optical Path

B G

A
2 7

1
(a) Bidirectional Ring

4 3 2 2 3 4

D C B B C D

1 A 1

E F G G F E

5 6 7 7 6 5
(b) Directed Tree

Figure 3.29 Embedded star on a bidirectional ring.

the star, with its LDC set to do full combining and broadcasting for all λ-channels on a
specified waveband. All other nodes serve to complete the tree configuration, joining the
NASs to the root node. In this case, node A has been chosen as the root node so that the
optical paths can be redrawn in the form of a directed tree as shown in Figure 3.29(b).
Depending on the selectivity of the LDCs, the waveband allocated to the star may be
“thin,” holding only a single λ-channel, or “thick,” containing the full usable optical
spectrum. (The latter case would hold if the LDCs had no waveband selectivity.)
In either case, the combined capacity of all allocated λ-channels is available for
sharing among the seven stations. (A negative feature of the embedded star is that
there is no spectrum reuse, so the total throughput can never be more than the combined
capacity of all allocated channels.) If only one λ-channel is used, all stations can share the
 Network Connections 147

channel using TDM/TDMA, and this requires only a single transceiver per station. This
is the most economical realization in terms of station cost. In this case, the total network
capacity is R 0 . If more than one λ-channel is used, full utilization of the channels requires
either multiple transceivers or rapidly tunable transceivers in the stations. For example,
if seven λ-channels are available, then the TDM/T-WDMA approach of Section 3.2
could be employed in an FT-TR arrangement with one fixed transmitter and one tunable
receiver in each station. This would yield a total network capacity of 7R 0 bps.
The same approach can be applied to the mesh network of Figure 3.26. For example,
node E in the figure can be taken as the root of an embedded star. Except for the different
physical topology, all aspects of the discussion of the bidirectional ring apply here as
well. (In fact, the seven stations accessing the embedded star have no way of knowing
in what physical topology their MPS is embedded.)

3.4.3.3 LLN Realization: Hypernet Logical Topology

The point-to-point and fully shared logical topologies represent two extremes for real-
izing full connectivity among seven stations. Compromise solutions are also possible.
Figure 3.30(a) shows a seven-node hypernet logical topology in the form of a hyper-
graph25 with seven hyperedges. Each hyperedge represents an MPS that contains three
stations. Note that all stations in a common hyperedge can communicate with each other
because they are members of the same MPS. Because there is full logical connectivity
among all stations (vertices) within an MPS (hyperedge), and all LCs share a common
capacity, there are in effect six LCs “hidden” in each hyperedge, all sharing the same
capacity. In this example the logical topology is arranged so that each station can reach
any other station via one of the three hyperedges to which it is connected. This is a way
of providing full logical connectivity among all stations without placing them in the
same MPS.
Just as the point-to-point links comprising an LCG are represented conveniently
as a bipartite graph linking source and destination stations, the MPSs (multivertex
hyperedges) comprising a logical connection hypergraph (LCH) can be represented
by a tripartite graph, as shown in Figure 3.30(b). Here each hyperedge of the LCH is
represented as a central vertex connecting a transmitting set on the left to a receiving set
on the right. For example, hyperedge E 1 connects the transmitting set {2, 5, 7} to itself.
When each hyperedge’s transmitting and receiving sets are identical, the hypergraph is
called undirected, which is analogous to an undirected graph.
So far we have a top-down view of the network. Now, suppose this logical topology
is to be realized in the mesh network of Figure 3.26 as a set of embedded multicast
stars. A routing tree must be chosen for each of the seven MPSs, and wavebands must
be assigned to the trees so that the routing constraints are satisfied. A solution to this
problem is shown in Table 3.1. The tree for each hyperedge is given, together with its
waveband assignment. (The details of tree realization were discussed in Section 3.4.2.
The tree for hyperedge E 3 is displayed as an illustration in Figure 3.27.)

25
See Appendix A for the definition of a hypergraph.
148 Multiwavelength Optical Networks

E7
E6
2 5

4
E2
E1
E4 E3
E5
1 3
7
(a) Hypergraph Logical Topology

T E1 R
1 1

E2
2 2

E3
3 3

E4
4 4

E5
5 5

E6
6 6

E7
7 7

(b) LCH Tripartite Graph

Figure 3.30 Seven-node hypernet.

Note that the assignments are made so that no waveband is used more than once on
a link, ensuring that the DCA condition holds on all links. The routing and waveband
assignments were accomplished with a total of two wavebands, which is the minimum
number possible. As can be seen, it was possible to achieve a minimum waveband assign-
ment using shortest path routing of all trees except that supporting E 7 , which required
three links. As in the previous example, this realization can be implemented with various
numbers of λ-channels per waveband and various numbers of transceivers per station.
 Network Connections 149

Table 3.1 Tree routing on mesh.

Hyperedge Stations Tree Waveband

1 2, 5, 7 b 1
2 1, 2, 6 h 1
3 2, 3, 4 f, g 1
4 1, 4, 5 a, d 1
5 1, 3, 7 a, c 2
6 4, 6, 7 d, e 2
7 3, 5, 6 b, g, h 2

Suppose, for example, that we allocate three λ-channels per waveband with the channels
accessed using TDM/T-WDMA operating in FT-TR mode. This could be implemented
using a single transceiver in each station to access each MPS, requiring three transceivers
per station, one for each MPS to which it is connected. The total network capacity is
then 21R 0 , with a capacity of 3R 0 shared among six logical connections in each MPS.

3.4.3.4 Multistar Realization

In the previous example, we chose a logical topology (a seven-hyperedge hypergraph)
and embedded it on a given physical topology, in this case a mesh. There was no relation
between the chosen LT and the prescribed PT. However, as mentioned, one is sometimes
free to design the PT to suit the chosen LT. In the case of hypernets, a multistar realization
is a natural design approach, in which the tripartite graph representation of the LCH
can be viewed as defining a multistar physical topology. For example, in the LCH of
Figure 3.30(b), suppose each hyperedge corresponds to a broadcast star, in which the
central vertex represents the star coupler, directing signals from the transmitting set to
the receiving set. In this example, each station is connected to three stars, allowing it to
reach all other stations.
As seen from the stations, this multistar realization has the same characteristics as the
case of multiple broadcast stars embedded on the mesh physical topology of Figure 3.26.
The only difference is that all stars are realized on separate fibers, so we no longer need
two wavebands to support the MPSs. To conform to the previous example, a single
waveband containing three λ-channels can be reused on each star, with TDM/T-WDMA
operation on the star using FT-TR. This would require three transceivers per station and
would give a total network capacity of 21R 0 .
By this time the reader should have realized that the multistar topology used in
the earlier WRN example was derived from the hypernet development presented here.
(Compare the multistar WRN of Figure 3.28[a] with the hypernet LCH of Figure 3.30[b].)
The only difference is that the nodes in the WRN case perform wavelength permutation
instead of simple combining and broadcasting in the hypernet.

3.4.3.5 Comparisons
Table 3.2 summarizes the main features of the various seven station realizations pre-
sented earlier. It compares them based on spectrum utilization, reuse factor, transceivers
150 Multiwavelength Optical Networks

Table 3.2 Seven-station comparisons.

No. of
Type of Physical Logical Reuse transceivers Network
network topology topology factor per station Spectrum capacity

WRN Ring Pt.-Pt. 7 6 1200 42R 0

WRN Mesh Pt.-Pt. 10.5 6 800 42R 0
WRN Multistar Pt.-Pt. 21 6 400 42R 0
LLN Any Fully shared 1 1 200 5R 0
LLN Mesh Hypernet 3.5 3 400 21R 0
LLN Multistar Hypernet 7 3 200 21R 0

Pt.-Pt. = point-to-point.

per station, and total network capacity. Because there are so many variables in these
examples, we must be careful not to compare apples with oranges! For the purpose
of comparing spectrum usage, some assumptions must be made about wavelength and
waveband spacing. Figure 3.31 shows the assumed values. For the wavelength-routed
case, Figure 3.31(a) shows a 200-GHz spacing between centers of wavebands, with a
guard band of 75 GHz between wavebands and a single λ-channel in each waveband.
Figure 3.31(b) shows the case of the LLN, in which the 125-GHz usable bandwidth of a
waveband supports up to five λ-channels on 25-GHz spacings. This λ-channel spacing
could easily accommodate optical signals running at OC-48 speed (2.5 Gbps) on each
channel. Thus, for example, the hypernet realization on a mesh requires two wavebands
and uses three λ-channels in each waveband. Assuming that a full waveband is dedicated
to the hypernet regardless of whether all of its channels are used, we assign a spectrum
allocation of 400 GHz to this case.

200 GHz

λ-channel

f (GHz)
125 GHz 75 GHz 125 GHz

(a)

f
25 GHz

(b)

Figure 3.31 Assumed channel spacings.

 Network Connections 151

How does the spectrum utilization compare? In going from the ring to the mesh
to the multistar, spectrum utilization tends to decrease and the reuse factor increases,
as is to be expected with the improved physical topologies. The LLN realizations are
more economical of spectrum and transceiver usage than the WRN case, with the fully
shared LLN realization requiring the least resources. (Reuse factors in WRNs and LLNs
are difficult to compare because WRNs use point-to-point optical connections, whereas
LLNs use multicast connections.) Because network capacity is directly proportional to
the number of transceivers, the WRN realizations have the most capacity. Of course, there
are other factors to take into account, such as rapid tuning requirements in multichannel
LLN realizations and lack of channel sharing in WRNs.
Channel sharing has some significant advantages that are not immediately apparent.
When a pooled capacity is shared among many logical connections, that capacity can
be allocated to match unbalanced traffic distributions and can be reapportioned as
traffic demand changes. Furthermore, in a shared medium it is possible to provide
dynamic capacity allocation by implementing packet switching in the optical layer (see
Section 3.2.2). For example, in the case of a fully shared logical topology, all 42 LCs
could be operated in a packet-switched mode so that each connection receives the
capacity it needs just when it needs it. In the WRN case, the fixed capacity allocations
make it impossible to adapt to unbalanced traffic. Consequently, traffic imbalances result
in reduced network throughput. Finally, as mentioned at the beginning of this chapter,
shared-medium multipoint optical connections act as a natural support for multicast
logical connections.

3.5 Logically-Routed Networks

The examples in the previous sections illustrate various alternatives for implementing
high logical connectivity in transparent optical networks while keeping resource utiliza-
tion under control, limiting station complexity to a few optical transceivers, and using
a few hundred GHz of optical bandwidth. This was possible because the networks were
small, containing at most seven stations. When we move to larger networks, the transpar-
ent optical approach soon reaches its limit. For example, we found that six wavelengths
and six optical transceivers per station were necessary and sufficient to achieve full log-
ical connectivity among seven stations on a bidirectional ring using wavelength-routed
point-to-point optical connections. But suppose we increase the number of stations to
22. Then the number of wavelengths increases more than ten-fold to 61, requiring many
terahertz of optical bandwidth at the assumed wavelength spacings, and the number of
transceivers per station increases to 21 (see Section 6.3.4.2). Economically and techno-
logically, this begins to push the limits of the current state of the art. When the limits
of optics are exceeded we must turn to electronics; i.e., logically-routed networks. In
Section 2.6 the concept of a logical network overlay was presented, wherein logical
switching nodes interconnect a set of logical links, forming an electronically switched
logical topology (i.e., an LRN). The example of Figure 3.2(b), which consists of a uni-
directional ring logical topology embedded in a star physical topology, is a very simple
152 Multiwavelength Optical Networks

illustration of how switching in the logical layer can remove some of the connectivity
burden from the purely optical portion of the network. We shall now extend that approach
to larger networks. To fix ideas, in the discussion that follows it is useful to focus on a
specific type of logical network overlay, such as an IP network. In this case the LSNs are
IP routers, and the links are carrying asynchronous (packet) traffic.
In this case the associations between end systems are virtual connections, and the
logical layer is supported directly by the multiwavelength optical layer. This is IP over
WDM. However, the development applies equally well to synchronous traffic, in which
the LSNs might be SONET digital cross-connect systems or similar devices, and the end-
to-end connections would consist of synchronous bit streams in the form of dedicated
circuits running at DS1 or DS3 bit rates (1.544 or 44.736 Mbps, respectively) or higher.
These would typically be used as supports for higher layers of virtual connections, which
might be made through telephone Central Offices (for POTS) or IP routers (for Internet
data). Often several of these layers are stacked. An example is IP over ATM over SONET
over WDM. This constitutes a division of the virtual connection layer in Figure 2.1(a)
into several virtual sublayers.
The essential feature of the LSN is that it performs a sorting and routing function for
traffic delivered to it from an NAS accessing an underlying optical network. The traffic
arriving on the input ports interfaced to the NAS (usually in the form of a continuous
bit stream) is demultiplexed into small units (for example, IP packets, ATM cells, or
DS3 circuits) routed to the appropriate output ports and remultiplexed for delivery to
an NAS or an associated end system. This is illustrated in Figure 3.32, which shows an
LSN interposed between an NAS (accessing a node in an optical network) and two end
systems.
The difference between logical connections in a purely optical network and an LRN
is illustrated nicely by Figure 3.2. In a purely optical network, the end systems connect
directly to the external ports of the NASs, as exemplified by Figure 3.2(a), so transport

End
systems

LSN

NAS

ONN

Figure 3.32 A logical switching node in an optical network.

 Network Connections 153

between a pair of end systems is supported by the logical connections originating and
terminating at the corresponding NAS ports. Assuming that all logical connections are
point-to-point, this means that the transport network seen by the end systems can be
described by a logical connection graph as shown in Figure 3.2(a), in which the vertices
represent the NASs terminating those connections. In the case of an LRN, however,
the LSNs create an extra layer of connectivity between the end systems and the NASs.
Thus the end systems access the logical network through the LSNs, the LSNs access the
transparent optical network through the NASs, and the NASs are interconnected through
optical network nodes. The vertices of the connection graph describing an LRN – for
example, that of Figure 3.2(b) – are the LSNs instead of the NASs. If the underlying
optical infrastructure is a wavelength-routed network, each link in the logical topology
is supported by its own point-to-point optical connection. But each optical connection
requires one optical transceiver, so the number of logical connections incident on each
LSN (a vertex of the logical connection graph), called its degree, is limited by the
number of optical transceivers in the NASs serving the corresponding LSN. Because
this number is generally small for reasons of cost, the connectivity among the LSNs is
sparse.
Because of this sparse logical connectivity, most paths between pairs of end systems
require multiple logical hops, traversing several LSNs. At each LSN the traffic is sorted
and routed on the next logical link toward its destination. A key performance measure
for any LRN is the average number of logical hops experienced by its traffic. Because
each unit of traffic uses a network link on each hop, the same unit of traffic reuses the
network a number of times equal to its hop count. Therefore, the average link load is
proportional to the injected traffic multiplied by the average hop count. Because links
have limited capacities, the maximum possible injected traffic, or network capacity,
is inversely proportional to the average hop count. In addition to improving network
capacity, lower hop counts improve many other measures of network performance. In the
case of packet/cell traffic, fewer hops means less processing in the switches, less delay,
less probability of packet/cell loss, and better survivability, because each connection is
less likely to encounter a link or node fault. Similar remarks apply to circuit-switched
traffic except that delay and cell loss are replaced by call-blocking probability.
In this section we explore some of the properties of LRNs embedded on transparent
optical infrastructures. Point-to-point logical topologies are considered first, followed by
hypernets. The issue of the design of the logical topology is not considered here. However,
this is an important factor in achieving high performance. The logical topology must
be “matched” to the expected traffic distribution to keep the average hop count low,
and at the same time it must be matched to the underlying optical infrastructure so that
cost-effective embedding is possible. The design problem for LRNs is discussed in more
detail in Chapter 7.

3.5.1 Point-to-Point Logical Topologies

As mentioned earlier, one of the objectives of using logical switching on top of a
transparent optical network is to reduce the cost of the station equipment (particularly
154 Multiwavelength Optical Networks

1 5 1

2 6 2

3 7 3

4 8 4

(a) Logical Topology

T R
1 1
2 2
3 3
4 4
5 5
6 6
7 7

8 8
(b) LCG

1 2 3 4 5 6 7 8

(c) Folded Bus

Figure 3.33 Eight-node ShuffleNet.

by reducing the complexity of the optics) while maintaining high network performance.
Thus, we are interested in logical topologies that achieve a small average number of
logical hops at a low cost (meaning small node degree and simple optical components).
An interesting class of regular logical topologies called multihop networks was proposed
with this objective in mind. An example is ShuffleNet. The idea of the ShuffleNet design
was to use simple station equipment (a small number of transceivers requiring no rapid
tuning) and yet achieve good performance in networks scalable to large sizes.
Figure 3.33(a) shows an eight-node ShuffleNet logical topology assumed to be oper-
ated in a packet-switched mode. It consists of two stages with four LSNs each, joined by
unidirectional links. There are 16 links, compared with a requirement of 56 for full logi-
cal connectivity. To clarify the flow of traffic, the first stage is repeated, and connections
from the second stage back to the first are indicated by dashed lines, showing that the
 Network Connections 155

logical topology forms a cylinder. As in the end system interconnection example at the
beginning of this chapter, each LSN performs two functions: It exchanges traffic with
externally connected end systems, and it sorts and routes traffic in transit through the
network to the destination end systems. A connection between any pair of end systems
can be established over a suitable logical path, and it is easily seen that any node can
communicate with any other node using a number of logical hops not exceeding three.
It turns out that with uniform traffic and shortest path routing, the average number of
hops in this network is two. Figure 3.33(b) presents the LCG in our standard bipartite
form. These networks are scalable to large sizes by adding stages and/or by increasing
the degree of the nodes (see Section 7.2.1).
Up to this point nothing has been said about the underlying optical network. As
originally proposed, a suggested implementation of ShuffleNet was on a folded bus, as
shown in Figure 3.33(c). This PT is equivalent to a broadcast star in the sense that it allows
no spectrum reuse, and therefore if a WDM/WDMA system is used, 16 wavelengths
are required for the 16 point-to-point LCs. To improve spectrum utilization, let us now
assume that the ShuffleNet is to be embedded in a wavelength-routed bidirectional ring
using single fiber pair access links and two transceivers in each station. Figure 3.34 shows
the locations of nodes on the ring. (The arrows on each link simply define a reference

d 4 e

3 8

c f

6 7

b g

5 2

a 1 h

Figure 3.34 ShuffleNet embedding.

156 Multiwavelength Optical Networks

Table 3.3 Wavelength assignments for ShuffleNet on a ring.

Source node
Destination
node 1 2 3 4 5 6 7 8

1 — — — — 1 2 — —
2 — — — — — — 1 2
3 — — — — 2 1 — —
4 — — — — — — 2 1
5 1 2 — — — — — —
6 — — 1 2 — — — —
7 2 1 — — — — — —
8 — — 2 1 — — — —

direction for the fibers according to the labeling convention introduced in Section 3.4.2.
Each link consists of a pair of bidirectional fibers.) Note that the node/station placements
are not arranged in numerical order. They were, in fact, chosen to minimize the required
number of wavelengths. A possible wavelength assignment is shown in Table 3.3, which
indicates that the 16 LCs can be realized using only two wavelengths, for a reuse factor
of 8 (routing is always on the shortest path on the ring).
Figure 3.35 shows the details of node 1 in this realization. As shown in Figure 3.35(a),
node 1 receives traffic from nodes 5 and 7 (in addition to injected traffic) and forwards
nonexiting traffic to nodes 5 and 6. Packets from node 5 (7) are received on wavelength
λ 1 (λ 2 ) and delivered to the LSN, where they are buffered and sorted as indicated in
Figure 3.35(b). Those destined for the local end systems exit the network at this point,
and the others are routed to the two transmitters. Packets whose next node is 5 (6) are
delivered to O T 1 (O T 2 ). The total capacity of this network is 8R 0 bps, where R 0 is the
transmitter bit rate. This represents one half of the combined capacity of all transmitters,
with the factor of 12 appearing because the average hop count is 2. For comparison, if
full logical connectivity (56 LCs) was to be realized purely optically on the ring, this
would require eight wavelengths with seven transceivers in each station and would give
a network capacity of 56R 0 .

3.5.2 Multipoint Logical Topologies: Hypernets

The seven-station example in Section 3.4.3 illustrated how high connectivity may be
maintained in transparent optical networks while economizing on optical resource uti-
lization through the use of multipoint connections and hypernet logical topologies. But
that example was a special case of a hypernet wherein each station can reach any
other station in one logical hop, so that no logical switching was required. These ideas
are even more potent in larger networks, when multipoint connections are combined
with logical switching. We now extend them to this more general setting while at the
same time outlining a general approach for embedding logical topologies on prescribed
 Network Connections 157

5 1 1 5

1
2 2

7 6
(a)

1 2 1 2

OT1 OT2 OR1 OR2

TP TP RP RP

(b)

Figure 3.35 Details of ShuffleNet node.

physical topologies. Our illustrative example in Figure 3.36(a) shows a logical connection
hypergraph that consists of 22 vertices (logical network nodes) and 10 undirected hyper-
edges, with each hyperedge of size four (i.e., containing four vertices).
To provide more generality to this example, two types of nodes are included: LSNs
(shown as hexagons) and logical terminal nodes (LTNs) (shown as triangles). The
terminal nodes have no logical switching function and serve only as access points to
interface end systems to the LRN. For example, in an ATM network the terminal nodes
might execute ATM end node functions (such as call setup and admission control),
and the LSNs would execute end node functions as well as call-routing functions (VP
switching). The degree of a vertex in an undirected hypergraph is defined as the number
of hyperedges to which it belongs.
In this example all LSN vertices are of degree two, whereas the terminal vertices are
of degree one. Vertices within the same hyperedge can communicate with each other
over a single logical hop, whereas those in different hyperedges require several hops in
tandem. For example, a packet moves between nodes U and C in one hop but requires
three hops between nodes U and K . The sorting and routing functions in the LSNs
158 Multiwavelength Optical Networks

permit full connectivity among all pairs of nodes in this network, albeit over multiple
logical hops in most cases. Using shortest path routing, the longest logical path between
a pair of nodes is four hops (e.g., from U to S via G–J–M). This is called the diameter
of the hypergraph. (The seven-node LCH in Figure 3.26 is an example of a hypergraph
in which every vertex has degree three, every hyperedge is of size three, and any vertex
can reach any other in a single logical hop. Thus the diameter is one, implying that no
logical switching is required.)
We assume that the hypergraph has been chosen to match the traffic requirements
of end systems attached to the nodes. However, it bears no relation to the physical
topology that will support it. Figure 3.36(b) shows the location of the logical nodes
on a prescribed physical topology to be called Atlantis. It is made up of 29 links and
18 ONNs. The ONNs contain LDCs; thus the optical infrastructure is an LLN. Each LSN
or LTN accesses the optical network through its own NAS, connected to an ONN by a
single fiber pair access link. We wish to embed our hypergraph into Atlantis, making
economical use of optical resources.
The general approach is shown in Figure 3.37(a), which shows several intersecting
hyperedges of an arbitrary hypernet topology. They are to be embedded in the physical
topology shown below them. As in our previous examples, each hyperedge represents a
set of logical network nodes that are to be fully connected via an MPS realized in the
purely optical network. For example, hyperedge E 1 in Figure 3.37(a) contains four LSNs
and a terminal node, all of which must be embedded as an MPS. The dashed vertical
lines indicate access links connecting the logical nodes to ONNs, where, for example,
the NAS supporting logical node A is connected via link a to optical node a.
Once more using the embedded broadcast star as a vehicle for realizing the MPS,
we select a tree joining the logical nodes in the MPS (bold lines in Figure 3.37[a]),
designating one ONN, say b, as its root. Because this is an undirected embedded star,
the LDCs at the various nodes are set to route all optical signals from the transmitters
in the five NASs to node b and then multicast the combined signals back to the five
NASs. In this way, all 20 logical connections within the MPS can be arranged to share
the common multicast medium. (As usual, this can be done with or without loopback
connections.) The resultant routing tree is shown in detail in Figure 3.37(b), indicating
the fibers used for each connection.
The general embedding procedure now consists of two parts: (1) find a set of trees for
the MPSs and (2) assign wavebands to the trees to satisfy the DCA constraint.
The tree selection might be made to minimize tree size (shortest path routing), to
avoid links loaded heavily with traffic from other users of the same infrastructure, or to
satisfy any other criteria. If minimization of optical spectrum utilization (i.e., assigned
wavebands) is the priority, the trees should be routed to minimize their intersections.
This may lead, at times, to using some circuitous routes. Because the degree of each
LSN is two, there will always be a pair of trees that intersects on each access link. With
single fiber pair access links, this means that a lower bound on the number of required
wavebands is two.
It turns out that it is possible in this case to find trees and waveband assignments
to support the 10 hyperedges using a total of three distinct wavebands, for a reuse
 Network Connections 159

N B

E3 P E1

O E4 E2 C

L Q D
E

E9 E7
K F
T U
E10 E8
I G
J E5 H

E6
M R
S V
(a )

L
l
J K 11 12 A
j S M
10 k a
27 13
r 28 1
9 14 15
m 29 B
R
I q Q 16 17 b
i V
8 23
T
24 p
h P 18
H
25
N 2
7
22 n 26
G g O
21
o
U e 19 c
6
5 C
E 20
f 4 3
d
F
D
(b)

Figure 3.36 Twenty-two node hypernet.

160 Multiwavelength Optical Networks

E4

B
A
C E F
D

E2
E1 E3
b
a e
d
c

a 2 b
1
3
c 4 e f

(a)

LSN including NAS

LTN including NAS

c
C c c C
1
a
A a a A
2
b
B b B
3 d

D d d D
4
e
E e e E

(b)

Figure 3.37 Hypernet embedding.

 Network Connections 161

factor of 3.33. For example, a possible choice of trees for E 1 , E 2 , E 5 , and E 6 is shown
as bold links in Figure 3.36(b). Note that they are nonintersecting on all links, including
their access links, and hence the same waveband can be reused on all of them. It is left
as an exercise for the reader to find routing trees and waveband assignments for the
remaining MPSs.
As in our other examples, either thin or thick wavebands can be used to implement
the 12 required logical connections within each MPS. For example, if a single λ-channel
is used in each waveband, the LCs can be realized using TDM/TDMA, with a single
transceiver needed for an NAS to access one MPS. Because each LSN accesses two
hyperedges, its supporting NAS needs two transceivers. The terminal node NASs only
require one transceiver. However, if higher network capacity is desired, a TDM/T-WDMA
system can be used over thick wavebands. Using 4 λ-channels in each waveband, for a
total of 12 channels, and operating in FT-TR mode, each MPS now supports four times
as much traffic as in the TDM/TDMA case, without increasing the number of optical
transceivers. However, rapid receiver tuning is now required.
Again using the wavelength-routed bidirectional ring as a baseline for comparison,
we recall from the beginning of this section that 61 wavelengths are required for purely
optical full connectivity among 22 stations, and 21 transceivers are required in each
NAS. The improvement in spectrum usage in Atlantis compared with the ring is only
partly attributable to a richer physical topology in this case. (Using a wavelength-routed
network with single fiber pair access links, the least possible number of wavelengths
and transceivers for full connectivity with any physical topology is 21.) The key to
optical resource conservation in this case is the hypernet logical topology. The number
of transceivers is kept small by the combination of multipoint optical paths and logical
switching, which means that each LSN’s transmissions “fan out” to six other logical
nodes using only two transmitters, with the same “fan in” advantage for the receivers.
Spectrum usage is kept small by using a single routing tree to support 12 LCs in each
MPS.
The hypernet model is not limited to LRNs realized on an optical infrastructure. Any
large network characterized by highly connected clusters of nodes, with intercluster
communication realized through specially equipped relay points, has the basic structure
of a hypernet. For example, a cellular radio network is conveniently modeled as a
hypergraph in which each hyperedge represents a cell, the LSNs correspond to base
stations, and the terminal nodes correspond to mobile users. In another context, an
internet can be modeled as a hypernet. In this case, the hyperedges represent constituent
networks (subnets of the internet), the LSNs represent gateways between the subnets, and
the terminal nodes represent hosts accessing the subnets. What these networks have in
common is that each is made up of constituent parts wherein the capacity of each part (a
wireless cell or a subnet) is shared among its included nodes.26 End-to-end connectivity
is maintained in any of these hypernets by finding a “good” logical path through the
various relay points.

26
In the case of an internet, complete sharing within a subnet may not always be a good model of what is
actually happening.
162 Multiwavelength Optical Networks

3.6 Summary

This chapter has given a largely descriptive picture of the four basic classes of multiwave-
length networks in order of increasing complexity: uncontrolled or static, wavelength
routed, waveband routed, and logically routed. In each case the underlying physical layer
was a transparent optical network. The objective was comparative analysis. In drawing
comparisons, realistic, specific parameter values were used as much as possible so that
quantitative comparisons could be made.
We found that the larger the network, the more complex the supporting architecture
must be. The hardest part of scaling to large sizes is connectivity, because the number of
possible end-to-end connections in a network grows as the square of the number of end
systems. Connectivity in uncontrolled networks is limited by the inability to reuse the
optical spectrum. Connectivity in wavelength-routed networks is limited by the fact that
an optical transceiver and a λ-channel are required for each logical connection. This limit
is circumvented in waveband-routed networks (LLNs) by using multipoint optical paths
sharing the capacity of a common waveband. Multipoint paths are in turn limited by the
fact that their combined throughput cannot exceed the capacity limit of the waveband
they share. LRNs break the connectivity bottleneck but do it by implementing high
connectivity electronically, incurring concomitant high processing loads in electronic
switching equipment.
The easiest aspect of scaling networks to large sizes is the throughput or network
capacity requirement because optical fibers have so much capacity to start with and
because each time we add an NAS to the network we add transceiver capacity. In
moving from simpler to more complex architectures, we find that the efficiency of
optical resource utilization improves, with more optical spectrum reuse and less optical
transceivers required per station.
The illustrative examples in this chapter, even the 22-node case, were chosen intention-
ally as “toy” problems – small enough and simple enough to be solved by inspection,
by trial and error, or by a small amount of hand calculation. They are to be consid-
ered as points of departure for a deeper exploration of the various issues introduced
here: topological considerations at the physical and logical levels; routing, channel,
and waveband assignment; methodologies for network design; and the relations among
performance, control, and resource utilization. Each section points to a later chapter:
The static networks discussed in Section 3.2 are studied in more depth in Chapter 5,
wavelength-/waveband-routed networks discussed in Sections 3.3 and 3.4 are described
again in Chapter 6, and LRNs (discussed in Section 3.5) are described in detail in
Chapter 7.
The most important concept to retain from this chapter is the view of the overall net-
work as many superimposed layers of connections. These connections can be configured
to be one-to-one, one-to-many, many-to-one, and multicast. The end systems typically
see a set of virtual connections that provide end-to-end transport on demand. The vir-
tual connections are in turn supported by a network of logical connections matched
to the needs of a community of end users. The links in the logical layer are realized
 Network Connections 163

as transparent optical connections, which are in turn supported by optical paths laid
out on a fixed fiber topology. Except for the fibers, everything else in the network is
reconfigurable, with each layer (to a large extent) independent of the others.
Virtual connections can be routed over many alternate logical paths, and logical topolo-
gies together with the optical connections/paths that support them can be reconfigured
independently. A large, purely optical network may serve as a common infrastructure
for several independently managed logical networks, each one tailored to the needs of a
different user community. Conversely, a logical network covering a geographic expanse
extending beyond the reach of purely optical connections may be supported by several
concatenated, independent optical networks, in which the end-to-end logical paths are
formed as a sequence of separate parts joined together at logical gateways between net-
works. This extraordinary degree of flexibility produces opportunities and challenges
that are explored in the rest of the book.

3.7 Problems

1 Show that if full virtual connectivity is required among the five stations in
Figure 3.2(b), and if shortest path routing is used, there will be 10 virtual connections
multiplexed on each LC. Generalize this result to the case of n stations.
2 The network of Figure 3.20 is operated as a WRN with each internodal link equipped
with a single fiber pair.
(a) Explain why at least two wavelengths are required to support full optical connectivity
among the five stations.
(b) Assuming single fiber pair access links, find a possible RCA for the 20 optical
connections using a minimum number of wavelengths.
(c) Repeat part (b) using nonblocking access links and only two wavelengths.
3 The network of Figure 3.26 is operated as a WRN with each internodal link equipped
with a single fiber pair and each station having a nonblocking access link.
(a) Explain why at least four wavelengths are required to support full optical connectivity
among the seven stations.
(b) Find a possible RCA for the optical connections required for full connectivity using
a minimum number of wavelengths.
4 Modify the seven-station example of Section 3.4.3 and Table 3.2 as follows. Let
guard bands between wavebands be increased to 100 GHz, with the usable bandwidth
maintained at 125 GHz. Assume that in the case of the WRN the optical signals can
run at twice the bit rates of the LLN (e.g., 5 Gbps instead of 2.5 Gbps) because of their
increased wavelength spacing. Recalculate the spectrum utilization and total network
capacity in this case for each row in the table.
5 Complete the routing and waveband assignment for the hypernet of Figure 3.36.
(Check your answer against Table 7.6.)
164 Multiwavelength Optical Networks

Bibliography

[Bernstein+04] G. Bernstein, B. Rajagopalan, and D. Saha. Optical Network Control: Architec-

ture, Protocols, and Standards. Boston, MA: Addison-Wesley 2004.
[Maeda98] M. W. Maeda. Management and control of transparent optical networks. IEEE J.
Select. Areas Commun., 16(7):1008–1023, 1998.
[Schwartz87] M. Schwartz. Telecommunication Networks: Protocols, Modeling, and Analysis.
Reading, MA: Addison-Wesley 1987.
[Wei+98] J. Y. Wei, C.-C. Shen, B. J. Wilson, M. J. Post, and Y. Tsai. Connection management for
multiwavelength optical networking. IEEE J. Select. Areas Commun., 16(7):1097–1108, 1998.
4 Enabling Technology
Coauthored by Neophytos Antoniades
The City University of New York/College of Staten Island

Throughout this book the approaches taken to system design and performance evalu-
ation are based on the constraints of the enabling technology. Available fiber capacity
is assumed to be limited by the constraints and imperfections of optical transceivers,
amplifiers, and cross-connects. These constraints affect maximum available spectrum,
wavelength spacing, and maximum bit rates per channel. Optical connections are as-
sumed to have limited reach, both geographically and in terms of the number of optical
cross-connects they may traverse. Sizes of switches as well as their speed, complexity,
and functionality are also assumed to be limited by cost and performance constraints,
ultimately going back to the limits of the underlying technologies. Trade-offs between
optical and electronic methods of implementing connectivity and routing are suggested,
in which the optimal design point depends again on relative cost and performance of the
enabling technologies.
Although emphasizing that these technological constraints are paramount, we pur-
posely keep as much of a separation as possible between the architectures discussed in
the book and the limitations of any specific technology. The reason is obvious: Today’s
technology is likely to be obsolete tomorrow. After more than a decade of gestation in
the laboratory, photonic and optoelectronic technology has matured to the point where
a wide range of technological choices are available for implementing each function in a
network, so that cost-effectiveness and viability in the field are the primary issues now
rather than proof of concept, which was the issue in the network testbeds just a few
years ago. After a period of explosive growth at the turn of the millennium followed by
consolidation as the economy contracted and then stabilized, a broad industrial base now
exists worldwide to respond to the need for components for the optical communications
infrastructure.
Keeping this background in mind, we present in this chapter the essential features of
representative devices and systems used in the implementation of multiwavelength opti-
cal networks. The objective is to convey an understanding of the principles that underlie
the functioning of the basic components of these networks, some notion of the perfor-
mance of typical current devices, and a methodology for evaluating the performance of
devices in network applications.
It is impossible in a few pages to give more than a glimpse of the state of the art of
current photonic and optoelectronic technology as it applies to optical networking. For
166 Multiwavelength Optical Networks

a more complete treatment the reader is referred to books completely devoted to the
subject (e.g., [Kaminow+02a, Kaminow+02b]).

4.1 Evolution of Transmission and Switching Technology

Advances in optical networking thus far have been largely in the area of point-to-point
optical transmission. Transmission system performance is typically measured in terms
of bit-rate-distance products, where distance is measured from the source to the point at
which a signal must be regenerated (i.e., “cleaned up”).1
The evolution of telecommunications from the early days of the telephone to cur-
rent optical fiber systems shows about three orders of magnitude of increase in bit-
rate-distance product each time a new technology is introduced – from megabits
per second-kilometers for microwave relays to gigabits per second-kilometers for un-
amplified fibers to terabits per second-kilometers for single-wavelength amplified links.
Interest in single-wavelength optical transmission was soon replaced by a focus on
multiwavelength transmission in the mid 1990s. Progress in the field can be judged
anecdotally by comparing “hero” experiments reported in the first edition of this book
to current record performance. A 360-Tbps-km experiment was performed in 1997 at
KDD using 16 2.5-Gbps channels over 9000 km with wavelength separations of 0.5 nm
[Otani+97]. Results presented the following year by the same group increased the bit-
rate-distance product to approximately 530 Tbps-km using 60 channels at 5.3 Gbps
over 1650 km [Murashige+98]. Recently, a bit-rate-distance product of approximately
36,000 Tbps-km has been obtained using 149 channels running at 40 Gbps with 50-
GHz spacing over 6120 km [Charlet+05]. The resultant optical spectral efficiency of
0.8 bps/Hz is more than an order of magnitude higher than the previous results. For some
time there has been nothing left to prove as far as transmission distance is concerned.
Soliton demonstrations both in the laboratory and in the field suggest that the reach
of the amplified, but unregenerated, optical fiber transmission link is for all practical
purposes unlimited (see Section 4.3.3).
Advances in optical switching technology have followed the scenario traced by earlier
electronic networks. Developments in digital transmission (e.g., the T1 carrier system)
made it cost-effective to move from analog to digital (i.e., time division) switching
and therefore drove the development of digital switching technology. In a similar way,
developments in multiwavelength transmission are a driving force in the development of
multiwavelength switching technology. Because current activity is centered on network
deployment in the metropolitan and regional area sectors, switching systems are being
developed most actively for those applications. However, large switches for long-haul
networks have also been successfully deployed.

1
Digital signal regeneration involves reshaping and retiming the pulses, operations that are almost always
done electronically. However, all-optical approaches are now being demonstrated and have the potential to
replace electronic techniques, especially at high bit rates.
 Enabling Technology 167

This chapter is organized as follows. Sections 4.2 to 4.9 focus on the point-to-point
transmission link. We begin with a brief overview of the components of the commu-
nication path, followed by a discussion of optical fiber transmission principles and
transmission impairments. Amplifiers, transmitters, and receivers are then considered in
conjunction with related technologies: modulation formats, equalization, and forward
error correction. A discussion of performance impairments in a network environment
completes the transmission link picture. Sections 4.10 to 4.12 deal with the devices
and subsystems necessary to combine transmission links into networks: coupling and
switching devices, filters, multiplexers, wavelength converters, signal regenerators, and
optical cross-connects. We conclude in Section 4.13 with discussion of physical layer
performance evaluation methodology, applying it to some case studies of current interest.
The objective is to show how simulation can be used to obtain insight and quantitative
results based on large complex network models.

4.2 Overview of the Optical Connection

A point-to-point optical connection (OC) in a multiwavelength network consists of an

optical transmitter (OT) and receiver (OR) joined by an optical path consisting of fiber
links traversing one or more optical network (switching) nodes (ONNs), as shown in
Figure 4.1. In the example shown, a laser is modulated externally by an information
source and is followed by a power amplifier that raises the power level launched on the
fiber. Additional line amplifiers between fiber sections boost the power to compensate
for attenuation along the line, and a preamplifier raises the power level at the input of the
receiver to improve the signal-to-noise ratio (SNR) at the input to the electrical detector.
A tunable optical bandpass filter is also shown, making this a tunable receiver. In digital
systems, performance is measured in terms of bit error rates (BERs), whereas in analog
systems the criterion is electrical SNR at the output of the receiver.
The performance is affected by imperfections in the optical transmitter (for exam-
ple, laser drift, linewidth, and modulator performance), transmission impairments in

Information
Source
Line PD
Preamplifier
Amplifier
Optical Electrical
Modulator Detector
ONN
Power Bandpass
Laser
Amplifier Filter

OT OR

Figure 4.1 A point-to-point optical connection.

168 Multiwavelength Optical Networks

the fibers (attenuation, dispersion, and nonlinear effects such as four-wave mixing),
amplifier characteristics (nonflat gain profile, amplified spontaneous emission noise),
switch imperfections (cross-talk), and receiver imperfections (nonideal filtering and
tuning, noise figure). Our discussion is limited for the most part to intensity-modulated
direct-detection (IM/DD) systems. That is, the laser is intensity modulated either directly
or externally, and the receiver uses direct detection of optical intensity.

4.3 Optical Fibers

This section begins with an exposition of how light propagates through optical fibers
under ideal conditions. This is followed by a survey of the principal impairments in
fiber propagation: attenuation, dispersion, and fiber nonlinearities. We conclude with a
brief description of soliton propagation, which provides a promising approach to very
long-distance optical transmission.

4.3.1 Principles of Guided-Wave Propagation

The end-to-end optical path in a communication system consists of a series of guided-
wave structures: the fibers themselves, as well as the various components used for
coupling signals in and out of the fibers and switching them from one fiber to another.
Thus, we begin with a brief discussion of how light propagates through waveguides.
In cases when geometries have large dimensions compared with a wavelength (for
example, in multimode fibers or in free space), geometric optics is sufficient to explain
the phenomena of interest. In other cases (for example, single-mode fibers), a wave
picture is necessary.

4.3.1.1 Rays: Geometric Optics

The wave-guiding properties of an optical fiber are easily understood in the context of
a multimode step-index fiber. (However, for reasons to be explained later, multimode
fibers are rarely used in high-speed long-distance communications.) As shown in the
cross-section in Figure 4.2(a), a step-index fiber has a core of radius a with a con-
stant refractive index n 1 , and a surrounding glass cladding of outside radius b with a
slightly lower index n 2 .2 The air surrounding the cladding has a refractive index n 0 .
Single-mode fibers typically have cores with diameters of 8 to 12 µm and a cladding
diameter of 125 µm, whereas multimode fibers have cores of approximately 50 µm in
diameter.
Reflection and refraction at the boundary of two media are illustrated in Figure 4.3,
which shows the case n 1 > n 2 . The relations between the angle of incidence θi , the angle
of reflection θr , and the angle of the transmitted ray (angle of refraction) θt are
θr = θi (4.1)
2
The refractive index of a medium is the ratio of the speed of light in free space to the speed of light in the
medium.
 Enabling Technology 169

Step-Index Fiber Graded-Index Fiber

Jacket
b b
Cladding
a a

Core

n1 n1
n2 n2
a a
Index

Index
b b
n0 n0

Radial Distance Radial Distance

(a) (b)

Figure 4.2 Refractive index profiles for fibers. (From [Agrawal97, Figure 2.1]. Copyright
1997.
c
Reprinted by permission of John Wiley & Sons, Inc.)

and

n 1 sin θi = n 2 sin θt , (4.2)

where the latter is called Snell’s law.

What makes fiber optics work is total internal reflection, which was first demon-
strated by John Tyndall in 1854. Equation (4.2) shows that there is a critical angle

Incident Ray n1 n2 < n 1

θi

θr θt

Reflected Ray

Refracted Ray

Figure 4.3 Snell’s law.

170 Multiwavelength Optical Networks

Unguided Guided
Ray Ray
Acceptance
Cone

θc
θc
θa

Figure 4.4 Ray propagation in a step-index fiber. (From [Saleh+91, Figure 8.1-3a]. Copyright


c 1991. Reprinted by permission of John Wiley & Sons, Inc.)

θc = sin−1 n 2 /n 1 , where the transmitted ray lies right on the boundary. At angles of
incidence greater than θc , all energy is totally reflected, resulting in a guided ray. It is
these rays, guided within the core of the fiber, that carry our optical signals. Figure 4.4
illustrates guided and unguided rays. A ray entering the fiber at a sufficiently small angle
of incidence (shown as the acceptance cone in the figure) is totally reflected, bouncing
back and forth between the internal walls of the fiber as it propagates, whereas a ray
incident outside the acceptance cone is partially refracted at each bounce. For rays en-
tering the fiber from air, the largest possible angle of incidence for guided rays is the
acceptance angle θ a , which equals

sin−1 n 21 − n 22 .

The numerical aperture of the fiber is defined as


N A = sin θ a = n 21 − n 22 .

Note that numerical aperture increases with the fractional refractive index change,
 = (n 1 − n 2 )/n 1 . In a typical cladded fiber,  ≪ 1; thus, only a narrow cone of light
is accepted as a guided ray. However, in an uncladded fiber, both  and NA are large.
(For an uncladded silica glass fiber, NA > 1, meaning that rays from all directions are
accepted.)
It is important to note that guided rays entering the fiber at slightly different angles of
incidence take shorter or longer paths from end to end. Thus, if a pulse of energy incident
on the fiber is spread throughout the acceptance cone, the arriving energy is dispersed
in time due to the different path lengths so that the pulse is “smeared out” in time. This
multipath or intermodal dispersion becomes worse for large numerical apertures. Thus,
a high light-gathering ability does not make the fiber a good communication medium.
To reduce the effects of intermodal dispersion, fibers are also manufactured with a
graded-index profile. The ideal graded index for reducing dispersion turns out to be
one in which n decreases parabolically from the center of the core to the cladding; see
Figure 4.2(b). With a graded index, the rays are bent as they approach the cladding, as
shown in Figure 4.5. This bending, together with the fact that rays farther from the core
travel faster (due to a lower n), can reduce intermodal dispersion by several orders of
magnitude.
 Enabling Technology 171

n2

a
n1

Figure 4.5 Ray propagation in a graded-index fiber. (From [Agrawal97, Figure 2.3]. Copyright


c 1997. Reprinted by permission of John Wiley & Sons, Inc.)

4.3.1.2 Modes: The Wave Picture

Geometric optics fails to predict the behavior of light when the dimensions of the
confining medium are comparable with the wavelength, which is the case in single-
mode fibers and in most photonic devices. In this case, the electromagnetic wave picture
is required. For this we must start with Maxwell’s equations, the fundamental equations
governing all electromagnetic phenomena.
An electromagnetic wave is defined at any point in space and time by its electric and
magnetic field vectors E(r, t) and H(r, t) respectively, where r is a position vector in
some arbitrary coordinate system. These quantities are related by
∂D
∇ ×H = (4.3)
∂t
∂B
∇ ×E = − (4.4)
∂t
∇ ·D = 0 (4.5)
∇ · B = 0. (4.6)

These are Maxwell’s equations for a charge- and current-free medium, where the
quantities D and B are the electric and magnetic flux densities, respectively. In free
space, the flux densities are directly proportional to the corresponding fields. However,
in general we have

D = ǫ0 E + P (4.7)
B = µ0 H + M, (4.8)

where ǫ0 and µ0 are the electric permittivity and the magnetic permeability of free space,
respectively, and P and M are the polarization and magnetization densities, respectively.
In almost all situations encountered in optical transmission, the medium is non-
magnetic (M = 0) – a condition that is assumed henceforth. The P vector can be
related in complex ways to the corresponding E field, and it is this complex relationship
in the glass medium that is responsible for most of the impairments in fiber transmission.
Furthermore, by modifying this relationship in special ways through external control, it
172 Multiwavelength Optical Networks

is possible to construct the many devices that are the building blocks of optical networks:
switches, isolators, filters, and so forth.
A wave analysis of fiber propagation normally begins with an assumption of an
“ideal” fiber, after which nonideal conditions are added as small perturbations on the
ideal model. Let us assume the following ideal conditions. The medium is
r Linear: P is a linear function of E.
r Nondispersive: The medium is nondispersive in time if P at any point in time is a
function of E at the same point in time (in other words, the medium is memoryless).
It is spatially nondispersive if P at any point in space is a function of E at the same
point in space.
r Homogeneous: The relation P(E) is independent of r.
r Isotropic: The relation P(E) is independent of the direction of E. In isotropic media,
P and E are parallel.

Under these assumptions, Equation (4.7) can be written in the simpler form

D = ǫE, (4.9)

where the scalar constant ǫ is now the permittivity of the medium. Maxwell’s equations
can then be combined to yield the vector wave equation

1 ∂ 2u
∇ 2u − = 0, (4.10)
c2 ∂ 2 t
√
where c = 1/ ǫµ0 is the speed of light in the medium. The wave equation is satisfied
by each component of both E and H.
Because of linearity, any solution of the wave equation can be represented as a
linear combination of other solutions. This suggests representing the optical fields that
satisfy that equation as sums of sinusoids; that is, moving from the time domain to
the optical frequency domain. Let E(r, t) = Re{E(r)e jωt } and H(r, t) = Re{H(r)e jωt },
where E and H are the complex envelopes of E, and H, respectively, and ω is the optical
frequency in radians per second. Each of these functions represents a monochromatic
signal. Substituting either of these into the wave equation, we obtain the Helmholtz
equation

∇ 2 U + k 2 U = 0, (4.11)

where k = ω/c is called the wavenumber.

The story becomes much more involved from now on. Solutions of the Helmholtz
equation must be found that match the boundary conditions of the fiber for the core and
the cladding. Working in the cylindrical coordinate system of the fiber (Figure 4.6), the
spatial variables are radius r , axial distance z, and angle φ. The classic approach uses
separation of variables, in which case we seek solutions in the form

U (r, φ, z) = u(r )e− jlφ e− jβz (4.12)

 Enabling Technology 173

Er
a Cladding

r φ
Ez

Core z
y
Eφ

Figure 4.6 Cylindrical coordinates. (From [Saleh+91, Figure 8.1-4]. Copyright

c 1991. Reprinted
by permission of John Wiley & Sons, Inc.)

for an integer value of l.3 These represent traveling waves along the axis z of the fiber,
with propagation constant β.
Solutions of Equation (4.11) that satisfy the required boundary conditions exist only
for certain permissible (characteristic) values of β. These values must satisfy the char-
acteristic equation associated with Equation (4.11), which embodies the constraints of
the fiber geometry. Each characteristic value of β yields a solution for u(r ) – a Bessel
function in the case of cylindrical geometry – and for each value of the index l in
Equation (4.12), the characteristic equation has a discrete set of solutions: βlm , for m, a
positive integer. A solution of Equation (4.11) for each value of βlm represents a mode
in the fiber.
Note that there is a dependence on optical frequency (or wavelength) in Equa-
tion (4.11), through the wavenumber k = ω/c, so that the propagating (guided) modes
and the constants βlm depend on the frequency of the optical signal. The function con-
necting each propagation constant βlm to frequency ω is called a dispersion relation. It is
this relation that determines the speed at which power in the mode propagates: its group
−1
velocity vlm . The group velocity for a given mode is related to β by vlm = dβlm /dω.
When vlm is frequency dependent, monochromatic waves of different frequencies prop-
agate at different speeds. This phenomenon is known as waveguide dispersion. Because
a pulse of light contains energy distributed over a range of frequencies, each of these fre-
quency components propagates at a different speed, resulting in a deformation (typically
spreading) of the pulse as it propagates down the fiber (see Section 4.3.2.3).
There are only a finite number of guided modes in the fiber at any given frequency,
corresponding to a finite number of values of β satisfying the boundary conditions.
Each mode except the fundamental mode (l = 0, m = 1) has a cutoff frequency below
which it cannot be sustained in the fiber. Thus, as the optical frequency is decreased
(wavelength is increased), a point is eventually reached when only one guided mode
exists. A fiber is single mode when it operates over the range of wavelengths sustaining
only the fundamental mode.
These concepts have a more intuitive interpretation in the case  ≪ 1, which corre-
sponds to typical fibers. In this case, the wave propagating down the fiber has only small
3
Each component of the electric and magnetic fields is in this form.
174 Multiwavelength Optical Networks

field components in the z direction, so that it approximates a transverse electromagnetic

(TEM) wave (i.e., a plane wave propagating along the axis of the fiber). The condition
for single-mode operation of a fiber of radius a is given by this equation:4
2π (a/λ0 )NA < 2.405, (4.13)
where λ0 is the free space wavelength of the signal. Equation (4.13) indicates that for
single-mode operation, a fiber must have a radius on the order of its wavelength of
operation and/or a small numerical aperture.
Like plane waves, modes in waveguides have polarization states. For each mode (i.e.,
each value of βlm ), there are two linearly independent solutions of the wave equation,
which correspond to two orthogonal polarization states. For example, linear polarization
in the two Cartesian directions x and y represents two orthogonal states. The polarization
states are actually two “degenerate” modes with the same propagation constant. In a
fiber with a circular cross-section, the polarization state (i.e., the relative amplitudes
of each polarization mode) tends to change randomly along the fiber due to minor
defects in the fiber. This is because the symmetry of the fiber and the fact that the two
polarization modes propagate down the fiber at the same speed make it easy for power
to be transferred from one mode to another. Sometimes it is desirable to eliminate this
random drift in polarization, which can be done by forcing the two polarization states
to have different propagation constants. In practice this is accomplished in polarization-
maintaining fibers by making the cross section of the fiber noncircular or by making the
medium anisotropic.5

4.3.2 Optical Fiber Technology: Transmission Impairments

As demands for high performance increase, it becomes more and more important to
minimize the effects of transmission impairments in fibers, either through improvement
and redesign of the fiber itself or by compensating for deleterious effects. This sec-
tion explores the limits imposed on transmission distance and bandwidth imposed by
attenuation, dispersion, and certain nonlinear effects in fibers.

4.3.2.1 Fiber Geometry and Fiber Cables

Optical fibers for communications are normally made of silica and are manufactured
with a variety of geometries. Certain dopants (e.g., GeO 2 and P2 O5 ) are used in the
core to increase the refractive index, and others (e.g., B2 O3 ) are used in the cladding
to decrease the refractive index. The refractive index profile influences the waveguide
dispersion and can be designed to manage the overall dispersion characteristics of the
fiber (see Section 4.3.2.3).
The cables containing the fibers must be designed for mechanical and environmental
protection as well as strength. Large cables may contain well over 100 individual fibers.
Figure 4.7 shows some typical cable designs. One illustration depicts a ribbon cable, in
which as many as 12 ribbons containing 12 fibers each are packed in one cable.
4
See [Saleh+91, p. 286].
5
Anisotropy can be induced by stressing the fiber in a particular direction.
 Enabling Technology 175

Mylar
Kevlar Polyurethane Tape Steel Rod
Polyethylene Polyethylene

Steel Fiber Ribbons

Fiber Fiberglass Core Fiber Polyethylene
Tube

Figure 4.7 Commercial fiber cables. (From [Agrawal97, Figure 2.20]. Copyright

c 1997. Reprinted
by permission of John Wiley & Sons, Inc.)

4.3.2.2 Attenuation
Although the ideal fiber discussed earlier is a lossless medium, real fibers have losses
due to a number of mechanisms. The simplest way to express these losses is through a
relation of the form

d P/dz = −α P, (4.14)

where P is the optical power propagating down the fiber at some point z, and α is a
positive attenuation coefficient. Integrating Equation (4.14) gives

PR = e−αL PT , (4.15)

where PT is the power launched into the fiber and PR is the power received at the end of
a fiber of length L. The attenuation coefficient is generally expressed in units of decibels
per kilometer; that is,
10 PR
αdB = − log10 , (4.16)
L PT
where L is in units of kilometers. In this case, the exponent in Equation (4.15) becomes
−0.23αdB L.
The coefficient α is a function of wavelength, as shown in Figure 4.8. As can be
seen, the attenuation minimum occurs near 1550 nm and is approximately 0.2 dB/km in
currently used fibers. A secondary minimum of approximately 0.5 dB/km occurs near
1300 nm. Most current systems operate near 1550 nm.
As indicated in Figure 4.8, losses are due primarily to three effects: material absorp-
tion, Rayleigh scattering, and waveguide imperfections.
r Material absorption occurs because of resonances of the silica molecules as well as
the impurities in the fiber. Resonance absorption is indicated in the figure as ultraviolet
(UV) absorption (due to electronic resonances at the short wavelengths) and infrared
absorption (due to vibrational resonances at the longer wavelengths). The most serious
impurity effect is due to OH ions (from traces of water trapped in the fiber), which
cause the major peak at 1390 nm as well as several minor peaks.
176 Multiwavelength Optical Networks

3
αdB (dB/km)

1
OH
Rayleigh Absorption
Scattering

0.3
UV Absorption
Band Tail Infrared
Absorption

0.1
0.6 0.8 1.0 1.2 1.4 1.6 1.8
Wavelength λ (µm)

Figure 4.8 Attenuation as a function of wavelength. (From [Saleh+91, Figure 8.3-2]. Copyright

c 1991. Reprinted by permission of John Wiley & Sons, Inc.)

r Rayleigh scattering occurs because the medium is not absolutely uniform, which
causes small fluctuations in the refractive index. This causes the light to be scat-
tered, attenuating the propagating wave. Because the scattering is proportional to λ−4 ,
Rayleigh scattering is the dominant loss factor at the short wavelengths. This, together
with infrared absorption at the long wavelengths, limits the usable optical spectrum
to a range of approximately 800 to 1700 nm (excluding the OH attenuation peaks),
which represents approximately 20 THz of bandwidth.
r Waveguide imperfections are caused by nonideal fiber geometries, which occur due to
manufacturing imperfections and small bends and distortions in the fibers. Normally
these contribute a relatively small additional component to the loss.

As can be seen from Figure 4.8, the various loss mechanisms limit the low loss trans-
mission window to a neighborhood of 1550 nm. Most current high-speed and medium- to
long-haul transmission systems operate near the center of this window: 1530–1565 nm,
designated as C- (conventional) band by the International Telecommunications Union
(ITU). As demand has grown for more usable bandwidth, and technology has been
developed to exploit a wider transmission window, additional bands have been defined:
the longer wavelength bands, the L- (long) band, 1565–1625 nm, and the U- (ultra long)
band, 1625–1675 nm; and the shorter wavelength bands, the S- (short) band, 1460–
1530 nm, the E- (extended) band, 1360–1460 nm (where the OH peak occurs) and the
O- (original) band, 1260–1310 nm, where the first generation of relatively short-range
lightwave systems operated. In 1998 Lucent Technologies introduced ALLWave fiber,
in which the “water peak” loss is effectively eliminated. Other manufacturers followed
with similar products.
 Enabling Technology 177

Dense wavelength division multiplexing (DWDM) with wavelength spacings down to

12.5 nm was standardized by the ITU in 2002 for long-haul applications. However, it was
clear that the economics of metropolitan area optical networks were quite different from
long-haul networks (link lengths of 100s or 1000s of kilometers). Metro networks (diam-
eters of about 100 km or less) can tolerate more severe fiber transmission impairments
than long-haul networks, which means they can operate outside of the C band. However,
they cannot justify the expensive equipment necessary to maintain the tight requirements
for wavelength stability in DWDM. These considerations, together with the fact that the
problem of the water peak had been eliminated led to the coarse wavelength division
multiplexing (CWDM) standard designed for Metro networks. It specifies a grid at 20-nm
spacing providing 18 channels covering a range from approximately 1300 to 1600 nm.
Although fiber loss was the major limiting factor in optical fiber transmission until
the 1990s, this limitation was eliminated to a large degree in C-band systems, virtually
overnight, with the introduction of the EDFA in 1989 (see Section 4.4). More recently,
improvements on the EDFA, Raman amplifiers (Section 4.4.2), and other types of fiber
amplifiers have extended the usable long-haul transmission window into the longer and
shorter wavelength bands.

4.3.2.3 Dispersion
With optical amplification greatly mitigating the effects of attenuation, the effects of fiber
dispersion become more important. A narrow pulse launched on a fiber tends to smear
out as it propagates along the fiber, with its width broadening as it progresses. When a
pulse broadens to the extent that it overlaps neighboring pulses, the resultant intersymbol
interference (ISI) sharply increases the BER. This fiber dispersion phenomenon imposes
a limit on the bit rate that can be supported on a dispersive fiber of a given length.
There are two basic dispersive effects in a fiber: intermodal dispersion and chromatic
dispersion.

Intermodal Dispersion
The most serious form of dispersion, intermodal dispersion, occurs in multimode fibers
because different modes have different group velocities. (This was illustrated in the geo-
metric optics picture of propagation, in which different rays traveled different distances.)
Because the pulse power is distributed (generally unequally) over different modes, repli-
cas of a pulse, one for each mode, arrive at the destination with different propagation
delays, spreading out the received energy. Figure 4.9 illustrates pulse-broadening effects
in different types of fibers.
As shown in Figure 4.9(a), the pulse arrivals for different modes in a step-index
fiber are spread out in time, with different energy in each. Figure 4.9(b) depicts how
received pulses are bunched closer together in a graded-index fiber. In fibers greater
than a certain critical length, mode coupling due to material imperfections can cause a
reduction of the intermodal dispersion, as shown in Figure 4.9(c). Because of intermodal
dispersion, multimode fibers are limited to very low bit-rate-distance products (on the
order of tens of megabits per second-kilometer for step-index fibers and a few gigabits
per second-kilometer for graded-index fibers).
178 Multiwavelength Optical Networks

0 t 0 t
(a) Multimode Step-Index Fiber

0 t 0 t
(b) Graded-Index Fiber

0 t 0 t
(c) Multimode Step-Index Fiber (coupled modes)

Figure 4.9 Broadening of pulses due to dispersion. (From [Saleh+91, Figure 8.3-8]. Copyright

c 1991. Reprinted by permission of John Wiley & Sons, Inc.)

Chromatic Dispersion
Although intermodal dispersion is absent in single-mode fibers, there is another effect –
chromatic dispersion, also called group velocity dispersion (GVD) – that occurs due
to the frequency dependence of the group velocity. There are two sources of chromatic
dispersion: (1) waveguide dispersion and (2) material dispersion.
A simple way of understanding chromatic dispersion without going into the exact
details of pulse shapes is to work in the frequency domain. Assuming that the transmitted
pulse has a spectral width σω in optical (radian) frequency, (or, equivalently, σλ =
λ2 σω /2πc in wavelength), each spectral component of the pulse arrives with delay
T = L/vg , where vg is the group velocity for that component. The received pulse width
is then broadened by an amount
d −1
σT ≈ L v σω = |D|Lσλ , (4.17)
dω g
where the dispersion parameter D (given in units of picoseconds per kilometer–
nanometer) is defined as6
2π c d 2 β
D=− . (4.18)
λ2 dω2
As indicated in Section 4.3.1.2, the propagation constant β in a single-mode fiber
depends on frequency through the dispersion relations, which in turn depend on the
6
When operating at a zero-dispersion wavelength (D = 0), a more exact expression for pulse broadening,
taking higher derivatives of β into account, is necessary.
 Enabling Technology 179

Cladding
a a
Core 0 0
n n

Dispersion
Coefficient
0 0

λ λ
(a) Dispersion-Shifted (b) Dispersion-Flattened
Fibers Fibers

Figure 4.10 Dispersion coefficients as a function of frequency. (From [Saleh+91, Figure 8.3-6].
Copyright

c 1991. Reprinted by permission of John Wiley & Sons, Inc.)

geometry of the fiber. This is the source of waveguide dispersion. In addition, the material
in fibers is time dispersive; that is, the polarization density P is not a memoryless
function of E in the fiber medium. This time-dependent effect can be modeled as a
frequency-dependent refractive index n(ω), which produces an additional, material-
dependent frequency variation in β, so that

D = D M + DW , (4.19)

where D M is the material dispersion and DW is the waveguide dispersion. In general,

both components are frequency dependent.
The material dispersion in standard silica fibers is negative at short wavelengths
and positive at the longer wavelengths, passing through zero at approximately 1300 nm.
Because DW is a function of fiber geometry, it is possible to produce fiber designs, called
dispersion-shifted fibers, in which the contribution of DW causes the zero-dispersion
point to shift to the 1550-nm range, thereby making minimum dispersion occur at the
same wavelength as minimum loss. In dispersion-flattened fibers, the design produces
a dispersion profile that is close to zero over a wide spectral range. These two cases
are illustrated in Figure 4.10. A typical refractive index profile for a dispersion-shifted
fiber is shown in Figure 4.10(a), together with the dispersion coefficient as a function
of wavelength. The material dispersion is indicated by a dashed line and the combined
material and waveguide dispersion is shown by a solid line. Figure 4.10(b) shows the
same quantities for dispersion-flattened fibers.

Polarization Mode Dispersion

As mentioned earlier, the two polarization states of the fundamental mode may propagate
at slightly different group velocities due to asymmetries in the fiber. This is known as
polarization mode dispersion (PMD), which also produces pulse broadening.
The effects of PMD are considered to be negligible on a digital transmission link if
the average time differential, τPMD , between the two modes is less than one-tenth of
180 Multiwavelength Optical Networks

the bit period of the signal. For example, this gives an allowable value of average time
differential of 40 ps for a 2.5-Gbps OC-48 signal or 10 ps for a 10-Gbps OC-192 signal.
As opposed to chromatic dispersion, a phenomenon that can be calculated in a de-
terministic fashion, PMD is typically produced by independent randomly distributed
effects in a fiber. Because of this the PMD effects accumulate in a “square root” fashion.
Thus, for a link of length L km, we have
√
τPMD = D PMD L, (4.20)
√
where D PMD is the PMD dispersion coefficient in ps/ km. In a normal (not polarization-
maintaining) fiber, this effect is usually negligible compared to chromatic dispersion.
However, because of its deterministic nature chromatic dispersion is relatively easy √ to
manage, whereas PMD is not. Typical values of PMD are less than√ D PMD = 0.2 ps/ km
7
√ and D PMD = 2 ps/ km for a dispersion-
for newly installed standard single-mode fiber
shifted fiber but can be as low as 0.15 ps/ km. However, PMD can be more serious
in polarization-sensitive switching devices. For example, for a lithium niobate switch
(see Section 4.10.1), a typical value of average time differential might be τPMD =
2 ps. Thus, cascading of a large number of these can present dispersion problems at high
bit rates. See Section 4.9 for a more detailed discussion.

Polarization-Dependent Loss and Gain

The variation of the insertion loss/gain with the different input signal polarization states
is called polarization-dependent loss/gain (PDL/PDG). Due to these effects, a polarized
signal applied to a polarization-sensitive component may be attenuated or amplified
differently than the unpolarized noise, and the signal-to-noise ratio will be modified
accordingly. Such SNR modifications lead to performance degradation. PDL appears
in optical components such as directional couplers, isolators, filters, and band-splitters,
whereas PDG is an effect found in optical amplifiers and certain types of switches. The
system impact of PDL is discussed in Section 4.9.2.

Effects of Dispersion on Maximum Bit Rates

As stated in the beginning of this section, it is the ISI due to pulse broadening that
limits bit rates in dispersive fibers. A good estimate of the maximum permissible bit
rate is obtained by assuming that the bit rate Rt cannot be allowed to exceed a value
Rmax = k/σT = k/|D|Lσλ , where a reasonable value of k is 0.25, indicating a pulse
overlap into the next bit interval of approximately 25%.
It remains to connect pulse spectral width σλ to other system parameters. Unfortu-
nately, σλ depends in a complex way on the laser characteristics, the way in which it
is modulated (direct or external), and the waveform of the modulating pulse. However,
some simplified relations can be obtained in two limiting cases: (1) wide laser linewidth
and (2) narrow laser linewidth.
In the first case, the laser linewidth (in optical frequency) is assumed to be much larger
than the width of the Fourier transform of the pulse itself. The latter is on the order of

7 √
Although there is still older fiber in the ground with PMD values of D PMD = 0.5 ps/ km.
 Enabling Technology 181

T p−1 , where T p is the pulse width. This is typical of directly modulated or multimode
lasers modulated at relatively low bit rates. In this case, σλ ≈ σl , where σl is the laser
linewidth (in wavelength units), so that
0.25
Rmax ≈ . (4.21)
|D|Lσl
In the second case, the laser linewidth is assumed to be small compared with the
pulse spectral width. This is called the transform-limited case. It applies, for example, to
externally modulated single-mode lasers running at high bit rates (gigabits per second).
We now have σλ ≈ λ2 /cT p . Assuming that Rt ≈ T p−1 , we have
1 0.25c
Rmax ≈ . (4.22)
λ |D|L

Analog Modulation: Subcarriers

Effects of dispersion manifest themselves significantly but differently in analog-
modulated systems. For example, consider an optical carrier, intensity modulated by
a sinusoidal signal. This produces upper and lower sidebands with phases relative to the
carrier that are modified by dispersion as the signal propagates down a fiber. This phase
change can produce a situation in which the intensity modulation is almost completely
converted to phase modulation. Using direct detection, the phase modulation will not be
reproduced at the output of the photodetector, so that the modulating signal will be lost.
These phase shifts due to dispersion limit the maximum possible modulation frequency.
In an example of this effect, analyzed in [Meslener84], it is shown that a total dispersion
of 2000 ps/nm limits the modulation frequency to a maximum of approximately 5 GHz.
(Dispersion of this magnitude might typically occur over 100 to 200 km of standard
single-mode fiber [SMF].) This is a serious limitation in subcarrier systems, in which
the subcarrier frequencies can be in the 10-GHz range (see Section 5.2.2).

Dispersion Management
The history of fiber deployment throughout the world closely reflects the development
of the related transmission system technology. Until the late 1990s, most of the fiber in
the ground in the United States and Europe was standard SMF, with properties that are
consistent with the state of the transmission art in the mid-1980s. The dominant trans-
mission systems at that time were “second-generation” systems (using the terminology
in [Henry+88]). These systems, operating in the 1300-nm band, became feasible when
suitable laser and detector technology became available in that band. (At that time, the
technology for the 1550-nm band had not yet been developed.)
Dispersion was a particularly serious impairment in second-generation systems be-
cause single-mode lasers were not available at that time. Thus, the fact that 1300 nm is
close to the zero-dispersion point in an SMF made it a logical band in which to work
even though attenuation is considerably higher than at 1550 nm.
With the development of single-mode lasers (see Section 4.5.1) – the current “third-
generation” systems – operating in the 1550-nm band largely replaced the second gen-
eration. However, to deal with the nonzero dispersion in this band, the fiber deployed to
182 Multiwavelength Optical Networks

optimize transmission performance was dispersion-shifted fiber (DSF). Because

Japanese deployment followed that in the United States, DSF is the dominant fiber
deployed in Japan today.
The advent of the EDFA, which operates in the 1550-nm band, extended enormously
the reach of these third-generation systems. And with that extension came the appear-
ance of nonlinear transmission impairments; in particular, four-wave mixing (FWM).
Nonlinear effects had been negligible before the EDFA because the maximum transmis-
sion distances between signal regenerators were on the order of hundreds of kilometers,
compared with thousands of kilometers for amplified systems. This was not long enough
for nonlinear impairments to become significant.
It was quickly realized that minimization of dispersion in the fiber made the four-wave
mixing problem worse. Thus, and fortuitously, the older SMF turned out to be better
for combatting nonlinearities (see Section 4.3.2.4). It remained, however, to deal with
dispersion.
It can be seen from Equation (4.22) that in the best possible case – the transform-
limited regime – the dispersion-limited reach of a connection decreases with the square
of the bit rate. Thus, to extend this limit, something must be done about the dispersion
coefficient D. In attacking this problem, the most important observation is that it is a
linear phenomenon. Thus, its effects are reversible by cascading suitable compensating
linear components. One of the first approaches to dispersion management was to use
lengths of dispersion-compensating fibers (DCFs) in cascade with a dispersive trans-
mission link. The DCFs are designed to have a large negative dispersion in the 1550-nm
band to compensate for the 10 ps/nm-km to 20 ps/nm-km positive dispersion in SMFs.
Another approach uses a nonzero dispersion fiber (NDF) with small dispersion (posi-
tive or negative), tailored to reduce the effects of nonlinearities without significant pulse
broadening. By using fiber spans with alternating positive and negative dispersion, the
net dispersion coefficient can be made almost zero over a long span. At the same time, the
nonzero dispersion over each span tends to reduce the effects of FWM [Chraplyvy+95].
In one experiment a 400-Tbps-km bit-rate-distance product was achieved using combi-
nations of positive-dispersion NDF and DCF [Srivastava+98].
Spectral inversion is still another approach to dispersion compensation [Gnauk+97].
In midsystem spectral inversion, the signal spectrum is inverted in the middle of a long
span of dispersive fiber. In this way, the frequency dependence of the group velocity
(which produces dispersion), experienced by the signal over the second segment of the
span, is just the opposite of that experienced over the first half of the span. This is
because the frequency components of the inverted signal spectrum “see” relative phase
shifts that are just the opposite of those seen by the uninverted spectrum. The net effect
is to cancel out the dispersion over the total span. Several processes have been used for
implementing spectrum inversion, including four-wave mixing and difference-frequency
conversion (see Sections 4.3.2.4 and 2.3.3).
Finally, in a fourth approach, chirped fiber Bragg gratings (FBGs) are used for disper-
sion compensation. In the work reported in [Garrett+98], six long (1-m) gratings were
cascaded over 480 km of conventional fiber to achieve eight-channel transmission with a
38.4-Tbps-km bit-rate-distance product. We leave the details of FBGs to Section 4.10.4.3.
 Enabling Technology 183

With the development of various means of controlling dispersion the next issue that
appeared was dispersion slope. In a WDM system carrying many wavelengths, the slope
of the dispersion curve causes different wavelengths to experience different dispersion.
To improve performance, compensation is now applied to slope as well as dispersion
itself. A monolithically integrated AWG-based slope equalizer has been demonstrated
with 8- and 16-channel operation at 40 and 20 Gbps, respectively [Takiguchi+03].
At this writing, dispersion compensation is a rapidly developing activity. It is clear
that as systems and networks move to higher bit rates (OC-768 and beyond) dispersion
management becomes a critical task for the network designer. Dispersion slope compen-
sation also becomes necessary for long-haul systems. Several commercial dispersion
compensation modules have been announced that have a high dispersion slope com-
pensation rate (80–100%) that is suitable for standard single-mode fiber. At the above
high data rates, small changes in dispersion across the system can be very deleteri-
ous. These changes may come from various factors such as temperature changes in the
environment, network reconfiguration, and/or fiber imperfections. Tunable optical dis-
persion compensators are an attractive solution to this problem although their economics
are not yet justified. (A MEMS-based tunable dispersion compensator is described in
[Madsen+99].) A more economical approach is settable optical dispersion compensa-
tion modules, which offer the designer the ability to change their dispersion value on
the day of system installation but remain fixed thereafter. In addition to compensation
at the level of fibers, there are several new and promising approaches for pre- and post-
compensation within the transmitter and receiver, respectively. Examples are prechirp
techniques (Sections 4.5.3.1 and 4.9.3) and electronic equalization (Section 4.7.3). We
are close to the day when dispersion will be managed at will, and the dispersion problem
will be largely vanquished.

4.3.2.4 Nonlinear Effects

Just as dispersion effects became significant when the attenuation problem was solved,
nonlinear effects became dominant when dispersion was reduced. Although the nonlinear
effects in fibers are very small, they can become important over long, amplified but
unregenerated transmission spans. The following nonlinear effects are of concern in
multiwavelength systems:
r Stimulated Raman scattering (SRS)
r Stimulated Brillouin scattering (SBS)
r Four-wave mixing (FWM)
r Self- and cross-phase modulation (SPM/XPM)

Stimulated Raman Scattering

In each of the two scattering effects just listed, light interacts with the fiber medium
producing inelastic collisions, during which the wavelength of the scattered photon is
longer than that of the incident photon, meaning that energy is lost in the process. In each
case, a signal present at the wavelength of the scattered photons can produce stimulated
emission of another photon at the same wavelength and hence amplification.
184 Multiwavelength Optical Networks

Stimulated Raman scattering involves energy loss to vibrational waves in the medium,
with all waves interacting with each other as they propagate in the same direction in the
fiber. In the case of a single-channel system, the threshold power at which SRS becomes
important is very high: approximately 500 mW. Thus, as opposed to Rayleigh scattering,
it is unimportant as a loss mechanism in communication systems. It does have important
consequences in introducing cross-talk (intermodulation) in multiwavelength systems.
Because SRS involves the transfer of power from the wave at the higher frequency to
that at a lower frequency, it can be used to advantage for optical amplification, in which the
higher frequency light serves as a pump, amplifying the lower frequency wave carrying
the signal to be amplified. This is a very broadband process (approximately 200 nm),
which was proposed for fiber amplification as long ago as 1973 [Stolen+73]. Although
it has been largely eclipsed by the EDFA, there has recently been renewed interest in
Raman amplification in both the 1300- and 1550-nm bands. (See Section 4.4.2.)

Stimulated Brillouin Scattering

In SBS, the power lost in the scattering process is transferred to an acoustic wave. In this
case the scattered wave and the acoustic wave both propagate in the backward direction.
As opposed to SRS, SBS is a very narrowband process: The downshift in frequency for
systems operating in the 1550-nm band is approximately 11 GHz. For this reason, there
is no significant cross-talk effect, but the threshold for SBS is much lower than that for
SRS – a few milliwatts for fibers longer than 20 km.

Four-Wave Mixing
FWM and phase-modulation effects can be explained in terms of a nonlinear relation be-
tween P and E, producing nonlinear refraction. In silica glass, there is a small cubic term
in this nonlinearity, related to a third-order nonlinear susceptibility χ (3) (see [Agrawal95,
p. 16]). The cubic relation between P and E can be expressed as a square-law nonlinearity
in the refractive index:
n(ω, |E|2 ) = n 0 + n 2 |E|2 , (4.23)
where the first term represents the linear, frequency-dependent part of n, accounting for
material dispersion, and the second term accounts for various nonlinear effects.
The relation in Equation (4.23) produces intensity-dependent and lossless8 effects on
the fiber, leading to pulse distortion when only a single signal is present and cross-talk
when several signals are present, as in multiwave systems. Although these effects are
small, they can accumulate substantially on long, amplified fiber links. These effects are
not always deleterious. They can sometimes be used to advantage (see Sections 2.3.3
and 4.3.3).
Four-wave mixing can occur if three signals are present at neighboring optical frequen-
cies ω1 , ω2 , and ω3 . The cubic nonlinearity then potentially produces additional signals
at the sum and difference frequencies – for example, ω4 = ω1 + ω2 − ω3 . (Another
possibility exists with only two signals present at frequencies ω1 and ω2 ,
8
Energy may be exchanged between propagating waves, but no energy is absorbed by the medium, as it is in
SRS and SBS.
 Enabling Technology 185

giving ω4 = 2ω1 − ω2 .) These signals occur only if certain “phase-matching” con-

ditions are satisfied. Phase matching of two propagating waves occurs when their prop-
agation constants are equal. In this case, approximate phase matching happens when
the frequencies of all four signals are sufficiently close together. This condition occurs
in the case of dense WDM systems running at moderate power levels. We then have
ω1 ≈ ω2 ≈ ω3 ≈ ω4 , where ω4 = ω1 + ω2 − ω3 . With phase matching, FWM accumu-
lates along a long fiber link, imposing a severe distance limit on multiwave transmission.
For example, a 32-channel system running at a power of 0.5 mW per channel with
50-GHz channel spacing on DSF is limited to approximately 100 km.
The problem of FWM on a DSF is exacerbated by the fact that the near absence
of dispersion causes different frequency waves to propagate at nearly the same group
velocity, producing approximate phase matching. A way to avoid FWM is to break the
phase-matching condition by intentionally introducing dispersion. For example, in an
amplified SMF, the system described in the previous paragraph has negligible FWM
for as long as 5000 km! Thus, it is possible to solve both the dispersion problem
and the FWM problem simultaneously by introducing controlled alternating (positive
and negative) dispersion throughout a long fiber link to disrupt phase matching while
maintaining the net end-to-end dispersion near zero [Chraplyvy+95].

Self- and Cross-Phase Modulation

Refractive index nonlinearity can also produce small changes in the propagation constant
as a function of intensity. Over a length of fiber, this produces accumulated phase shifts,
which vary with intensity. The phase shifts represent SPM, producing a corresponding
broadening of an optical pulse as it propagates down the fiber. To a first approximation,
this can be explained by the fact that the pulse is phase shifted by a different amount
at each point in time, depending on its intensity at that time. When more than one
signal is present, the nonlinear interactions between the signals can produce a related
phenomenon – XPM – that depends on the aggregate power in all signals. This becomes
more troublesome as more signals are superimposed on each other in multiwavelength
systems.

Effective Interaction Length

Because the various nonlinear effects depend on intensity, they become less significant
as the intensity of a wave decays. Thus, the nonlinear effects tend to disappear as the
wave progresses down the fiber. This results in a limited effective length over which the
fiber nonlinearity influences the wave. For a fiber with attenuation coefficient α, this
effective interaction length is defined as
1 − e−αL
L eff = , (4.24)
α
where L is the actual fiber length. For long fibers, we have L eff ≈ 1/α, which gives
L eff ≈ 20 km in the 1550-nm window.
It is important to note that the concept of effective interaction length applies to
unamplified systems. In systems containing cascaded line amplifiers, each amplifier on
186 Multiwavelength Optical Networks

the fiber typically restores the signal intensity to the value originally launched on the
fiber, so that an additional length L eff must be associated with each amplified span.

Nonlinear Effects in Multiwavelength Systems

As indicated, most of the nonlinear impairments in fibers increase with increasing
numbers of simultaneous channels. Thus, it is of interest to study these effects in systems
with many closely spaced channels. Figure 4.11 shows relations between maximum
permissible power per channel and number of channels in a typical system, considering
the various nonlinear effects individually. The case considered is a DSF of length
22 km operating in the 1550-nm band with 10-GHz channel spacing. The criterion
for determining upper limits is a 1-dB power penalty; that is, at the limiting power a
typical channel will have its SNR degraded by 1 dB at the receiver due to the nonlinearity
being examined.
In the case of SRS with many closely spaced channels, power is transferred from the
higher frequency channels to the lower frequency ones; that is, the higher frequency
channels act as pumps, amplifying the lower frequency channels. Because this pumping

1000
Raman

100
Maximum Power per Channel (mW)

10 XPM

Brillouin

1 FWM

0.1

0.01
1 10 100 1000
Number of Channels

Figure 4.11 Limitations due to nonlinear effects in multiwavelength systems. (From [Chraplyvy90,
Figure 10]. Copyright
c 1990 IEEE. Used by permission of The Institute of Electrical and
Electronics Engineers, Inc.)
 Enabling Technology 187

depends on the intensity at a particular point in time, the power transfer depends on the
bit patterns of the various waves, thus introducing a cross-talk effect. The Raman power
limit in a multiwavelength system is illustrated in Figure 4.11. Another curve in the
figure shows the power limit imposed by XPM, as a function of the number of channels.
In contrast to SRS and XPM, Figure 4.11 shows that the power limits imposed by FWM
and SBS are independent of the number of channels. However, they are the critical ones
for systems with a small number of channels.

Large Effective Area Fibers

Much effort has gone into the design of fibers with a large “effective core area.” The
objective is to increase the area over which the optical signal power is distributed
and yet keep the desired wave-guiding properties of the fiber.9 The advantage of these
fibers is that spreading the total optical power over a larger area reduces the optical
field intensity in the core at any one point. This, in turn, reduces the various nonlinear
effects discussed previously. Various designs involving special (e.g., fluoride) doping,
as well as special index profiles, produce effective core areas exceeding 100 µm2 ,
which is roughly twice the area of standard fibers, while maintaining the desired fiber
properties [Kato+98].

4.3.3 Solitons
Well before the advent of fiber optic communications, it was known that a special type
of solitary wave or soliton could exist in certain types of media that are both dispersive
and nonlinear. Although each of these effects by itself tends to distort and broaden a
propagating pulse, the right combination of dispersion and nonlinearity produces a nar-
row, stable pulse that propagates over long distances without any distortion whatsoever,
with one effect compensating for the other. This, of course, is the ideal situation for
long-distance communication.
The form of a soliton can be deduced by modifying the wave equation, Equation (4.10),
to include a nonlinear and time-dispersive refractive index. For a wave propagating in
the z direction, we can write
∂2 E 1 ∂ 2 (n 2 E)
2
− 2 = 0, (4.25)
∂z c ∂ 2t
where n in Equation (4.25) is of the form of Equation (4.23).
By assuming a quadratic dependency of β on ω (corresponding to a linear dependence
of group velocity on ω), assuming that the nonlinearities and dispersive effects are weak,
and dropping “small” terms, it can be shown (e.g., see [Kazovsky+96]) that the complex
envelope u(z, t) of a soliton satisfies the nonlinear Schrodinger equation:
∂u j ∂ 2u
= sgn(β2 ) 2 − j|u|2 u. (4.26)
∂z 2 ∂t

9
A challenge in designing large effective core area fibers is to keep the bending losses low. Larger core
diameters tend to increase losses due to the bending of the fiber.
188 Multiwavelength Optical Networks

t − β1z
T0

Figure 4.12 Soliton.

A solution of Equation (4.26), called the fundamental soliton solution, is10


jaz t − β1 z
u(z, t) = U0 e sech , (4.27)
T0

where a = |β2 |/T02 , T0 is the pulse width, and β1 and β2 are, respectively, the first and
second derivatives of β with respect to ω. Equation (4.27) represents a soliton with a
shape shown in Figure 4.12.
The fundamental soliton is stable in the sense that if a pulse approximating a soliton
in shape and amplitude is launched on a fiber it tends toward a soliton as it propagates
and thereafter retains the soliton form. Furthermore, solitons propagating in opposite
directions pass through each other “transparently.”
To maintain solitons over long distances, fiber amplification is required. Many soliton
experiments using EDFAs and picosecond pulses were carried out in the 1990s. Typical
experiments used a recirculating fiber loop containing several EDFAs to simulate long
transmission links. An example is a demonstration of soliton transmission over a 15,000-
km distance at 5 Gbps. The distance was simulated by a loop with 27-km EDFA spacing
[Mollenauer+92].

4.3.4 Photonic Crystal Fibers

To this day, virtually all optical fibers used in telecommunications are variations on
the theme of the core/cladding structure of Figure 4.2. However, in recent years, the
feasibility of a radically new fiber structure, the photonic crystal fiber (PCF), has been
demonstrated. Also known as a holey or microstructured fiber, it guides light using an
arrangement of air holes running along its z axis. Figure 4.13 shows three configurations,
where the white areas are air holes in the silica fiber. The holes are inserted in the fibers
by stacking capillary tubes in the fiber preform. Three different methods of guiding
light are used in these fibers. The one most resembling ordinary fiber propagation is
the effective-index PCF, shown in Figure 4.13(a). Here the array of holes surrounding
the solid core (made by omitting the center hole in the regular array) play the role of
the low index cladding in an ordinary fiber. However, by properly dimensioning the
holes and their spacing, the effective index of the array can be made considerably lower

10
An infinite number of higher order solutions also exist.
 Enabling Technology 189

(a) (b) (c)

Figure 4.13 Three types of microstructured fibers. (From [West+01, Figure 1]. Copyright

c 2001
IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

than the cladding index of an ordinary fiber and strongly wavelength dependent: low
at long wavelengths and approaching the core index at short wavelengths. As a result
the condition for single mode operation (Equation [4.13]) is satisfied at all wavelengths,
producing an “endlessly single-mode” fiber.
A second version of the PCF is the air-clad core fiber of Figure 4.13(b), where a solid
center is surrounded by a single ring of large diameter holes, which produce the effect
of a solid rod hanging in air, meaning that the core-cladding index ratio is extremely
large. In this case, propagation effects are quite different: the effective area seen by the
propagating mode is exceedingly small, meaning that the fiber becomes highly nonlinear
because of the increased power density seen by the material.
Finally, the third version is the photonic band gap (PBG) fiber. An example is shown
in Figure 4.13(c), where an air core is surrounded by a regular array of holes. In this
case, the holes act much like a crystal lattice, interacting with the lightwave to produce
photonic band gaps: wavelength bands where no propagation takes place down the fiber.
One application of this type of PCF is that the guided wave can be essentially confined
to the hollow core, resulting in extremely low attenuation.
As with many new technologies, applications are often serendipitous: the predicted
ones (e.g., broadband low loss transmission) have not materialized, but other appli-
cations, especially at the component level, are being proposed at a rapid rate. For
example, by properly tailoring the hole configuration, fiber designs can be optimized for
dispersion flattening, dispersion shifting and dispersion compensation [Shen+03]. Po-
larization maintaining PCFs have been proposed using hole arrays that produce highly
birefringent fibers [Libori+01]. A number of optical signal processing applications
make use of fiber nonlinearities, which can be controlled and enhanced using PCFs (see
Section 4.11). An example is pulse compression using a highly nonlinear microstructured
fiber [Petropoulos+03].
The array of holes forming the photonic crystal can also be used on a slab to produce
waveguides in a planar configuration. In [Sugimoto+03] a directional coupler is built
using this technique, where the guides are fabricated by omitting rows of holes from
190 Multiwavelength Optical Networks

the array where the guides are located. Another application of a planar geometry is
a PCF-based purely optical switch [Stapleton+07]. A directional coupler composed of
photonic crystal waveguides on a silicon-on-insulator wafer is operated as a switch under
optical control. Light traversing the coupler is directed to one or the other of the output
ports by illuminating the coupler with an optical control signal from an 850-nm laser.
The same basic structure can also be configured as a tunable filter.
A particularly intriguing potential application involves the combination of PCF
and semiconductor technology. The goal is to fabricate semiconductor devices within
the holes of microstructured fibers so as to produce all-fiber optoelectronic devices
[Badding+06]. The vapor deposition technology used in semiconductor fabrication has
been adapted for this task by using pressures that are high enough to make the gases act
like fluids. In this way, the chemical vapors flow into the holes of a PCF producing coat-
ings on the walls of the holes that have the desired electronic properties. A field effect
transistor has been fabricated within a fiber as a proof of the utility of these techniques.
These methods open the door to many applications of microstructured fibers that extend
their use beyond purely optical waveguiding devices to the optoelectronic domain; for
example, embedding semiconductor pumping within a fiber laser or enhancing and tai-
loring nonlinearities within a fiber by inserting appropriate nonlinear semiconducting
material – possible uses would be in wavelength conversion, regeneration, and other
optical signal processing operations.
At this writing, PCFs for practical applications are still on the beginning of the
learning curve, and their success or failure will depend on how practical issues such
as manufacturability, reliability, and cost are resolved. However, the prognosis is good,
considering the rate at which demonstrated applications are appearing on the scene.

4.4 Amplifiers

Because fiber attenuation limits the reach of a nonamplified fiber span to approximately
200 km for bit rates in the gigabit-per-second range, wide area purely optical networks
cannot exist without optical amplifiers. Amplifiers are typically used in three different
places in a fiber transmission link. As power amplifiers, they serve to boost the power of
the signal before it is launched on the line, extending the transmission distance before
additional amplification is required. Amplifiers operating as line amplifiers are located at
strategic points along a long transmission link to restore a signal to its initial power level,
thereby compensating for fiber attenuation. Finally, as a preamplifier, the device raises
the signal level at the input of an optical receiver, which serves to improve signal detec-
tion performance (i.e., the receiver sensitivity). In each case, the desired properties are
different. For power amplifiers, the important feature is high gain; preamplifiers require
a low noise figure, and line amplifiers require both. Optical amplifiers are also employed
at various other points in a network (for example, within an optical switching node to
compensate for losses in the switch fabric). Semiconductor optical amplifiers (SOAs)
were developed in the 1980s but they never had a serious impact on long-distance
transmission because of a number of negative features discussed in Section 4.4.3. In
 Enabling Technology 191

the case of fiber amplifiers, especially the EDFA and the Raman amplifier (RA), how-
ever, the situation was quite different. The first articles on EDFAs appeared in 1987
[Desurvire+87, Mears+87]. Within a few years of that time, 9000-km unrepeatered
transmission was demonstrated. Shortly thereafter, soliton experiments showed that
transmission distances could be extended almost indefinitely. All of these experiments
used EDFAs. It is not an exaggeration to say that these devices have revolutionized optical
communications.
Although the EDFA played a fundamental role in extending the reach of optical
transmission systems it still had some drawbacks, including operation confined to a
limited band of the optical spectrum and a nonflat gain profile. In contrast, RAs, which
were first demonstrated well before the EDFA [Stolen+73] and then virtually ignored for
three decades, have more recently attracted renewed interest. This stems mainly from
their ability to increase both the reach and the aggregate bit rate carried on a fiber; that
is, the usable fiber bandwidth.
As cost has replaced performance as the driving force in technology development,
other more cost-effective amplifers have come to the fore, collectively known as amplets.
We begin this section with a discussion of EDFAs, followed by descriptions of RAs and
SOAs and concluding with a brief look at amplets.

4.4.1 Erbium-Doped Fiber Amplifiers

The EDFA belongs to a family of rare-earth-doped fiber amplifiers, the class of other pos-
sible dopants, including praseodymium (used for amplification in the 1300-nm range),
neodymium (originally used for very high-power lasers), ytterbium (which has been used
as a codopant with erbium), and thulium (amplifying in the S band). The important place
of the EDFA in optical communications is due primarily to the fact that the properties
of erbium produce amplification in a fairly wide band (approximately 35 nm) within
the 1550-nm low-attenuation window in fibers. Furthermore, the EDFA has many other
desirable features that will become apparent as we proceed.
Three different EDFA structures are shown in Figure 4.14. In each case, the amplifier
is of the traveling wave type, consisting of a strand of single-mode fiber, typically on the
order of tens of meters long, doped with erbium. (The points S in the figure represent
fiber splices.) The EDFA is an optically pumped device, so energy is supplied by an
optical source (an LD), which injects power into the doped fiber at a wavelength matched
to the characteristics of erbium (980 or 1480 nm). Pumping can be forward, backward, or
bidirectional (see Figures 4.14[a] through 4.14[c], respectively). The pump is typically
coupled into the transmission fiber via a wavelength-selective coupler (WSC). (EDFA
modules used in the field typically include other components, such as optical isolators to
eliminate reflected power, and various devices for signal power monitoring, stabilization,
and control.) Amplification occurs by transfer of power from the pump wave to the signal
wave as it propagates down the doped fiber.
Like many other forms of amplifiers of electromagnetic radiation, the EDFA op-
erates via a three-energy level system. The model representing this process is shown
in Figure 4.15. Levels E 1 , E 2 , and E 3 are the ground, metastable, and pump levels,
192 Multiwavelength Optical Networks

Pump
Signal Signal
In WSC Out
S S EDF S
S

Pump LD

(a) Forward Pumping

Pump
Signal Signal
In WSC Out
S EDF S S
S

Pump LD

(b) Backward Pumping

EDF

Pump Pump
Signal Signal
In Out
WSC WSC
S S S S
S S

Pump LD Pump LD

(c) Bidirectional Pumping

Figure 4.14 Basic erbium-doped fiber amplifier structures. (From [Desurvire94, Figure 5.4].
Copyright

c 1994. Reprinted by permission of John Wiley & Sons, Inc.)

N3 E3

N2 E2
980 nm

1530 nm

N1 E1

Figure 4.15 Energy levels in EDFA.

 Enabling Technology 193

respectively. The populations (fractional densities) of erbium ions in the three energy
levels are denoted N1 , N2 , and N3 , where N1 > N2 > N3 when the system is in ther-
mal equilibrium (no pump or signal present). When pump and signal are present, these
populations change as ions move back and forth between levels, accompanied by the
emission or absorption of photons at frequencies determined by the energy-level dif-
ference. The wavelengths associated with the dominant transitions are indicated in
Figure 4.15. The wavelength λ for each transition is given by the quantum relation
λ = hc/E, where h is Planck’s constant and E is the difference in energy levels.
In actuality, the three levels in the simplified diagram of Figure 4.15 are narrow bands,
so each transition is actually associated with a band of wavelengths rather than a single
line.
Two pump wavelengths are typically used for EDFAs: 980 and 1480 nm. As shown
in Figure 4.15, by absorbing energy from a 980-nm pump, Er3+ ions in the ground state
are raised to state E 3 . The rate at which these transitions occur is proportional to N1 Pp ,
where Pp is the pump power. These excited ions decay spontaneously to the metastable
state E 2 , and this transition occurs at a rate much faster than the rate from level E 1 to level
E 3 .11 This means that in equilibrium under the action of the pump, the ion population in
the ground state is reduced and accumulates largely in state E 2 . This process is referred
to as population inversion because we now have N2 > N1 , the reverse of the situation in
thermal equilibrium. The transition rate from level E 2 to level E 1 is very slow compared
with the other transitions, so that the lifetime τ , in the state E 2 (the reciprocal of its
transition rate to E 1 ) is very long (approximately 10 ms). Similar pumping action can
occur at 1480 nm, in which case the ions are raised directly to the upper edge of the E 2
band. Reliable semiconductor laser pump sources have been developed for EDFAs at
both the 980- and 1480-nm pump wavelengths.
The wavelength band for transitions from state E 2 to the ground state is in the 1530-
nm range, making it ideal for amplification in the lowest attenuation window of fibers.
The dominant transitions from E 2 to E 1 are radiative, which means that they are of
two types: spontaneous emission and stimulated emission. In the former case, an ion
drops spontaneously to the ground state, resulting in the emission of a photon in the
1530-nm band, and this appears as additive noise. Spontaneous emission noise is an
unavoidable by-product of the amplification process, predicted by quantum theory. Its
phase, direction, and polarization are independent of the signal. In the case of stimulated
emission, an incident photon in the 1530-nm range stimulates the emission of another
photon at the same wavelength in a coherent fashion (with the same direction, phase,
and polarization). If the incident photon is from a signal, this produces the desired
amplification of the optical field. However, the incident photon could also have originated
as a spontaneous emission “upstream” on the fiber, in which case this is called amplified
spontaneous emission (ASE), which represents the major source of noise in amplified
fiber transmission systems.

11
The transition is dominantly nonradiative, so energy is lost to the fiber medium rather than emitted in the
form of radiation.
194 Multiwavelength Optical Networks

Gain

1500 1550 1600

λ (nm)

Figure 4.16 EDFA gain profile.

4.4.1.1 Gain Profile

The fairly large amplification bandwidth of the EDFA is due to the finite width of
the energy bands. The width of the energy bands is caused by a number of physical
phenomena, including the Stark effect [Desurvire94, p. 8], which splits the main energy
levels into many sublevels. Because the population is not distributed uniformly within the
E 2 band, the gain is not flat. A typical plot of gain as a function of wavelength is shown
in Figure 4.16. The uneven gain profile, with a peak at approximately 1530 nm, produces
significant problems in a multiwavelength system when many amplifiers are cascaded
over a long transmission span. Not only does uneven gain amplify different wavelengths
unequally, but it also causes a large accumulation of ASE at the peak of the gain
profile, which can eventually saturate the amplifier. Because amplifier cascading on long
links accentuates these effects seriously, gain flattening is an important consideration
in EDFAs. Several solutions to this problem are currently in use. One approach is to
modify the design of the amplifier itself by using different materials such as fluoride
glass [Clesca+94]. Other approaches use gain equalization via controllable attenuators
(see [Ford+98] for a MEMS-based example) or inverse filtering. (See [Vengsarkar+96]
for an example using fiber gratings.)

4.4.1.2 Gain Saturation

The gain of an EDFA is approximately independent of the signal power as long as the
pump power is made high enough so that the pumping rate is much larger than the
stimulated emission rate. This is called the unsaturated gain or small-signal regime. The
small-signal gain under these conditions is an increasing function of pump power. For
a given fiber structure and doping, and a given pump power, there is an optimal fiber
length that maximizes gain. For lengths smaller than the optimum, the pump power is not
maximally utilized, and for larger lengths, pump power is exhausted somewhere along
the fiber, and attenuation takes over. Typical optimal lengths are in the range of tens of
meters. Maximum small-signal gains for EDFAs are typically 30 to 40 dB.
All amplifiers eventually exhibit gain saturation as the signal power is increased. In the
saturated case, the signal extracts so much power from the pump as it propagates down
the fiber that the stimulated emission rate becomes comparable with the pumping rate.
 Enabling Technology 195

The larger the input signal, and the higher the unsaturated gain, the sooner saturation is
out
reached. As saturation increases, the gain decreases. The saturation output power Psat
is defined as the output power at which the gain is compressed by 3 dB. The values of
out
Psat for typical EDFAs are in the hundreds of milliwatts. It should be noted that ASE
also contributes to saturation in an EDFA. When input signals are very small, it is the
ASE that saturates the amplifier first. This is known as amplifier self-saturation.
Because saturation is a nonlinear effect, it produces a number of complications when
multiple signals are being amplified. One problem is that the saturated gain for any one
signal depends on the aggregate power of the other signals as well as its own power.
Thus signals (as well as accumulated ASE) tend to “steal” power from each other. Useful
analytical models for these effects appear in [Jopson+92, Saleh+90]. An advantageous
effect of saturation is that a small amount of it in each amplifier in a cascade of several
amplifiers tends to produce a self-regulating effect.
Several other nonlinear effects are a consequence of this power-stealing phenomenon
but on a shorter time scale. The amplifier gain at any instant in time is a function of the
excited state population N2 , which is depleted momentarily by stimulated emission when
a signal is present. One manifestation of this occurs when an intensity-modulated digital
signal changes from a 0 to a 1. The resultant fluctuation in N2 causes corresponding gain
fluctuations, which are most pronounced in the saturated regime and in the presence of
large signals. Another manifestation occurs when beats from two signals spaced closely
in optical frequency cause gain fluctuations at the beat (difference) frequency.
The gain fluctuations affect all signals being amplified and thus can potentially produce
undesirable cross-talk, with one signal’s intensity fluctuations changing the gain for the
others. These effects are significant only when the gain dynamics are such that gain can
vary on a time scale as fast as that of the signal fluctuations. A simplified interpretation
of gain dynamics in an EDFA is based on the assumption that the maximum speed for
gain fluctuations is on the order of the reciprocal of the lifetime in the excited state,
which is approximately 10 ms. However, actual gain transients in EDFAs can occur on
time scales of hundreds of microseconds, which cannot be predicted using the lifetime
alone. (For a more detailed explanation see [Desurvire94, p. 417].)
In any case, these numbers indicate that signals fluctuating on time scales more rapid
than, say, 100 µs will cause no significant cross-talk in EDFAs. This corresponds to
a minimum bit rate of approximately 10 Kbps to avoid cross-talk (or a WDM signal
separation of approximately 10 KHz to avoid beat frequency effects). The lack of this
cross-talk effect for bit rates higher than 10 Kbps is one of the important advantages of
the EDFA over the SOA (see Section 4.4.3).

4.4.1.3 Noise and Noise Figure

The ASE noise generated in an EDFA can be the limiting performance factor in an
optical transmission link. It is therefore important to quantify this effect.
For an amplifier with gain G, the ASE noise power spectral density at the output at
optical frequency ν (in each polarization state) is [Agrawal97]

Ssp = n sp (G − 1)hν, (4.28)

196 Multiwavelength Optical Networks

∆f
Pin Filter/
Detector

PD
(a)

∆f
Filter/
Detector

(b)

Figure 4.17 Illustration of noise figure.

where n sp , the spontaneous emission factor, is a function of the state population and
approaches its minimum value of 1 with full population inversion. The ASE noise
spectrum for an EDFA is roughly the same shape as the gain profile.
The significance of the ASE noise is most clearly expressed in terms of SNRs and the
amplifier noise figure Fn . These quantities are defined in terms of electrically detected
signals in an ideal system, as illustrated in Figure 4.17. The noise figure is defined as
SNRin
Fn = , (4.29)
SNRout
where SNRin is the electrical SNR seen when a signal of power Pin is converted to a
photocurrent at the output of an ideal photodetector (PD) – see Figure 4.17(a). The noise
in this case is shot noise due to the fact that the ideal detector is counting photons,
which arrive randomly at the detector. (The detection process must be an integral part
of any noise calculation, reflecting the quantum limits of lightwave transmission.) The
numerator in Equation (4.29) is given by
(R Pin )2 (R Pin )2 Pin
SNRin = 2
= = , (4.30)
σ 2q R Pin  f 2hν f
where (R Pin )2 is the square of the average photocurrent, σ 2 = 2q R Pin  f is the shot
noise power (the variance of the photocurrent), R = q/ hν is the responsivity of an ideal
detector, q is the electron charge, and  f is the bandwidth of the electrical detector.
The quantity SNRout is the electrical SNR seen with the amplifier inserted before the
photodetector – see Figure 4.17(b). To find SNRout , we compute the variance of the
photocurrent after amplification with gain G. Because the detector acts as a square-law
device, the photocurrent variance contains terms due to shot noise and ASE noise by
themselves, as well as signal-spontaneous emission beat noise because of the mixing
between the signal and the ASE in the photodetector. It turns out that the latter is the
dominant term, provided that G ≫ 1, and most of the ASE noise is filtered out at the input
of the detector. This can be done by making  f small enough to exclude extraneous
noise but include the desired signal. Then we have

σ 2 ≈ 4(RG Pin )(RSsp ) f (4.31)

 Enabling Technology 197

so
G Pin
SNRout ≈ . (4.32)
4Ssp  f
Using Equations (4.28), (4.29), (4.30), and (4.32), the noise figure is given by

Fn ≈ 2n sp (G − 1)/G, (4.33)

which corresponds to at least a 3-dB SNR degradation in the high-gain case. (In real
systems, Fn is typically at least 4 dB.)

4.4.1.4 Amplifier Chains

Over a long transmission link, it is necessary to use several EDFAs interconnected
by fiber sections to compensate for fiber attenuation.12 The gain of each amplifier is
normally adjusted so that it compensates for the attenuation on one section of fiber. The
question of optimal amplifier spacing then arises. It turns out that this is a fairly complex
issue that depends, among other things, on the way in which the amplifiers are pumped,
effects of fiber nonlinearities, and practical issues such as amplifier accessibility, cost,
and so forth.
We examine a fairly simple model here, in which a fiber of length L is divided into N
sections of spacing s = L/N . An amplifier is placed after each section, with a saturated
gain that just compensates for the fiber attenuation on one section: G = eαs.13 The total
accumulated noise power spectral density at the end of this chain (taking into account
both polarization states) is then

Sn = 2N Ssp = 2n sp (G − 1)N hν = 2n sp (eαs − 1)N hν. (4.34)

Note from Equation (4.34) that for a fixed amplifier spacing the effect of accumulated
noise in the cascade grows linearly with the length of the link but decreases as the
amplifier spacing decreases (i.e., as the number of amplifiers increases). Thus, the
optimal strategy in this case is to place a very large number of low-gain amplifiers
very close together, with the limiting case being one long, distributed amplifier. Cost,
however, dictates the opposite strategy! In current practice, a compromise is reached,
with spacings ranging from 20 to 100 km, typically giving an SNR at the receiver of at
least 15 dB. The spacings are based on constraints such as maximum permissible power
on a fiber,14 effects of fiber nonlinearities, and receiver sensitivity.

12
It is also possible to design the EDFAs as distributed amplifiers in which the doped fiber is long (on the
order of kilometers) and the gain is distributed throughout its length [Desurvire94, p. 122].
13
Exact compensation is not necessary. In a long chain, if the amplifiers are adjusted to have a small-signal gain
slightly larger than the section attenuation, saturation tends to stabilize gains at a cascaded value that just
compensates for the end-to-end fiber loss. This happens because if the signal amplification overcompensates
for loss on the upstream stages, the resultant increased signal power drives the downstream amplifiers into
saturation, thereby tending to stabilize the overall saturated gain.
14
The maximum power is determined not only by the technological limitations – laser and amplifier output
powers and fiber nonlinearities – but also by eye safety as well. The latter consideration limits power to
well under 100 mW to limit risks to service personnel (see [ANSI88]).
198 Multiwavelength Optical Networks

1.2

copolarized
Normalized gain coefficient 1.0
orthogonal
0.8

0.6

0.4

0.2

0.0
0 5 10 15 20 25 30 35
Pump-signal frequency difference [THz]

Figure 4.18 Raman gain coefficient in bulk silica as a function of frequency shift. (From
[Bromage04, Figure 15]. Copyright
c 2004 IEEE. Used by permission of The Institute of Elec-
trical and Electronics Engineers, Inc.)

4.4.2 Raman Amplifiers

The discussion of the EDFA provides a useful framework for describing the Raman am-
plifier: They are both fiber amplifiers, with important similarities as well as differences,
so they can often complement each other in applications.
As mentioned in Section 4.3.2.4, stimulated Raman scattering (SRS) can cause trans-
mission impairments in fibers, but it can also be used for amplification. When SRS is
used for amplification, pump power is introduced into a fiber carrying an optical signal,
with the pump operating at a frequency higher than the signal frequency, just as in the
EDFA (and other rare-earth-doped fiber amplifiers). The pump photons interact with the
material in the fiber through inelastic collisions, producing scattered photons at lower
energy (and frequency) than the pump photons, with the remaining energy imparted
to the fiber medium in the form of vibrational waves, called optical phonons. If the
frequency of the scattered photon is the same as that of a signal photon propagating
in the fiber, it can stimulate the emission of a second signal photon, thereby amplify-
ing the signal, a process identical to that which occurs in the EDFA. The performance
of the RA can be expressed in terms of a Raman gain coefficient (RGC). An illustration
of the form of a normalized RGC as a function of frequency shift between the pump and
signal appears in Figure 4.18. As the figure shows, RA gain is polarization dependent.
The gain coefficient for copolarized pump and signal waves is an order of magnitude
higher than in the orthogonally polarized case. Polarization dependence is mitigated by
the averaging effect of the polarization mode dispersion in the fiber medium and can
be circumvented by using either polarization diversity pumping or a single depolarized
pump [Bromage04].
An important difference between the RA and the EDFA is that the energy levels of
Er3+, which determine the gain profile of the EDFA (Figure 4.16) are fixed, thereby
 Enabling Technology 199

fixing the position of the amplification band of the device, as well as the possible pump
frequencies. The amplification band for the EDFA is fixed in the vicinity of 1530 nm –
the middle of C-band – which is a primary reason for its importance in optical commu-
nication but which limits its flexibility in exploiting other transmission bands in optical
fibers. In contrast, for the RA it is only the pump/signal frequency difference (a band
centered around 13 THz) that is fixed by the physics of the process, and any pump
frequency can be used. Changing the pump frequency automatically shifts the waveband
where amplification occurs. Thus the amplification band of an RA can be centered at
any desired frequency in the optical fiber transmission window by adjusting the pump
frequency appropriately. Furthermore, for a single pump the amplification bandwidth is
large (about 6 THz in Figure 4.18), and this band can be extended by superimposing
several pumps at different frequencies. This makes the RA an excellent tool for widening
the usable bandwidth of long-haul WDM transmission systems beyond C- and L-bands
into the S- and U-bands and beyond.
The RA can be configured either as a distributed or discrete (lumped) amplifier.
A typical distributed RA (DRA) consists of a long transmission fiber into which a
counterpropagating (backward) Raman pump is injected. (Backward pumping reduces
the effect of pump noise, as explained below.) The distributed amplification results in
reducing the perceived loss along the span, which effectively improves the reach of
the span and/or increases its capacity. In a discrete RA, the amplifier consists of a coil
of fiber together with pump(s) and ancillary equipment for monitoring, control, and
perhaps other purposes such as dispersion compensation, gain flattening, or adding and
dropping channels. Isolators are used to keep the pump power from escaping into the
line. The fiber medium used in the discrete case is shorter than in the DRA, but it still is
typically of the order of kilometers – two orders of magnitude longer than the EDFA. A
significant advantage of the discrete RA is that the amplifying fiber can be chosen at will
to suit a number of criteria. For example, a dispersion compensating fiber can be used to
provide dispersion compensation for the transmission fiber, with the additional benefit
of improving the Raman gain coefficient [Gruner-Nielson+01]. The primary purpose of
discrete RAs is generally to expand the usable bandwidth of a transmission link, whereas
the primary purpose of a DRA is to improve the reach of a fiber span.
When several amplified spans are placed in tandem, with lumped line amplifiers
placed at the junction points, the result is a hybrid arrangement as shown in Figure 4.19.
The advantage of this arrangement is illustrated by a comparison of signal powers along
the line with and without the DRAs (see Figure 4.20). Without the DRAs the signal
level drops linearly along each span. Due to the overall span loss a high signal power
must be launched into each span, which tends to produce nonlinear impairments at the
beginning of the span. But at the far end of the span, the attenuation drops the signal into
the noise level. Clearly the span is too long for discrete amplification alone. However,
by adding distributed amplification throughout the span, the signal initially attenuates at
almost the same rate as without RA but then increases in power toward the far end of the
span as it encounters stronger pump power. The DRA pump signal power is also shown,
decreasing from right to left in the figure, which is why distributed gain is highest toward
the far end of the span. The net effect of the distributed amplification in the spans is to
improve the overall system performance by reducing noise as well as nonlinear effects.
200 Multiwavelength Optical Networks

Transmission Line Fiber

Optical Line
(Raman Gain Fiber)
Amplifier
Pump Signal(s)

Coupler Coupler Coupler

Amplified Transmission Signal

Raman Pump
Unit

Figure 4.19 Hybrid distributed-discrete amplification. (From [Islam02, Figure 4]. Copyright
c
2002 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

Distributed amplification keeps the signal above the noise level at the far end of each
span, so the optical SNR at the input of each line amplifier is improved. Furthermore,
distributed amplification makes it possible to launch the signal into each span at a lower
power level, thereby reducing nonlinear impairments due to high signal levels.
There are a number of additional considerations that work for and against the RA. On
the positive side, it operates in ordinary silica fibers, requiring no special materials or
dopants. This makes it ideal as a means of adding distributed amplification to existing
long transmission links. Furthermore, it has better ASE noise properties than the EDFA.
The latter has an ASE noise power spectral density given by Equation (4.28), where
n sp > 1 due to incomplete population inversion, whereas the RA noise spectrum has
the same form with n sp = 1; i.e., it acts like an EDFA with full population inversion.
However, there are additional sources of noise in RAs that can be more serious than
ASE: in particular, multipath effects caused by reflections and double Rayleigh scatter-
ing. Rayleigh scattering causes forward propagating signals (or noise) to be scattered

Original Transmission DRA Pump Signal

Signal

Nonlinear Effects
Signal Power (dB)

High Noise

DRA Amplified Distance

Transmission Signal

Figure 4.20 Signal and pump power in hybrid system. (From [Islam02, Figure 5]. Copyright
c
2002 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)
 Enabling Technology 201

backward, but when a signal encounters this phenomenon twice, the doubly scattered
signal propagates in the forward direction, recombining with the original signal af-
ter a multipath delay. Discrete double reflections due to imperfections, splices, and
connectors in the fibers cause similar multipath effects. Because the Raman effect is
weak, long fibers are required in RAs, which tend to increase the multipath effects.
Unintended reflections and Rayleigh scattering are present in all fiber systems, but they
are attenuated in a passive fiber. However, when the fiber is pumped the Raman gain
magnifies these effects to the point where the multipath interference places a limit on
the usable gain in an RA [Bromage04].
RAs are normally less efficient than EDFAs in converting pump power to output
signal power. However, their efficiency improves, exceeding that of the EDFA, at the
large aggregate signal powers that occur in long-haul WDM systems with high channel
counts [Islam02]. Furthermore the gain in the fiber medium depends strongly on the
type of fiber being used. Because gain is proportional to pump intensity, it increases
when a given amount of pump power is confined to a small fiber core. Thus, fibers with
smaller cores such as DCFs produce significantly higher Raman gain. This is a particular
advantage in discrete RAs, where there is some choice in the type of fiber being used.
Another drawback of the RA is that it has a very fast (fs) response to pump fluctuations.
This can lead to coupling of pump noise into the amplified signals. These effects can be
mitigated by using a counter-propagating pump, in which case the effects of the pump
fluctuations are averaged out over the length of the pumped fiber. Otherwise, they require
the use of “quiet” pumps; i.e., pumps with very low relative intensity noise.
In deploying RAs as distributed amplifiers, there are some other practical concerns
due to the high pump powers employed. Connectors should be minimized in favor of
splices to reduce reflections and attenuation, and when connectors are required they must
be designed to survive the high pump powers. Also, to protect personnel, automatic laser
shutdown systems must be employed.

4.4.3 Semiconductor Optical Amplifiers

The structure of an SOA is similar to that of a semiconductor laser (see Section 4.5.1).
It consists of an active medium (a p–n junction) in the form of a waveguide, with a
structure much like the stripe geometry laser described in Section 4.5.1. The mobile
carriers (holes and electrons) now play the role of the Er3+ ions in the EDFA.
The energy levels of the electrons in a semiconductor are confined to two bands: the
conduction band, containing those electrons acting as mobile carriers, and the valence
band, containing the nonmobile electrons. A hole, representing the absence of an electron
in the valence band, also acts as a mobile carrier. Mobile electrons and holes are abundant
(i.e., are majority carriers) in n-type and p-type material, respectively.
The two energy bands in semiconductors play a role analogous to band E 2 and
E 1 in the EDFA, but they are much broader than the EDFA bands. A band gap, E g ,
separates the lower edge of the conduction band from the upper edge of the valence
band so that the energy change involved in moving from one band to the other is at
least E g . Transfer of an electron from the valence band to the conduction band (with
202 Multiwavelength Optical Networks

the absorption of energy) results in the creation of an electron-hole pair. One way in
which this occurs is through the absorption of a photon, as in a photodetector (see
Section 4.6.1). The reverse phenomenon, electron-hole recombination (with release of
energy), occurs either nonradiatively (by transferring energy to the crystal lattice) or
radiatively, with the emission of a photon.
The radiative case is of interest to us here for applications in light sources as well as am-
plifiers. Radiative electron-hole recombination occurs either spontaneously or through
stimulated emission involving interaction with an identical photon. These two processes
are analogous, respectively, to the spontaneous and stimulated emission processes in an
EDFA. By proper choice of the semiconductor materials (e.g., InGaAs or InGaAsP),
bandgaps that yield emission and/or absorption wavelengths in the ranges desired for
optical communications (e.g., 1300 or 1550 nm) can be produced.
For photon emission to occur by electron-hole recombination at an optical frequency,
ν, an electron-hole pair must be present with energy levels separated by an amount
E = hν. Furthermore, if the recombination is by stimulated emission, a photon of the
same frequency must be present to interact with the electron-hole pair. The conditions
for these effects to occur depend on the various carrier concentrations and the photon
flux in the active region (the layer around the p–n junction).
In an unbiased p–n junction, a “depletion layer” exists around the junction caused by
diffusion of majority carriers across the junction and subsequent recombination on the
other side. This creates a net charge on each side of the junction and hence a retarding
electric field, preventing further diffusion and draining carriers from the layer around the
junction. The depletion layer can be broadened by reverse-biasing the junction, thereby
augmenting the retarding field. This is the condition for operation of the p–n junction
as a photodetector.
On the other hand, by forward-biasing, the retarding field is reduced, allowing more
majority carriers to cross the junction, becoming minority carriers on the other side.
This creates a condition favorable to recombination in the active region because once
the mobile electrons from the n side cross over to the p side (at which point they
become minority carriers), they encounter a large concentration of holes with which
to recombine. A similar situation occurs for the mobile holes moving in the opposite
direction. This effect, which increases the population of minority carriers in the active
region on each side of the junction, is called minority carrier injection.
The current flow through the forward-biased junction acts as an electrical pump,
supplying the energy necessary to produce an inversion of the carrier population in
the active region. This is analogous to the Er3+ ion population inversion in the EDFA
produced by optical pumping. The light-emitting diode (LED) is a simple application of
radiative recombination. It is a forward-biased p–n junction producing its radiation by
spontaneous emission. This effect is called injection electroluminescence.
Now suppose an optical signal is introduced into a waveguide embedded in a forward-
biased p–n junction, which we now want to use as an amplifier. By applying sufficient
injection current, conditions can be established in which stimulated emission dominates
spontaneous emission and absorption in the guide. At this point, optical gain is produced,
and the device becomes a semiconductor amplifier. Because the energy bands are broad
in a semiconductor, the SOA amplifies over a much wider band than an EDFA.
 Enabling Technology 203

Although its broadband gain characteristic is a positive feature, the SOA has a number
of negative features. First, the carrier lifetime in the high-energy state is very short (on
the order of nanoseconds). As indicated earlier, this means that signal fluctuations at
gigabit-per-second rates cause gain fluctuations at those rates, producing cross-talk
effects between simultaneously amplified signals. These effects do not occur in EDFAs
until the bit rate drops into the 10-Kbps range. Second, because of its asymmetrical
geometry, the SOA is polarization dependent. The EDFA, with its cylindrical geometry,
is not. Third, the coupling losses between the fibers and the semiconductor chip reduce
substantially the usable gain and output power. Fourth, the noise figure of a typical SOA
is slightly higher than that of a typical EDFA due to fiber-chip coupling losses, although
advances in packaging technology have improved that.
Because of recent improvements in broadband SOAs [Park+05], polarization-
dependent gain (PDG) and noise figure (rather than gain flatness and saturation-induced
cross-talk) are becoming the predominant limiting performance factors. The best com-
mercial SOAs can be specified having PDG as low as 0.5 dB over the C-band (30-nm
bandwidth). However, CWDM-capable SOAs typically exhibit PDGs of 1 dB or more
over a 70-nm band (four CWDM channels).

Quantum Dot SOAs

Another improvement on the basic SOA, which shows promise in a broad range of
applications is the quantum dot SOA [Akiyama+01]. Quantum dots (QDs) are extremely
small semiconductor particles (of the order of 10 nm). SOAs using arrays of QDs in their
active region exhibit important changes in their gain characteristics because of the way
in which the QDs concentrate the injected electron-hole pairs, modifying their energy
states. Some properties that have been demonstrated are extremely fast (picosecond)
nonlinear gain response, very large bandwidth (a few hundreds of nanometers) high
output power, and low noise figure. A typical QD SOA described in [Akiyama+06]
uses InAs QDs on an InP substrate in a self-assembled array. (Electrolytic and pyrolytic
synthesis techniques have been developed in which the QDs assemble themselves into
highly ordered arrays.) The ultrafast gain response produces very low distortion in
amplifying pulses at very high output powers (23 dBm). (Conventional and quantum
well SOAs do not have the high-speed nonlinear gain response necessary to produce
clean pulses at very high powers.)
These QD SOAs have application as amplifiers in CWDM systems, which can benefit
from their broadband characteristics, and as optical signal regenerators, which can use
their high-speed response to advantage in reshaping rectangular pulses. Other potential
applications are in tunable and mode-locked lasers, in wavelength conversion using
FWM, and in optical gates.
SOAs lend themselves to applications in which they can be combined with other
semiconductor components. This includes monolithic integration with switch fabrics or
WMUX/DMUXs fabricated as semiconductor planar lightwave circuits, use as on-off
switches in an integrated switch fabric, and as components of optoelectronic integrated
circuits (OEICs); for example, as a preamplifier in an optical receiver or a power am-
plifier in a laser transmitter. In addition, SOAs have found application in WDM optical
networking at the metro level as described in Section 4.4.4.
204 Multiwavelength Optical Networks

4.4.4 Amplification Trends in Metro Optical Networks: Amplets

Deployment of metropolitan and regional WDM networks is accelerating, and amplets
based on either fiber or waveguide technologies are potentially key components in
these cost-critical applications. The term amplet is used to describe low-cost, modest
performance amplifiers, which are usually not used as line amplifiers but rather as
devices that compensate for loss due to other network elements. As a result amplets
become enablers for new network architectures. Their characteristics include low-cost,
small size, low power consumption, and high reliability. Some of their potential metro
network applications include

1. Range extension on the access network side, which is usually unamplified

2. Subband aggregation where they provide valuable gain as band amplifiers to minimize
signal degradation, allow incremental growth [Antoniades+04b], and compensate for
dispersion compensation device losses
3. Flexible add/drop loss compensation where they can be used as wavelength agile
single-channel amplifiers to overcome network element losses and provide indepen-
dent control of channel launch power
4. Transmitter booster and receiver preamplifier functions
5. Compensation for excess loss in optical switch fabrics

Typical EDFA amplets are intended for single-channel or band operation with the
following performance considerations: output power, power stability, gain and gain flat-
ness, polarization-dependent gain, and multipath interference. Because these amplifiers
provide gain among a small set of channels they can significantly improve system perfor-
mance on those channels. The challenges center around reducing cost (by reducing pump
power, fiber length, and component count and using uncooled pumps). These devices
are fabricated with reduced size, reasonable output powers (+15 dBm total output power
is typical), and relatively wide bandwidths, so they can be used in various C-, L- and C-
+L-band configurations. In addition, low gain ripple and low noise figures are desirable.
Available 80-µm-diameter fibers, coolerless 980 pumps, and miniature components
are some of the enabling technologies making low-cost EDFA amplets possible. Next-
generation amplet array devices may further reduce costs by sharing uncooled pumps.
The erbium doped waveguide amplifier (EDWA) is another proposed amplet device,
which offers the additional potential of integrability with other devices such as optical
attenuators and dispersion compensators. The size of EDWAs is similar to that of SOAs,
and advances in wafer processing enable lower costs at high volumes. An example of
an EDWA is a 9.8-cm-long buried channel waveguide fabricated with a two-step ion ex-
change process in Er-/Yb-doped phosphate glass [Philipsen+00]. Commercial devices
have been demonstrated involving full integration of EDWAs with pump and associated
passive devices. They are small and have output powers of the order of 10–15 dBm,
noise figures of the order of 6–7 dB, and gains of 13–20 dB. Cascades of such devices
have been demonstrated in transparent WDM metro rings in [Reichmann+01].
SOAs incorporating improvements to solve some of the problems mentioned in
Section 4.4.3 are also candidates for applications in metro and regional networks. The
 Enabling Technology 205

main reason for this is their compact size and competitive pricing. In addition, their
broad bandwidth serves the requirements of CWDM, which has established itself as a
cost-effective alternative to DWDM in business access and metro networks. A 240-km
CWDM transmission system using SOAs and Raman amplification in a hybrid configura-
tion was demonstrated in [Reichmann+06], and a discrete hybrid SOA-Raman amplifier
was reported in [Iannone+06], using a highly nonlinear fiber as the gain medium.
Recent articles have reported the impact of PDG on cascades of SOAs in CWDM
systems and presented methodologies for deriving PDG specifications for SOAs in
such systems [Antoniades+06, Roudas+07]. Several methods have also been proposed
for combating the patterning effects (cross-talk) mentioned in Section 4.4.3 and these
are described in detail in [Spiekman03]. Two of the main ones are (1) operating in the
linear regime of the SOA and (2) gain clamping. The latter was initially demonstrated
in [Francis+01] in what is called a linear optical amplifier (LOA). An LOA is
essentially an SOA integrated with a vertical-cavity surface-emitting laser (VCSEL)
(see Section 4.5.2). The VCSEL shares the same active region with the amplifier and is
used to maintain a constant gain in the amplifier. In [Zimmerman+04] and references
therein, a number of recent demonstrations of the use of amplets in metro applications
are presented.

4.5 Optical Transmitters

The light sources for optical transmission systems have evolved from LEDs operating in
the 850-nm range in the first-generation systems (in the 1970s) to semiconductor lasers,
which are still evolving today. With some exceptions for short-distance and low-bit-rate
applications such as fiber to the home,15 virtually all optical transmission systems use
semiconductor lasers as their light sources. This section is confined to that topic.

4.5.1 Lasers
The laser (for “light amplification by stimulated emission of radiation”) was invented by
Schawlow and Townes in 1958, shortly after the maser (for “microwave amplification by
stimulated emission of radiation”), which performed the same operation in the microwave
domain. The first solid-state (ruby) laser was demonstrated by Maiman in 1960. Lasers
exist in myriad forms, including gas lasers, dye (liquid) lasers, and fiber lasers. (One
type of fiber laser consists of a ring of erbium-doped fiber [i.e., an EDFA closed on
itself].16)
A semiconductor laser, also called a laser diode or injection laser, is basically a
device for converting electrical energy to monochromatic light. The conversion process
requires two things: (1) an “active” medium with properties that facilitate conversion
from electrical to optical energy and (2) a geometry suitable for generating and emitting
a monochromatic light beam.
15
See, for example, [Iannone+98].
16
Recall that this type of configuration can arise in optical networks through the inadvertent creation of closed
optical paths. See Figure 2.13.
206 Multiwavelength Optical Networks

Cleaved
Surface

+ −
p n

Cleaved
Surface

Figure 4.21 Fabry–Perot laser. (From [Saleh+91, Figure 16.3-1]. Copyright

c 1991. Reprinted by
permission of John Wiley & Sons, Inc.)

The principle of operation of a laser is the same as that for any other oscillator: gain
plus feedback. In the case of a laser diode, the active medium is a forward-biased p–n
junction. It has optical gain as a consequence of current injection, producing electrical-
to-optical energy conversion. Furthermore, the active medium is enclosed in a geometry
that provides optical feedback; in other words, a portion of the light that it amplifies is
fed back to be reamplified, producing sustained oscillation when the net loop gain is
greater than unity.
A basic laser structure, known as a Fabry–Perot (FP) laser, is shown in Figure 4.21.
The active region is contained in an optical cavity of rectangular geometry, with partially
reflecting mirrors on two sides. The cleaved surfaces (facets) in the figure act as mirrors,
providing partial internal reflection because of the difference between the refractive index
of the cavity material and air. With the exception of the partially reflecting mirrors this
is very similar to the semiconductor optical amplifier described in Section 4.4.3. In an
FP cavity, light is reflected back and forth between the mirrors, with certain wavelengths
being reinforced by constructive interference when the successive reflections of the wave
are in phase with each other. These resonances occur whenever the round-trip distance
between the mirrors is a multiple of the wavelength in the material.
As in the case of SOAs (and LEDs), energy is converted from electrical to optical
form through radiative recombination of holes and electrons in the active region. The
 Enabling Technology 207

necessary population inversion is achieved by current injection in the forward-biased

junction. The longest wavelength of the photon-emission process is determined by the
bandgap of the material, but the possible wavelengths of oscillation in the laser are
limited to those reinforced by the geometry of the cavity.
Lasing is achieved by increasing the optical gain sufficiently to overcome losses
(including the transmitted light) from the cavity. The minimum current required for
lasing is called the threshold current. It should be noted that if the cleaved surfaces
in Figure 4.21 are given antireflection coatings, the device becomes a semiconductor
amplifier. Light enters through one facet, is amplified, and leaves through the other.17
Practical laser structures use more complex geometries than that shown in Figure 4.21
and require special materials to produce light at the wavelengths needed for optical
communication. To reduce the threshold current and to control the radiation pattern,
either a stripe geometry or buried heterostructure (BH) is used. Each confines the light
to a narrow region, which reduces the required threshold current and reduces the light-
emitting area to a small spot instead of the broad beam shown in Figure 4.21. The
optical properties of the material are “engineered” by choosing the materials (typically
indium gallium arsenide phosphide [InGaAsP] for 1300-nm and longer wavelengths)
and adjusting the thickness of the active layer. Multiple quantum well (MQW) structures
are designed with several very thin active layers stacked up, achieving high gain at high
powers [Tsang87].

4.5.1.1 Single-Frequency Operation

The early laser diodes were multimode, multifrequency devices. In the FP laser, for
example, the cavity can support many modes (and frequencies) of oscillation. To obtain
the monochromatic light that is necessary for high bit-rate-distance products, the geom-
etry must be modified to achieve single-frequency operation. By narrowing the active
region, as in a BH laser, all cavity modes of an FP laser are suppressed except the
longitudinal modes in the narrow, active strip. In this way, it operates much like a single-
mode waveguide, supporting longitudinal modes associated with different frequencies.
To reduce operation to a single frequency, all but one of the longitudinal modes must be
suppressed, resulting in single longitudinal mode operation. A number of approaches
are used for this, including cleaved-coupled cavity (C3 ) lasers, external cavity lasers,
distributed Bragg reflector (DBR) lasers, and distributed-feedback (DFB) lasers. In each,
a resonant element is introduced to single out the desired mode.
The most commonly used light sources for high bit-rate, long-distance transmission
are the DBR and DFB lasers. Figure 4.22 shows their structure. Based on the Bragg effect,
a grating such as a corrugated reflecting surface or a periodic variation of refractive index
in an optical medium acts as a wavelength-selective mirror. A distributed Bragg reflector
(DBR) reflects light with a wavelength corresponding to twice the grating period. In
effect, the grating operates as a narrow-band reflecting filter. The distributed Bragg
gratings at the two ends of the DBR laser in Figure 4.22(a) replace the mirrors in an
17
In cases where the injection current is high enough to make stimulated emission dominate spontaneous
emission, and yet the light loss through the facets is so high that the lasing threshold is not attained, the
device emits light as a superluminescent LED.
208 Multiwavelength Optical Networks

p p n p p p n

Active Active
Λ Layer
Layer
Diffraction
Gratings
Guiding
Layer

(a) (b)

Figure 4.22 Single-frequency lasers. (From [Saleh+91, Figure 16.3-10]. Copyright

c 1991.
Reprinted by permission of John Wiley & Sons, Inc.)

FP laser, reflecting light only at wavelengths related to the Bragg grating period. In this
way, the DBR laser operates in a single longitudinal mode (and frequency) defined by
the Bragg grating. Similarly, in Figure 4.22(b), a periodic frequency-selective structure
is introduced as a corrugation along the active layer in the DFB laser. The periodic
structure produces distributed reflections (feedback) that reinforce each other only at
wavelengths related to the periodicity of the corrugations. Although these devices are
costly compared with multimode lasers, they are currently the workhorses of long-
distance optical communications.

4.5.1.2 Mode-Locked Lasers

In a multimode laser, the light that is produced has a wide spectrum, which typically
manifests itself in the form of a wave with slowly and randomly varying amplitude
and phase. However, light with a wide spectrum can also manifest itself in the time
domain as a sequence of very narrow pulses. This happens if the relative amplitudes and
phases associated with the different modes of a multimode laser are “locked” to values
representing the Fourier coefficients for a periodic pulse train. This is the principle of
the mode-locked laser. These devices are often used as sources of periodic picosecond
pulses for soliton transmission. Pulse widths as short as 30 fs have been generated in this
manner (for example, see [Saleh+91, p. 535]). Mode locking is usually implemented
by modulating the laser gain periodically by driving it with a microwave oscillator and
using a saturable absorber or a modulator inside the resonator.

4.5.1.3 Supercontinuum Generation

Supercontinuum light generation is an interesting application of mode-locked lasers.
A supercontinuum is a very broad continuous spectrum formed by propagation of
 Enabling Technology 209

high-power pulses through nonlinear media. First observed in the 1970s, it occurs in
various nonlinear materials such as gases and semiconductors. The term supercontin-
uum does not cover a specific phenomenon but rather a plethora of nonlinear effects,
which, in combination, lead to extreme pulse broadening. Such effects can be self-phase
modulation, cross-phase modulation, and parametric four-wave mixing (FWM) as de-
scribed in detail in Section 4.3.2.4. The widest spectra are obtained when pump pulses,
typically generated by mode-locked lasers, are launched close to the zero-dispersion
wavelength of a highly nonlinear medium. Supercontinuum light has been generated in
the visible wavelength range using Ti:Sapphire femtosecond mode-locked lasers to drive
nonlinear photonic crystal fibers with zero-dispersion wavelengths in the desired wave-
length range. In the infrared range of interest in optical communications, supercontinua
have been generated using mode-locked EDFA fiber lasers [Westbrook+05], as well as
injection-locking of a mode-locked laser diode [Mori+05]. Mode-locked laser diodes
are used to pump the nonlinear material because of their low-noise characteristics and
because they can easily be used to generate optical pulses with high repetition frequency
(50 GHz) and short pulse width (≤ 5 ps). Polarization maintaining, dispersion flattened
fiber is one of the media used to reduce polarization fluctuations and control dispersion,
which is extremely important for the generation of the supercontinuum [Mori+01].
One of main applications of the supercontinuum has been as a white-light source
used in many characterization schemes such as interferometer-based dispersion mea-
surements, broadband attenuation characterization, and numerous spectroscopy and mi-
croscopy procedures. Broadband sources are also needed for the low-coherence-based
imaging-technique optical coherence tomography (OCT). OCT has been used exten-
sively in a large wavelength range for both biomedical and nonbiomedical applications.
However, for biomedical applications, wavelengths in the 1300- to 1500-nm wavelength
range are highly attractive as they provide high penetration depth in tissue.
One of the key WDM applications for supercontinuum light is the generation of a
multiwavelength comb of frequencies anchored to the center frequency of the ITU grid,
with well-controlled spacings. In conventional WDM sources, all wavelengths must be
controlled independently, a sizable task as the number of WDM channels increases.
However, a single multiwavelength source combined with a WDM demultiplexer, which
can generate multiple precisely controlled optical carriers simultaneously, is an effective
way to solve the issue of channel management. A supercontinuum generated by a mode-
locked laser is ideal for this purpose, because the spectrum of the mode-locked pulse
train inherits a periodicity at the frequency of the oscillator used for mode locking.
As an example, the system described in [Mori+05] generates a full grid of carriers at
25-GHz spacing locked to the 193.1-THz center frequency and seamlessly covering the
S-, C-, and L-bands from 1460 to 1625 nm. It uses a mode-locked laser diode (ML-
LD) driven by an oscillator to produce a 25-GHz pulse train. To enhance the signal
quality by reducing its relative intensity noise, a synchronizing signal from a precisely
stabilized master laser diode is injected into the ML-LD. This markedly improves the
noise qualities of the generated carriers. At this writing, work is continuing in the area
of using supercontinuum light in WDM applications.
210 Multiwavelength Optical Networks

4.5.1.4 Tunability
Laser tunability is important in many multiwavelength network applications. Slow trans-
mitter tunability (on a millisecond time scale) is required for setting up connections in
wavelength-/waveband-routed networks, and rapid tunability (on a micro- or nanosec-
ond time scale) is required for time-wavelength division multiple-access applications.
Furthermore, the local oscillator in a coherent receiver must also be tunable to track the
transmitter frequency.
Slow tunability over a range on the order of 1 nm is effected easily via temperature
control, which is used mainly for frequency stabilization (e.g., for locking a laser fre-
quency to a frequency standard). Rapid tunability is achieved in DBR and DFB lasers by
changing the refractive index, which can be done by changing the injected current in a
local region around the grating. The design of the laser allows this to be done without dis-
turbing the active region affecting the gain. This technique is used in multiple-electrode
lasers to achieve tuning ranges as high as 51 nm [Amann94]. In these lasers, tuning may
be either continuous or discrete. An example of the discrete case is a sampled-grating
DBR laser, which can be tuned to one of 101 wavelengths spread over a 44-nm band
[Delorme+98].
Another approach to rapid tunability is to use multiwavelength laser arrays with
wavelengths that span the desired tuning range. One or more lasers in an array can be
activated to produce simultaneous transmissions at selected wavelengths. An integrated
array of 21 lasers with 3.72-nm wavelength spacing, combined and coupled into two
optical amplifiers, is shown in Figure 4.23. The chip is 1 × 4 mm with the DFB laser
array split into two groups of 10 and 11 lasers, visible as horizontal stripes in the figure.

Figure 4.23 Laser array. (From [Soole+94, p. 59]. Copyright

c 1994 Telcordia Technologies, Inc.
Reproduced with permission.)
 Enabling Technology 211

The light from the lasers is fed to a star coupler (the dark area on the right), with two
output waveguides fed to optical amplifiers on the far right of the chip.
Photonically integrated laser arrays are cost-effective in the sense that the per-
wavelength cost of packaging, fiber pigtailing, and control circuitry is reduced. However,
the manufacturing cost of an array is higher than that of an equivalent number of single
lasers because of reduced yield with arrays.

4.5.2 Vertical Cavity Surface Emitting Lasers

Vertical-cavity surface-emitting lasers (VCSELs) are a rapidly emerging transmitter
technology offering a number of potential advantages over the edge-emitting lasers de-
scribed in Section 4.5.1. Their designs allow the chips to be manufactured and tested
on a single wafer. Large arrays of devices can be created on a single chip operating
at many different wavelengths. Using this multiwavelength array, VCSELs can easily
and inexpensively provide wavelength tuning over a relatively large range. The VCSEL’s
uniform, single mode beam profile is desirable for coupling into optical fibers. Further-
more, they are compact, high speed, have low thresholds, and require low drive current.
During the 1990s, the feasibility of long-wavelength (1.3 and 1.5 µm) VCSELs suitable
for WDM applications has been demonstrated. A review of VCSEL development can
be found in [Hasnain00].
Although there are a variety of VCSEL designs, they all have certain aspects in
common. A generic layered semiconductor structure is shown in Figure 4.24. Two
distributed Bragg reflectors of opposite doping are stacked with a cavity layer between
them. The active region resides within this cavity and is composed of a number of
quantum wells. Multiple quantum wells are grown within the cavity to increase the
attainable output power, which is emitted vertically (up or down). The position of the
quantum wells is crucial for maximizing the gain of the device. Current is injected into
the active region through a current guiding structure. To achieve an array of VCSELs
with different wavelengths on the same chip, a thickness gradient at the bottom or top of
the substrate creates a cavity thickness variation that translates into wavelength variation.

Light out
Contacts

p-DBR

Cavity layer
Active layer

n-DBR

Figure 4.24 Typical VCSEL structure.

212 Multiwavelength Optical Networks

Tunability of a VCSEL device can be achieved by mechanically varying the length of the
VCSEL cavity, something that has been demonstrated using a micromechanical structure
that essentially varies the position of a vertically suspended DBR [Hasnain00].
VCSELS have considerable potential in the cost-sensitive metropolitan networking
environment. Their relatively low power can be offset by the use of low-cost SOAs. In
[Filios+03], transmission experiments using 2.5-Gbps VCSELs have demonstrated the
effective use of such devices in WDM metro applications.

4.5.3 Modulation Technology

Lasers are modulated either directly, by varying the injected current, or externally, by
passing the emitted light beam through a controllable device that changes the ampli-
tude and/or the phase of the light. We consider directly modulated lasers (DMLs) and
externally modulated lasers (EMLs) separately.

4.5.3.1 Direct Modulation

The behavior of the laser under small-signal and large-signal modulation is of interest
in different contexts. When modulated with an analog signal, including subcarrier sig-
nals, the modulation index is likely to be small to reduce nonlinear distortion, among
other reasons. In that case, a small-signal model that assumes sinusoidal modulation
is appropriate. The key performance parameter in this case is the small-signal modu-
lation bandwidth, which is the modulation frequency at which the response drops by
3 dB. It can be shown that the modulation bandwidth is proportional to the frequency
of relaxation oscillations in the laser, which increases with the square root of the output
power under biased and unmodulated conditions [Agrawal02, p. 112]. Lasers designed
specifically for high-speed modulation have reached bandwidths of 24 GHz. This wide-
band response is particularly useful for multichannel subcarrier applications, in which
the aggregate modulation bandwidths of all subcarrier channels is limited by the laser
modulation bandwidth when direct modulation is used (see Section 5.2.2).
In digitally modulated systems, it is generally the large-signal modulation behavior
that is of interest. For example, in an intensity-modulated system, the current may vary
from near threshold for a 0 to twice the threshold for a 1. But because changes in
current also produce changes in refractive index, they also create phase modulation
(or, equivalently, frequency modulation) along with the intended intensity modulation.
This effect has been termed chirp, and it produces a significant broadening of the pulse
spectrum. (The parameter involved in intensity-phase modulation coupling is called
the linewidth enhancement factor because the phase modulation increases the effective
spectral width of the emitted light.)
It has been shown [Koch+86] that the instantaneous angular frequency deviation
.
(i.e., chirp) φ(t) of a directly modulated DFB laser is approximately related to its output
optical power P(t) by:


. α 1 d P(t)
φ(t) ∼
= + κ P(t) . (4.35)
2 P(t) dt
 Enabling Technology 213

4 40

3 30

Frequency Chirp (GHz)

2 20
Power (mW)

1 10

0 0

−10
0 0.5 1 1.5
Time (ns)

Figure 4.25 Pulse and accompanying chirp. (From [Agrawal97, Figure 3.23]. Copyright

c 1997.
Reprinted by permission of John Wiley & Sons, Inc.)

In Equation (4.35), α is the linewidth enhancement factor and κ the adiabatic chirp
coefficient. The first term is called transient chirp, whereas the second term is called
adiabatic chirp. A transient chirp-dominated DML will exhibit ringing at the bit tran-
sition points and usually the frequency of the 1s will be similar to that of the 0s
elsewhere. An adiabatic chirp-dominated DML will exhibit in general two separate
levels for the frequency of the 1s and the 0s with ringing during the duration of the
bits.
An example is shown in Figure 4.25, in which a laser is modulated with a rectangular
pulse of width 500 ps (the solid line), corresponding to a 2-Gbps bit rate. The optical
frequency (the dashed line) increases on the leading edge of the pulse and decreases
on the trailing edge due to the transient chirp, with a total excursion of approximately
20 GHz. (The limited modulation bandwidth produces the pulse distortion.) This spectral
broadening can lead to serious dispersion problems on long fiber spans. Thus, DMLs
are generally avoided in favor of the more expensive externally modulated transmitters
on long- and/or high-bit-rate transmission links.
The situation is quite different in metropolitan and regional-area WDM networks,
where transmitter costs are a paramount concern. DMLs have attracted considerable
attention for 2.5-Gbps operation in both 1300- and 1550-nm bands in these applications.
They have the advantages of small size, low-cost, low-driving voltage, and high avail-
able output power. Although their chirp characteristics can significantly limit the maxi-
mum achievable transmission distance over standard single-mode fibers, the interplay of
214 Multiwavelength Optical Networks

transmitter chirp and negative fiber dispersion can be used to offset the deleterious
effects of chirp (see Section 4.9.3).
Negative dispersion fiber in conjunction with DMLs can expand the dispersion-limited
transmission distances significantly. A figure of merit used commonly in the industry to
characterize the performance of a transmitter in the presence of chirp/dispersion is laser
rating. It is basically the total dispersion that can be tolerated on a link using that trans-
mitter before the chirp/dispersion Q penalty reaches 2 dB. The laser rating is determined
experimentally using back-to-back and transmitter plus fiber BER vs. received power
measurements, which are then converted to Q penalties. (See Equation [4.55] for the
relation between Q and BER.) Q penalties determined by simulation can also be used to
approximate this parameter. At 2.5 Gbps typical commercially available 1550-nm DMLs
are rated at 1800 ps/nm. This translates into a maximum transmission distance of about
90 km over SMF. In [Tomkos+01b], it is shown that this distance can be expanded to
over 600 km using negative dispersion fiber. At 10 Gbps, DMLs are rated very poorly,
tolerating only about 170 ps/nm (a transmission distance of only 10 km over SMF)
and thus making them unlikely candidates for Metro network applications. However,
the recent emergence of 10-Gbps DMLs and applications such as 10-Gb Ethernet has
prompted the investigation of other technologies to extend reach. For example, elec-
tronic dispersion compensation (EDC) (discussed in Section 4.7) can double the reach
of these devices where there is a large installed base of standard SMF, ruling out the use
of negative dispersion fiber [Feuer+03].
The effects of chirp are also important in intensity-modulated analog systems when
operated at sufficiently high modulation indices. A thorough analysis of the limits
on analog transmission due to dispersion, using directly modulated lasers, appears in
[Meslener84]. The effect of dispersion on a sinusoidally modulated optical carrier is ex-
amined, showing that frequency modulation effects due to direct modulation, combined
with fiber dispersion, can produce serious harmonic distortion in systems that would be
essentially distortion free if frequency modulation was absent.

4.5.3.2 External Modulation

External modulation, although requiring extra (and costly) components, avoids the chirp
problems discussed earlier. It is, therefore, the technique of choice in high-performance
long-distance systems. In external modulation, the laser output is constant, and it is
modulated by inserting a controllable attenuator (or phase shifter) between the laser and
the fiber link. A typical external modulator uses a Mach–Zehnder (MZ) interferometer
structure implemented on a LiNbO3 substrate, as shown in Figure 4.26. In this device,
the light is split equally between two parallel waveguides on the surface of the substrate
and is then recombined at the output. A control voltage, V, applied to the electrodes
shown in the figure creates a small change in the refractive index through an electro-
optic effect. This causes a change in the propagation constant and a relative phase shift
at the point where the two waves combine. When no voltage is present, the relative phase
shift is zero and the recombined signal exits the device without attenuation (except
for the losses in the waveguides). When a voltage is applied that produces a phase
shift of π between the two arms, the signal is extinguished. Thus, the device acts as
 Enabling Technology 215

Input
Light
Ii

Modulated
Light
Io

Figure 4.26 Mach–Zehnder interferometer. (From [Saleh+91, Figure 18.1-5]. Copyright

c 1991.
Reprinted by permission of John Wiley & Sons, Inc.)

a voltage-controlled switch. These devices operate at high speeds (e.g., 40 GHz) and
can be fabricated using various materials. A modification of the MZ interferometer of
Figure 4.26 allows the application of control voltages on both waveguides, resulting
in a device called a dual-drive MZ external modulator. This enables the generation
of arbitrary quadrature signals [Ho05] that find application in transmission procedures
such as optical suppressed carrier modulation. External modulators can also be used for
phase modulation in coherent systems using a single waveguide with an electro-optically
modulated refractive index.

4.5.3.3 Electro-absorption Modulated Lasers

Transmitters that incorporate an electro-absorption modulator (typically based on InP)
and a DFB laser in one package (EA-DFBs) have become attractive technologies in
WDM systems and networks. The most common of these are those that involve a mono-
lithic device, incorporating both laser and modulator on the same indium phosphide chip.
Another alternative is a hybrid device, with distinct laser and modulator parts integrated
in one package. Compared to transmitters using separate external modulators, typical
EA-DFBs are smaller in size and potentially low cost but have relatively lower output
powers, although higher output power devices have recently been reported [Moodie+03].
Demand is emerging for very compact EA-DFBs with low-drive-voltage requirements
for 80- to 100-km, 10-Gbps metro applications and new short-reach (2-4 km) 40-Gbps
systems. Figure 4.27 shows the operating principle of an EA-DFB transmitter. On the
left is a CW DFB laser followed by the EA modulator. The active part of the modulator
consists of InGaAsP epitaxially grown on top of the InP wafer. The operation of the EA
216 Multiwavelength Optical Networks

Modulation Signal

DCbias
Transmitted
Light

P type
CW DFB
laser E field
N type
Active region

Figure 4.27 Typical structure of an EA-DFB transmitter.

modulator is essentially based on the fact that the reverse bias voltage creates an electric
field in the active region (the device operates like a reverse-biased p–n junction). When
no electric field is present the InGaAsP material is transparent, whereas when an electric
field is present the material absorption coefficient increases dramatically turning the
device opaque to incoming light, effectively acting as a modulator. Passive waveguides
connect the DFB laser region to the modulator and then to the output fiber.
We have seen that chirp is an issue in modulator performance. EA modulators have an
interesting feature allowing adjustment of chirp to zero or even negative values through
changes of the reverse bias voltage.
The linewidth enhancement factor α, defined in Section 4.5.3.1, for DMLs, can now
vary as a function of the reverse bias voltage. EA-DFBs exhibit only transient chirp
as defined in Equation (4.35). Figure 4.28 shows measured absorption and linewidth

20 2

15 1
Absorption (dB)

10 0
Alpha

5 −1

0 −2

−5 −3
0 0.5 1 1.5 2
Reverse bias voltage (V)

Figure 4.28 Absorption and chirp (linewidth enhancement factor) parameters versus reverse bias
voltage for a typical EA-DFB transmitter. (From [Roudas+01, Figure 1a]. Copyright
c 2001
IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)
 Enabling Technology 217

enhancement factor characteristics as a function of the reverse applied voltage for a

typical EA-DFB laser. A negative linewidth enhancement factor will provide better
performance characteristics over positive dispersion fiber, whereas an EA-DFB with
positive α will provide better performance characteristics over negative dispersion fiber,
as explained in Section 4.9.
The use of EA-DFB lasers in conjunction with negative residual dispersion is prefer-
able to their use with positive residual dispersion because higher output powers (typically
by 8 dB) are achieved. This is seen in Figure 4.28 where the laser would be biased for a
positive average α parameter; i.e., lower values of reverse bias voltage V. This produces
lower EA absorption values resulting in higher transmitter output powers. However, the
extinction ratio in this case will be worse because the absorption curve in that region is
rather flat.
In [Roudas+01], a model is presented that allows the calculation of system reach
using an EA-DFB.

4.6 Optical Receivers in Intensity-Modulated Direct-Detection Systems

The basic transmission link considered in this book involves an intensity-modulated

optical transmitter and a direct detection receiver as shown in Figure 4.1. In this section,
we focus on the receiver structure, consisting of the photodetector, a “front-end” amplifier
(possibly incorporating equalization), and a detection circuit (in the case of digital
transmission).
The photodetector converts the optical signal to a photocurrent, using direct detection,
and the amplifier raises the power of the photocurrent to a level sufficient for further
electronic processing. In the digital case, this processing is primarily clock recovery,
sampling, and threshold detection to extract the digital bit stream from the received
signal. We consider each of these components next.

4.6.1 Photodetectors
The photodetectors used in optical transmission systems are semiconductor photodiodes.
In the simplest device, the p−n junction, the operation is essentially the reverse of a
semiconductor optical amplifier. As shown in Figure 4.29, the junction is reverse biased,
enlarging its depletion layer, and in the absence of an optical signal only a small,
minority carrier current, called the dark current, flows. A photon impinging on the

− +

i(t )
Light p n

Figure 4.29 Photodiode.

218 Multiwavelength Optical Networks

surface of the device and entering the semiconductor can be absorbed by an electron in
the valence band, transferring the electron to the conduction band. Each electron/hole
pair produced this way in the depletion region contributes to the photocurrent i(t). For
this absorption to take place, the photon energy hν must be at least as great as the
bandgap energy of the material, E g . Thus, photon absorption, and hence optoelectronic
conversion, occurs only at wavelengths less than a cutoff wavelength, λc = hc/E g .
Typical materials with sufficiently large λc for optical fiber transmission systems are
various forms of InGaAs and InGaAsP. These are wideband devices, in which the
absorption and conversion process takes place over the entire spectrum of interest in
optical fiber communications.
In a photodiode, the relation between the incident optical power Pin and the photocur-
rent I p is given by I p = R Pin , where the responsivity, R (in amperes per watt), is in turn
given by

R = ηq/ hν. (4.36)

In Equation (4.36), η is the quantum efficiency of the device, which depends on the
fraction of incident photons absorbed. The absorption coefficient of the semiconductor
materials currently in use is high enough so that a very thin slab of material (on the order
of 10 µm) is sufficient to make η ≈ 1.

4.6.1.1 PIN Photodiodes

Electrons (holes) are also produced by photon absorption in the p (n) regions, but
the component of photocurrent due to these effects tends to distort the photodetector
response. This is because the minority carriers in the p and n regions diffuse slowly
into the depletion region, resulting in a delayed response compared with the dominant,
depletion region drift response. (They may also be lost by recombination before reaching
the depletion region.) The PIN photodiode (for p-type, intrinsic, n-type) solves these
problems by increasing the width of the depletion region, thereby causing a larger fraction
of the optical power to fall on the depletion region. With the PIN photodiode, an extra
layer of intrinsic material (either undoped or lightly doped) is sandwiched between the
p and n regions. Its effect is to improve the responsivity of the device by capturing most
of the light in the depletion region. The PIN photodiodes in a more complex double-
heterostructure geometry, which involves two extra layers of doped semiconductor, can
be designed with close-to-ideal properties: near 100% quantum efficiency, absence of
diffusion current, wide optical bandwidth, and large electrical bandwidths (in the tens
of gigahertz).

4.6.1.2 Avalanche Photodiodes

The photocurrent generated by the photodiodes discussed thus far is limited by the
fact that, at best, each photon produces one electron of photocurrent. This may not
be sufficient when the incident optical power is very low. The circuitry following the
photodiode, including the front-end amplifier, contains unavoidable sources of thermal
noise. If the level of the photocurrent is too low, the photocurrent signal is lost in
the thermal noise. This is where the avalanche photodiode (APD) comes in. The APD
 Enabling Technology 219

resembles a PIN photodiode with an extra gain layer inserted between the i and n layers.
Its purpose is to provide a current gain.
If the photoelectrons produced in a photodiode are accelerated to sufficient speeds,
they can produce additional electrons by colliding with the semiconductor material in
a chain reaction, multiplying the number of electrons produced by each photon. This
impact-ionization effect produces a current gain similar to that in a photomultiplier
vacuum tube. The APD produces this result by using relatively large applied voltages to
provide the necessary accelerating electric field (on the order of 105 V/cm) in the gain
region of the APD.
The current gain in the APD produces a much larger responsivity than in conventional
photodiodes and thus is an effective means of overcoming thermal noise limitations
in lower optical power systems. It does, however, introduce its own noise due to the
unavoidable random nature of the electron multiplication process.

4.6.2 Front-End Amplifier: Signal-to-Noise Ratio

Two types of amplifiers are typically used following the photodetector: a high-input
impedance voltage amplifier and a transimpedance amplifier, which is actually a current-
to-voltage converter. (The transimpedance amplifier is most frequently used in current
systems.) A circuit using the latter is shown in Figure 4.30(a). The photodiode is coupled
to the input of a high-gain inverting operational amplifier with gain A as shown. A small-
signal equivalent circuit model for this arrangement is shown in Figure 4.30(b). The
current source Isig represents the signal current, and R and C represent, respectively,
the resistance and capacitance seen looking into the amplifier. The resistance is the
parallel combination of Rbias and R a /(A + 1), where R a /(A + 1) is the equivalent input
impedance of the amplifier. The bandwidth of the circuit is B = 1/2π RC. Because
A ≫ 1, the resistance R can be made small, so that the bandwidth B can be made large
if desired.

+
Ra

−A

Rbias

(a)

ish id +
Isig R C υ
−
(b)

Figure 4.30 Transimpedance amplifier.

220 Multiwavelength Optical Networks

The additional current sources, i sh and i d in Figure 4.30(b), represent the shot noise
produced by the signal and the dark current, respectively. Using these sources, we find
that the SNR at the output of this circuit (assuming that amplifier noise is dominated by
the thermal noise in the equivalent resistance R) is given by
2
Isig
SNR = 2
. (4.37)
i sh + i d2 + i th2
In Equation (4.37), we have
2
Ish = 2q Isig B
Id2 = 2q Id B (4.38)
Ith2 = 4KTB/R,

where K and T are Boltzmann’s constant and absolute temperature, respectively. The
shot noise expression is given for a PIN photodiode. A modification, including an “excess
noise factor,” is necessary for APDs (for example, see [Agrawal97, p. 167]).
Because all of the noise terms here are proportional to the receiver bandwidth B, this is
normally chosen as small as possible, to exclude noise, but large enough to pass the signal
spectrum. Clearly, to improve the SNR it is desirable to make Isig as large as possible,
which can be done by using either an APD rather than a PIN photodiode or an optical
preamplifier before the photodetector. In both cases the SNR will be improved because
of the gain of the APD or the amplifier, until the noise generated in the amplifying device
(APD or optical amplifier), together with other noise and interference accumulated on
the transmission link, dominate the receiver noise. In the case of the APD, there is an
optimal value of the APD gain beyond which the SNR starts to deteriorate due to the
excess noise factor (see Problem 7 at the end of this chapter).
It is important to recall that the SNR discussed here is defined on the electrical
side of the e/o interface as opposed to the OSNR defined on the optical side, and the
assumption is that the signal current is due to an uncorrupted optical signal impressed
on the photodetector. The other noise and interference effects accumulated on a fiber
link have not been included in Equation (4.37). They include accumulated ASE noise
from amplifiers along the transmission link; waveform distortion due to fiber dispersion;
distortion and cross-talk effects due to fiber and amplifier nonlinearities; and various
fiber, filter, and switch cross-talk effects. These issues appear in subsequent sections,
leading up to a complete treatment of end-to-end optical connections in a network setting
in Section 4.13.
The shot noise terms in the previous SNR expressions are based on a Gaussian
approximation of the shot noise probability density. This is an approximation that is
valid (in digital systems) for optical power levels corresponding to a large number of
photons transmitted in each bit interval. As the bit rate is increased and/or the optical
power level is decreased, we eventually arrive at a point, called the quantum limit,
where only a few photons are transmitted in each bit interval. In this case, the Gaussian
approximation is no longer valid, and the signal detection process must be viewed as a
photon-counting process.
 Enabling Technology 221

Suppose a digital signal is intensity modulated by on-off keying (OOK), at a bit rate
R b , with a 1 represented by an optical power 2P, for an average power of P, assuming
equiprobable 1s and 0s. This gives an average of N p = 2P/ hν R b photons for each 1
transmitted, and no photons when a 0 is transmitted, assuming no other sources of photons
(noise). The optimal detector in this case would declare that a 0 was transmitted if no
photons are counted in the bit interval and declare that a 1 was transmitted otherwise.
In this idealized model, there will never be a case when a 0 is mistaken for a 1. The
opposite can happen, however. Because the counting statistics are Poisson, there is a
finite probability, P(0|1) = e−N p , that a 0 is declared when a 1 was transmitted. The
BER for the system in the quantum limit is, therefore, BER = 21 e−N p . For example, for
a BER of 10−9 we need approximately 20 photons for each symbol 1 or an average of
10 photons per bit. This results in approximately −60 dBm for a 1-Gbps signal in the
1550-nm range. Powers in practical systems are typically well above −40 dBm, two
orders of magnitude higher than the quantum limit. The approach to the quantum limit
can also be interpreted in terms of its effect on information rates and channel capacity
rather than BER. For an information-theoretic treatment of the quantum limit, see
[Stern60].

4.6.3 Digital Signal Detection: Noise, Interference, and Bit Error Rate
In a digital transmission link, the final step in the receiver is signal detection. We restrict
ourselves here to the simplest (and most common) type of system: an IM/DD binary
communication system using an OOK modulation format. As shown in Figure 4.31,
the photocurrent is amplified and filtered, and the filter transfer function is chosen to
preserve or enhance the signal waveform while suppressing noise. It typically has a
bandwidth, B (in hertz), somewhat less than the signal bit rate, R b (in bits per second).
A common choice is B = 0.65R b . The amplified and filtered signal is then sampled and
threshold detected. Typical waveforms illustrating this process are shown in Figure 4.32.
The data signal that intensity modulates the laser transmitter is assumed to be an ideal
train of rectangular pulses:
∞

s(t) = al u(t − lT ), (4.39)
l=−∞

Preamplifier PD
Optical y(t)
Amplifier/ Threshold Detected
Bandpass
Filter Detector Data
Filter a^l

Sample

Clock
Recovery

Figure 4.31 Binary receiver.

222 Multiwavelength Optical Networks

s(t)

1 0 1 1

t
Bit period
(a)

y(t) y1 y3

D
y4
y2
t
(b)

a^l

1 0 1 0 Error

t
(c)

Figure 4.32 Typical waveforms in an IM/DD system.

where

1 −T /2 < t ≤ T /2
u(t) = (4.40)
0 otherwise
and {al } represents a sequence of data symbols (1s and 0s) with bit period T .
After traversing the transmission link, the receiver front end, and the filter, the trans-
mitted data signal s(t) [see Figure 4.32(a)] appears at the input of the sampling/threshold
detection unit as a distorted and noisy waveform, y(t) – see Figure 4.32(b). The clock
recovery unit, shown in Figure 4.31, processes y(t) to determine at what points in time
the signal should be sampled.18
Using this information, the received signal y(t) is sampled periodically at a point
t0 within the bit period, chosen to provide a high probability of correct detection. The
signal samples yl are compared with a decision threshold D, shown as a dashed line in
Figure 4.32(b), and a decision is made on the detected data symbol: âl = 1 if yl > D

18
The clock recovery unit is essentially a variable frequency oscillator, phase locked to the data signal (i.e.,
a phase-locked loop). For it to work properly, there must be frequent transitions between 1s and 0s in the
data signal, so that the circuit is able to discern the bit interval boundaries. This is generally ensured by
“scrambling” the transmitted data stream to convert long strings of repeated symbols into interspersed 1s
and 0s.
 Enabling Technology 223

y1

y0

t
0 Bit Period 2T
t0
T
(a)

t
0 T 2T
(b)

Figure 4.33 Eye diagram.

and 0 otherwise. Figure 4.32(c) shows the detected data symbols âl corresponding to the
transmitted data stream in Figure 4.32(a). Note the presence of a bit error in the fourth
symbol.
The received waveform in Figure 4.32(b) is often represented in terms of an eye
diagram, as shown in Figure 4.33. The eye diagram superimposes time traces corre-
sponding to successive bit periods. This is what is seen on an oscilloscope synchronized
to a submultiple of the bit rate of the transmitted signal. Figure 4.33(a) shows what
a typical eye diagram would look like if all signal impairments were deterministic in
nature. Each trace represents a distorted bit waveform (representing a 1 or a 0) over
two bit periods, where the distortion is a result of dispersion, limited bandwidth, and
deterministic nonlinearities, as well as the effects of square-law detection. All of these
effects produce ISI, which is deterministically (but nonlinearly) related to the bit pattern.
The ISI comes from the interfering effects of neighboring pulses that overlap in time
because of dispersion.
Eye diagrams are used frequently for test purposes, during which the bit stream is
generated as a periodic pseudorandom test pattern with a long but finite period. In this
case, individual traces contributing to the eye diagram repeat during each period of the
test pattern. The cross within the “eye opening” in Figure 4.33(a) indicates a good point
for sampling and setting the threshold. The traces below the cross are all due to 0s, and
those above are due to 1s, which corresponds to an “open eye.” The quantity y1 − y0
224 Multiwavelength Optical Networks

indicates the size of the eye opening. The vertex of the cross is positioned at the chosen
sampling time, t0 , and threshold setting, D. If sampling and thresholding are executed
as shown, the BER will be zero. Notice, however, that as the eye opening (y1 − y0 )
decreases, the allowable range of sampling and threshold settings diminishes. For small
eye openings, a small amount of timing jitter in the transmitter or the receiver or a slight
misplacement of the threshold can produce a catastrophic BER even in the absence of
any random noise.
Figure 4.33(b) shows the open eye with random noise added. Now, instead of a
well-defined deterministic pattern, we have a “fuzzy” picture because the noise is not
repetitive. Furthermore, the fuzziness produces occasional excursions of the waveform
across the threshold, with corresponding bit errors. The more the eye closure, the more
frequently an error occurs. Also, the wider the band of “fuzziness” (corresponding
roughly to the random noise power), the higher the BER.
An exact determination of BER requires a detailed model of the processes along the
transmission path. The relation between the waveform y(t) and the original waveform
s(t) is extremely complex due to a succession of linear and nonlinear operations on the
signal. The original data signal intensity modulates a laser that, if modulated directly,
produces an optical field containing frequency modulation (chirp) as well as the desired
intensity modulation. The optical field also includes intensity and phase noise, which are
responsible for the finite linewidth of the unmodulated laser. The signal then experiences
linear dispersion combined with the various nonlinear distortion and cross-talk effects,
discussed in Section 4.3.2.4, and these effects accumulate with the length of the fiber.
The signal also accumulates ASE noise from line amplifiers and cross-talk from any
optical cross-connects it traverses before reaching its destination. The resultant optical
field presented to the photodetector undergoes another nonlinear operation (square-law
detection) and is further contaminated by shot noise and amplifier noise before reaching
the sampler and detector.
Imperfections in transmitter clocking and receiver clock recovery also introduce
timing jitter, which degrades the detection process, as indicated earlier. It is clearly
impossible to do an exact analysis of these effects. Fortunately, in any properly designed
system all of these transmission and reception impairments must be small. For if they
are not, the system will not be operating at the low BER required for digital optical
transmission systems. This allows us to make many simplifying approximations and still
come up with some useful results on BER.
We first reduce the mathematical model for the detection process to a pair of functions,
p(y|0) and p(y|1), representing the probability density functions (PDFs) for the sampled
amplitude at the input of the decision unit, conditioned on the transmission of a 0 or a 1 in
the corresponding bit interval.19 An illustration of typical PDFs appears in Figure 4.34,
which shows an ideal detection process.
The basic assumption in this model is that the PDFs are conditioned only on one
transmitted symbol, so the deterministic ISI effects represented by the eye diagram of

19
The reception bit interval is delayed in time with respect to the transmission bit interval due to propagation
delays.
 Enabling Technology 225

p(y | 0)
p(y | 1)

Figure 4.34 Ideal detection.

Figure 4.33(a) can be ignored, as well as any statistical dependencies of the random
noise effects on other symbols (or other signals).20
Assuming that 1s and 0s are equiprobable, it turns out that the setting of the decision
threshold D that minimizes the probability of error (BER) is at the crossover of the two
PDFs, as shown in Figure 4.34. This corresponds to maximum likelihood (ML) detection
(see, for example, [Lee+94]). The ML detector declares that a 0 was transmitted when
p(y|0) > p(y|1) and that a 1 was transmitted when p(y|1) > p(y|0). The total probability
of error in this case corresponds to the dark area in Figure 4.34, where the probability
of declaring a 1 when a 0 was sent (or the reverse) is shown as the dark area to the right
(or the left) of D, respectively.
The picture becomes still simpler if we approximate the PDFs as Gaussian, so
(y − yi )2


1
p(y|i) = √ exp − i = 0, 1, (4.41)
2π σi 2σi2
where yi is the mean value of y when symbol i is transmitted and σi2 is the noise variance.
It can be shown that the probability of error Pe using the optimal threshold setting is
then given by
 ∞
Q2
 
1 Q 1 2 1
Pe = erfc √ =√ exp(−α ) dα ≈ √ exp − , (4.42)
2 2 π Q/√2 Q 2π 2
where
y1 − y0
Q= (4.43)
σ0 + σ1
and erfc is the complementary error function. Plots of the exact expression for Pe as a
function of Q, and its exponential approximation, are shown in Figure 4.35. Note that
the exponential approximation is excellent for small values of Pe , which is the range of
20
Theoretically, it is possible to include ISI effects in the detection model by extending the conditioning
over the set of all neighboring symbols in the bit sequence that contribute to ISI. However, this requires a
detailed model for all of the transmission impairments producing the ISI, which is, in most cases, out of
the question.
226 Multiwavelength Optical Networks

Exponential
1 Approximation

.1
Pe
.01

10−3

10−4

10−5

10−6

Q
1 2 3 4 5

Figure 4.35 BER as a function of Q.

interest to us here. The square of the argument Q in Equation (4.43) can be interpreted
as an electrical SNR, as seen by the decision unit. Figure 4.35 illustrates the fact that
the BER is a very sensitive function of Q so that very small improvements in SNR have
very substantial effects on BER, with all of this being due to the fact that the noise is
(approximately) Gaussian. For example, defining

Q 2dB = 20 log10 Q (4.44)

we find that Q 2dB = 14.3 dB for a BER of 10−7 , and Q 2dB = 15.4 dB for a BER of 10−9 .
For the remainder of this chapter (unless otherwise noted), we will use Q 2dB as defined
in Equation (4.44) whenever we refer to Q performance in a system or Q penalty for
an impairment. It should be noted, however, that in the literature, including articles
referred to in this chapter, the quantity Q dB = 10 log10 Q is often used when referring to
Q penalty.
The Gaussian approximation is usually very good when there are many independent
sources of random noise (although this is not always the case). Deterministic effects
contributing to ISI can be lumped into the values of yi . For example, a worst-case BER
can be obtained by basing the argument y1 − y0 in Equation (4.43) on an eye diagram,
obtained either by simulation or measurement. The eye opening y1 − y0 in Figure 4.33(a)
corresponds to the worst case of ISI and thus provides an upper bound on BER for any
assumed noise variances.
The term receiver sensitivity is often used as a measure of receiver quality. We define
it here to mean the average optical power required at the input to the receiver to achieve
a given BER at a given bit rate R b . Typical receiver sensitivities for a BER of 10−9
and R b = 1 Gbps are −26, −36, and −45 dBm for PIN, APD, and optical preamplifier
receivers, respectively. The best of these (−45 dBm) corresponds to more than 1000
photons per bit – two orders of magnitude higher than the quantum limit.
 Enabling Technology 227

4.6.4 Analog Systems: Carrier-to-Noise Ratio

There are several applications in which the laser transmitter is intensity modulated by an
analog signal, which may be carrying any kind of information in either analog or digital
form. The following are some examples:
r Cable television (CATV) distribution networks
r Distribution networks for microwave wireless systems
r Optical networks using microwave subcarriers (see Section 5.2.2)

In analog systems, there is no decision mechanism, and the receiver performance is

measured in terms of carrier-to-noise ratio (CNR). The CNR is defined as
2
Isig
CNR = , (4.45)
σn2

where Isig represents the signal component of the photocurrent and σn2 is the total noise
at the receiver. The latter is given by

σn2 = σRIN
2
+ σsh2 + σs−ASE
2 2
+ σASE−ASE + σth2 , (4.46)

where the terms on the right side of Equation (4.46) are the laser relative intensity noise
(RIN),21 the shot noise, the two noise components caused by signal–ASE and ASE–ASE
beating in the photodetector, and the receiver’s thermal noise, respectively. A quantitative
treatment of these effects can be found in [Desurvire94].
The CNR for any particular system depends on a host of system parameters, which is
beyond the scope of the current discussion. However, one consideration worthy of note
is that when the information-bearing signal is in analog form (e.g., AM TV), it is often
much more vulnerable to noise than when it is in digital form. For example, AM CATV
systems require a 50- to 55-dB CNR, whereas the corresponding quantity, Q 2dB , for an
acceptable digital system is approximately 25 dB.
An example of the use of dispersion compensation in a high performance application
of analog DWDM appears in [Lu+04]. The objective was to demonstrate bidirectional
CATV transmission over 50 km of standard SMF in the face of combined dispersion
and nonlinear fiber impairments. At DWDM channel spacings, the main nonlinear
effect influencing transmission fidelity is XPM, producing composite second order
(CSO) and composite triple beat (CTB) interference. Chirped fiber-grating dispersion
compensators were used to reduce dispersion while also mitigating the nonlinear effects.
In the demonstration, four λ-channels were transmitted (two in each direction) at 200-
GHz spacings, with each wavelength carrying 77 channels on RF subcarriers. A CNR
better than 50 dB was obtained with CSO and CTB performance more than 72 and 69 dB,
respectively.

21
The RIN is caused by random fluctuations in the optical field generated by the laser transmitter.
228 Multiwavelength Optical Networks

4.7 The End-to-End Transmission Channel

Referring to the layered view of the network architecture in Figure 2.1, we have so
far in this chapter confined ourselves to the optical connection layer, which terminates
at the e/o interface; i.e., at the transmitter and receiver. However, there is additional
processing that usually occurs on the electrical side to turn an optical connection into a
transmission channel. Figure 4.36, which expands schematically on Figure 4.1, shows
some basic operations that occur in the electrical domain in transforming a bit stream
on a transmission channel to an optical signal and then operating on the received optical
signal to transform it back to a bit stream at the other end of the transmission channel.
These operations complete the picture of the physical layer in Figure 2.1.
Data are presented to the input of the transmission channel in the form of a stream of
bits di as shown in Figure 4.36. For purposes of FEC, redundant bits may be added in an
encoder, which outputs code bits c j as shown in Figure 4.38. These are then processed
electrically to produce a modulating signal adapted to a given modulation format (e.g.,
OOK), and this signal drives the optical modulator and laser transmitter. On the receiving
side, an optical receiver, exemplified by the block diagram of Figure 4.31, converts an
optical signal received from a λ-channel to an electrical waveform, represented by y(t)
in Figure 4.36. Again, depending on the modulation format, further processing will take
place before y(t) is converted into an output bit stream. In the simplest case of OOK,

TP TP OT
OT

eo
Modulating
Input data signal Opti
Optical
Code bits Electrical modulator/
Encoder Modulator/
di modulator
transmitter
transmitter

Optical
Op
channel

Optically
detected
Estimated waveform
Corrected data data
y (t) ti
Optical
Electrical
Decoder detector/
tector/
processor
receiver

RP
RP OR
OR

Figure 4.36 Transmission channel processing operations.

 Enabling Technology 229

1 0 1 1 0 0 1

(a) NRZ

(b) RZ

(c) DPSK
0
0 ?π 000 0 π
?? π 0
t

(d) Duobinary

Figure 4.37 Modulation formats.

clock recovery and thresholding is all that is necessary in the electrical processor to
produce a detected data stream, e j , which is the estimated version of the stream c j , and
may contain bit errors. However, if more elaborate modulation formats are used and/or
if equalization is applied to the waveform y(t), the electrical processor is more complex.
Finally, the decoder processes the detected data stream to convert it to the final output
stream f i , where some or all errors in the estimated data have been corrected. We now
describe some representative modulation formats and coding schemes, concluding with
a brief treatment of electronic equalization.

4.7.1 Modulation Formats

Various modulation formats have been used in WDM communication systems. They
have two key features: the way the data are represented in the optical signal, and the
keying technique. For example, in intensity modulated systems the data are represented
by the intensity of the optical signal and in PSK by the phase. The keying alternatives
are known as non-return-to-zero (NRZ) or return-to-zero (RZ). In the former case, the
information-bearing signal occupies a full bit period, whereas in the latter case the
information is contained in a pulse occupying less than the full bit period. Each can
incorporate various data representations. For example, OOK is an intensity modulation
representation that involves turning the transmitter power on for a 1 and off for a 0.22 It
can be implemented either as NRZ (Figure 4.37[a]) or RZ (Figure 4.37[b]). Comparing
the two, RZ occupies more bandwidth than NRZ because the pulses are narrower, and
22
Amplitude-shift keying (ASK) generalizes OOK to multiple levels of intensity.
230 Multiwavelength Optical Networks

RZ has a higher peak power than an NRZ signal running at the same average power
level.
In DWDM systems, the interplay of dispersion and nonlinear effects may or may not
favor RZ over NRZ, depending on bit rates [Hayee+99]. At moderate speeds
(10–20 Gbps) in dispersion-managed systems with alternating positive and negative
dispersion fibers, RZ tends to perform better than NRZ, because the latter is more
affected by nonlinearities. At higher speeds (40 Gbps) NRZ tends to perform better,
because the larger bandwidth occupied by the RZ pulse makes it more vulnerable to dis-
persion. However, the situation is more complicated than this. If the modulation depth is
high, then the spectrum is narrowed, and under those conditions RZ outperforms NRZ at
high speeds at transoceanic distances. Also, as we have seen, solitons (Section 4.3.3) are
RZ pulses (typically picoseconds in width) that preserve their shape due to the positive
effects of a combination of dispersion and nonlinearity.
So far in discussing optical transmission we have focused on intensity modulated
systems because of their simplicity and because they dominate in currently deployed
networks. However, NRZ-differential phase-shift keying (NRZ-DPSK) and RZ-DPSK
are alternative modulation formats that offer certain advantages compared to the tradi-
tional OOK formats. In conventional PSK, the phase of the carrier is modulated at the
transmitter and at the receiver phase-to-intensity conversion is needed before the signal
reaches the photodetector, because the photodetection process is phase insensitive. This
can be done in a coherent optical system, as explained in Section 4.8. For differential
PSK (DPSK) the data is represented by a phase change between successive bit periods.
For example, a π -phase change might represent a 0, with no change for a 1, thus pro-
viding the differential encoding feature (see Figure 4.37[c]). At the receiver, the phase
change between successive bits is measured using a delay interferometer (DI) which
converts phase changes into intensity levels. In a Mach–Zehnder DI, the lengths of its
arms differ by one bit period so that successive encoded bits are superimposed at its
output, with the earlier bit serving as a phase reference for the later one. There will be
either constructive or destructive interference at the output depending on whether a phase
change has occurred between adjacent bits. In this way the phase changes are converted
into intensity changes. Because the DI is actually a linear filter it can be considered
to be lumped together with the optical bandpass filter in the receiver block diagram of
Figure 4.31.23 To achieve maximum receiver sensitivity, a balanced receiver consisting
of a two-output MZ feeding two photodiodes is required. With this configuration, the
NRZ-DPSK has a 3-dB receiver sensitivity advantage over NRZ-OOK. DPSK has also
been shown to be more tolerant of fiber nonlinearity penalties due to its more evenly
distributed power. RZ-DPSK can offer an additional advantage compared to NRZ-DPSK
because of the inherent receiver sensitivity improvement of RZ formats, as mentioned
above. A comprehensive review of DPSK theory, implementation, and experimental
results can be found in [Gnauk+02].

23
Frequency-shift keying (FSK), where the bits are represented by optical frequency deviations, is an-
other modulation format wherein the receiver requires optical filtering in front of the photodetector. See
[Kaminow+87] for an early use of FSK in optical transmission.
 Enabling Technology 231

A number of other RZ modulation formats have been proposed such as chirped

RZ (CRZ) and carrier suppressed RZ (CSRZ). The former is generated by using bit-
synchronous phase modulation of the RZ signal, where the phase deviation is varied
over the bit period to provide optimum performance for the particular nonlinearities
and dispersion map of the system. The latter format (CSRZ) is generated from RZ by
suppressing the carrier and has a narrower spectrum, which makes it more tolerant to
dispersion effects. Other popular modulation formats are duobinary RZ (D-RZ) and
modified D-RZ (MD-RZ) [Cheng+02]. Duobinary is a form of intensity modulation
where the transmitted pulse intensity is based on the sum of two adjacent bits (see
Figure 4.37[d]). Therefore, the output can take one of three values. In duobinary the
intensity of each pulse is correlated with its neighbors (because sums of adjacent bits
are used), and this narrows its spectral width. For this reason, these formats were shown
to exhibit dispersion tolerance as described in more detail in [KaminowChap16+02].
An interesting approach to duobinary transmission [Chowdhury+06] combines op-
tical carrier suppression and separation (OCSS) with optical duobinary modulation to
obtain high channel bit rates (40 Gbps), high-density DWDM transmission (50-GHz
channel spacing), and therefore very high spectral efficiency (0.8 bps/Hz), over a dis-
tance of 200 km. The objective was to demonstrate high-speed transmission using lower
speed, low-cost electronic and optical technology for applications from the core network
down to metro and access networks. Carrier suppression is accomplished using a dual-
arm lithium niobate Mach-Zehnder modulator biased at its minimum intensity output
point and driven in its two arms by complementary 20-GHz CW signals. The duobinary
data then intensity modulate the resultant suppressed carrier optical signal.24
This is a good example of how optical technology has matured. Although suppressed
carrier techniques were commonly used for more efficient transmission since the early
days of radio, it is only recently that the approach has been used in the optical domain.
Furthermore, at the high bit rates used in this example, optical processing has been found
to be a more cost-effective approach than other techniques for achieving high-spectral-
efficiency DWDM transmission over long distances.25
It should be clear from the above that each modulation format has its advantages and
disadvantages. There are always trade-offs: using a complex modulation format to adapt
to particular conditions on an optical path can increase maximum achievable transmis-
sion distance and bit rates but will most likely raise the cost of the transmitter/receiver
equipment.

4.7.2 Forward Error Correction

We now focus on the error control part of the system of Figure 4.36, condensing the
other processing operations into the modulator and demodulator shown in Figure 4.38.
In a communications system that employs block error-correction coding, a digital
24
The dual-arm LiNbO3 modulator configured for OCSS in an application to optical label switching is shown
in Figure 10.47.
25
The spectral efficiency achieved in this demonstration is comparable to that which is common for electronic
transmission over copper but rare for optical fiber transmission.
232 Multiwavelength Optical Networks

Digital d1, d2, d3, ... , dk c1, c2, ... , cn

Information (n,k) Encoder Modulator
Source Data bits Code bits

Corruption
Channel

Digital f1, f2, f3, ... , fk e1, e2, ... , en

Information (n,k) Decoder Demodulator
Sink Data bits Estimates Code bits Estimates

Figure 4.38 FEC encoding/decoding functions.

information source sends a data sequence comprising k bits of data to an encoder.

The encoder inserts redundant (or parity check) bits, thereby outputting a longer se-
quence of n code bits called a codeword. On the receiving end, codewords are used by a
suitable decoder to extract the original data.
A code of the type just described is called an (n, k) binary block code, and its coding
efficiency is k/n. A systematic binary linear block code is one whose parity check bits
are calculated as a linear function (modulo 2) of its data bits. Error correcting codes have
a long history and have been refined to adapt to a wide variety of channels. A commonly
used family of codes are the Reed–Solomon (RS) codes. An RS code is specified as
RS(n, k) with s-bit symbols. This means that the encoder takes k data symbols of s bits
each and adds parity symbols to make an n symbol codeword. There are n − k parity
symbols of s bits each. An RS(n, k) code can correct any pattern of up to t symbols
containing any number of errors where 2t = n − k.
By correcting the most probable patterns of random errors, FEC provides a coding
gain (or system Q margin). A coding gain of QdB at a given BER means that the Q
factor required to achieve the specified BER with FEC is QdB less than that required
without FEC. A downside is that the channel bit rate must be increased to accommodate
the redundant bits, and another is that the codec is complex and costly at high bit rates.
However, there are many applications where FEC is worthwhile. As used in WDM long-
haul transmission and undersea transmission, FEC offers the system designer significant
performance margins. For example, a commonly used code is the RS(255, 239), which
provides a coding gain of about 6 dB at BER 10−15 at the expense of increasing the
channel bit rate by 7%. For 40-Gbps transmissions (OC-768), FEC is especially important
because of the high OSNR required at the receiver and the increased impairment penalties
at higher line rates. However, the only way to reach these speeds with current technology
is to demultiplex to several lower speed bit streams and implement FEC with parallel
processing; a costly solution. On the other hand, advances in FEC technology have made
it a cost-effective possibility in metropolitan WDM networks at lower speeds. A more
detailed analysis of error control techniques and their applications in optical transmission
is presented in [KaminowChap17+02].
 Enabling Technology 233

4.7.3 Equalization
The compensation at the receiver, of channel-induced distortion of the optical signal is
called equalization. Electronic equalizers are designed to process the optically detected
waveform y(t) in Figure 4.36 to cancel or mitigate intersymbol interference at the
sampling instant. Because the primary cause of ISI is dispersion, these devices are also
called electronic dispersion compensators (EDCs).
The most common type of electronic equalizer forms a weighted sum of the present and
past sampled values (taps) of the waveform y(t) taken at a sampling instant determined
by a locally generated clock, and this weighted sum is applied to a threshold device to
make each bit decision (see Figure 4.39). The weights wi are calculated and updated
based on previous bit decisions, making this an adaptive decision-directed equalizer. The
number of taps used in the equalizer should be enough to cover the time span over which
interfering pulses overlap. This is the well-known transversal filter equalizer, which has
a long history, having been used in telephone modems in the 1960s. However, the analog
telephone channel for which this was originally designed was, to a good approximation, a
linear additive noise channel, whereas the optical channel is nonlinear due to the square-
law action of the photodetector.26 Nevertheless, experimental results show that electronic
equalization still works in direct detection optical channels. One possible reason is the
robust nature of the weight-setting algorithm. Basically, it just tries to do the best it can
to open the eye of the receiver no matter what impairments are encountered.
A number of studies have focused on equalization technology and its applications
in the cost-sensitive metro WDM networking environment. In [Feuer+03], it was ex-
perimentally demonstrated that a four tap prototype EDC module roughly doubles the
dispersion-limited transmission reach for a channel with a DML source running at

y(t) yk yk−1 yk−2 yk−3

Sampler
Sampler T T T

w1 w22 w33 w4

+
Threshold
Th Estimated data
detector

Weight
We
update

Figure 4.39 Transversal decision-directed equalizer.

26
We assume a direct detection receiver here. If coherent detection were used and fiber nonlinearities were
negligible, the channel would be linear. Also, in a direct detection receiver, if equalization was done in the
optical domain, as a filtering operation in front of the detector, the channel seen by the equalizer would be
linear, but this would be much more difficult to implement.
234 Multiwavelength Optical Networks

10 Gbps over standard single-mode fiber. (As noted in Section 4.5.3.1, DMLs operating
at OC-192 are an emerging low-cost technology but have severe limits in reach due
to high chirp.) Transmission improvement was also demonstrated in [Woodward+03],
where EDCs were used in a WDM ring with EA-modulated transmitters. System margins
were significantly improved and error floors eliminated. Finally, electronic equalization
can be used in conjunction with many other techniques, such as advanced modulation
formats, forward error correction, and adaptive optical PMD compensators, to increase
the reach of a system or network. An example of the power of combined equalization
and FEC appears in [Winzer+05]. The objective here was to increase data rates in a
bidirectional CWDM system from 2.5 to 10 Gbps. Successful operation was achieved
on up to 80 km of standard SMF using DMLs rated for 2.5 Gbps.
Electronic equalizers can be produced in integrated form so they are potentially
low cost. Also, because of their robust performance characteristics, they can be used
for the simultaneous compensation of a variety of impairments: chromatic dispersion,
filter-induced distortion, polarization effects, and certain nonlinear effects. This may
allow the system designer to relax requirements on other devices, employing lower cost
transmitters, WADMs, and optical multiplexers/demultiplexers without compromising
performance.

4.8 Coherent Optical Systems

The basic structure of the coherent (heterodyne) receiver is shown in Figure 2.33, and
repeated in Figure 4.40 for convenience. The signal and local oscillator optical fields are

PD
ε (t)
iph(t) v (t)
d νLO Coupler BPF ED
c

LO

(a)

Si (f )

HBPF

f
0 fI ν2 − νLO

Sv (f )

f
Spectra
(b)

Figure 4.40 Heterodyne receiver.

 Enabling Technology 235

combined in the coupler, and the resultant field is converted to a photocurrent, i ph (t), in
the photodetector – see Figure 4.40(a). The photocurrent is then filtered in the bandpass
filter (BPF), of bandwidth Be , and is detected electrically in the electrical detector (ED).
With signal and local oscillator powers of Ps and PLO , respectively, the photocurrent is
given by
  
i ph (t) = R Ps + PLO + 2 Ps PLO cos (2π f I t + ϕs − ϕl ) , (4.47)

where R, f I , ϕs , and ϕl are the photodetector responsivity, the intermediate frequency,

the signal phase, and the local oscillator phase, respectively. The effect of the bandpass
filter and the electrical detector is to filter out any extraneous signals and noise and to
perform square-law detection of the remaining signal, giving a resultant signal power
2
of Isig = 2R 2 Ps PLO . The power spectral densities of i ph and the output voltage, v, are
shown in Figure 4.40(b). The unique property of the coherent receiver is that the signal
power at the output of the detector is proportional to the local oscillator optical power
PLO , whereas the thermal receiver noise is independent of PLO . The total noise power
2
accompanying Isig is a sum of shot noise and thermal receiver noise, where the shot
2
noise power σsh = 2R Be PLO is also proportional to PLO . This means that by making the
local oscillator power sufficiently large, the signal and shot noise terms can be made to
dominate the thermal noise. In this shot noise limit, the electrical SNR is given by
R Ps
S N Re = . (4.48)
q Be

With this SNR at the input to a decision circuit, the coherent receiver performs
considerably better than a typical direct-detection receiver. However, this is an ideal
model that neglects possible polarization misalignment and laser phase noise in both the
signal and the local oscillator waves.

4.9 Performance Impairments in a Network Environment

In studying the origin of performance impairments in this chapter, we focus primarily

on the various optical network devices and subsystems in isolation. The same can be
said for the various performance impairments. However, an optical signal accumulates
noise, distortion, and interference originating from many sources before it reaches its
destination, and their cumulative effects cannot always be deduced from looking at them
one by one. In this section, we examine the performance impact of several cumulative
or combined effects in optical channels, including cross-talk, signal power divergence,
interaction of chirp and dispersion, and PMD. As will be seen, the whole picture is more
than the sum of its parts.

4.9.1 Cross-Talk
The discussion of receiver performance thus far has ignored an important, perhaps the
most important, source of performance degradation: interference from unwanted signals
236 Multiwavelength Optical Networks

(i.e., cross-talk). The effects of cross-talk on receiver performance are considered next.
We focus on the digital transmission case. 
Suppose a signal s with optical field Es (t) = Re[ 2Is (t)e j2πνs t+φs ] is received at
a direct-detection receiver equipped with a wideband photodetector followed by an
electrical low-pass filter of bandwidth Be , chosen according to the bit rate (band-
 of s. Accompanying the signal is an interferer x with an optical field Ex (t) =
width)
Re[ 2Ix (t)e j2πνx t+φx ]. The photocurrent at the output of the photodetector can then be
written

Is Ix e j2π(νs −νx )t+φs −φx .
 
i(t) = R Is (t) + Ix (t) + 2Re (4.49)
In Equation
 (4.49), Is and Ix represent the signal and interferer intensities, respec-
tively, and Is Ix represents a signal-interference “beat” term that appears because
of the square-law detection process.27 There are three cases of interest, as shown in
Figure 4.41(a), depending on the value of |νs − νx |:
r ν = ν : interchannel cross-talk
s x
r ν ≈ ν : cochannel heterodyne cross-talk
s x
r ν = ν : cochannel homodyne (multipath) cross-talk
s x

In the first case, the nominal optical frequencies are different; in the second case, the
nominal frequencies are the same but the absolute frequencies are slightly different; and
in the third case, both frequencies are exactly the same (which only occurs if the two
signals come from the same source).
These are illustrated in Figures 4.41(b) and 4.41(c). In Figure 4.41(b), two optical
connections (S, S∗ ) and (X, X∗ ) are set up with a small amount of signal x “leaking” over
to receiver S∗ , as indicated by the dashed line. (This might be a result of switch cross-talk
in optical node A.) The categorization of the interference in this case is based on the
magnitude of the optical frequency difference. If the two connections are on different
assigned wavelengths λs = λx , then |νs − νx | > Be ; that is, the beat interference term in
Equation (4.49) lies outside the bandwidth of the receiver filter. In this case, it is only
the interferer intensity Ix that affects signal detection and BER. This is interchannel
cross-talk.28
If the two sources are transmitting on the same nominal wavelength (i.e., λs ≈ λx ,
and |νs − νx | < Be ), then the beat interference term in Equation (4.49) lies within the
bandwidth of the receiver filter. Now, both the interferer intensity and the beat frequency
term affect signal detection, with the beat frequency term constituting cochannel het-
erodyne cross-talk.29 Last, Figure 4.41(c) shows a case in which a small portion, s ′ , of
the signal s leaks through a parallel path (dashed line), rejoining the desired signal at
the receiver. As explained in Section 3.4.1, this violates the distinct source combining
27
This is the same effect that appears in a heterodyne receiver. As in that case, the magnitude of the beat term
depends on the polarization alignment of the two contributing fields. We have chosen the “worst case” of
perfect alignment.
28
The interference at a receiver caused by interchannel cross-talk can be eliminated by placing a narrowband
optical filter in front of the photodetector.
29
Cross-talk terminology varies considerably. Our cochannel heterodyne cross-talk is sometimes called
coherent common channel cross-talk.
 Enabling Technology 237

Cross-talk

Interchannel Cochannel
νs ≠ νx νs ≈ νx

Homodyne Heterodyne
νs = νx |νs − νx | < Be
(a)

x*

s
s A B s*
x

(b)

s
s A B s*
s′

(c)

Figure 4.41 Types of cross-talk.

condition. Because s ′ arrives with a delay and phase shift, it acts as an interferer. We call
this common-source interference cochannel homodyne cross-talk, multipath cross-talk,
or interferometric cross-talk [Gimlett+89].30 This case is the most difficult to evaluate
because the sign and magnitude of the cross-talk term depends on the relative phases
and polarizations of the two signals.
Referring to Equation (4.49), the cross-product term at the output of the filter is
√
Is Is ′ cos [φs (t) − φs ′ (t)], where an argument t has been included in the phases to
reflect the fact that s is not a constant-frequency optical signal but contains fluctuations
in phase (and amplitude) due to laser linewidth as well as the modulation impressed
on the laser. If the signals are exactly in phase at some point in time, the effect of the
beat term is to magnify the resultant signal (because the optical fields add, in this case),
whereas the opposite is true if they are out of phase. When the time delay between s
and s ′ is small (smaller than the signal’s coherence time), then the phase fluctuations are

30
Other sources of homodyne cross-talk are multiple reflections within a fiber, causing delayed replicas of
the transmitted signal to be superimposed on it.
238 Multiwavelength Optical Networks

correlated.31 For longer delays, they are uncorrelated. In the correlated case, the interferer
acts to produce a multipath fading effect, similar to that in radio transmission systems
in which the interference varies slowly from constructive to destructive, depending
on the instantaneous phase relations. In the uncorrelated case, the phase fluctuations
due to laser linewidth, as well as intensity fluctuations due to modulation, produce a
rapidly fluctuating noiselike cross-talk term. In predicting performance when significant
homodyne cross-talk is present, the best one can do in most cases is to assume a worst-
case scenario, in which polarizations are aligned.
A convenient measure of cross-talk, or for that matter any type of signal impairment,
is the impairment-induced power penalty. It indicates how much additional signal power
is required to maintain a specified BER in the presence of the particular impairment.
Expressed in decibels, it is defined as


Power required with impairment
P = 10 log10 . (4.50)
Power required without impairment
The power penalty incurred with homodyne interference illustrates the concept. Con-
sider the following model. A large number, N , of interferers are present with random
phases, each with an intensity Is ′ = ǫ Is . Assuming aligned polarizations, it has been
shown (see [Gimlett+89, Goldstein+94]) that the PDF for the resultant aggregate inter-
ference is approximately Gaussian, which leads to a power penalty given by

P = −5 log10 [1 − 4Q 2 N ǫ], (4.51)

where Q is the Q factor corresponding to the reference BER. For example, at a BER
of 10−9 , Q = 5.9 (Q 2dB = 15.4). Using Equation (4.51) the effect of the homodyne
cross-talk power penalty can be plotted as shown in Figure 4.42(a). The solid and
dashed curves show the BER without and with homodyne cross-talk, respectively. The

Nǫ = 0
Pe
Nǫ = 0.001 Nǫ = 0.009

10−9 Error Floor

10−7

0.32 dB Received Received

Power Power (dB) Power (dB)
Penalty
(a) (b)

Figure 4.42 Power penalty with homodyne cross-talk.

31
The coherence time tc for a laser is approximately equal to the reciprocal of its linewidth: tc = 1/ν. A
related quantity, coherence length, lc = ctc , defines the distance between two observation points of the
same signal required to decorrelate the two observed signals.
 Enabling Technology 239

horizontal spacing between the two is the cross-talk power penalty. The dashed curve
is plotted for N ǫ = 0.001. Another curve for N ǫ = 0.009 is plotted in Figure 4.42(b).
Note that in this case there is an error floor, corresponding to the BER (≈10−7 ), at which
4Q 2 N ǫ = 1, where the power penalty tends toward infinity. It is impossible to achieve
BERs smaller than the error floor because of the nature of the cross-talk.
Among the different forms of cross-talk described above homodyne is the most severe,
because the cross-talk is produced by the signal itself. Therefore, increasing the signal
power will not indefinitely reduce the BER, as reflected by the error floor.
The precise relations between cross-talk and BER depend a many factors, including
modulation format, phase, and polarization of optical signals. Procedures have been
developed to take into account these factors to evaluate cross-talk-induced penalties
semianalytically [Ho+99, Kamalakis+03].

4.9.2 Signal Power Divergence

For a variety of reasons, it is desirable to run a WDM system with equal powers
(or OSNRs) in each channel. However, signal power or OSNR imbalance can re-
sult from nonideal and uneven loss/gain functions of optical components, such as fil-
ters, MUXs/DMUXs, optical amplifiers, and couplers, from laser misalignment, from
polarization-dependent loss (PDL), and/or from dynamic reconfigurations of the network
that add/drop channels at different power levels. Signal power divergence accumulates as
the channels propagate through an optical path. The impact of this effect on the network
performance can be very significant, causing adverse effects in both the low-power and
high-power channels. The weaker channels are affected the most by receiver electrical
noise and interference within the network causing low OSNR, whereas the strongest
channels can potentially suffer from fiber nonlinear effects. The dynamic range limits
of the optical receivers can also affect performance.
Lightwave transmission systems and networks with a large number of optical com-
ponents (such as undersea systems or metro networks) are vulnerable to performance
degradations due to PDL, which was described in Section 4.3.2.3. The total PDL along
a particular optical path will depend on the initial signal polarization orientation and
on the relative orientation of the polarization axes of each component in its path. Due
to PMD the principal states of polarization vary randomly in a wavelength-dependent
way and, as a result, the interaction of PMD with PDL introduces among other effects a
wavelength-dependent variation of the loss that each signal will experience in the path.
One of the resulting effects is signal power divergence among the channels (known as
PDL-induced ripple) and OSNR deterioration at the receiver. The signal power diver-
gence becomes worse through long amplifier chains as the strongest growing channels
compete for gain with the weakest channels, and the unpolarized noise is attenuated
differently than the signals.
PDL is a probabilistic effect involving interactions of random polarization states with
components along a signal path. This suggests a Monte Carlo simulation as one way to
estimate the PDL-induced ripple magnitude in a typical system [Chowdhury+99]. (See
the case study of Section 4.13.2.2.) Each component is represented as a birefringent
240 Multiwavelength Optical Networks

element with random orientations of its principal axes. All signals propagate through
these elements starting with some random orientation of their polarization axes and the
PDL-induced ripple statistics are calculated at the receiver based on repeated simulation
runs over a large set of signal and component axis orientations. The mean+5σ value
of PDL-induced ripple can be used as a good figure of merit for system engineering.
In [Mecozzi+04], an analytic model is presented that deals with the effect of PDL on
OSNR in optical communications systems taking into account the effect of dynamic
gain equalization.

4.9.3 Chirp-Induced Penalty

As described in Section 4.5.3.1, on standard SMF, the chirp produced by direct modu-
lation of a laser interacts with fiber dispersion to distort and spread out the data pulses,
causing ISI and increased BER. The chirp-induced penalty can be estimated qualita-
tively by calculating the amount of eye degradation at the output of the receiver electrical
low-pass filter [Elrefaie+88] or, more accurately, by evaluating the error probability for
different receiver types. Commercially available DMLs exhibit significant variations in
the instantaneous power and frequency of their waveforms, among different manufac-
turers and among different samples from the same manufacturer.
Signal degradation occurs when positive chirp, resulting in a blue-shifted pulse leading
edge and red-shifted trailing edge, encounters positive fiber dispersion. However, if the
sign of the chirp or the dispersion (but not both) is reversed, pulse compression will
occur over a limited transmission distance, where the red trailing edge tends to catch up
with the blue leading edge. This effect actually produces a signal enhancement; i.e., a
negative power penalty. To obtain this desirable chirp/dispersion effect, either negative
chirping is generated in the transmitter (e.g., using an EA-DFB transmitter or an MZ
modulator tuned to produce the desired chirp), a process called prechirping, or negative
dispersion fiber is installed on the signal path.
The chirp/dispersion interaction was analyzed using a small signal model [Koch+84].
However, this does not appear to be applicable to the large excursions in a data pulse.
Instead, a phenomenological model was developed, based on experimental measurements
of the peak-to-peak chirp and the chirp duration parameters [Antoniades+02]. This
simple model describes the chirp/dispersion interactions and their effect on the level of
the received 1s and 0s and provides accurate results in simulations.

4.9.4 Optical Filter Concatenation: Distortion-Induced Penalty

A serious signal impairment that is unique to transparent optical networks is the penalty
due to WDM filter concatenation. This effect is relatively small in a point-to-point
optical system because a given signal usually passes through just two filters: a MUX
and a DMUX. However, in a transparent optical network, a signal may be demultiplexed
and remultiplexed at many network elements throughout its path before it is finally
received. Thus the signal experiences the concatenation of the entire set of filters in its
path. The effective spectral transfer function of the filter set is the product of all of the
individual filters’ transfer functions and can therefore be much narrower in spectral width
 Enabling Technology 241

than that of a single filter. Consequently, optical MUX/DMUX concatenation causes

signal attenuation and distortion leading to ISI, which eventually limits the maximum
number of optical network elements that can be cascaded. The resulting signal spectral
clipping due to the concatenation of MUX/DMUXs diminishes the signal power seen
by the network elements, producing changes in the operating points of any variable
optical attenuators (VOAs) and EDFAs along the signal path. Therefore, to deduce the
effects of filter concatenation on signal quality it is necessary first to take into account
the interaction among optical MUX/DMUXs, VOAs, and EDFAs. This information
can then be used to determine the power levels of the optical signal and amplified
spontaneous emission (ASE) noise in the network. A comprehensive model including
the aforementioned effects of optical filter concatenation appears in [Roudas+02]. It
is based on a semianalytic technique for the evaluation of the error probability at the
receiver. The error probability evaluation takes into account arbitrary pulse shapes,
arbitrary optical MUX/DMUX and electronic low-pass filter transfer functions, and
non-Gaussian photocurrent statistics at the output of a preamplified direct-detection
receiver.

4.9.5 Polarization Mode Dispersion Impact on System Performance

As described in Section 4.3.2.3, PMD is a random effect in fibers. It can be a significant
impairment in high-data-rate transmission systems and networks, especially for already-
installed older fiber spans. PMD is defined as the statistical average of differential group
delay (DGD) between the two principal states of polarization (PSP) of a signal on an op-
tical fiber. It has been shown that to a first-order approximation, the PMD of a transparent
link is the root-mean-square sum of the PMD contributions of individual components,
leading to the square root in Equation (4.20). Essentially, this model implies quadrature
addition of PMD for transmission fiber, dispersion compensation modules, and network
elements (including amplifiers) that have PMD. Simulations and experiments have cal-
culated the Q penalty versus DGD in a system. What is really important using this model
is the value of the accumulated system DGD as a percentage of the bit duration. For
example, in [Chowdhury+04] it was shown that a DGD accumulation of 33% of the bit
period (e.g., 33 ps for a 10-Gbps system) produces a Q penalty of roughly 0.5dB.32
Because PMD is a probabilistic effect, total allowed PMD in a commercially deployed
system is calculated to yield an acceptable outage probability. Following current ITU
standards a safety factor of 3.7 is included in the PMD to assure a 10−7 outage probability.
33
In the previous example, this translates into 3.7 = 8.92 ps tolerable PMD for the entire
system. In the case study of Section 4.13.2.2, the total PMD is calculated based on the
quadrature addition model described above factoring in the 3.7 safety term.

4.10 Optical and Photonic Device Technology

As discussed in Section 2.3, network nodes perform the operations of routing, switch-
ing, and wavelength conversion, and these functions are implemented in a variety of
32
The Q dB definition used in this study is 10 log10 Q.
242 Multiwavelength Optical Networks

ways in specialized network elements; static routers, optical cross-connects, wave-

length add/drop multiplexers, wavelength-selective cross-connects, and wavelength-
interchanging cross-connects. When deployed in operational networks these network
elements take the form of complex subsystems incorporating a host of ancillary mon-
itoring, control, and management functionality as well as the basic optical transport
functions. Although some of them may include electronics to a greater or lesser degree,
their essential building blocks are optical and photonic, including directional couplers
(either static or dynamic), switches, wavelength-selective elements (filters), polarizers,
and isolators. In some cases wavelength selectivity is combined with a switching or
tunability function in the same device to produce a multiwavelength switching element.
Each of these devices performs a basic linear operation required in a transparent optical
network (see Section 1.3).
Some optical elements are fabricated as all-fiber devices, making it easy to connect
them to transmission fibers with very small coupling losses. Others are fabricated in the
form of integrated configurations using techniques derived from electronic integrated
circuit fabrication. Examples of fiber devices are fused fiber directional couplers and
fiber Bragg grating filters. Examples of integrated assemblies are optical switch fabrics,
wavelength multiplexers/demultiplexers, and fully integrated optical receivers. Control-
lability is effected in devices such as switches, controllable attenuators, or tunable filters
either by using materials whose optical properties can be changed under external control
or by using systems whose physical geometry can be changed. Examples of the former
are liquid crystals whose refractive index can be changed under electrical or magnetic
control, thermo-optic materials whose optical properties are temperature dependent (e.g.,
silica or polymers), and electro-optic materials such as lithium niobate and semiconduc-
tors, whose refractive index can be controlled through an applied electric field. The most
interesting development in the category of systems with controllable physical geometry
is the optical microelectromechanical system (MEMS). In optical applications, MEMS
use free-space optics and electromechanically actuated moving parts to produce a wide
variety of optical systems, including gain and dispersion equalizers and large optical
switch fabrics.
In this section, we examine the principles of operation, performance characteristics,
and underlying technology of some representative optical and photonic devices, starting
with couplers and switches and concluding with wavelength-selective devices, which
include filters and multiwavelength switches.

4.10.1 Couplers and Switches

Controllable coupler and switch technology has come full circle over the past decade:
from free-space optomechanical technology to guided-wave devices and back again.
Typical early examples of optical switches consisted of bulk assemblies using free-space
optics – mirrors, prisms, collimators, lenses, and gratings (for wavelength-selective
devices). In these devices, the physical geometry is controlled either purely mechanically
or electromechanically. They generally have good switching performance (low cross-
talk and loss). However, because of their complex and potentially unreliable mechanical
 Enabling Technology 243

structures as well as their relatively slow speed (switching times of the order of milli-
seconds) these have been largely supplanted by guided-wave devices, which are easier
to build, more reliable, faster, and lend themselves well to manufacturing as optical
integrated circuits (OICs) using planar lightwave circuit (PLC) technology. Although
guided wave optical devices are the basic components of most state-of-the-art optical
switch architectures, as well as many other optical subsystems, it is possible that they
will eventually be overtaken by MEMS switches in applications requiring large-scale
integration. This represents a return to free-space optomechanical devices but using
much more sophisticated architectures, control, and fabrication techniques.
In evaluating and comparing switching devices there are many criteria to take into
account: speed, power consumption, cost, reliability, polarization and wavelength inde-
pendence, insertion loss, and cross-talk. Because wavelength-selective cross-connects
are composed of large numbers of switching devices, they are the most demanding
applications of these devices with respect to most criteria except speed.33 In particular,
overall system insertion loss and cross-talk are critical performance parameters, and
scalability is an essential consideration. Insertion loss in an OXC is the total attenuation
along an active path from an input port to an output port. Because most OXCs have
multistage switch fabrics, their overall insertion loss is several times the insertion loss
of one switching device. Similarly, the total cross-talk appearing at an output port of
a large switch fabric includes combined effects of many cross-talk terms generated in
each elementary switching device. These multiple effects make it very important to
minimize insertion loss and cross-talk at the device level, and we emphasize that issue
here. The effects of embedding many imperfect devices in switch fabrics are discussed
in Section 4.12.1.
It is useful to express the phenomena of cross-talk and insertion loss in terms of
the conceptual models used in this book, taking the directional coupler of Figure 2.8
as our generic 2 × 2 switching element. The power transfer matrix of a symmetric
lossless coupler is given by Equation (2.3). In real devices, it is impossible to make
α = 0 or α = 1. Instead, we have α = ǫ > 0 when the switch is set in the bar state
and α = 1 − ǫ < 1 in the cross state. As indicated in Equation (2.13), this means that
in either state a fraction ǫ of input signal power leaks into the wrong output port. The
leakage parameter ǫ, usually expressed in decibels as ǫdB = 10 log10 ǫ, is called the
cross-talk of the switching element. Note that because the device is passive and lossless,
the cross-talk is accompanied by an insertion loss of the same value from input P1 to
output P1′ .34 In many cases, a switching element is characterized by its extinction ratio
between a specified pair of input and output ports, which indicates the ratio of power
transferred through the switch when its state is changed from ON to OFF. For example,
the coupler of Figure 2.8 might be used as an ON/OFF switch from input P1 to output
P1′ , where α = ǫ in the bar state (the ON state) and α = 1 − ǫ in the cross state (the

33
Because OXCs are used for provisioning and recovering optical paths on a circuit-switched basis, high
speed is not required; switching times of milliseconds are sufficient. However, optical packet-switching
systems require much faster switching times – of the order of nanoseconds.
34
Real devices may exhibit different cross-talk values in each state; i.e, we may have α = 1 − ǫ ′ in the cross
state, where ǫ ′ and ǫ are different.
244 Multiwavelength Optical Networks

OFF state). The extinction ratio is then defined as the ratio of P1′ in the ON state to P1′ in
the OFF state, which in this case is (1 − ǫ)/ǫ. Thus, for switches with small leakage pa-
rameter ǫ, the extinction ratio in decibels is approximately the negative of the cross-talk.
The following sections describe some representative switching devices, beginning
with guided-wave devices, following with MEMS switches, and concluding with liquid
crystal architectures.

4.10.1.1 Guided-Wave Switches

Guided-wave photonic devices are frequently fabricated using PLC technology. Compo-
nents such as waveguides, couplers, filters, and switches are constructed on the surface
of a substrate, with the most common materials being silica on a silicon substrate, silica
on silica, ion-implanted silica, titanium diffused into lithium niobate, III-V semiconduc-
tors, or polymers. PLC technology offers the possibility of integrating large numbers of
similar devices on a single chip, which is useful in applications where components are
replicated many times; for example, switch fabrics and DWDM multiplexers. In this way,
a single module fabricated as an OIC replaces a set of bulk devices interconnected by
fibers. A major advantage of integration is the cost saving involved in eliminating the in-
terconnections, which are labor intensive, unreliable, and introduce significant coupling
losses. The downside of integration is the lower yields associated with manufacturing
large wafers.
Silica-on-silicon-based technology has also been used with advantage in fabricating
and packaging hybrid OICs composed of disparate components such as waveguides,
lenses, photodetectors, and lasers. The waveguides and other PLC devices such as
couplers, multiplexers, and switches are fabricated in silica and are interconnected with
other nonsilica devices, constructed as surface-mounted modules, to form a complete
subsystem such as an optical transceiver. This technique, known as the silicon optical
bench, has advantages in versatility, cost, and performance [Henry+89, Li+97].
The most common structures for guided-wave switching devices are the 2 × 2 di-
rectional coupler, the MZ or interferometric switch, and the Y-branch switch. Each can
be built using a variety of materials and fabrication technologies. Each material of-
fers a choice of control mechanisms. Possibilities in LiNbO3 include electro-optic and
acousto-optic control. Thermo-optic control is used in silica guides, either electro-optic
or thermo-optic control in polymers and electro-optic or carrier injection in semicon-
ductors. Given this wide range of choices, switching devices can usually be chosen with
cost and performance properties tailored to the needs of each specific application.

Controllable Directional Coupler

The operation of the controllable directional coupler is based on the fact that when
two waveguides are placed close together, coupling is possible between the two guides
because of the overlap in their fields. (The light is not confined completely to the guides,
with an evanescent field spreading out to the surrounding medium.) Provided that the
propagation constants of the waves in the two guides are the same (the phase-matching
condition), light entering one guide couples into the other and vice versa. Depending
on the length of the interaction region, the coupling may be partial or complete. In the
 Enabling Technology 245

Fibers

Figure 4.43 Controllable directional coupler. (From [Saleh+91, Figure 21.1-7b]. Copyright


c 1991. Reprinted by permission of John Wiley & Sons, Inc.)

configuration shown in Figure 4.43, a voltage applied to the electrodes increases the
propagation constant in one of the guides and decreases it in the other. In this way, the
phase matching is controlled electrically, thereby controlling the coupling ratio between
the ports. Referring to the conceptual model of Equation (2.3), varying the voltage varies
the control parameter α and places the coupler in the cross state, bar state, or any partial
coupling state between the two. The directional coupler can therefore be operated in an
analog mode, sweeping it through a continuous range of states by varying the control
voltage. These devices are typically fabricated in LiNbO3 using PLC technology. The
propagation constant is controlled via an electro-optic effect wherein an applied electric
field produces a change in the refractive index of the material. Because LiNbO3 has a
large electro-optic coefficient and subnanosecond switching time, it has been a preferred
material for controllable directional couplers as well as many other photonic switching
devices. LiNbO3 directional couplers can achieve cross-talk down to −30 dB and fiber-
to-fiber insertion losses of 2–3 dB.
Controllable directional couplers are also realized in other technologies. For example,
PLCs using polymer waveguides on silica have been fabricated using thermo-optic
control, where refractive index control is applied through a heating electrode placed
over one of the guides [Keil+94]. Polymer devices have been integrated into small
switch fabrics (five couplers) with a switching time of 1 ms, overall extinction ratios of
17–19 dB, and a fiber-to-fiber insertion loss of 10 dB, about half of which is due to the
fiber-to-chip interfaces.
Semiconductor devices are another alternative as discussed in [Renaud+96]. In one
approach, integrated switch fabrics are fabricated using InGaAsP waveguides, which
offer a number of control possibilities, either using an electro-optic effect or carrier
injection to control refractive index. A typical example is a 4 × 4 switch made up of six
246 Multiwavelength Optical Networks

Phase
Modulator

Guide 1 3-dB + 3-dB

Coupler Coupler
−

Guide 2 +

Figure 4.44 Mach–Zehnder switch. (From [Korotky+88, Figure 11.4]. Copyright

c 1988 Aca-
demic Press. Used with permission of Elsevier.)

couplers using carrier injection, with a 15-dB overall insertion loss and −15-dB cross-
talk. Although insertion loss is generally high in semiconductor-based switches, this can
be offset by the use of SOAs integrated on the same wafer as the switching devices.
However, attenuation followed by amplification necessarily augments the noise level.

Mach–Zehnder Interferometric Switch

The MZ interferometer, configured as a single input/output device, was described in
Section 4.5.3.2. An electro-optically controlled version of this device, used as a 2 × 2
switch, is shown in Figure 4.44, where 3-dB couplers terminate the two arms of the
device. In a symmetric device (both arms the same length) when no voltage is applied to
the phase modulator, signals entering guide 1 or 2 exit on guide 2 or 1, respectively. If we
designate this as the cross state, then it is switched to the bar state when a control voltage
shifts the relative phases of the waves traversing the two arms by any odd multiple of π .
One drawback of the MZ switch is that the couplers must have an exact 3-dB ratio to
achieve a perfect cross state. Imperfect fabrication leads to undesirable cross-talk in the
cross state. A typical switch of this type fabricated in LiNbO3 has fiber-to-chip-to-fiber
insertion loss of 4 dB and cross-talk of −25 dB [Pohlmann+91]. Its switching time is
less than 1 nanosecond. Other versions have been built using other technologies. For
example, typical parameters using the electro-optic effect in InGaAsP semiconductors
are 3 dB of insertion loss and −16 dB of cross-talk.
Mach–Zehnder switches have also been built in silica on a silica substrate using
thermo-optic control. Because heating is involved, switching times are in the millisecond
rather than subnanosecond range. In a typical device [Kasahara+02], a thin film electrical
heater is deposited on one of the arms of the MZ structure to control the switch state.
Applying heat to the arm induces a change in refractive index and a phase change of π in
the signal on that arm, thus changing the state of the switch. In thermally actuated devices,
especially those used in large arrays, it is important to reduce the required switching
power while maintaining good switching speed and low cross-talk. The switch in this
example achieves 5.2 ms switching time using 90 mW of power. It has over 30 dB of
extinction ratio (−30 dB cross-talk) and about 1 dB of insertion loss, which is almost
all due to fiber-to-chip coupling. It is essentially polarization independent.
The thermo-optic versions of the MZ switch have also been fabricated using polymer
waveguides on a silicon substrate. Polymers offer easy and low-cost fabrication and low
power consumption with relatively fast switching times. An example is a device using
 Enabling Technology 247

Thermo-optic
Intersection
heater ′
Input port 2 Output port 1

OF
F
′
Input port 1 Output port 2

Optical ON
Sig
nal

1st MZI 2nd MZII

3-dBcoupler

Figure 4.45 Two-stage Mach–Zehnder switch. (Adapted from [Goh+99, Figure 1]. Copyright

c 1999 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers,
Inc.)

asymmetric X-junctions to replace the 3-dB couplers in Figure 4.44 and employing
heaters on the two guides to change their index of refraction [Oh+98]. It has −20 dB
of cross-talk, 4.5 dB of insertion loss, and a 2-ms switching time while consuming only
10 mW of power.

Cascaded Mach–Zehnder Switches

Although −30 dB of cross-talk is very good performance for a single device, when many
such devices are combined into a large switch fabric the accumulated cross-talk may
be too high to keep the total interference to an acceptable level. Thus, it is important
to make the cross-talk produced by each device as low as possible. One approach is
space dilation, but it requires replacing each switch by a set of four switches (see
Section 4.12.1.4). A more economical approach that works in many applications (see
Section 4.12.1.1), but yields an incomplete switch, is the two-stage MZ switch shown
in Figure 4.45 [Goh+99]. It consists of two thermo-optic MZ switches fabricated using
silica-based PLC technology as in [Kasahara+02]. When both elementary switches are
in the bar state, the paths from 1 to 1′ and 2 to 2′ are active (called the OFF state in
this case). Assuming that the control parameter α has the same value for both MZs, the
overall power transfer matrix for the switch is:


(1 − α) 0
. (4.52)
α2 (1 − α)

To evaluate cross-talk, suppose α = ǫ when an MZ is in its bar state and α = 1 − ǫ

in the cross state. When the cascaded switch is in the ON state (both MZs in the cross
state) the path from 1 to 2′ is active, but this is an incomplete 2 × 2 switch, because there
is no path from 2 to 1′ in either switch state. Referring to Equation (4.52), when the
switch is in the OFF state the leakage onto the path from 1 to 2′ is ǫ 2 ; i.e., the cross-talk
in decibels is reduced by a factor of 2 compared to that of a single MZ. This is known as
second-order cross-talk (see Section 4.12.1.4). In the ON state, cross-talk from 1-1′ is ǫ,
that is, first order. Thus it is advisable to use this configuration in applications where it is
only the OFF state cross-talk that is of concern (see Section 4.12.1.1, for an application).
248 Multiwavelength Optical Networks

e1
tr od
elec
few mrad
ele
ctr
od
e2
Adiabatic region

Figure 4.46 Y-branch switch. (Adapted from [Renaud+96, Figure 8]. Copyright
c 1996 IEEE.
Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

Y-branch Switches
The Y-branch switch, shown in Figure 4.46, can be used as a 1 × 2 switch, in the reverse
direction as a 2 × 1 switch, or in a combination of four as a 2 × 2 switch. The principle
of operation of the 1 × 2 device is that the input wave propagates to the output arm
with the higher refractive index, which can be controlled by various methods, depending
on the materials used. The Y-branch has numerous advantages over switches like the
directional coupler and the MZ, which both involve controlled phase relations. The
advantages include polarization independence, relative insensitivity to manufacturing
tolerances, wide optical bandwidth, and a step-like switching curve under electro-optic
control. It has been called a “digital switch” because of this control behavior. A typical
electro-optically controlled device fabricated in LiNbO3 has less than −16 dB of cross-
talk with a control voltage of 60 V over the entire optical fiber transmission window.
Various versions of the Y-branch have also been fabricated using semiconductors
with optical bandwidths of more than 50 nm. For example, a carrier-injection controlled
InGaAsP device has −18 dB of cross-talk, requires 30 mA of switching current, and
has a 10-dB fiber-to-chip-to-fiber insertion loss when four devices are integrated into a
2 × 2 switch [Renaud+96]. Finally, thermo-optically controlled polymer devices have
also been used in Y-branch configurations, with −20 dB of cross-talk and a 2.5-dB
insertion loss.
The Y-branch is a very versatile device as a building block for large switch fabrics.
For example, switches containing a total of 88 Y-branch elements have been fabricated
in LiNbO3 on a single 8 × 80-mm wafer [Murphy+97].

Gate Arrays
All switching devices discussed thus far have been passive lossy devices: no power
is supplied to optical signals as they traverse the switch, and insertion losses can be
significant. A different approach is the active gate array switch, fabricated as a PLC with
active elements, supplying gain as well as control. Semiconductor gate array switches
have been demonstrated, with a typical architecture shown in Figure 4.47. In the figure,
which shows the actual layout on a chip, a passive splitting network is followed by an
array of amplifier gates (SOAs), which are in turn followed by a passive combining
network. In this realization, additional SOAs are used at the input and output ports to
 Enabling Technology 249

Figure 4.47 Gate array switch. (From [van Berlo+95, Figure 1]. Copyright

c 1995 IEEE. Used by
permission of The Institute of Electrical and Electronics Engineers, Inc.)

help compensate for splitting and combining losses as well as other excess losses in
the PLC. Note that this is essentially the same arrangement with the same functionality
as the generalized switch architecture of Figure 2.18. The monolithic structure uses a
combination of passive and active waveguides for the splitter/combiners and the SOAs,
respectively. The I/O amplifiers are run at a constant injection current, and each gate
is controlled by varying its injection current: using zero current to set it to the OFF
(absorbing) state and sufficient injection current in the ON (gain) state to yield an
overall fiber-to-chip-to-fiber insertion loss of less than 2 dB. The 4 × 4 configuration
demonstrated in [van Berlo+95] was designed to be largely polarization insensitive (less
than 1 dB polarization dependence on most paths), which is one of the challenges of
semiconductor gate array switches. It has a 40-dB extinction ratio on most paths. Cross-
talk in PLC gate arrays is generated by various sources, including waveguide crossings.
In this example, cross-talk varied from −45 dB to −17 dB, depending on the active
connections, with the higher value probably due to fabrication imperfections.

Bubble Switch
The bubble switch, a product of Agilent Technologies, is a novel variant of the PLC
switching fabric. Here, a crossbar grid like that of Figure 2.14 is fabricated in the form
of intersecting waveguides on a silica chip. In the realization shown in Figure 4.48,
the waveguide chip is sealed to a silicon controller chip, which contains an array of
controllable electric heating elements at the cross-points. As shown in the figure, each
cross-point has a trench carved out of it, which is filled with a fluid whose refractive index
matches that of the waveguides. As long as fluid is present, the cross-point is “open” and
the light signal in each guide takes a straight path through it. To close the cross-point, its
heater is activated causing the fluid to vaporize, producing a bubble, which lowers the
refractive index in the trench. As a result, the trench wall encountered by light entering
on the IN guide in the figure acts as a mirror, deflecting the IN signal to the OUT
guide through total internal reflection. For operation in the crossbar configuration of
Figure 2.14, the IN/DROP guide would correspond to a horizontal input bus and the
ADD/OUT guide would correspond to a vertical output bus. These switches have been
demonstrated in sizes up to 32 × 32 with overall insertion loss less than 9 dB, cross-talk
from −50 to −55 dB, and switching times of approximately 8 ms.
250 Multiwavelength Optical Networks

ADD Sig
nal
INSig
nal
F
luid
Trench

Center H
eater

Optical
aveg
W uide

Sidew
all
Metalization

u
Bbble
DROPSig
nal OU
T Sig
nal

Figure 4.48 Laser-activated bubble switch element. (From [Hengstler+03, Figure 1]. Copyright


c 2003 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

4.10.1.2 Microelectromechanical Systems: MEMS

These systems, fabricated using ingenious micromachining techniques, had their first
applications in the early 1990s as microscopic actuators, motors, sensors, and simi-
lar devices. In the optical communication field, they have been applied to dispersion
compensation (Section 4.3.2.3), dynamic gain equalization (Section 4.4.1.1), and re-
configurable wavelength add/drop multiplexing (Section 4.10.5.3). However, their most
ambitious application is in large-scale free-space integrated optical switch fabrics, the
focus of this section.
The first generation of MEMS-based switch fabrics are known as 2D switches. They
have a planar architecture, with an N × N switch consisting of an integrated array of
N 2 moving mirrors, used to make free-space optical beam connections between sets
of N input and N output fibers in crossbar fashion. Typically fabricated in polysilicon
on a silica substrate, they use either free rotating or torsion hinged mirrors actuated
electrostatically. An example of the free rotating case, used to realize an optical cross-
bar switch of the form of Figure 2.14, is shown in Figure 4.49(a), where a mirror in
a position (i, j) is set in the vertical position to deflect a light beam from input fiber
i to output fiber j. A translation plate, shown in Figure 4.49(b), is used to “flip” each
mirror up or down. Because of the small mirror size, switching times of submillisec-
onds are achieved. Cross-talk in these systems is better than −60 dB for an 8 × 8
switch.
Although 2D MEMS switches have many good features, including low cross-talk,
wavelength, polarization and bit-rate independence, compactness, relative ease of fabri-
cation, and low power consumption, there is a scalability problem due to insertion loss.
Because of the divergence of the optical beams in free space, the coupling loss between
 Enabling Technology 251

Switch
Mirror

Microlens

Silica Substrate

(a) Schematic Drawing of the Micromachined Free-Space Matrix Switch

Switch
Mirror

Hinge
Joint
Actuated
Translation Plate

Silica Substrate

(b) Working Principle of the Switch Mirrors

Figure 4.49 2D mechanical switch using micromachined mirrors. (From [Lin98, Figure 1].
Copyright
c 1998 IEEE. Used by permission of The Institute of Electrical and Electronics
Engineers, Inc.)

input and output fibers grows significantly with the size of the switch. For example,
fiber-to-fiber insertion loss on the longest path through a switch of this type increases
from 2.9 to 6.9 dB for 16 × 16 and 32 × 32 switches, respectively. Note also that the path
lengths through the switch are not uniform, so insertion loss varies from one connection
to another. Because of this, the 2D switch has a practical size limitation of about 32 × 32
to keep insertion loss below 10 dB. Although this may seem like a large switch, it is not
large at all in DWDM switch configurations. For example, consider a typical DWDM
switch using the opaque O-O-O architecture of Figure 4.88. It carries one wavelength
on each port of the core switch fabric, giving the overall switch complete nonblocking
WIXC functionality. With as few as 10 wavelengths on each link, a 32 × 32 switch fabric
in this configuration can serve no more than three input and output fibers.
252 Multiwavelength Optical Networks

optical
signal
Lens array Fi
Fiber array

MEMS mirror array

ME

Op
Optical path

MEMS
array

Fiber array
Fi
Lens array
Le

(a) (b)

Figure 4.50 3D MEMS switch. (From [Chu+02, Figure 3]. Copyright
c 2002 IEEE. Used by
permission of The Institute of Electrical and Electronics Engineers, Inc.)

The second generation, called a 3D MEMS-based switch, largely circumvents the

scalability problem, but the 3D system requires twice as many mirrors and a very
challenging multistage optical structure and control system. Figure 4.50 shows a
“Z-structure” for an N × N 3D switch. Light beams from an array of N input fibers
are focused on an array of tilting mirrors, with one mirror dedicated to each input fiber,
with a similar arrangement used for the output fibers. As shown in Figure 4.51, these
mirrors are supported by gimbals so they can tilt (under electrostatic or electromagnetic
control) in two directions. To set up a connection between input fiber i and output fiber
j, the input mirror i is tilted to focus its beam on output mirror j, which also has to
be tilted to direct the incoming beam to fiber j. Note that to accommodate all possible
connections, the angular position of each mirror must move through a wide range of two-
dimensional states. An advantage of this geometry
√ is that all paths are approximately the
same length and the path length grows as N instead of N , reducing the insertion loss
problem – insertion losses of less than 10 dB are possible with ports numbering in the
hundreds.
There are many design challenges. Some basic requirements include: high spring
stiffness and low mirror mass for fast mirror response time, large (of the order of 1 mm)
mirror diameter, high reflectivity, large radius of curvature, and large tilt angle (up to
10◦ ) together with stability in the face of temperature changes and vibrations. Mirror
arrays are typically fabricated using surface micromachined polysilicon or single-crystal
silicon (SCS) starting with a silicon-on-insulator (SOI) wafer. Fabrication techniques
well established for integrated electronics are used in each case. Whereas a metal coating
is required to achieve good reflectivity in polysilicon, the necessary mirror properties,
 Enabling Technology 253

Figure 4.51 3D gimbaled mirror. (From [Greywall+03, Figure 1]. Copyright

c 2003 IEEE. Used
by permission of The Institute of Electrical and Electronics Engineers, Inc.)

flatness, temperature stability, and high reflectivity, are achievable in SCS, together with
good spring properties, without metallization.
Although magnetic actuators have been used in MEMS switches [Bernstein+04],
the electrostatic approach is preferred because it is relatively simple to integrate elec-
trostatic actuators into a silicon-based structure. Whether electrostatic or magnetic,
a sophisticated control system is required to achieve reasonably fast switching times
(several milliseconds) while controlling power consumption and maintaining accurate
beam pointing in the face of external perturbations (temperature change, shock, and
vibration), as well as parameter drift and manufacturing imperfections. Feedback con-
trol is the best way to achieve these ends, as well as to combat the natural instability
of electrostatic actuators.35 However, feedback systems require mirror position sensors
to generate the feedback signals, as well as sophisticated control algorithms and con-
siderable real-time computing power capable of implementing these algorithms in a
system that may have thousands of mirror controllers operating simultaneously. Be-
cause of their significantly increased complexity, 3D MEMS-based switches are cur-
rently economically competitive with their 2D counterparts only in very large port count
switches.

35
A spring-restrained electrostatically controlled mirror tends to be unstable beyond a certain angular dis-
placement limit, because under a constant voltage (or charge) the electrostatic force increases as its tilt
angle increases, producing a “snap down” effect until it reaches a mechanical stop.
254 Multiwavelength Optical Networks

4.10.1.3 Liquid Crystal Switch Architectures

Liquid crystals are materials that have anisotropic refractive properties that can be
changed under the influence of an applied electric field. This gives them special polariz-
ing properties in response to an applied control voltage. In conjunction with other optical
elements, including mirrors, polarizers, lenses, and other birefringent (polarization-
sensitive) crystals, LCs can be used to spatially deflect light signals and modulate their
phase and intensity. The molecules in LC material tend to line up with each other,
producing their anisotropic optical properties, and the application of an electric field
changes the orientation of the molecules. Two types of LC materials that are commonly
used in display and communications applications are twisted nematic crystals and smec-
tic crystals. The former act as controllable polarization rotators that can be deactivated
by an applied electric field; response times are slow (of the order of milliseconds). The
latter exhibit ferroelectric polarization behavior, acting as polarizers whose axis of po-
larization can be switched between two states by the application of an electric field; they
are relatively fast, with switching times of the order of 10 microseconds. Liquid crystals
are most commonly used in spatial light modulators (SLMs) where a slab of LC material
sandwiched between other elements is arranged to produce an array of pixels that are
controlled through a grid of electrodes to produce prescribed light patterns. This is a
mature technology, having been widely used for many years in consumer applications
such as computer and television displays.
Applications in optical switching require a higher degree of manufacturing preci-
sion, but the material offers many advantages: low insertion loss across the full optical
transmission window, scalability because of the inherent parallelism of the SLM, low
power consumption, and low cost. An example of a recent application is the LC-based
holographic optical switch. A holographic phase-modulation pattern created on an SLM
can be used to steer a light beam to a desired target by diffraction. This opens up many
possibilities for free-space switch architectures operating much like MEMS switches
but without moving parts. One possible configuration is shown in Figure 4.52, where
a 1 × N switch is built around a reflective SLM consisting of an LC slab backed by a
mirror [Fracasso+03]. Light from a single input fiber is spread by the lens system onto
the surface of the SLM, which consists of individually controllable pixels; the mirror

positive lens
(achromat doublet) programmable
single-mode beam steerer
fiber array

f f

Figure 4.52 Liquid crystal holographic switch. (From [Fracasso+03, Figure 3]. Copyright

c 2003
IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)
 Enabling Technology 255

Figure 4.53 Two hologram N × N liquid crystal holographic switch. (From [Manolis+02, Figure 1].
Copyright
c 2002 IEEE. Used by permission of The Institute of Electrical and Electronics
Engineers, Inc.)

causes the light to make two passes through the slab. Each pixel is controlled to produce
a specified phase shift in the reflected light, which is refocused by the optical system into
a beam impinging on one of the output fibers. Deflection of the beam to a given output
fiber is achieved by electronically writing a precalculated hologram into the pixels of the
SLM. Thus, the SLM acts much like an electronically controllable mirror. A 1 × 8 switch
has been demonstrated using smectic LCs [Crossland+00]. It is polarization insensitive
with 17 dB of insertion loss, cross-talk ranging from −19 to −40 dB (depending on the
switch setting), and a switching time of less than 20 µs. The high insertion loss in this
case, is in large part because of properties of the LC material used here, which constrain
the hologram to binary phase modulation. The architecture of a more elaborate two-
hologram N × N switch is shown in Figure 4.53 [Manolis+02]. In this case, each input
fiber and each output fiber has a dedicated hologram, and the SLMs are transmissive
rather than reflective. To set up a connection between input fiber i and output fiber j, the
input hologram is programmed to deflect its beam onto output hologram j, which also
must be programmed to direct the incoming beam to fiber j. This structure is similar to a
3D MEMS switch architecture. A 3 × 3 switch was demonstrated with a 19-dB insertion
loss, and −35 to −40 dB of cross-talk. Although LC holographic switches have stringent
optic alignment requirements, errors in alignment can be corrected by building them
into a modified hologram. Comparing the cross-talk and insertion loss figures to MEMS
switches, the latter have the advantage.36 However, the LC switch has no moving parts
and can achieve speeds in the microsecond range. Wavelength selectivity can also be
built into LC switching devices (see Section 4.10.5.2).

4.10.2 Reciprocity
A fundamental property of Maxwell’s equations in linear isotropic media is reciprocity
[Ramo+94]. When applied to guided-wave optical systems, reciprocity has a simple
informal interpretation: If a wave propagates from one point (an input) in a system to

36
The comparison is somewhat unfair at this writing because the LC switch is at an earlier stage of development
than the MEMS switch. Current developments in the LC television display field may spill over to optical
switching applications.
256 Multiwavelength Optical Networks

a1
1
b1

a2
2
b2
..
.
an
n
bn

Figure 4.54 Illustration of reciprocity.

other points (outputs), then waves injected at those outputs will propagate backward to
the input in exactly the same way.
A more quantitative definition of reciprocity is illustrated by the n-port device shown
in Figure 4.54. Suppose it contains an arbitrary interconnection of linear components
all built from isotropic materials. In Figure 4.54, the quantities ak and bk represent
normalized complex amplitudes of incident and reflected waves, respectively. The nor-
malization is such that 12 Re[ak ak∗ − bk bk∗ ] is the net average power into port k. In the
case of passive devices, the net power into the device, 21 Re[ k (ak ak∗ − bk bk∗ )], must be


nonnegative, and in lossless devices it must be 0.

If a and b represent the vectors of incident and reflected waves, respectively, then their
relationship is defined by

b = Sa, (4.53)

where S is called the scattering matrix for the device. For reciprocal systems, S is
symmetric. This implies that the individual responses (reflected waves) to single ex-
citations (incident waves) are related by bi /a j = si j = b j /ai . Because these are re-
lations among complex amplitudes, they are defined for each optical frequency and
are generally frequency dependent. A frequency-dependent S represents any frequency-
selective element such as a filter, WMUX, or WDMUX. If the system is passive, we have
x∗t [I − S∗t S]x ≥ 0 for all x, which means that the expression in brackets is a Hermitian
matrix. If it is lossless, S∗t = S−1 ; that is, S is a unitary matrix. If it is lossless and
reciprocal, S∗ = S−1 .
For example, the scattering matrix S for the 2 × 2 coupler in Figure 2.8 is a 4 × 4
matrix, which can be related to the power transfer matrix in Equation (2.3) as follows.
Relabeling the output ports 1′ and 2′ as 3 and 4, respectively, we can partition S into
four 2 × 2 submatrices Si j , where S 21 is the matrix relating the two incident waves at
the input ports to the two reflected (outgoing) waves at the output ports. (For an ideal
 Enabling Technology 257

coupler, Sii = 0.) It can be shown that

√ √ 
1−α j α
S 21 = S 12 = √ √ . (4.54)
j α 1−α
The power transfer matrix of Equation (2.3) follows directly from Equation (4.54).
The most important consequence of reciprocity for our purposes is that if we know
the optical field (and thus power) relationships for signals traversing a system in one
direction, these relationships must also hold for signals traversing the same system in
the opposite direction. The “system” may be a single device, a network element, or a
whole network, as long as it is linear and reciprocal. It is for this reason that the standard
building blocks of our systems can be used for pairs of similar operations simply by
reversing inputs and outputs: splitters become combiners,37 and multiplexers become
demultiplexers.

4.10.3 Nonreciprocal Devices

One reason for introducing the concept of reciprocity here was to point out that, in
certain applications, nonreciprocal devices are needed. The simplest example is the
fiber amplifier. If an EDFA is inserted in a fiber to amplify signals propagating from left
to right, it will also amplify signals propagating in the opposite direction. Because there
are often undesired signals moving in the wrong direction in a fiber (for example, due to
small reflections at various points in a system), it is important to prevent this backward
propagation. This requires a nonreciprocal device, called an isolator.
The principle of the isolator is shown in Figure 4.55. Incident light is vertically
polarized by polarizer A , experiences a 45◦ clockwise rotation as it passes through the
Faraday rotator, and continues on through polarizer B, which is oriented to allow it to
pass unattenuated. A wave entering in the opposite direction through B starts off at the
45◦ polarization angle, is rotated counterclockwise another 45◦ , resulting in a horizontal
polarization, and is therefore blocked by polarizer A . These opposite directions of
polarization rotation can only occur in nonreciprocal devices. Faraday rotators are made
of special materials that have nonisotropic dielectric properties in the presence of a
constant applied magnetic field.
Another application of the Faraday rotator is in the circulator shown in Figure 4.59. The
directions of circulation of the forward- and backward-propagating waves, as described
in Section 4.10.4.3, can be obtained only in a nonreciprocal device.

4.10.4 Optical Filtering Technology

One of the factors determining how efficiently the fiber bandwidth can be exploited is the
maximum packing density of wavelengths and wavebands.38 This in turn is dependent
37
The inevitability of combining losses is demonstrated easily using reciprocity. See Problem 9 at the end of
this chapter.
38
The distinction between wavelengths and wavebands was explained in Section 2.2. There, the discussion
was generic and focused on wavebands. In this section, the focus is on specific devices that may be used
258 Multiwavelength Optical Networks

45°

45°

Transmitted
B 45° Wave
x

Polarizer B

Incident
y Faraday
Wave
Rotator
Polarizer A
(a)

45°
Reflected
Wave

B 45°

Polarizer B

y Faraday
Rotator
Polarizer A
(b)

Figure 4.55 Optical isolator. (From [Saleh+91, Figure 6.6-5]. Copyright

c 1991. Reprinted by
permission of John Wiley & Sons, Inc.)

on wavelength filtering and switching technology. The two critical areas of application
of wavelength-selective devices in multiwavelength networks are in optical receivers
(either fixed or tunable) and wavelength-selective switches (WSSs).
Tuned (or tunable) optical receivers must select signals on a given wavelength and
reject others on different wavelengths appearing on the same fiber. Undesired signals
at different wavelengths that reach the photodetector constitute interchannel cross-talk,
which degrades receiver performance just as cochannel (same wavelength) cross-talk
does. The ideal filtering devices for performing the selection at the receiver should
have a rectangular transfer function: flat passbands with linear phase shift (constant
delay) over the passband, and high attenuation in the stop band. An ideal filter transfer
function is impossible to achieve, and a real filter only approaches this ideal, with nonflat
passband, nonlinear phase, sloping “skirts,” limited attenuation in side lobes and other

either on a waveband or wavelength level. To conform to current usage, the term wavelength is usually
used inclusively here and encompasses wavebands as well as wavelengths. Cases where wavelengths and
wavebands are handled differently are exceptions to this rule.
 Enabling Technology 259

irregularities in its stop band. Flat passband and linear phase is needed to faithfully
reproduce the transmitted signal at the detector. Steep skirts and absence of side lobes
mean less spacing (guard band) is needed between the wavelengths sharing a fiber to
minimize interchannel cross-talk.
The passbands and the guard bands between them necessarily must be somewhat
wider than minimum values calculated under ideal conditions to allow tolerance for
laser and component misalignments, drifts, and imperfections. (See Section 4.13.2.1 for
an example of the laser misalignment problem.)
Similar issues arise in the design of WSSs. In most cases, WSSs use the three-
stage architecture of Figure 2.21, requiring wavelength demultiplexing, followed by
space switching, followed by remultiplexing. The criteria concerning flatness and guard
bands that apply to receiver filters also apply to the demultiplexing/multiplexing transfer
functions here, except that the requirements are usually more stringent. Because an
optical signal most likely traverses many demultiplexers and multiplexers on an end-to-
end optical path, as mentioned in Section 4.9.4 the overall transfer function seen by the
signal is the result of cascading many elementary filtering and switching components.
Each component has variable characteristics because of manufacturing imperfections,
aging, and temperature, so the bandwidth of the end-to-end transfer function seen by the
signal can be much narrower than the bandwidth of an individual filtering element and
far from the ideal rectangular form.
The problems caused by these imperfect filter characteristics are compounded by
laser misalignment, with the overall effect translating into waveform distortion, ISI,
and interchannel cross-talk.39 If a system is not designed with appropriate tolerances,
these effects eventually render purely optical paths traversing many WSSs unusable.
To allow tolerances for imperfections, passbands and guard bands for independently
routed wavelengths must be considerably larger than the spacings dictated by optical
receiver requirements alone. It is for this reason that waveband routing, which wastes
less bandwidth in guard bands, is generally more spectrally efficient than wavelength
routing (see Section 2.2).
The devices most widely used for optical filtering and WMUX/WDMUX applications
are Fabry–Perot (FP) filters, multilayer interference (MI) filters, fiber and waveguide
gratings, and arrayed waveguide gratings (AWGs). Although the FP filter has been known
since the 19th century, the remaining devices only became viable commercial products
since the mid-1990s. Important criteria for these devices are passband flatness, good side
lobe suppression, high out-of-band attenuation, polarization independence, low PMD,
low insertion loss, and insensitivity to environmental conditions; e.g., temperature,
shock, and vibration. Tunability properties are also of interest, including a wide tuning
range, simple and low power control mechanism, and high tuning speed. It goes without
saying that cost is primordial, and this is related to simplicity, manufacturability, and
reliability.

39
An analytical study of this problem, based on realistic filter characteristics, appears in [Roudas+97].
260 Multiwavelength Optical Networks

4.10.4.1 Fabry–Perot Filters

The FP filter (resonator), also known as an etalon, consists of a rectangular cavity with
flat parallel mirrors on opposite sides as shown in Figure 4.56(a). We have already seen
an example of this in the FP laser of Figure 4.21. With perfect mirrors and a lossless
medium in the cavity, lightwaves propagating normal to the mirrors will be reflected
back and forth, with constructive interference between waves in phase with each other
on successive passes through the cavity. This only occurs for certain wavelengths – the
resonant wavelengths, with values, λn = 2d/n, where d is the mirror spacing and n is
an integer. In a filter, the reflectivity of the mirrors is less than 100% so that some light
can enter and leave the cavity. Also the medium is lossy, so the waves are attenuated
by a factor r < 1 on each round-trip through the cavity. In this case, the discrete set of
resonant wavelengths, λn , or equivalently resonant frequencies νn = c/λn , is replaced
by a continuous periodic spectral distribution for the light intensity I in the cavity, as
shown in Figure 4.56(b). The spacing, c/2d, between adjacent frequencies νn is called
the free spectral range (FSR). The width of each resonance at its half power points, called
the full width at half maximum (FWHM), increases with increasing r . The input/output
frequency response of the filter has the same shape as the spectrum of Figure 4.56(b).
Note that the frequency response in one passband is far from the ideal rectangular form,
with rounded top and wide skirts. This is a disadvantage in DWDM applications, where
interchannel cross-talk is an issue. However, an imperfect passband is less important in
CWDM, where larger guard bands are used between channels.
In most WDM applications (e.g., tuned receivers or WADMs) a filter with a single
passband is desired for selecting one wavelength out of many in the WDM spectrum.
In such cases, the periodic shape of the FP frequency response can be a problem unless

In Out

Mirrors
(a) Filter Structure

I FSR
FWHM

ν
νn−1 νn νn+1
(b) Intensity Spectrum

Figure 4.56 Fabry–Perot filter and its spectral response.

 Enabling Technology 261

the FSR is large enough to contain the full usable WDM spectrum. The requirement
of a large FSR combined with the requirement of an FWHM small enough to select
a single channel out of a large number of closely spaced channels dictate a filter with
a large ratio FSR/FWHM, which is called the finesse of the filter. A way of achieving
effectively single-passband operation is to modify the basic FP resonator to greatly
increase its finesse. One approach is to make the mirrors in Figure 4.56(a) frequency
selective by replacing them with distributed Bragg reflectors, as is done in the DBR laser
of Figure 4.22(a).
Tunability can be imparted to the basic FP resonator by changing the mirror spacing
d. This is conveniently accomplished in fiber FP filters, where an FP resonator is formed
in the space between the ends of two fibers, which are coated to serve as the mirrors.
The cavity itself may be filled with air or fiber, and the spacing is typically controlled
through a piezoelectric actuator. A typical commercial device has 2.5-dB insertion loss,
35-GHz passband, and tunability of 60 nm.
Tunable FPs can also be constructed by filling the cavity with a medium whose
refractive index can be varied, thereby changing the resonant frequencies. A rapidly
tunable filter acts like a wavelength selective switch, which is ON when its passband is
aligned with a selected wavelength and OFF when the passband is detuned from that
wavelength. Tunable/switchable liquid crystal fiber FPs have been demonstrated using
an applied electric field to change the index of a smectic LC in the cavity. Tuning range
of 12 nm has been achieved with switching time less than 30 µs, finesse of 166, and
insertion loss of 4.1 dB [Bao+96].
A more sophisticated version of the FP filter that incorporates both single-passband
operation and tunability has been built using InGaAs/InP technology and a MEMS-based
actuator [Irmer+03]. An FP cavity is enclosed between two DBRs, which are fabricated
using micromachining techniques to form layers of InP separated by air gaps. One DBR
is integrated into a membrane, which is displaced under electrostatic control to provide
tuning. The resultant device can be tuned over 140 nm with a 3- to 5-nm passband.
It blocks all other wavelengths in the entire optical fiber transmission window. Tuning
times are in the 20-ms range.

4.10.4.2 Multilayer Interference Filters

The operation of the MI filter is illustrated in Figure 4.57(a). When used as a WDMUX
(or a drop filter), several input signals at different wavelengths enter the device at
point 1, with the signal to be dropped exiting at point 4. The filter consists of a glass
substrate on which are deposited several layers of dielectric thin films (on the right
side of the substrate in the figure). The reflections between the layers allows a selected
wavelength to be transmitted to the right and reflects the remaining wavelengths back
through the glass and out the left side (point 2). In this case, wavelengths λ1 and λ2 enter
on the left, λ1 is transmitted and dropped on the right, and λ2 is reflected and exits on
the left.
Because this is a reciprocal device, it acts as a WMUX if operated in the reverse
direction, as shown in Figure 4.57(b). Now, channels on λ1 and λ2 enter at points 4 and
2, respectively, and are multiplexed, exiting at point 1.
262 Multiwavelength Optical Networks

DMUX
Input Signals
1 3 Operation
λ1, λ2

2 4 Transmitted Signal
Output Signal
λ1
λ2

(a) Demultiplexing Mode

MUX
Output Signals 1 3 Operation
λ1, λ2

2 4
Input Signal Input Signal
λ2 λ1

(b) Multiplexing Mode

Figure 4.57 MI filter.

The MI filter is a cavity resonator similar to the FP filter. It is built using multiple
layers of transparent dielectric materials of differing refractive index, deposited on a
glass substrate. A cavity consisting of a film of lower index material is sandwiched
between two dielectric reflectors consisting of several alternating layers of high and low
index thin films. The result is a filter with essentially the same spectral characteristics
as the FP with DBR mirrors described in Section 4.10.5. One of the attractive features
of the MI filter as opposed to the FP filter is that by using a stack of several cavities
a multicavity device with prescribed characteristics can be obtained. The filter center
frequency and passband shape can be adjusted by the appropriate choice of layering
geometry and dielectric material, producing a close-to-ideal optical frequency response:
flat passband and steep side slopes.
Tunable MI filters can be built by using materials whose refractive index is control-
lable. In [Domash+04], a class of filters is demonstrated using amorphous silicon as
the high index film in single- and multicavity devices. This material is useful for a
tunable filter, because its high thermo-optic index coefficient facilitates thermal control.
Dual cavity filters have been demonstrated with tunability over 40 nm and less than
1 dB of insertion loss. Using the same technology, a class of low-cost switchable filters
has also been demonstrated. A conventional fixed multicavity MI filter is modified by
adding to it a thermo-optically tunable cavity of the type just described. By adjusting the
resonances of the various cavities properly, the device is arranged so that at room tem-
perature the cavities are mismatched, so the filter reflects all optical wavelengths. When
a control is applied to raise the tunable cavity to a “resonant” temperature, the complete
stack becomes transmissive in its design passband. In effect this becomes an ON/OFF
switched filter, which can be used in various multiwavelength switching applications.
 Enabling Technology 263

GRIN Lens GRIN Lens Channel 1

GRIN Lens Channel 3

ns
Le GRIN Lens Channel 5
GRIN

Channel 2 ns
Le GRIN Lens Channel 7
IN
GR
ns
Channel 4 Le
G RIN
Fused Silica Block
ns
Channel 6 Le
IN
GR

Channel 8

Figure 4.58 MI filter array. (From [Scobey+96, Figure 5]. Copyright
c 1996 IEEE. Used by
permission of The Institute of Electrical and Electronics Engineers, Inc.)

Extensions of this idea produce “hitless” tunable filters, where the MI filter is designed
to be thermally switched from one center wavelength to another without passing through
any undesired intermediate filter states.
An integrated array of MI filters deposited on both sides of a substrate can be used
as a multiwavelength WDMUX/WMUX as shown in Figure 4.58. The figure shows a
schematic representation of the operation of the device as a WDMUX. Graded refractive
index (GRIN) lenses are used to collimate the light before directing the signal at a slight
angle to the MI filter. A first wavelength (channel 1) is dropped and collected by another
collimating lens. All other wavelengths are reflected to the left. In this way, the remaining
light is reflected from filter to filter with a single channel removed at each reflection.
Typical parameter values are wavelength channel spacing of 100 GHz and insertion loss
of 7 dB for 16 wavelengths.

4.10.4.3 Fiber and Waveguide Gratings

We have already seen the Bragg grating as a component of a DBR laser in Section 4.5.1.1,
where it is used as a wavelength-selective reflecting filter. Although gratings have been
used as basic wavelength-selective elements in free space optical devices since the be-
ginnings of optical communications, it was not until the advent of the fiber Bragg grating
(FBG) that grating-based devices found wide application in filters, WMUX/WDMUXs,
WADMs, and similar apparatus. The FBG is manufactured by “writing” the grating into
a fiber containing photosensitive material. This is done by exposing the fiber to a peri-
odic pattern of UV light, which acts on dopants in the fiber to produce a corresponding
periodic variation of its refractive index. The FBG can be designed to have very good
264 Multiwavelength Optical Networks

Circulator
Bragg Grating
Port Port
Input 1 2 Λ3

λ1, λ2, ..., λ8 λ1, λ2, λ4, ...

Port λ3
3
λ3

Figure 4.59 FBG used as a drop filter.

reflective filtering properties: flat response in the reflection band, 30- to 40-dB rejection
of out-of-band signals, and insertion losses of the order of tenths of a dB.
The use of the FBG as a drop filter is shown in Figure 4.59. Multiple wavelengths
propagate from left to right in the fiber, passing from port 1 to port 2 in the circulator.
Wavelength 3 is reflected from the FBG, entering port 2, after which it exits at port 3.
Imperfect reflection of the dropped signal will cause leakage of a portion of its power
onto the fiber outbound from port 2. This may cause interchannel cross-talk if a channel
at wavelength 3 is added downstream, so it is important to keep this leakage small. One
way of building a WDMUX with FBGs is by using a passive splitter, with each arm
containing an FBG drop filter for one wavelength. This, however, incurs a splitting loss
proportional to the number of wavelengths being demultiplexed, so it is practical for
only a small number of wavelengths. Typical FBG drop filters are available for systems
with wavelength channel spacings on the ITU grid (50 or 100 GHz) with insertion loss
of 2 to 3 dB (including circulator loss).
By adding a second circulator to the right of the grating in Figure 4.59 the drop filter
is turned into a WADM, adding and dropping the wavelength selected by the grating.
The FBG can be made tunable by changing the pitch of the grating through either
stretching or compression. Tunable FBG filters have been built using piezoelectric actu-
ators with tuning times of the order of milliseconds over a range of 45 nm [Iocco+99].
Bragg gratings can also be written directly into planar waveguides using the same UV
fabrication principles. Thus, for example, an MZ grating-based WADM can be fabricated
using PLC technology, as shown in Figure 4.60. The basic MZ switch is modified by
writing identical Bragg gratings into its two parallel waveguides. The grating reflects

Channel 3
1 waveguide
waveguide Add
Input

Drop Output
2 4
Coupler Grating Coupler

Figure 4.60 A Mach–Zehnder WADM. (From [Kaminow+97, Figure 7.14a]. Copyright

c 1997
Academic Press. Used with permission of Elsevier.)
 Enabling Technology 265

a selected wavelength λ entering on the Input port so that it exits on the Drop port.
Other wavelengths in the input signal are transmitted through the grating and exit on the
Output port. If the wavelength λ is present at the Add port, it will be reflected by the
grating and combined with the other input wavelengths at the Output port. These devices
have been fabricated using both silica and polymer waveguides. WMUX/WDMUXs
can be constructed by cascading MZ WADMs in an arrangement similar to the MI
filter WDMUX/WMUX in Figure 4.58. A four-wavelength polymer WDMUX/WMUX
designed for a 400-GHz grid has uniform 2.5-dB insertion loss with very good channel
isolation.
Another grating application is the chirped FBG, where the pitch of the grating is varied
along its length, to produce wavelength dependence of the group delay of the reflected
signal. With appropriately designed chirp, these gratings can be used as dispersion
compensation devices (see Section 4.3.2.3).
Still other applications involve long period fiber gratings, which operate on a principle
different than the Bragg effect. A grating with a period of tens of microns, many times the
wavelength of the transmitted light, does not reflect the light in the fiber. Instead, it acts
to couple the light from the core into the cladding where it is lost. Long-period gratings
designed to have specified transmission spectra have been used to advantage for EDFA
gain equalization where the grating transmission spectrum approximates the inverse of
the nonuniform EDFA gain profile (see Section 4.4.1.1). Using two long-period gratings
written side by side in a fiber, the flattened gain profile is reduced to variations of no
more than 0.45 dB over a 37-nm band [Harumoto+02].

4.10.4.4 Arrayed Waveguide Gratings

The AWG or phased array (PHASAR) has the functionality of the static wavelength
router, shown conceptually in Figure 2.10, but it operates using interference principles
in an elegant fully intergrated PLC. A signal entering a given input port on a given
wavelength is routed to an output port determined by its wavelength. If only a single
input port is used with n output ports, the device acts as a WDMUX, or if operated in the
reverse direction, it acts as a WMUX (reciprocity again). In an m × n system, it creates
a static routing pattern determined by the geometry of the device.
Typically fabricated in silica on silicon technology, the AWG is a phased array of
multiple waveguides with path length differences between neighboring guides arranged
to create the desired phase relations at the output coupler Co [Dragone+91] (Figure 4.61).
The input signals pass through a combining coupler Cc and then through the waveguides
and exit at the output coupler. The phase differences between signals routed through the
different paths cause different wavelengths from the same input port to be focused on
different output ports. By proper design of the waveguide geometry, an AWG can be
arranged to multiplex/demultiplex a given number of λ-channels on a given grid, say
16 channels on the 100-GHz ITU grid. In idealized form, the power transfer matrix of
the AWG has the properties of a Latin router as described in Section 2.3.1.2. When
a multiwavelength signal is present on a given input port, each of its λ-channels is
routed to a different output port. The routing rule for multiwavelength signals from each
different input port produces a different permutation of the constituent λ-channels at
266 Multiwavelength Optical Networks

Cc
Co

λ1
λ1, λ2, ..., λn
λn

Figure 4.61 Arrayed waveguide grating.

the output ports, with the property that all channels appearing at a given output have
distinct wavelengths. The channel and wavelength permutation rules are determined
from the geometry of the waveguides and couplers. Of course, the filtering properties of
the AWG are imperfect, so the usual effects of insertion loss and interchannel cross-talk
are present. In addition the AWG has a finite FSR as in FP filters. Nevertheless, AWGs
have been demonstrated with impressive characteristics, and much progress has been
made since their invention in 1990. This includes solving problems of temperature and
polarization sensitivity as well as achieving flat passband shapes, low cross-talk, and
low insertion loss. Devices have been demonstrated for 32 channels at 100-GHz spacing
with 2.1-dB insertion loss, 49-GHz channel FWHM, and −28-dB cross-talk and for
256 channels at 25-GHz spacing with 2.7-dB insertion loss, 14.4-GHz channel FWHM,
and −33-dB cross-talk [Hibino00].
As with the FBG, the AWG is versatile enough to be adaptable to many different
fabrication technologies and applications. In addition to silica-on-silicon fabrication,
AWGs have been built as semiconductor PLCs, an approach that facilitates monolithic
integration with SOAs, lasers, and photodiodes [Yoshikuni02]. Applications include
integrated wavelength-selective switches (Section 4.12.2), dispersion slope equalizers
(Section 4.3.2.3), integrated WDM receivers, and multiwavelength modulators.

4.10.5 Multiwavelength Switch Technology

Transparent waveband-space switches can be realized in different ways depending on
how and where wavelength selectivity (filtering), tuning, and switching are included in
the switch fabric. The most straightforward approach is an interconnection of separate
bulk wavelength-selective devices (e.g., filters and WMUX/WDMUXs) with photonic
space switches (e.g., the architecture of Figure 2.21). Other approaches, which are
discussed in this section, conform closely to our definition of a generic multiwavelength
switch, involving combining, filtering, tuning, and switching in one package. These
range from integrated monolithic devices to combinations of full integrated devices
with other components.
 Enabling Technology 267

We have seen that devices designed as static filters, or WMUX/WDMUXs, can often
be converted to tunable or switched devices by controlling their parameters externally.
For example, the reflection band of an FBG can be displaced by stressing it to change
its grating period. In this way, it can be switched under external control to reflect power
in different bands, acting as an ON/OFF switch in those bands.
Other devices are designed from the start to be MWSs, switching several wavelengths
independently. We describe several of these in the next sections and conclude with a
discussion of wavelength dilation, a means of improving the performance of MWS-based
wavelength-selective switches.

4.10.5.1 Acousto-Optic Tunable Filters

As illustrated in Figure 4.62, the acousto-optic tunable filter (AOTF) described in
[Smith+90] is a two-input/two-output device constructed on a LiNbO3 substrate, which
acts as a 2 × 2 controllable and wavelength-selective directional coupler. The basic
AOTF implements a switching operation by “flipping” the polarization of the input wave
and hence is inherently polarization dependent. The device shown in the figure is a
polarization-independent version of the elementary AOTF, implemented by polarization
diversity techniques using polarization beam splitters (PBSs). An input is present on the
upper left fiber, and it is split by the PBS into its two polarized components as shown.
(The transverse magnetic [TM] and transverse electric [TE] polarizations correspond
to the vertical and horizontal components with respect to the substrate surface.) The
TM component then traverses the upper waveguide and exits on the right. An electric
interdigital transducer (IDT) formed by electrodes on the surface, driven by a radio-
frequency source, creates a surface acoustic wave that propagates in the same direction
as the lightwaves. Through an acousto-optic effect in the material, this forms the equiv-
alent of a moving grating, which can be phase matched to an optical wave at a selected
wavelength by choosing the correct frequency of the electrical excitation. A signal that
is phase matched (the selected signal) is “flipped” from the TM to the TE mode, so the

Acousto-optic
Polarization-Flipping
Filter

Input Output
PBS PBS

TM TE Unselected
Input
Output

Selected
TE TM Output

IDT Acoustic Beam

in Acoustic Waveguide

Figure 4.62 Acousto-optic tunable filter.

268 Multiwavelength Optical Networks

output PBS directs it to the lower output. The unselected signal exits on the upper output.
A similar polarization-flipping and -selecting process takes place for the TE component
on the lower guide. A signal entering on the lower port of the input PBS on the selected
wavelength undergoes the same operations, exiting on the upper output port. Thus the
device is in the cross state for the selected wavelength. By superimposing excitations of
several frequencies, several wavelengths can be selected simultaneously, which makes
this an MWS, as shown conceptually in Figure 2.22. A normally unwanted by-product
of the interaction is that the optical frequency of a selected signal is Doppler shifted by
an amount equal to the acoustic excitation frequency, typically of the order of 100 MHz
in LiNbO3 devices.
The switch states of AOTFs can be controlled independently for each wavelength over
a continuum of values so they can be configured to act as 2 × 2 wavelength-selective
LDCs. They can be tuned over very wide ranges of wavelengths (up to 200 nm), and
switching times are of the order of microseconds. The original versions had problems
with insertion loss and side-lobe suppression. However, they were improved to have
30-dB side-lobe suppression and low loss [Jaggi+98]. Typical parameter values are:
about 1 nm wavelength spacing, and 2 dB fiber-to-fiber insertion loss.
AOTFs have also been built using other structures and materials. In [Sapriel+02], a
TeO2 crystal is used as the acousto-optically controlled medium. A 1 × 2 AOTF switch
was demonstrated using a piezoelectric transducer driven by an RF signal to launch an
acoustic wave into the crystal. This generates a moving grating, propagating collinear
to an input lightbeam. The device has an input fiber and two output fibers for bar state
and cross state outputs. As in LiNbO3 devices, a wavelength is selected by adjusting
the frequency of the excitation so that a phase-matching condition exists between the
acoustic wave and the selected input wave. Unselected wavelengths exit the device on
the bar state fiber, whereas the selected λ-channel is deflected by diffraction to the cross
state fiber. This is a polarization-dependent effect, but the device is made polarization
independent by inserting polarization splitters and half-wave plates at the input and
output ports. Devices have been demonstrated with 0.75-nm passband, 53-µs switching
time, and worst-case cross-talk at 4-nm channel spacing about −25 dB. Performance is
poorer at smaller channel spacings, however.
Another approach is the all-fiber AOTF. These operate by using the acousto-optic
excitation either to switch a lightwave between modes in a two-mode fiber [Park+02] or
to couple light between core and cladding modes in a single-mode fiber [Satorius+02].
In both cases an acoustic flexural wave is launched along the fiber using a piezoelectric
transducer. In [Satorius+02], a core block is inserted into the fiber to prevent signals
in the core from passing the blocking point. A phase-matching condition between the
flexural wave in the fiber and a selected wavelength propagating in the core causes the
selected signal to be converted to a cladding mode while the unselected signals remain
in the core and are blocked. The selected signal propagating in the cladding bypasses the
core block and is then converted back to the cladding mode through the inverse of the
original mode conversion process. An interesting by-product of these two conversions is
that the optical frequency shift incurred in most AOTFs is cancelled out in this device.
Prototypes have been built with 3.5-nm passband, −17-dB sidelobes, 29-dB extinction
ratio, and 5-dB insertion loss. By sandwiching a single AOTF between a circulator and a
 Enabling Technology 269

Faraday rotator/mirror a two-pass device is formed with superior characteristics: 2.5-nm

passband, −29-dB sidelobes, and 38-dB extinction ratio, but the insertion loss increases
to over 12 dB because of the two passes through the device and the additional losses in
the accompanying elements.

4.10.5.2 Liquid Crystal Multiwavelength Switch

Multiwavelength switches can also be realized using liquid crystals as described in
[Patel+95]. The simplest polarization-dependent form of a 1 × 2 liquid crystal switch
is shown schematically in Figure 4.63. An input signal (indicated as input 1) is dispersed
spatially (demultiplexed) using a grating device. Each wavelength component is directed
through one pixel of an electrically controlled liquid crystal SLM, onto a polarization-
sensitive deflective element (e.g., a calcite crystal). The SLM is a twisted nematic LC
slab that rotates the polarization of a beam traversing it when no electric field is applied
but leaves the polarization unchanged in the presence of an applied electric field. The
controlled polarization rotation in combination with the polarization-selective deflection
produces a controlled beam displacement, with the displaced beam remultiplexed onto
output 2. If undisplaced, it is remultiplexed onto output 1. In the figure, a multiwavelength
signal is applied to input 1, and one of its wavelengths (the dark beam) is shown deflected
to output 2. In this fashion, an array of beam deflectors under independent electrical
control produces an MWS from input 1 to outputs 1 or 2. A similar operation is executed
on a second signal entering on input 2 (not shown in the figure), so the two operations
result in a 2 × 2 MWS or reconfigurable WADM. These devices can be built with flat

Grating

Grating

Output 2
Input 1 Output 1

Liquid Crystal Beam Displacer

Array Element (e.g., calcite)

Figure 4.63 Liquid crystal MWS. (From [Patel+95, Figure 1]. Copyright

c 1995 IEEE. Used by
permission of The Institute of Electrical and Electronics Engineers, Inc.)
270 Multiwavelength Optical Networks

passbands and can accommodate a large number of channels. [Ranalli+99] describes a

40-wavelength device with 50 GHz of channel separation, −30 dB of interchannel cross-
talk, and less than 7 dB of insertion loss. It is polarization insensitive using polarization
diversity techniques similar to those used for the AOTF in Figure 4.62.

4.10.5.3 MEMS-Based Wavelength-Selective Switches

A different approach to wavelength-selective switching uses free-space optics to combine
the wavelength selectivity of a grating with optical beam switching using micromirrors.
The result is a reconfigurable WADM whose block diagram is shown in Figure 4.64(a)
[Ford+99]. The device has multiwavelength inputs on its IN and ADD ports, and
multiwavelength outputs on its PASS and DROP ports. The combination of WMUX,
WDMUX, and single-channel reflecting/transmitting switches places the WADM in
either the pass-through state (switch reflection) or the ADD/DROP state (switch trans-
mission) for each wavelength. A λ-channel entering the IN port is passed through the
upper circulator to the PASS port when a switch reflects power at that wavelength.
When a switch is in the transmission state (which is bidirectional), a λ-channel on the
corresponding wavelength entering the IN port is transmitted via the switch and the
two circulators to the DROP port, and similarly a λ-channel entering the ADD port is
transmitted to the PASS port. The optics are shown in Figure 4.64(b). Inputs from the IN
and ADD ports are directed via the circulators and collimating lenses to a grating, which
is used as both a WDMUX and WMUX. (The IN signal enters on the upper beam in the
figure and the ADD signal on the lower beam.) The demultiplexed channels pass through

Circulator

IN PASS Grating

ADD
WDM Mux
C2
Single-Channel Switches Fold
(reflect/ transmit) Mirror Collimation
DROP
Lenses
WDM Mux
Focus Lens
IN C1 PASS
ADD DROP
Circulator

(a) Block diagram (b) Free-space optics

IN PASS IN PASS
DROP
ADD

v v

(c) Mirror states

Figure 4.64 A MEMS-Based WADM. (From [Ford+99, Figures 1–3]. Copyright

c 1999 IEEE.
Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)
 Enabling Technology 271

an additional lens system to be focused on the device plane where an array of tilting
mirrors is located. (The λ/4 plate is used to reduce PDL in the grating.) Figure 4.64(c)
shows the operation of the tilting mirrors. For each wavelength in the PASS state, the
mirror, which is hinged in the middle, is tilted to the left by an electrostatic actuator.
This reflects the IN beam backward along its same path through the lens system and
grating (this time acting as a WMUX) to circulator C1 . The ADD beam is deflected
and dumped. In the ADD/DROP mode, the tilting mirror deflects the ADD beam to the
PASS port and the IN beam to the DROP port. A WADM has been demonstrated for 16
wavelengths at 200-GHz spacing. It has a 0.7-nm passband at FWHM, a 5-dB (8 dB)
insertion loss on the PASS (DROP) path, a 32 dB extinction ratio, −30 dB of cross-talk,
a 0.2-dB PDL, and a 20-µs switching time, which is extremely fast for a MEMS device.
MEMS-based 1 × K multiwavelength switches have been constructed using a similar
approach. In this case, the optical system directs an N wavelength signal from a single
input fiber via a grating demultiplexer to an array of N single-axis micromirrors with
electrostatic control. Each mirror now requires analog control to steer its beam to one
of K output fibers. Another pass through the same lens and grating system is used for
multiplexing and focusing the beam onto the selected output fiber. A 1 × 4 prototype
has been demonstrated, designed for 128 wavelengths on a 50-GHz grid [Marom+02]. It
has an insertion loss of less than 5 dB and wide flat-top passbands with low interchannel
cross-talk, making it suitable for operation at relatively high bit rates (10 Gbps) under
realistic transmission conditions.

4.10.5.4 MI Filter-Based Switches

As explained in Section 4.10.4.2, the MI filter can be designed to be switchable from a
completely reflecting state (OFF) to a bandpass transmission state (ON) using thermal
control. With this technique, the WMUX/WDMUX arrangement of Figure 4.58 can be
employed as a building block for a WADM by replacing its static MI filters by switchable
filters. Figure 4.65 illustrates the concept [Domash+04]. When all filters are in the OFF
state, all wavelengths entering on the IN port are passed on to the OUT port, and no
wavelengths present on the ADD fibers are transmitted. When a filter in the DMUX
is switched ON, its wavelength is extracted from the multiwavelength signal on the IN
port and exits on the corresponding DROP fiber. Similarly, when a filter in the MUX
is switched ON, and a signal is present on its ADD port, that channel is added to the
multiwavelength signal on the OUT port. In this way, switching, wavelength selectivity,
and variable attenuation can all be combined in the same device. (The insertion loss
through the MI filter passband can be varied by varying its temperature.)

4.10.5.5 AWG-Based Switches

Three related examples of monolithic integration of the AWG with various types of
switching devices have been developed at NTT laboratories. The first is a WADM
realized as a PLC using silica on silicon, combining MZ switches with four AWGs
on a chip [Okamoto+96]. Its logical architecture is the three-stage arrangement of
Figure 2.21, with two input and two output fibers. Two of the AWGs serve as the first-
stage WDMUXs and two as the third-stage WMUXs. Each 2 × 2 second-stage switch is
272 Multiwavelength Optical Networks

DMUX MUX

Express

?λii λi
IN OUT
λ?i+1 λ?i+1

λ?i+2 λi+2

λ?i+3 λi+3

Switchable/Variable
Add/Drop

Drop λi − λi+3 Add λi − λi+3

Figure 4.65 An MI filter-based WADM. (From [Domash+04, Figure 16]. Copyright

c 2004 IEEE.
Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

realized as a double-gate switch using four thermo-optically controlled MZ switches in

the space-dilated configuration of Figure 4.80. In this way fabrication imperfections can
be tolerated in the MZ switches while maintaining low cross-talk. WADMs are reported
for 16 channels with −28.4 dB of cross-talk, 7.8–10.3 dB of on-chip insertion loss, and
a switching time of less than 1 ms.
In [Suzuki+98] a 2 × 2 multiwavelength switch with the same logical architecture
as in the previous example is described, but it is built around a single AWG monolith-
ically integrated with an array of 2 × 2 switches, this time realized as single thermo-
optically controlled MZ switches. The AWG does fourfold duty as two WDMUXs and
WMUXs, requiring a total of 2N + 2 input and output ports in an N wavelength system.
Multiwavelength signals on the two input fibers are fed to two ports of the AWG, which
demultiplexes them to 2N output ports. Pairs of these demultiplexed signals on the same
wavelength are fed back to other ports on the AWG via 2 × 2 switches. This second
pass through the AWG remultiplexes them onto two output ports connected to the out-
put fibers. A demonstration device was reported for four wavelengths on a 100-GHz
grid, fabricated using silica on silicon PLC technology. Its performance was as follows:
cross-talk of less than −10 dB (largely due to manufacturing imperfections in the MZ
switches), fiber-to-fiber insertion loss of 7.3–8.5 dB, and a switching time of 1 ms.
A third example is a 1 × 1 (ON/OFF) multiwavelength switch called a channel selec-
tor. It uses semiconductor PLC technology to integrate AWGs with SOA gate switches
[Yoshikuni02]. The logical architecture is the N wavelength 1 × 1 case of Figure 2.21,
with an array of N ON/OFF SOA gates placed between a WDMUX and a WMUX,
each implemented as an AWG. Reported performance of an eight-wavelength device
on a 200-GHz grid is cross-talk of less than −40 dB, an extinction ratio greater than
 Enabling Technology 273

50 dB, and no net insertion loss because of the gain of the SOAs. However, the SOAs
are polarization sensitive, introducing a PDG of 1–2 dB.

4.10.5.6 Wavelength Dilation

The three-stage architecture of Figure 2.21 for a WSS requires m space switches for
a system operating on m wavelengths together with a pair of WDMUX/WMUXs for
each input/output port. The space switches can be of any form, either generalized or
permutation. In the latter case, the WSS becomes a wavelength-selective cross-connect.
As pointed out in Section 2.3.2.4, any space-switch fabric can be converted to a WSS
fabric by replacing each elementary switching device with an equivalent MWS. In this
approach, no WDMUX/WMUXs are required. In the case at hand, the m-fold replication
of space switches is replaced by a fabric of MWSs operating on m wavelengths.
Unfortunately, many types of elementary (2 × 2) MWSs are either costly or impossible
to realize when m is large and/or the wavelength spacings are small. A compromise
solution is a wavelength-dilated architecture, as shown in Figure 4.66 [Sharony+92].
At first glance, this appears to have the worst qualities of both the three-stage and the
MWS approach. It is a three-stage architecture similar to Figure 2.21, but each space-
switching layer is replaced by an MWS fabric. The advantage is that there are fewer
MWS layers in the wavelength-dilated architecture than there are space switches in the

λ1, λ4, λ7

MWS1
WDF
1
2
8×8
1
In 1 2 Out 1
λ1 − λ8 2, 5, 8 2, 5, 8 λ1 − λ8
3
3, 6 3, 6

λ2, λ5, λ8

In 2 Out 2
MWS2

In 8 λ3, λ6 Out 8

MWS3

Figure 4.66 Wavelength-dilated switch.

274 Multiwavelength Optical Networks

configuration of Figure 2.21. Furthermore, each MWS fabric operates on fewer and
more widely separated wavelengths than the single MWS fabric would process. This is
accomplished by “combing out” subsets of wavelengths with spacings that are multiples
of those in the original channel set.
In the example shown in Figure 4.66, an 8 × 8 switch operating on eight wavelengths
is shown, with a factor of three wavelength dilation. It requires three 8 × 8 MWS fabrics,
MWS1 , MWS2 , and MWS3 . The wavelengths on each input fiber are demultiplexed using
a tree of wavelength-dropping filters (WDFs) and are remultiplexed using a similar tree
operated in the reverse direction. The wavelengths dropped/added by each WDF are
indicated with an arrow under the WDF. The set of wavelengths λ1 , λ4 , and λ7 is
directed from each input fiber to MWS1 ; the set λ2 , λ5 , and λ8 is directed to MWS2 ;
and λ3 and λ6 are directed to MWS3 . Note that each MWS must operate on two or three
wavelengths with spacings that are three times as large as the original set. Thus, large
guard bands are present between adjacent wavelengths, reducing interchannel cross-talk
as well as the cost of the components. Because the WDFs are static devices (e.g., fiber
Bragg gratings) their cost is considerably less than the switching devices, which have
been reduced in number.

4.11 Wavelength Conversion and Signal Regeneration

Wavelength converters can be used as components of WIXCs and (in combination

with static wavelength routers) to implement WSXCs (see Section 2.3.3). The use of the
WIXC as a means of removing bottlenecks in wavelength-routed networks is illustrated in
Figure 2.28. This improves performance in the higher layers of the network architecture,
reducing blocking in dynamic routing, increasing overall load carrying ability, and
making protection and restoration simpler. Although the wavelength conversion function
is not an inherently nonlinear operation, a wavelength converter may be either transparent
or opaque depending on whether it is implementated using “weakly” or “strongly”
nonlinear devices.
Signal regeneration is an operation that cleans up digital signals that have been
distorted and contaminated with noise as they traverse a path through a network. There
are three levels of regeneration: 1R (reamplification), 2R (reamplification and reshaping),
and 3R (reamplification, reshaping, and retiming). 1R is really not regeneration at all
and will not be discussed further. The other two are highly nonlinear operations; they
cannot be performed on superimposed channels. Regeneration increases the maximum
bit rate that can be supported on an optical channel for a given BER or, equivalently,
decreases the BER at a given bit rate, without necessitating a reduction of the signal
impairments along the optical path. The result, which is felt at the physical layer, is
to effectively extend the reach of the transmission path. Although linear amplification
can improve transmission performance by increasing the OSNR, it is the nonlinearity of
the regeneration process that enables it to improve performance in a way that no linear
device can do. Thus, for example, the OSNR at the receiver, and resultant BER at the
end of a long optical path can be improved by placing optical amplifiers at frequent
 Enabling Technology 275

o e e o

PD
Amplifier
Input
λ1

Laser Modulated
Modulator Output
λ2
λ2

Figure 4.67 Optoelectronic wavelength converter.

intervals along the path (see Section 4.4.1.4). However, the resultant improvement in
BER (compared to no amplification) will not be as much as if regenerators are used at
the same points.
Given the current state of electronic as compared to photonic technology, the simplest
and most cost-effective way of implementing wavelength conversion and regeneration
is in the electronic domain, first converting the signal to electrical form, doing the signal
processing electronically, and then reconverting it to optical form. This OEO approach
is shown conceptually in Figure 4.67. An intensity-modulated signal at wavelength λ1 is
converted to electrical form in a photodetector, amplified, and used to modulate a laser
operating at a different wavelength λ2 . (This is basically a bare-bones transponder, as
described in Section 2.5.) It will accept a signal at any wavelength within the range of
the detector, and output a signal at wavelength λ2 , independent of the input wavelength.
Because the photodetector is a square-law device this arrangement is inherently non-
linear and hence opaque: two superimposed signals at different wavelengths cannot be
converted simultaneously. An advantage of this device is that it can easily be modified to
perform regeneration. By inserting electronic pulse-shaping circuits after the amplifier,
this becomes a 2R regenerator, and if clock recovery and gating circuits are added, 3R
regeneration results. However, this makes the device signal format dependent; the range
of acceptable bit rates and modulation formats is constrained by the electronic circuitry.
Furthermore, the electronic approach has a speed limitation and it is power hungry, with
its cost increasing rapidly as the electronic speed limits are approached. Given these cost
and performance limitations, all-optical approaches, where signals remain in the optical
domain, are becoming an increasingly viable alternative to the OEO method.

4.11.1 All-Optical Wavelength Conversion

All-optical wavelength conversion can be performed as either an opaque or transparent
(nearly linear) operation. The methods of implementing opaque conversion typically
276 Multiwavelength Optical Networks

depend on electro-optic phenomena in materials such as semiconductors, involving

strong nonlinearities. As in the OEO converter, the output wavelength is independent of
the input wavelength. On the other hand, transparent conversion, known also as coher-
ent, wave-mixing, or parametric conversion, depends on relatively weak nonlinearities.
Viewed from the point of view of the optical channel transparent conversion behaves
linearly, so several superimposed λ-channels can be converted in the same device. As
opposed to the opaque converters, the wavelengths of the input and output signals of
a transparent converter are related. The parametric approach is typically an order of
magnitude faster than the electro-optic approach (subpicosecond rather than tens
of picoseconds), but it consumes more power. Because opaque conversion depends
on strong nonlinear effects, it lends itself naturally to regeneration as well. We con-
sider the transparent case here (for conversion only), followed by opaque systems for
conversion and regeneration in the next section.
Transparent devices for wavelength conversion are based on FWM or difference
frequency conversion, which are produced by nonlinear effects in fibers and waveguides.
Although we have defined transparency as being the same as linearity, the weakly
nonlinear effects considered here together with filtering produce a system that behaves
linearly for all practical purposes and is thus a transparent converter.
The conversion process results from the mixing of signals at different optical fre-
quencies in a device having a power-law nonlinearity. If two signals at frequencies ω1
and ω2 are combined in a device with an nth-order nonlinearity, in the expansion of
the resultant output terms there is one of frequency (n − 1)ω1 − ω2 . This phenomenon
produces the cross-talk problems on fibers that were discussed in Section 4.3.2.4. In that
case, it was indicated that when two signals at neighboring optical frequencies ω1 and
ω2 are impressed on a fiber with a cubic nonlinearity, an undesired cross-talk signal at a
third frequency 2ω1 − ω2 can appear, which is called four-wave mixing. This effect can
be turned to an advantage for wavelength conversion. If we let ω1 = ω p , the pump fre-
quency, and ω2 = ωs , the signal frequency, and call ωc = 2ω p − ωs the converted signal
frequency, this becomes a frequency converter [Inoue+93]. Interchanging the pump and
signal frequencies, another possible frequency for the converted wave is ωc = 2ωs − ω p .
Although this is a transparent conversion process and therefore potentially very useful, it
is usually a very weak (inefficient) effect in fibers. However, high-efficiency conversion
has been achieved using highly nonlinear fibers (HNLFs) and high pump powers. Con-
version efficiencies up to −5 dB have been achieved with 26-dBm pump power, using
0.2 m of low-dispersion slope HNLF to shift the wavelength of a 160-Gbps picosecond
pulse train from 1556 to 1578 nm [Igarashi+04].
Tunability is an important property of a frequency converter. Using a bismuth-oxide
PCF as the HNLF in an FWM device, a widely tunable converter has been demonstrated
[Chow+07]. With 1 meter of PCF a converter tunable over 35 nm was built, operating on a
10-Gbps NRZ signal with a power penalty of 1.2 dB and conversion efficiency of −19 dB.
The difference frequency converter (DFC) was discussed in Section 2.3.3 [Yoo+95,
Yoo+96]. This is based on a square law rather than cubic nonlinearity. The efficiency
of the effect depends on how well phase matching is realized. It can be enhanced by
special design of the waveguide in which the interaction occurs. Figure 4.68 shows
 Enabling Technology 277

λ1 λ2 2λp λ2′ λ1′

Relative Power (dB) 5

−5

−15

−25

−35
1520 1530 1540 1550 1560
Wavelength (nm)
(a)

−16
Conversion Efficiency (dB)

−19

−22

−25

−28
1490 1520 1550 1580 1610
Wavelength (nm)
(b)

Figure 4.68 Performance of a difference frequency converter. (From [Yoo+96, Figure 2]. Copyright


c 1996 AIP. Used by permission of The American Institute of Physics.)

the performance of a DFC based on a periodically domain-reversed aluminum gallium

arsenide (AlGaAs) waveguide. Figure 4.68(a) illustrates the simultaneous conversion of
two signals at 1528 and 1534 nm to 1555 and 1549 nm, respectively. Note the spectrum
“reflection” based on the pump wavelength (771 nm), as described in Figure 2.24. The
various processes involved in fabrication render the device polarization independent,
with a conversion efficiency of approximately −17 dB over a band of approximately
90 nm, as shown in Figure 4.68(b). (The solid and dashed curves in the figure represent
points for two different polarization states.) More recently, using periodic poling (pro-
ducing the domain reversal effect mentioned above) in a LiNbO3 waveguide, better than
−10 dB conversion efficiency has been achieved for wavelength shift of 60–70 nm at
data rates up to 160 Gbps [Brener+00].
Because the parametric converters reflect the signal spectrum, they reverse its phase
as a function of frequency (phase conjugation). This makes them useful for dispersion
278 Multiwavelength Optical Networks

compensation (see Section 4.3.2.3). For example, dispersion compensation using a DFC
placed halfway down a link has been demonstrated successfully at 100 Gbps over
160 km of dispersive fiber. The performance of DFC and FWM wavelength converters
may improve as highly nonlinear PCFs become available for this application.

4.11.2 Opaque Wavelength Conversion and Signal Regeneration

The family of opaque devices and subsystems for wavelength conversion and regener-
ation is large and diverse, but a conceptual picture that encompasses most of them is
shown in Figure 4.69. It is assumed throughout that the data format is intensity modu-
lation. A data signal at wavelength λ1 and a probe signal at wavelength λ p are applied
to the input(s) of a gating device. The data act as a control, turning the gate on or off to
allow or prevent the passage of the probe signal. In the case of 2R regeneration, the tap,
timing recovery, and clock are missing. The probe is then a CW signal, which is gated
on when a data pulse is present and off when a pulse is absent. For 3R regeneration, a
small portion of the data signal is tapped off to be used for timing recovery. Using the
timing information an optical clock signal synchronized to the data is generated in the
form of a pulse train at wavelength λ p . Sometimes the synchronization signal is applied
as a third input to the gate. After gating by the incoming data signal this becomes the
regenerated bit stream. (Note that a similar clock recovery arrangement in an optical

Data Signal λ1 Op
Optical
Gating Output λp
Tap device

Timing
Ti
Clock
Clock Probe signal λp CW/pulsed
recovery

Data
Threshold

Probe:
CW
Pulsed

Output

Figure 4.69 Opaque conversion and regeneration.

 Enabling Technology 279

HNLF

in out
SA OBPF

Figure 4.70 SA-based regenerator. (From [Matsumoto+02, Figure 1]. Copyright
c 2002 IEEE.
Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

receiver is shown in Figure 4.31. However, that is performed on the electrical side, after
detection of the optical signal by the PD.)
A commonly used configuration uses an SOA as the gating device, where the data
signal and CW probe are superimposed as a common input. The data pulses representing
1s are inserted at high-enough intensity to drive the SOA into saturation, and thus
as a consequence of cross-gain-modulation (XGM), the intensity of the probe signal
appearing at the output is low. With no input pulse present, the probe is amplified,
and thus a replica of the data signal appears at the output (with inverted polarity) at
wavelength λ p . (The data signal at wavelength λ1 must be removed by filtering at the
output of the device.) Because XPM is also present in the SOA, the output signal is
chirped, and this can be used to advantage to reformat the output pulse using appropriate
filters [Leuthold+04]. Furthermore, using a pulsed probe signal this configuration can
be used for 3R regeneration [Chayet+04].
Similar results are obtained if the gating device in Figure 4.69 is a saturable absorber
(SA). If the probe signal is eliminated and pulse-reshaping devices are included at
the output of the SA, this becomes a 2R regenerator without wavelength conversion
[Matsumoto+02] (see Figure 4.70). The SA has the property that it is opaque at low
input intensities and transparent at high intensities. Thus, it has the effect of blocking the
low-intensity noise fluctuations in the data signal when no pulse is present while passing
the pulses without attenuation. Although this eliminates noise when pulses are absent,
it does not reshape the data pulses. To do reshaping the pulses at the output of the SA
are passed through a HNLF followed by an optical bandpass filter (OBPF). The HNLF
produces SPM that, in combination with anomalous dispersion in the fiber, results in
pulse compression. The narrowband filter has the effect of equalizing pulse amplitudes
coming out of the HNLF, all of which results in 2R regeneration.
Better performance in regeneration is obtained with interferometric gating devices
such as the MZ configuration of Figure 4.71. In one arrangement, an SOA is used as
the nonlinear device (NL) in one arm of the MZ, and the input data and probe signals
are applied to the SOA at wavelengths λs and λ p , respectively. As a result of XPM, a

Input (λs)
NL

Probe (λp) OF Output (λp)

MZ

Figure 4.71 Nonlinear Mach–Zehnder regenerator.

280 Multiwavelength Optical Networks

4.5 km
DCF

3 dB
input signal coupler
PC

3 dB coupler

probe modulated probe

Figure 4.72 Nonlinear optical loop mirror regenerator. (From [Yoo96, Figure 6(b)]. Copyright

c 1996 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

pulse present in the data will cause a phase shift of the probe signal in the SOA arm.
The filter OF at the output removes the data signal and leaves the gated probe carrying
a replica of the data signal. If the MZ is designed for noninverting regeneration, there is
no transmission when no data pulse is present, and maximum transmission at the peak
of a pulse. This is a 2R regenerator and wavelength converter, with output wavelength
λ p . It can be upgraded to 3R regeneration using a pulsed probe. These systems can be
designed with counter propagating data and probe signals so that no filtering is required
to remove the data signal at wavelength λs from the output.
Another interferometric regeneration and conversion device is the nonlinear optical
loop mirror (NOLM). As shown in Figure 4.72 a probe is launched into the loop via a
3-dB coupler, splitting into equal clockwise and counterclockwise components. These
normally interfere destructively at the output and the probe is extinguished there. If,
however, the phase relation between the two components is altered, the interference is
no longer destructive and the probe appears at the output. This differential phase shift
can be induced under the control of a data signal, introduced in the coupler shown in the
figure. (Polarization controllers, PC, are shown because this is a polarization-dependent
device.) Fiber nonlinearity in the loop (typically several kilometers long) causes XPM
between the data signal and the probe thereby changing the phase relations between the
two counter-circulating components. This produces a gating effect on the probe, causing
a regenerated pulse at the probe wavelength to appear at the output. There is currently
renewed interest in NOLM devices for wavelength conversion and regeneration, because
of the possibility of using highly nonlinear microstructured fibers in the loop, greatly
reducing the necessary fiber length.
Another configuration of the optical loop mirror is the SLALOM (semiconductor laser
amplifier in a loop optical mirror), where the fiber nonlinearity is replaced by an SOA
[Eiselt+95]. This device is capable of various nonlinear operations on optical signals,
including 3R regeneration, two-pulse correlation, time division demultiplexing, and all-
optical gating, where one optical pulse controls the switching of another. An application
of the SLALOM to all optical header processing for optical packet switching appears in
Section 10.2. An advantage of the SLALOM over the NOLM is that it does not require
a long nonlinear fiber.
 Enabling Technology 281

Comparing the various approaches to wavelength conversion, FWM and the DFC have
performance advantages in WDM networks because they are transparent and hence can
handle multiple wavelengths simultaneously. They also have the best noise properties.
However, fabrication is difficult. Among regenerators/converters, interferometric devices
depending on XPM have a performance edge over devices depending on XGM or
attenuation saturation (SOAs and SAs) due to the better extinction ratios in the former.
On the other hand, SOAs using XGM have some practical advantages: conversion gain
and a simple structure. Whether wavelength conversion and/or signal regeneration is
appropriate in any particular application depends, of course, on various trade-offs. Does
the importance of maintaining transparency outweigh the signal quality improvements
obtained with regenerators? Do the performance improvements gained at the physical
level by regenerators and at the network level by wavelength converters justify the cost
of these devices? The answers to these questions continue to change as the technology
evolves.

4.12 Optical Switch Architectures

The conceptual view of switch architectures presented in Section 2.3 focused on generic
functionality while intentionally avoiding details of specific physical implementations.
That is, it gave us an abstract concept of a switch without being overwhelmed with the
details of implementation. However, with the description of the underlying switching de-
vices behind us we can now move on to physical implementation, which requires a more
concrete picture. In the next sections, we describe some alternative physical realizations
of switches that produce the desired functionality taking into account technological
constraints and cost-performance trade-offs. Space switch fabrics are considered first,
followed by wavelength-selective switches. Because space switches usually form the
core of WSSs, the performance of the former are a key to the performance of the latter.

4.12.1 Space Switches

Some representative space switch architectures were described in Section 2.3.2, char-
acterized as permutation switches, generalized switches, and linear divider-combiners
(LDCs). In this section, we revisit the space switch, focusing on physical layer issues:
losses, cross-talk, and fabrication complexity. We begin with crossbar architectures,
followed by multistage switches and a discussion of switch fabric cross-talk. Although
some of these structures can be used as generalized switches and LDCs, we focus the
discussion here on permutation switch functionality.

4.12.1.1 Crossbar Switches

The crossbar switch of Figure 2.14 dates back to the electromechanical telephone switch-
ing systems of the 1950s. A connection between an input-output pair was made using
electromagnetic actuators to move the corresponding horizontal and vertical metal cross-
bars to make contact at the cross-point. The characteristic feature of a crossbar switch
282 Multiwavelength Optical Networks

Input

Output
1’ 2’ 3’ 4’

Figure 4.73 Optical crossbar switch.

is that each connection is made by closing a single cross-point, which means that n 2
cross-points are required for an n × n (nonblocking) switch. There are many ways of
implementing this concept in the optical domain, with significant differences in perfor-
mance. Three possible realizations of an optical 4 × 4 crossbar are shown in Figures 4.47,
4.73, and 4.74, which we refer to as OXB1, OXB2, and OXB3, respectively. All of these
lend themselves to PLC technology. Although at first glance they appear quite different,
their common characteristic is that a single elementary switching device is set for each

1’

2’

2
Input Output
3

3’

4’

Figure 4.74 Path-independent loss crossbar switch.

 Enabling Technology 283

connection through the switch. The version of OXB1 realized with SOA gate arrays
(on-off switches) and two-stage splitters and combiners has already been discussed in
Section 4.10.1.1. The splitting and combining losses, which are its main drawback, are
compensated by amplifiers in this case, but large switches of this type are not feasible
because the combining and splitting losses grow as n 2 .
OXB2 and OXB3 are based on arrays of 2 × 2 switches. In OXB2, if the switch at a
cross-point is OFF, light from an input guide continues on its horizontal path through the
cross-point. If the switch is ON, light is deflected from the horizontal guide, downward
on the vertical guide as indicated by the bent arrow symbols at the cross-points. When
all couplers are in the OFF state, there are no active connections through the switch.
Note that the 2D MEMS switch of Figure 4.49 does in free space what OXB2 does
using waveguides, where a MEMS mirror set in the horizontal (vertical) position is
equivalent to an OXB2 switch in the OFF (ON) state. Furthermore, note that what is
required at each cross-point in OXB2 is an incomplete 2 × 2 switch of the type shown in
Figure 4.44, where the OFF state allows the two guides to cross without interchanging
signals, and the ON state diverts power from the input (horizontal) guide to the output
(vertical) guide while dumping any power that might be incoming on the vertical guide.
(A [one-sided] MEMS mirror has exactly the same characteristics.) No connection is
ever required from the vertical guide to the horizontal guide. Insertion loss in OXB2
is better than in OXB1, but its losses are not uniform, the longest path in Figure 4.73
traversing seven couplers, and the shortest traversing only one. This inequality of path
length also produces unequal levels of cross-talk on each path.
In contrast to OXB2, all paths in OXB3 are the same length and traverse the same
number of switches, so that it has path-independent insertion loss [Shimoe+87], a
significant advantage. As laid out in Figure 4.74, OXB3 bears more resemblance to a
multistage switch than a crossbar. However, it operates like a traditional crossbar in
that only one cross-point is activated to establish each connection. In the 4 × 4 example
shown in the figure, light propagates from left to right in input and output waveguides
that are coupled to each other by an array of sixteen 2 × 2 switches, just as in OXB2.
Now, each guide has a bend in it to provide the left-to-right signal flow. As in OXB2,
light remains in its guide when passing through a cross-point whose switch is set to the
OFF state, but light is diverted from an input guide to an output guide when the switch
is in the ON state. Each path traverses four switches. Once more, the incomplete 2 × 2
switch is useful here, as suggested by symbols at the cross-points indicating the direction
of signal flow in the ON state. We leave it to the reader to verify that if incomplete 2 × 2
switches are used at the cross-points of OXB2 and OXB3, there is a unique path, with
one ON switch setting, for each input-output connection, and the switch is strict sense
nonblocking. (The OXB3 configuration is also adaptable to 2D MEMS.)
It is interesting to note that if complete 2 × 2 switches are used at the cross-points
in OXB2 and OXB3, more than one possible path may exist for a given input-ouput
connection, but badly chosen paths will lead to blocking. Hence, these switches are
wide-sense nonblocking when built with complete 2 × 2 switches.
High-performance OXB3-type switch fabrics have been constructed as monolithically
integrated structures using silica PLC technology [Goh+99, Kasahara+02]. An 8 × 8
284 Multiwavelength Optical Networks

#8
#3

85 mm
#4 #7
#2 #5

#6
#1

85 mm

Figure 4.75 Circuit layout for 8 × 8 optical crossbar switch. (From [Kasahara+02, Figure 15].
Copyright
c 2002 IEEE. Used by permission of The Institute of Electrical and Electronics
Engineers, Inc.)

example is shown in Figure 4.75. It is fabricated on an 85 × 85-mm silica chip, where

each rectangle in the figure contains one stage of eight couplers. Each coupler is a
thermo-optic MZ switch. Using two-stage switches to improve extinction ratio and
reduce cross-talk, 8 × 8 units have been demonstrated with the following characteristics:
switching time of 2 ms, aggregate cross-talk of −43 dB (assuming equal input powers),
average insertion loss of 5.1 dB, and extinction ratio of 60.3 dB.

4.12.1.2 Routers/Selectors
Another switch fabric, shown in Figures 4.76 and 2.19, is known as a router/selector.40
These figures show identical topologies, but the switching elements are grouped dif-
ferently, with the former having fewer connection crossovers. Thus, in Figure 4.76, all
switches can be fabricated as an integrated component on a single substrate with only
12 waveguide crossovers. These structures can be implemented with both the splitting
and combining stages, made up of switching devices, or with one or the other of the stages
implemented with static devices, saving half of the switches [Spanke87]. However, the
latter case exhibits poorer insertion loss and cross-talk properties (see Section 4.12.1.4).
Because all required switches are 1 × 2 or 2 × 1, this architecture is adapted to Y-branch
implementation. A monolithic 4 × 4 semiconductor PLC device based on Y-branch
switches has been demonstrated using current injection. It is polarization insensitive
and has −13 dB cross-talk for each switch. In their unbiased state, the switches act as
3-dB couplers, so this fabric can be operated as an LDC, with broadcast and multiaccess
capability.

40
The δ−σ architecture, shown in Figure 2.19(a) as a 4 × 4 LDC, is identical in form to the router/selector.
The only difference is that the elementary switches in the LDC are allowed to run through a continuum of
states, whereas those in the router/selector are used only in the cross and bar states.
 Enabling Technology 285

Figure 4.76 Router/selector.

4.12.1.3 Benes Switches

So far we have focused on strictly nonblocking switch configurations, which typically
have more controllable devices than necessary but have good cross-talk properties.
Rearrangeably nonblocking switches, exemplified by the Benes family, are more eco-
nomical with respect to elementary switching devices but usually at the cost of increased
cross-talk.
An n × n Benes switch has 2 log2 n − 1 stages with n/2 2 × 2 switches per stage.
Figure 4.77 shows a 4 × 4 Benes fabric. The 8 × 8 version appears in Figure 2.17. As
indicated in Section 2.3.2.1, the Benes fabrics can be converted to strictly nonblocking
Cantor switches by using log2 n-fold space dilation. Space dilation is also useful for
reducing cross-talk. For example, if only one of the two ports on each input element
is used and only one of the two output ports on each output element is used, the 8 × 8
Benes switch of Figure 2.17 becomes a twofold dilated 4 × 4 switch with better cross-talk
properties (see [Hinton+93]). Benes optical switches have been fabricated in various
technologies. An 8 × 8 dilated Benes switch is described in [Watson+90]. Fabricated
as a LiNbO3 PLC using directional couplers, it was designed for operation in time
division-switching applications, with a switching time of 2.5 ns and cross-talk of less
than −20 dB.
Another more ambitious example is a 64 × 64 Benes switch implemented using
liquid crystal technology [Noguchi98]. This uses the polarization-deflection approach

Figure 4.77 Benes switch.

286 Multiwavelength Optical Networks

employed in the liquid crystal MWS of Figure 4.63, in an 11-stage Benes fabric. The
64 inputs are arranged in an 8 × 8 array and transmitted through a series of 11 liquid
crystal SLMs interposed between calcite crystal beam routers. The SLMs implement
the switching function by rotating the polarization of the beams under electrical control,
and the beam routers together with arrays of lenses direct the signals between stages
based on their polarization states. The complete switch fabric is polarization insensitive,
with average insertion loss of 9.5 dB, average cross-talk of −25 dB, and average PDL
of 0.21 dB.

4.12.1.4 Switch Cross-Talk

As indicated in Section 4.9.1, cross-talk can occur in a network when interfering signals
travel on unintended “leakage” paths and interfere at a receiver. But a switch fabric is
itself a connection network, and because of the many paths in a typical switch, many
cross-talk terms can be generated.
Switch cross-talk terms can be categorized by their order. As indicated in
Equation (2.13), when any real 2 × 2 switching element is in the bar or cross state,
a fraction ǫ of the input signal power leaks into the wrong output port. (This also holds
for 1 × 2 elements such as the Y-branch switch.) A leakage path through a switching
fabric that passes through k of these “wrong-way” hops attenuates the power of a signal
traversing it by an amount kǫdB . This is called kth-order cross-talk. Clearly, the higher
its order, the less troublesome the cross-talk; so an important figure of merit for a switch
fabric is the lowest possible order of cross-talk that can exist in it.
Examples of cross-talk terms of various orders in three different realizations of 4 × 4
switches are shown in Figures 4.78 and 4.79. Intended paths are shown in bold and
leakage paths are shown dashed. All connections are on the same nominal wavelength.
In each case, a small portion of a signal applied to an input port leaks over to an un-
desired output port. In the Benes switch of Figure 4.78(a), a portion of the signal on
input port 3 leaks to output port 1, passing through one wrong-way hop, which makes
this a first-order heterodyne cross-talk. In addition, there is a second-order homodyne

1 1
2 2 3

3 3 4
4 4

(a) Benes Architecture (b) Router/Selector Architecture

Figure 4.78 Orders of cross-talk.

 Enabling Technology 287

Cross-talk
Reduction
Active Switches Active
Dump
Splitters Combiners Active
Switches
Passive Combiners
Splitters

Figure 4.79 Enhanced performance switch.

(multipath) cross-talk term due to the leakage path from input port 1 to output
port 1. A second-order heterodyne cross-talk path is shown in the router/selector switch
of Figure 4.78(b), where the three switches on the dashed path are assumed to be set in
the bar state, and a third-order heterodyne path is shown in the enhanced performance
switch of Figure 4.79. In each case, the switch architecture is structured so that the orders
shown in these illustrations are the lowest possible for the switch.
In general, increasing the order of cross-talk requires adding additional switching
elements. For example, the 4 × 4 enhanced performance switch is constructed with
44 controllable, active switching elements, as well as four static splitters, to limit cross-
talk to third-order and above [Murphy+97]. For comparison, the Benes switch has
first-order cross-talk but requires only six elements, whereas the router/selector switch
has only second-order cross-talk and higher and requires 24 elements. In the case of
the router/selector, one might economize on switching elements by replacing the first
two stages of switches in Figure 4.78(b) by a fixed dividing network consisting of 3-dB
splitters. However, this would raise the level of cross-talk on the dashed path to first
order. Even though there would be an additional attenuation of 6 dB on that path, this
is significantly poorer cross-talk performance than in the version with 24 switching
elements.
It is worth noting that the fabrics that are most economical in element usage generally
produce multipath cross-talk, which has the most damaging effect on signal quality
(see Section 4.9). Conversely, switches designed with extra elements usually have no
multipath cross-talk. For example, we have seen that the Benes switch, as shown in
288 Multiwavelength Optical Networks

Figure 4.80 Space dilation.

Figure 4.78(a), has a multipath cross-talk path in parallel with the desired path between
input 1 and output 1. No such paths exist in the router/selector and enhanced performance
switches.
One way of increasing systematically the order of cross-talk in any switch fabric is by
space dilating each 2 × 2 switch. As shown in Figure 4.80, a set of four 2 × 2 switches
can be configured to produce a single space-dilated 2 × 2 switch with second-order
cross-talk. (Only one input is used on each switch on the left, and only one output is
used on each switch on the right, so that Y-branch switches can be used instead of 2 × 2
switches.) A complete fabric of these space-dilated switches would have the order of its
cross-talk increased by a factor of 2, but with a component count increased by a factor
of 4. For example, using space dilation for each elementary switch, the 4 × 4 Benes
switch requires 24 elements, which is the same number as the 4 × 4 router/selector
switch.
In addition to having less cross-talk, switch fabrics with extra elements are usually
more versatile. Thus, for example, we have seen that the router/selector configuration can
be used as an LDC with one-to-many (multicast) as well as many-to-one connections,
provided that its elementary switches are allowed to assume intermediate states. (The
gate array crossbar of Figure 4.47 has this property using only two-state elementary
switches.) Furthermore, cross-talk levels can often be reduced by appropriately setting
unused switching elements; these have “don’t care” states. For example, with the switches
set as shown in Figure 4.78(b) cross-talk leaks along a path from input port 2 to output
port 1. It is second order because it passes through two wrong-way hops. However, if
the second (unused) switch on this path is set in the cross state, the cross-talk is reduced
to third order.

4.12.2 Wavelength-Selective Switches

The generic wavelength-selective switches and wavelength-interchanging switches dis-
cussed in Section 2.3 are commonly known as optical cross-connects (OXCs). We
examine these in more detail now, beginning with the wavelength add/drop multiplexer,
also commonly known as an optical add/drop multiplexer (OADM) (a 2 × 2 special case
of the OXC) and concluding with fully configured OXCs.41

41
As networking evolves, terminology evolves with it. For consistency we generally use the term WADM in
referring to a generic wavelength add/drop multiplexer. However, in recent applications exemplified by the
cases that follow, usage has shifted from WADM to OADM. We use WADM and OADM interchangeably
henceforth.
 Enabling Technology 289

4.12.2.1 Optical Add/Drop Multiplexers

The OADM has emerged as one of the key building blocks for networks at the metro
and regional level. Used for adding and dropping wavelengths from intermediate points
along a transmission link or a ring, the role of OADMs in a typical network is illustrated
in Figure 4.81. They offer a simple and effective way to access a general mesh network
along links or within ring subnets. The OADM can be static or reconfigurable, and in
the latter case it is sometimes called an ROADM.
Initially developed in the early transparent optical network testbeds of the 1990s,
OADMs are now deployed in the current network infrastructure. Two fundamental
designs are the parallel and the serial architectures shown in Figures 4.82(a) and 4.82(b),
respectively. The 2 × 2 controllable couplers in the parallel realization make this a
reconfigurable OADM. The variable optical attenuators shown in the figure are used to
equalize the individual λ-channel power levels, which may vary due to upstream signal
power divergence or losses within the OADM. Because of the insertion loss involved
in the demultiplexing process, the parallel design is more suitable for high add/drop
channel counts, whereas the serial design is limited to applications with low add/drop
channel counts. The add/drop modules in the serial design are normally fixed devices,
but a practical advantage of this arrangement is that individual modules can be added as
needed. Both designs are well known and widely utilized [Wagner+96].
An OADM design based on hierarchical, multistage optical multiplexing and demul-
tiplexing of wavebands and wavelengths is shown in Figure 4.83. The motivation behind
distinguishing between bands and wavelengths is that in many cases it is not necessary
for every wavelength to be demultiplexed/multiplexed at every node if it carries no traf-
fic for that specific node and consequently all or some of the wavelengths should have
the ability to bypass any optical node. Optical node bypassing is a very cost-effective

OADM

OXC

Figure 4.81 OADMs in a network.

290 Multiwavelength Optical Networks

λ1

ED
EDFA λ2 EDFA
D

MUX
M

DMUX
M
U
U
X
X
λm

(a) Parallel

Wavelength Wavelength
add/drop add/drop

λ1 λm

(b) Serial

Figure 4.82 Parallel (a) and serial (b) OADM architectures with capability for m wavelength
add/drops.

WWavelength access
adding/dropping/switching
…. ….

wavelength
Band module n
DMUX .
. .
Band module 2
Band module 1

Band
DMUX
Band access
access

working fiber
West
Fiberraccess
Fibe access East
protection fiber

Figure 4.83 Functional diagram of an OADM based on wavebands and wavelengths. (Adapted from
[Saleh+99, Figure 12]. Copyright

c 1999 IEEE. Used by permission of The Institute of Electrical
and Electronics Engineers, Inc.)
 Enabling Technology 291

solution because (1) it avoids the use of more expensive equipment that might not be
needed (i.e., individual channel MUX/DMUXs) and (2) saves on amplification costs
because it results in reduced insertion losses on an optically transparent path to its
destination. At the same time, the band-level demultiplexing/multiplexing capability
allows for a cost-effective modular upgrade approach.
In this device, a band WDMUX/MUX combination first separates and combines sig-
nals at the waveband level. In combination with a manual patch panel this allows for
selective waveband add/drop or pass-through. Each dropped waveband goes through an
additional level of WDMUX/MUX at the wavelength level implemented in an individual
band module, and these modules can be added on an as-needed basis enabling modular
node expansion. Each band module can have its own channel-conditioning capability
such as amplification (dark triangles), attenuators for power equalization, and possi-
ble optical dispersion compensation modules. Switching at the wavelength level can
be performed manually or controlled electronically. The number of band modules that
are required can be minimized by concentrating the wavelengths being add/dropped at
the node into the smallest possible number of wavebands. A grooming process for this
purpose is discussed in Section 4.12.2.2 and Chapter 7.
Because protection is essential when OADMs are deployed in the field, the configura-
tion shown in Figure 4.83 is shown with two fibers: working and protection, along with
line amplifiers. The box labeled fiber access allows for bypassing the OADM as well
as implementing (1 + 1) protection in a ring network. The multiplexed signal outbound
from the band WMUX is bridged between the eastbound working fiber and the west-
bound protection fiber. Also, the signals inbound on the working and protection fibers are
monitored, and when a failure is detected on the working fiber the WDMUX is switched
over to the protection fiber, which carries a copy of the signal on the working fiber.
Another OADM architecture called Broadcast & Select (B&S) [Boskovic+02]
is shown in Figure 4.84. Essentially it replaces the WDMUX/MUX and switching
functions in Figures 4.82 and 4.83 by tunable receivers and transmitters. A reconfig-
urable wavelength-selective device can selectively pass or attenuate the power of any
λ-channel on the pass-through path under external control, providing power equalizing
and blocking functionalities in a single unit. A device performing this function using
liquid crystal technology [Ranalli+99] has been manufactured under the name dynamic
spectrum equalizer (DSE). Its dynamic capabilities make the above design a fully re-
configurable OADM. All incoming traffic is split into the drop and the pass-through
directions using conventional 3-dB splitters. On the drop side power is further split
using 1:N passive devices where N is the number of accessing stations (clients). Optical
filters can then separate the intended traffic at each station with the possibility of access-
ing a number of different channels in a dynamic way if the filters are made tunable. The
DSE is then used on the pass-through path to block each channel that has been dropped
and selected on the first stage, whereas at the same time equalizing all the pass-through
channels. On the adding side, channels can simply be inserted by combining the powers
from a set of N transmitters (tunability is also possible). Signal conditioning is provided
using input/output amplifiers, the dynamic spectrum equalizer, and optical dispersion
compensators (not included in Figure 4.84).
292 Multiwavelength Optical Networks

DSE
DSE

A C
DSE
DSE

RX TX TX RX

F - Optical filters
S - 1 to N power splitter
A - Optical amplifier
C - 3 dB coupler
DSE - Dynamic spectrum analyzer

Figure 4.84 Typical B&S OADM architecture. Assumes (1 + 1) protection. (Adapted from
[Boskovic+02, Figure 2]. Copyright
c 2002 IEEE. Used by permission of The Institute of
Electrical and Electronics Engineers, Inc.)

When used in a ring, (1 + 1) protection can be provided using a bridge-and-select

scheme much like that described for the OADM of Figure 4.83, assuming that one
of the fibers is a working fiber and the other is a protection fiber (carrying the same
information in the opposite direction). Bridging of the output signals to the two output
fibers is accomplished by feeding the two transmitters with the same electrical signal,
and selection is done by monitoring the outputs of the two receivers and switching from
one to another in case of a failure.
Another B&S architecture is shown in Figure 4.85. It is similar to the previous one
except that the DSEs are replaced by 4 × 1 wavelength selective switches, which select
either pass-through channels from inputs 1 and 2 or add channels from inputs 3 and 4.

In 1 Out 1
4 x1 1
4x

In 2 Out 2
4x1
4x1

Out 3 Out 4 In 3 In 4

RX
RX RX
RX TX TX
TX TX
RX
RX RX
RX TX
TX TX
TX

Figure 4.85 Typical 4 × 4 ROADM based on 4 × 1 wavelength selective switch and B&S-type of
architecture. (Adapted from [Antoniades+04a, Figure 15]. Copyright
c 2004 IEEE. Used by
permission of The Institute of Electrical and Electronics Engineers, Inc.)
 Enabling Technology 293

In the figure the fibers are oriented to make this look more like a 4 × 4 switch than
an ROADM, and indeed it is. In an ROADM working into a bidirectional fiber pair, as
in Figure 4.84, there is no path from an input fiber to the other output fiber, only a
pass-through path to the same output fiber. This limited connectivity characterizes a
typical ROADM. In the architecture of Figure 4.85, however, all paths are possible
because of the 4 × 1 switches. Signals from input ports 1 and 2 are distributed to the
four output ports using 1 × 4 splitters. Outputs 3 and 4 can selectively receive λ-channels
from each fiber using 1 × N splitters and tunable filters just as described in the basic
B&S design of Figure 4.84. The number of passive splits/combines on the drop/add
sites is dictated by how many channels need to be accessed from each fiber. The higher
that number, the more insertion loss (and consequently amplification costs) will be
incurred. A realization of this architecture has been proposed using a 1 × 4 MEMS
switch described in [Marom+02] (see Section 4.10.5.3).
The ROADM architecture of Figure 4.85 is essentially a two-stage switch with a pas-
sive (static) first stage and a controllable second stage. This arrangement can be reversed.
A configuration has been proposed consisting of a controllable wavelength-dropping
stage, followed by a passive combining stage for the passed-through wavelengths and
the added wavelengths [Doerr+03]. The wavelength-dropping stage is realized using a
silica-based monolithic PLC.

4.12.2.2 Optical Cross-Connects

When examining OXC architectures it is important to distinguish between functionality
and implementation. Figures 4.86, 4.87, and 4.88 show typical implementations of an
OXC. The first is a transparent architecture, suitable for a node of a transparent network.

... ...
...
...

...
...

Multiple transparent
optical switch matrices

Figure 4.86 Transparent OXC. (From [Ellinas+04, Figure 1]. Copyright

c 2004 IEEE. Used by
permission of The Institute of Electrical and Electronics Engineers, Inc.)
294 Multiwavelength Optical Networks

... ...

WDM Transponder

...
...

Electrical
Electrical
Fabric (O-E-O)
(OEO)

...
...

Transceiver cards

SONET
SONET IP
IP ATM
ATM

Figure 4.87 Opaque O-E-O OXC. (From [Ellinas+04, Figure 3a]. Copyright

c 2004 IEEE. Used
by permission of The Institute of Electrical and Electronics Engineers, Inc.)

It is based on multiple layers of O-O-O switch fabrics, one for each wavelength. The
terminology O-O-O as used here means that the signals enter the switch fabric in
optical form, remain in the optical domain in the switch, and exit in optical form. (That
generally means that the signal is not modified in any significant way in passing through
the switch, although there may be exceptions if optical signal processing occurs.) Three
links are shown incident on the switch, each represented as a single fiber pair carrying

... ... WDM Transponders

...

...

Transparent
Transparent
Optical
Optical
Switch Matr
Switch ix
Matrix
(OOO)
...

...

(O-O-O)

Transparent optical line cards

SONET IP ATM SONET/SDH line-terminations

Figure 4.88 Opaque O-O-O OXC. (From [Ellinas+04, Figure 3b]. Copyright

c 2004 IEEE. Used
by permission of The Institute of Electrical and Electronics Engineers, Inc.)
 Enabling Technology 295

multiple wavelengths. They are demultiplexed/multiplexed on a wavelength basis and

each layer switches signals on the same wavelengths. (We assume for simplicity that
no wavelength conversion, either optical or electrical, occurs in the switch.) This is the
three-stage architecture of Figure 2.21, but with additional input/output ports on the
switch fabric layers for access stations (transceivers) connected to the network via this
OXC. Because signals remain in the optical domain at all times, with no wavelength
conversion, compatibility of all links and access stations connected to this OXC requires
that all λ-channels must be on a common set of wavelengths; for example, the ITU grid.
Compatibility may entail additional requirements as well, depending on the degree of
transparency. There is no such thing as an absolutely transparent network! For example,
the modulation format, power level, and signal bandwidth or bit rate may also be specified
or constrained in some way. Power and bandwidth constraints are obviously necessary
to prevent interchannel cross-talk. Signals inserted into the network on the access links
must comply to these constraints, and for this reason the interface been the access station
and the OXC in this case is called a compliant interface. In our generic terminology,
this is a WSXC. For an OXC with a total of n input/output ports (including the access
links), carrying m wavelengths, this has m switch fabric layers, each consisting of an
n × n space switch.
The next two implementations are opaque. Figure 4.87 shows an O-E-O switch fabric
serving the same set of links and access stations, but this time there is no continuity of
the optical signal. Signals entering the switch fabric all connect to it via short-reach fiber
interfaces operating in a standard format at a standard wavelength (usually 1.3 µm).
Signals entering on the links, which would normally be on the ITU wavelength grid,
must be demultiplexed on a wavelength basis, after which each λ-channel is converted
from its own wavelength to the standard format and wavelength by a transponder, which
is a receiver/transmitter combination including some degree of electronic signal regen-
eration. The inverse operation takes place in the transponders on the outbound links.
The presence of the transponders makes this an opaque node; because electronic pro-
cessing is involved here, the signals on the links must be transmitted and received by
the transponders at standard speeds and formats. It should be noted that the receiving
side of the inbound transponder and the transmitting side of the outbound transpon-
der are the long-reach transmitters and receivers discussed in Sections 4.5 and 4.6,
respectively. If they are working into very long or poor-quality links they must be high
performance (read expensive) devices. Input signals on the access links are assumed here
to be exchanged with the switch fabric in the standard short-reach format. The switch
fabric itself is electronic, which means that another O/E (E/O) conversion is needed
at each input (output) port of the switch fabric, necessitating a transceiver at each I/O
port. An opaque OXC with a total of n I/O ports carrying m wavelengths requires a
single mn × mn electronic switch fabric, and each port must run at the speed of the
λ-channel it is serving. This becomes difficult, power-consuming, and expensive in the
electronic domain as bit rates move beyond 10 Gbps toward 40 Gbps and beyond. It
should be recalled that an mn × mn switch is considerably more complex than m n × n
switches. However, a non-blocking mn × mn switch fabric allows any input wavelength
to be translated to any output wavelength on any output port, giving the opaque switch
296 Multiwavelength Optical Networks

WIXC functionality as compared to the lesser WSXC functionality of the transparent

switch.
A different opaque approach shown in Figure 4.88 eliminates the electronic fabric.
Now the switch is O-O-O, where the signals in standard short-reach optical format are
interfaced directly to an mn × mn optical switch, which could be realized in any of the
technologies discussed earlier in this chapter. The transponders remain, so that this is
still opaque, but there is no need for any high-speed electronics in the switch or at its
edge.
At higher speeds and large wavelength count, transparent switches are simpler, more
flexible, more reliable, and less costly than the opaque version. For example, no changes
are required if channel speeds or modulation formats are changed. To compare switch
complexity, suppose there are 4 I/O ports with 64 wavelengths per fiber. The opaque
switch fabric is then 256 × 256. In a Benes realization (rearrangeably nonblocking)
this requires 1920 2 × 2 switching elements, while the transparent version requires 384
(see Section 4.12.1.3). If a crossbar realization is used, the opaque switch requires over
65,000 switch points as compared to 1024 for the transparent switch. Furthermore,
the transponders add a significant extra cost in the opaque case. The advantage of the
transparent architecture is even more impressive if switching is done on a waveband
rather than wavelength basis, in which case the number of layers in the switch fabric
is proportionately reduced. Nevertheless, opaque switches have a number of practical
advantages. They automatically regenerate all signals passing through the switch thereby
preventing the accumulation of optical signal impairments, they can monitor the signal
overhead bits for control and management purposes, and their wavelength conversion
properties facilitate efficient use of network capacity and fault recovery. At this writing,
the cost and complexity associated with large high-speed opaque switches is not an issue
because demand has not yet materialized at the higher speeds and larger sizes. However,
that can change in the future. One way of exploiting the advantages of both types of
switches is in a translucent network, composed of transparent nodes mixed with opaque
nodes at strategic locations. Properly designed this approach can achieve the best of both
worlds.
A way of improving the effectiveness of an OXC architecture is by doing the switching
at both the waveband and wavelength levels. Figure 4.89 shows an example of a hybrid
switch, sometimes called a multigranular switch. Traffic on the incident links is grouped
into three wavebands, w1 , w2 , and w3 , and terminating traffic is on wavelengths λ13
and λ23 in band w3 . The switch has a hierarchical structure with a waveband-selective
top layer (WSXC) switching w1 and w2 and dropping w3 down to the wavelength
layer (λSXC). There, it is demultiplexed and switched at a λ-channel granularity. The
wavelengths λ13 and λ23 are dropped and added at the local access stations. To put this
in the context of a nonhierarchical architecture, imagine that the top layer, WSXC in
Figure 4.89, represents a transparent OXC of the form of Figure 4.86 with three optical
switch matrices, each operating on wavebands rather than wavelengths. (The individual
links are not shown.) Now suppose the add/drop ports for band w3 are connected to the
wavelength layer λSXC, which passes the demultiplexed terminating wavelengths on to
the access stations. In general this switch requires one λSXC for each band being dropped
 Enabling Technology 297

w1 w1
w2 WSX C
WSXC w2
w3 w3

w3 w3

Figure 4.89 Hybrid waveband/wavelength switch.

to the wavelength layer. To make this architecture efficient, the traffic must be properly
assigned to λ-channels and wavebands. Normally, most of the traffic passing through a
typical OXC is transit traffic. If the traffic assignments are grouped into wavebands in
such a way that the terminating traffic is contained in just a small number of bands and
the transit traffic occupies a larger number of bands then only a few bands need to be
dropped and wavelength demultiplexed at the node, and the complexity and cost of the
OXC is diminished. This traffic assignment technique is an example of grooming at the
network level. We take up the grooming problem in a network context in Chapter 7.

4.13 Performance Evaluation: Methodology and Case Studies

In the discussion of transmission impairments such as dispersion, noise, and fiber non-
linearities, we began by treating each in isolation to obtain some meaningful results on
performance constraints and to yield some insight on the problem. Typically we found
an approximate limit on transmission link length and/or bit rate due to each type of
impairment. In Section 4.9, we took the next step of examining and modeling the accu-
mulated effects of certain impairments, but still in a relatively simple context, focusing
on one impairment at a time. In a wide-area multiwavelength network, however, a purely
optical connection traverses many different components (fibers, amplifiers, WMUXs,
WDMUXs, and cross-connects), accumulating distortion, noise, and interference as it
propagates. As noted, the problem of accumulated transmission impairments on long,
nonregenerated paths is one of the negative features of optically transparent networks.
Thus, the prediction of the combined effect of all these impairments is of crucial interest
to the network designer. As we shall see, the complexity of the problem suggests that the
best approach to performance evaluation in this case is simulation. As optical networks
have come of age and new photonic and optoelectronic technologies have proved their
viability in the real world, computer modeling and simulation are playing an increasingly
298 Multiwavelength Optical Networks

important role. It is a cost-effective way of performing virtual prototyping and system

engineering and testing before the actual network buildup and deployment. However, a
“brute force” approach to simulation is doomed to failure because of the sheer number
of components involved and their complexity.
In the following sections, we describe an efficient method of physical-layer
simulation and apply it to some case studies. The general objectives are
r To suggest a way of improving the computational efficiency of physical-layer simula-
tions
r To illustrate how the reach of an optical path is constrained by combinations of
technological limitations
r To illustrate how computer simulation and modeling can be used as a tool for network
design and engineering
r To use the case studies as an illustration of the current trends in network architectures

4.13.1 Physical-Layer Simulation: Three-Step Approach

Computer simulation can be used effectively for engineering the physical layer of WDM
optical networks, producing computational accuracy while maintaining speed and topo-
logical flexibility. A number of simulation packages intended for the network optical
layer exist today. The operation of the majority of these tools is based on the well-known
time- and frequency-domain simulation principles [Jeruchim+92]. Unfortunately, these
techniques are inefficient when dealing with very wideband systems, which is the case
in WDM networks. Simulation of WDM networks requires a large simulation band-
width, mainly because of the large aggregate bandwidth of the optical signals and the
wide spectral range of the ASE noise. Moreover, such networks contain a large num-
ber of optical paths over which simulations need to be performed. In this context,
complete time- and frequency-domain simulation is a very time-consuming process. As
a result, a multistep computationally efficient simulation methodology was introduced
in [Antoniades+02] that represents a semianalytic method of performance modeling.
The strategy of the overall simulation is to begin with a wavelength-domain simulation
of a network (presented in detail in [Roudas+00]) to find pertinent signal, noise, and
interference power levels only. This information is then used as a basis for a more refined
time-domain simulation of specific signals of interest. This two-step approach enables
handling systems with sizes that would render them intractable with other methods.
The above approach can then be expanded with a third step that enables efficient itera-
tions of performance calculations that help with the engineering and design of specific
topologies.
The first step involves a wavelength-domain simulation of the entire network being
analyzed. This is basically a low-resolution frequency-domain approach, and it is limited
to linear systems. Figure 4.90 illustrates the general methodology. In a (high-resolution)
frequency-domain simulation, shown in Figure 4.90(a), a multiwavelength signal is
represented by sets of samples of the complex signal spectra, with a resolution bandwidth
that is a small fraction of the bit rate in the case of digital signals. However, in wavelength
 Enabling Technology 299

WDM
Signal
ASE
Power
Noise
Spectra
... ...
Resolution Frequency
Bandwidth
(a) Frequency Domain

Total ASE Total Signal

Noise Power Power

... ...
Resolution Wavelength
Bandwidth
(b) Wavelength Domain

Figure 4.90 Wavelength-domain simulation.

simulation a much coarser resolution is used, as shown in Figure 4.90(b).42 In this case,
the spectral components of the quantities of interest (WDM signal and ASE noise
powers) are grouped into larger bins so that the fine structure of the frequency spectrum
is lost. However, the loss of resolution is offset by a gain in computational speed.
Typical effects studied in the wavelength domain are evaluations of EDFA gain profile,
ASE noise accumulation, and cross-talk in various network elements. The resolution
bandwidth (or bin size) appropriate to multiwavelength networks is on the order of
0.1 nm to ensure that adjacent signals fall in different bins. The rationale for using
a coarse resolution is as follows. First, the optical frequency characteristics of typical
optical components generally vary very slowly over the bandwidth of an individual
signal, so their action results in a simple scaling of each signal by a constant gain and
a change of carrier phase. Second, noise spectra are also slowly varying functions of
frequency. Third, insofar as the various interactions within the network are concerned,
it is usually sufficient to characterize the signals themselves as constant, unmodulated
carriers.43 A result of the undersampling performed in the wavelength-domain simulation
is that switching between time and frequency domains is not possible. Thus, this first
simulation step is used to calculate the average signal and cross-talk powers and OSNR
and not to evaluate the waveform evolution of channels through the optical components
of the network. The cross-talk-induced penalties, nonlinear effects in the fiber and other
optical components, and polarization effects are not captured in the wavelength-domain

42
Similar approaches exist under different names (for example, “power-budget simulation” [Rotolo+95]).
43
This still allows for the study of effects such as the influence of aggregate signal power levels on EDFA
gain and ASE noise because the signal variations due to modulation are much too rapid to affect these
properties of the EDFA.
300 Multiwavelength Optical Networks

simulations. These are studied only on individual paths of the network under study
in the time-domain part of the simulation (step 2). As a result, the time-consuming
time/frequency approach is executed only on selected worst-case paths in the network.
Wavelength-domain simulations and analysis identify these worst-case paths.
The second part of the simulation is a conventional time- and frequency-domain
computation that is run to study the performance of the optical paths that were identified
in the previous step. Having used the wavelength-domain simulation to determine the
pertinent signal, noise, and interference power levels, the rest of the network is now
suppressed, with each selected path viewed as an equivalent channel. Each equivalent
channel is characterized by its dispersion and attenuation, together with the noise and
interfering signals accumulated along the path, with fiber nonlinearities and polarization
effects included. The time-/frequency-domain simulations are used to calculate the error
probability (Pe ) for the selected channels as well as to obtain eye diagrams.
Once Pe is obtained the Q factor can be used as a convenient parameter for
system performance characterization. Based on a Gaussian approximation for noise
and interference, the relation between BER (equivalent to Pe ) and Q was defined in
Equation (4.42) as
 −Q2
1 Q e 2
BER = erfc √ ≈ √ (4.55)
2 2 Q 2π

so Q is determined from BER by inverting this expression. Consecutive time-/frequency-

domain simulations on a selected path can be run to determine the Q penalty for each
impairment. For example, in the case of optical cross-talk, cross-talk-induced Q penalty
is defined as the difference in Q 2dB at the optical receiver, observed at a given error
probability Pe for the case of no cross-talk in the system, and Q 2dB observed under
the effects of cross-talk. This approach helps the system/network designer quantify the
impact of impairments such as fiber nonlinearities, polarization effects as well as linear
optical cross-talk and distortion due to optical filter concatenation.
Summing this up, the two-step approach essentially separates macroscopic first-order
effects examined in the first step (e.g., power levels and noise spectra) from microscopic
second-order effects examined in the second step (e.g., dispersion and signal waveform
distortion). The first step uses a linear model of the network excluding any nonlinearities
and operates at low resolution, whereas the second-step model may include both linear
effects (e.g., cross-talk) and nonlinear effects (e.g., SPM). This essentially treats the
results of nonlinearities as perturbations on the dominant linear effects. It is normally
valid if nonlinearities are small compared with linear effects, which is the case in most
network applications. This leads to a third simulation step that has been applied in
metro network examples in [Antoniades+04a]. It is a budgeting approach based on
Q penalties for the various impairments being studied, obtained either from the second
step of the simulation or through stand-alone impairment models. A simplified baseline
Q performance estimator is run where the Q factors of all the channels being studied on
the selected worst-case paths are calculated. This is done by keeping track of the signal
and ASE noise powers from the optical transmitter, through the various components to the
 Enabling Technology 301

optical receiver. Channel power divergence, dispersion/chirp, ASE noise and receiver
noise terms are taken into account, but all other impairments whose Q penalties are to
be determined are excluded at this step. Finally at the receiver the other impairments
being studied (fiber nonlinearities, PMD, cross-talk, and so on) are budgeted one at a
time in the form of Q penalties, as discussed in Section 4.9. It must be noted that this
is not a simplistic Q budget because the margin allocated for each impairment is not
predetermined but is calculated from a corresponding impairment model obtained in the
second step of the simulation.
The motivation behind the third step is to make the design problem tractable. Ideally,
the network engineer should be able to explore all possible combinations of component
parameters in the network to optimize the design. In practice this is impossible because
of “parameter space explosion”: there are just too many parameters to explore. Thus
we divide the problem into two parts. The first part (steps 1 and 2) uses a baseline
model that can be simulated efficiently with several design iterations if necessary. The
second part (step 3) refines the first step using Q penalties computed based on the results
of the baseline simulation to take into account effects that are assumed not to change
much from iteration to iteration. Further iterations can be run at this level to refine the
design. In this way, the designer can “tweak” network performance, derive component
specifications, determine the possible dependencies of specific parameters on network
size, and in general better explore the performance versus cost trade-offs.

4.13.2 WDM Network Simulation Case Studies

The following three case studies illustrate the implementation of the simulation approach
described above. The first case study pertains to the MONET testbed ([see Wagner+96])
and focuses on issues in long-haul transmission, whereas the second and third are inspired
by more recent work in the industry that focused on the metro environment, reported in
more detail in [Antoniades+04a].

4.13.2.1 Study of a Wavelength Add/Drop Multiplexer Chain

The effects of filter imperfections and laser misalignments were studied in a system
consisting of a chain of 50 WADMs [Antoniades+97]. This might correspond, for
example, to a 50-node multiwavelength national scale (long-haul) network. The objective
was to determine how many WADMs can be cascaded before the OSNR drops below
an acceptable level. The system that was simulated is shown in Figure 4.91. It is an
eight-wavelength system, in which one channel is continued end to end, and the others
are added and dropped periodically at various WADMs along the path. Each WADM
is constructed as shown in Figure 4.92(a). [The same design as Figure 4.82(a).] The
WDMUXs and WMUXs consist of cascades of MI filters, with each signal passing
through eight filters in each WADM. The variable attenuators are servo driven to equalize
the signal levels, which are amplified nonuniformly by the EDFAs. Figure 4.92(b) is a
block diagram that represents the modules and their interconnections used in simulating
the WADM. The filter characteristics used in the simulation were based on measured
302 Multiwavelength Optical Networks

Fiber Fiber Fiber

EDFA
M
U WADM WADM ... WADM ...
X
Stage 1 Stage 2 Stage 50

Figure 4.91 WADM chain.

transmittance and reflectance of actual components. In the simulation, the EDFA gains
and ASE noise contributions are calculated based on aggregate signal power, using the
models described in [Roudas+99] and references therein. Some typical results are shown
in Figure 4.93, illustrating combined effects of laser misalignment and filter cascading.
In Figure 4.93(a), the power spectral densities (PSDs) of the signal and ASE noise are
shown after 24 WADMs, with a laser misalignment of 40 GHz in channel 5, which is
the end-to-end channel being studied. (The powers shown are totals in a bandwidth of
0.1 nm, which is the bin size.) Despite the misalignment, which places it at the edge
of the filter passbands, the variable attenuators are able to maintain uniform signal
levels but at the expense of increasing the noise level in channel 5. This reduces the
OSNR in that channel to approximately 20 dB, which is more than sufficient to sustain
a BER of 10−9 . After 50 WADMs, it is impossible to equalize the signal level, and it
drops below the noise level. (Similar results for laser misalignments of 30 GHz or less
show satisfactory OSNRs through the full chain of 50 WADMs.) It must be noted that
the results presented above pertain to wavelength-domain simulations (step 1) focusing
on OSNR-related effects only. Fiber nonlinearities and PMD are studied and budgeted
separately in the second and third simulation steps to ensure proper system operation.

4.13.2.2 Modeling Performance Impairments in a Metro WDM Network Composed

of Interconnected Rings
As commercial deployment of optical networks has taken hold, the design focus has
shifted from government-funded consortia (like MONET), to for-profit networking

EDFA D 2 EDFA
M
M
U
U
X
X
8

DMUX Switches MUX

EDFA EDFA
Attenuators
(a) Actual (b) Simulation

2 × 2 Switch

Variable Attenuator

Figure 4.92 WADM structure and simulation model.

 Enabling Technology 303

10

0
PSD (dBm/0.1 nm)

1540 1545 1550 1555 1560 1565 1570

Wavelength (nm)
(a) 40 GHz Misalignment after 24 WADMs

10

0
PSD (dBm/0.1 nm)

1540 1545 1550 1555 1560 1565 1570

Wavelength (nm)
(b) 40 GHz Misalignment after 50 WADMs

Figure 4.93 Simulation results for the WADM chain.

304 Multiwavelength Optical Networks

A Edge 1

Edge 2
Edge 6

Feeder (IOF) ring

B

Edge 5 Edge 3

Edge 4

Add/drop node
Interconnect node

Figure 4.94 Ring interconnect network architecture. Worst-case paths between A and B are indi-
cated. (From [Antoniades+04a, Figure 3]. Copyright
c 2004 IEEE. Used by permission of The
Institute of Electrical and Electronics Engineers, Inc.)

activities. A significant part of this work involves investigating the feasibility of trans-
parency in metropolitan optical networks. These are typically in the form of intercon-
nected optical rings, which have evolved historically from opaque SONET rings and their
electrical ancestors. Ring topologies were preferred due to their efficient fiber sharing
and resiliency. As shown in Figure 4.94, this type of network has feeder and distribution
sections, referred to also as interoffice (IOF) or core, and collector or edge, respectively.
The feeder sections aggregate traffic originating from the distribution sections and inter-
connect them while the edge rings distribute and collect traffic from individual business
or campuswide networks (shown as buildings in Figure 4.94). Feeder rings can also have
transfer points (points of presence) to the long-haul backbones and may also include
direct customer connection sites.
The network of Figure 4.94 models six edge rings attached to an IOF ring. Each edge
ring, with a typical circumference of about 15–30 km, can interconnect N users (access
nodes) that can be small businesses, campus networks, or Internet service providers
(ISPs). Edge rings are interconnected with the IOF ring using hub nodes that provide
the cross-connect capabilities. However, individual customers may directly connect to
the network through the hubs. Traffic aggregation occurs at either the access or the
hub node sites and can involve a wide range of service types such as Internet Protocol
(IP), asynchronous transfer mode (ATM), or frame relay (FR) over SONET/SDH and
 Enabling Technology 305

Ethernet. For the architecture of Figure 4.94, it was assumed that client services are
assigned the entire optical wavelength even if the demand does not reach OC-48 or
OC-192 and that a mixture of access-to-access or access-to-hub connections exist.
The physical layer model for each hub node used in this case study is the node
shown in Figure 4.83, involving hierarchical optical multiplexing and demultiplexing
of bands and wavelengths (see Section 4.12.2.1). Thirty-two channels are considered
in the network and these are partitioned into two major bands (C- and L-bands). Each
band is further partitioned into two-to-six other subbands with individual wavelengths
spaced 200 GHz apart and with a guard spacing between bands. Optical dispersion
compensating modules (i.e., a length of negative dispersion fiber) are assumed on each
band module to compensate for the dispersion on the IOF ring. More details of the
actual network design as well as typical component and system parameters used in
the simulations are presented in [Antoniades+02]. Access nodes are assumed to be
implemented using a serial design that provides for the add/drop of the appropriate band
on the first level (band MUX/DMUX) and then selects the appropriate channel on the
second level (channel MUX/DMUX).
The goal here is to derive a network design that can be upgradable to higher
bit rates without the need for costly equipment modifications, so both OC-48 and
OC-192 bit rates were investigated. In designing and engineering the proposed network
the three-step simulation process was employed. Figure 4.94 illustrates the worst-case
path (between nodes A and B) for the case study. The path from A to B is shown in light
gray and the return path in black. The path happens to be the longest as well as the one
that passes through the most optical components. The optical signal is added on the first
node of an edge ring, traverses the whole edge and IOF rings, and is dropped at the most
distant node of the last edge ring. A general design assumption is that each connection
is bidirectional with both directions on the same nominal frequency. This will become
important during the study of cross-talk contributions presented below.
In addition to the performance engineering issue, the question of how big a transparent
network can be supported is of great importance. Focusing on the OC-192 (10 Gbps)
bit rate, which needs to be supported as a direct upgrade for the OC-48 (2.5 Gbps) case,
three important network parameters are identified
r Power at the receiver (depends on channel launched power and loss on the final access
ring)
r OSNR at the receiver
r Total uncompensated fiber length on the worst-case path

The interplay of the above three parameters limits the size of the network. In the above
case study, IOF and edge rings of circumference 100 and 30 km, respectively, were used;
the above numbers were derived as a result of a network parameter study presented in
[Antoniades+02].
Wavelength-domain simulation is first performed to calculate OSNR, channel power
level, and cross-talk information for the network. Wavelengths are routed in such a way
that worst-case cross-talk is produced. In particular, the focus is on a given wavelength λ1
that travels the worst-case path shown in Figure 4.94. It is assumed that λ1 is being reused
306 Multiwavelength Optical Networks

30

2nd order
25 HUB node’s MUX/DMUX
Number of cross-talk Terms

20

15
Beating mixed terms

10
1st order HUB node’s MUX/DMUX

Dominant term
5
from access node

0
120 110 100 90 80 70 60 50 40 30 20 10
Individual Term cross-talk level (dB)

Figure 4.95 Histogram of all cross-talk terms accumulated at receiver B for the worst-case path of
Figure 4.94. (From [Antoniades+02, Figure 11]. Copyright
c 2002 IEEE. Used by permission
of The Institute of Electrical and Electronics Engineers, Inc.)

at each access ring in the network as well as for the return path from node B to node A.
The received signal at node B will contain a large number of cochannel cross-talk terms
that have accumulated due to imperfections in filtering and switching at the different
hub nodes as well as at access node A. (This is a cochannel heterodyne cross-talk effect
as explained in Section 4.9.) Figure 4.95 presents a power level histogram of all the
cochannel cross-talk terms present at node B that are obtained from wavelength-domain
simulation. The strongest cross-talk contribution is generated at the access node A and is
due to the return path signal. Part of this dropped signal will leak and combine with the
added signal (same nominal frequency) due to imperfections in the isolation of the band
MUX/DMUX used at each access node. The power level of the added signal as well as
the filter isolation performance will be the determining factors for this effect. In this case
study, this cross-talk term is 21 dB below the signal level and dominates all other terms,
which are of the order of −40 to −50 dB. Based on the above and assuming OC-192
connections using EA-modulated transmitters, time-domain simulation is performed on
the worst-case path to determine Q penalties for various impairments. For example, the
power levels of the 30 dominant cross-talk terms are used in time-domain simulation on
the worst-case path and the Q penalty is calculated based on simulated BER, converted
to Q via Equation (4.55) as explained in Section 4.13.1. Figure 4.96 shows the Q cross-
talk penalty versus the different levels of the dominant cross-talk term. It is clear that
a tighter cross-talk level should be maintained at each access node to achieve a 2-dB
 Enabling Technology 307

3.6

3.2
Cross-talk-induced penalty

2.8

2.4

2.0

1.6

1.2

0.8
50 45 40 35 30 25 20
Dominant cross-talk term level (dB)

Figure 4.96 Cross-talk-induced Q penalty in dB versus dominant cross-talk term power level.
(Adapted from [Antoniades+02, Figure 12]. Copyright
c 2002 IEEE. Used by permission of The
Institute of Electrical and Electronics Engineers, Inc.)

penalty as a target cross-talk budget in the network design. This cross-talk performance
can be realized by tighter cross-talk filter specifications on each band MUX/DMUX or
by additional filtering at each access node, with the side effect of increasing the pass-
through insertion loss of each one of these nodes. If the latter solution is chosen, only
five access nodes (as shown in Figure 4.94) along with 19 km of fiber can be supported.
Each node can add and drop one band and thus each access ring can potentially support
up to 20 wavelengths of the total of 32, which provides adequate network connectivity.
Additional penalties must be included for other effects. For filter concatenation, it
has been shown that a 2.0-dB distortion-induced Q penalty is an achievable budget
value in the case of EA-modulated and externally modulated transmitters at OC-192
[Tomkos+01a]. Calculations based on the model for PMD described in Section 4.9.5
yield a PMD-induced Q penalty of 0.2 dB (negligible).
Finally the Q penalty because of nonlinearities obtained for the above 32-channel
system is approximately 0.2 dB, which is close to negligible. This is due to the fact that
the channel launched power is limited to the region where fiber nonlinearities are not
dominant (about 0 dBm per channel is launched into the fiber).
The last step of the simulation consists of simple budgeting calculations based on the
Q penalties obtained above, focusing on the best and worst performing channels on the
identified path of Figure 4.94. Power equalizers should be an integral part of the hub
node design because they help mitigate the power divergence among the channels. One
of the sources of power divergence is component PDL. Simulations based on the model
in Section 4.9.2 have shown that in this network up to 0.5 dB PDL-induced channel
308 Multiwavelength Optical Networks

22

20
Q (dB)

18

BER 10 12

16

14

1520 1530 1540 1550 1560 1570 1580 1590 1600 1610
Wavelength (nm)

Figure 4.97 Q-channel performance for the worst-case path of Figure 4.94 assuming OC-192 bit
rate and EA-modulated transmitters. (Adapted from [Antoniades+02, Figure 14]. Copyright
c
2002 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

power divergence will accumulate between each hub. The power equalizers are therefore
designed to handle the above effect in addition to other sources of power divergence
such as dynamic add/drops due to network reconfigurations.
Figure 4.97 shows the Q parameter versus wavelength obtained using the simulation
methodology on the metro ring composed of six interconnected access rings each having
five add/drops (Figure 4.94). In the entire design, the channels will traverse a total of
38 km of dispersion uncompensated fiber (19 km in each access ring) and 85 km
of dispersion compensated fiber on the IOF ring. An additional 2-dB Q margin for
component aging is included in the results. Typical chirp/dispersion transmitter rating
parameters are used in the simulations as described in [Antoniades+02].

4.13.2.3 Engineering the Performance of a DWDM Metro Network

Deployment Scenario
In the next study, the focus moves to a different scenario in which four subnets composed
of interconnected rings resembling those in the previous example are joined via a large
ring or mesh. The transparency of each set of interconnected rings is now discarded
in favor of a more traditional SONET/SDH architecture as shown in Figure 4.98. This
represents a current evolutionary scenario in a highly populated metro region [Elby+02].
All the subnet rings shown in the figure are based on (opaque) SONET/SDH equipment
and the move to transparency involves only the superhubs that are to be joined in a
ring or a mesh connectivity pattern. In typical densely populated metropolitan areas,
superhubs number only a few and are typically spaced 30 to 50 km apart. They can be
implemented using ROADMs of the type described in Section 4.12.2.1. Eliminating the
 Enabling Technology 309

SONET OC-12/48
(collector rings)

SONET OC-48/192
(inter-Hub rings)

Central Office
(End Office)

Hub
Ring DWDM using ROADM
(inter-Superhub wavelength paths)
Superhub
(only in some locations)

Figure 4.98 A DWDM metro network deployment scenario. All rings represent typical SONET
OC-12/48/192 designs. DWDM is deployed only between the superhub nodes (dark squares)
in ring (solid) or possible mesh (dotted) configurations. (From [Antoniades+04a, Figure 11].
Copyright
c 2004 IEEE. Used by permission of The Institute of Electrical and Electronics
Engineers, Inc.)

details of the subnets, the remaining transparent interconnection network is shown in

Figure 4.99. (This is the mesh interconnection alternative.)
In this case, assuming minimum hop routing, the longest route between node pairs
is 70 km. Thus paths in this study are much less affected by fiber impairments than
those in the other case studies of Section 4.13.2.2. As a result, the first part of the
simulation procedure is easily skipped and the problem becomes an exercise in how best
to take advantage of performance versus cost trade-offs: essentially a budgeting approach

A 25 km B

5 km 5 km
45 km
C
D
50 km 45 km

20 km 20 km

E 20 km F

Figure 4.99 DWDM metro network case study based on the network deployment scenario presented
in Figure 4.98. Nodes represent only superhub stations with typical distances (not drawn to scale).
(From [Antoniades+04a, Figure 12]. Copyright
c 2004 IEEE. Used by permission of The Institute
of Electrical and Electronics Engineers, Inc.)
310 Multiwavelength Optical Networks

(simulation step 3). We start by focusing on a connection from node A to node D.

Path A-B-D is a minimum hop path for this connection, with length 30 km. For the
protection path, we select path A-F-D, which spans 70 km. In case this is not available, the
reprovisioned protection path will be A-C-E-F-D which is 65 km long. For the purposes
of this study the focus is on the worst-case path A-F-D (70 km), and OC-192 services will
be engineered based on that. Superhubs are assumed to have similar ROADM equipment
whose basic architecture is shown in Figure 4.85 for a typical 4 × 4 design.
For this study, 40 channels were used with 100-GHz spacing. Access to 26 channels
is assumed on each fiber at the drop side (and similarly for the add side). The 4 × 1
wavelength-selective switches on the output of these devices include the switching,
equalization, and blocking capabilities described in Section 4.12.2.1. Insertion loss of
5 dB and in-band cross-talk of −50 dB were used in the simulation models for this
ROADM based on work presented in [Marom+02]. For reasons stated above, effects like
filter concatenation, cross-talk, and nonlinearities will have significantly less importance
than in the previous case study. For example, the number of optical filters through which
the signal passes is now only 3, which results in negligible penalty (see [Downie+03]).
The cochannel cross-talk Q penalty is budgeted at only 1.0 dB based on the model
presented in Section 4.9.1. Fiber nonlinearities are almost negligible for the channel
spacing, bit rate, and per-channel launched power in this case (1 dBm/channel) and
are budgeted at 0.2 dB. PMD is budgeted at 0.4 dB following the modeling approach
presented in Section 4.9.5, and PDL-induced ripple effects and PDL/PMD interactions
are ignored considering the small number of optical components that the signal traverses
before being regenerated. A safety margin of 2.0 dB is allocated for component aging
and factored into the budgeting model.
This case study uses OC-192 DML transmitters (Section 4.5.3.1), which are a cost-
effective choice in metro networks, but which are limited in transmission distance due to
dispersion/chirp interaction. Hence, optical dispersion compensators (Section 4.3.2.3)
and electronic dispersion compensation modules (Section 4.7.3) are included on the
links. Budgeting simulations are run for (1) the case where only optical dispersion
compensation is used, and (2) the case where the system designer uses EDCs at the
receiver of path A-F-D of Figure 4.99 and engineers all 40 possible channels through it.
Figure 4.100 presents the results for the two simulation runs described above. In the case
where EDCs are used at the receiver, a phenomenological model for the EDC, derived
from [Feuer+03], is used in the simulation procedure. It turns out that EDCs without
optical dispersion compensation cannot prolong DML reach beyond 20 km. However,
EDCs were shown in this study to provide a 3.2-dB Q margin in dealing with the residual
dispersion left by basic optical dispersion compensation modules (no dispersion slope
compensation) (see Figure 4.100). This could be allocated to mitigate other impairments
such as PMD, which can be significant in networks using older fiber. The performance
margin might also be used to relax specifications on individual components. An example
would be gain ripple of amplifiers and/or other optical components.
The performance evaluation studies presented in this section illustrate the role that
simulation can play in estimating the performance of large systems subject to complex
transmission and switching impairments. Because simulation is accomplished rapidly
and efficiently, the results can be used to provide considerable insight to the network
 Enabling Technology 311

20

with EDC
19

Performance
18 margin

17

16 without EDC

15

14

Figure 4.100 Simulation results for path A-F-D in Figure 4.99 comparing Q-channel performance
with and without EDC. (Adapted from [Antoniades+04a, Figure 14]. Copyright
c 2004 IEEE.
Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

designer. To paraphrase Richard W. Hamming [Hamming62], the purpose of simulation

is insight, not numbers.

4.14 Problems

1 What is the critical angle, the acceptance angle, and the numerical aperture of a
step-index fiber with a core refractive index of 1.47 and a cladding refractive index of
1.44?
2 (Intermodal dispersion) Consider a multimode step-index fiber of length L and
refractive indices of n 1 and n 2 for the core and the cladding, respectively. The broadening
of a pulse injected in such a fiber can be estimated by considering the time difference
in the propagation of rays following the shortest and the longest paths to the receiver. If
this time difference must be limited to a fraction of the bit period, find the maximum bit
rate that can be transmitted through a fiber with the following parameters: n 1 = 1.47,
n 2 = 1.44, and L = 10 km.
3 (Resolution of the Helmholtz equation) Consider a step-index fiber of radius a
and refractive indices of n 1 and n 2 for the core and the cladding, respectively; assume
unbounded cladding. Using separation of variables to solve Equation (4.11), show that
any field component can be written in the form

A Jl (κr ) exp(− jlφ) exp(− jβz) r ≤ a
U (r, φ, z, ω) =
B K l (γ r ) exp(− jlφ) exp(− jβz) r > a,
312 Multiwavelength Optical Networks

where A and B are arbitrary amplitude constants, β is the propagation constant, κ 2 =

 kn 2 
(k 2 − β 2 ), γ 2 = n 12 − β 2 , and Jl (.) and K l (.), respectively, are the Bessel function
of the first kind and the modified Bessel function of the second kind of integer order l.
(Note that U is a function of ω through the dependence of the solution parameters on
the wavenumber k.)
4 For a cascade of N independent sources of PMD, where the ith source contributes
an average time differential of τPMD i , the overall average time differential is

 N

2
τPMD =  τPMD i.
i=1

Suppose an optical path passes through several WSXCs, each one containing a cascade
of four lithium niobate switches (τPMD = 2 ps each). What is the maximum number
of WSXCs that can be traversed before the PMD becomes significant, if the signals are
OC-48? OC-192?
5 A transmission link is composed of fiber with an attenuation of 0.3 dB/km, with
EDFAs spaced every 60 km. All EDFAs are identical, with amplifier gain equal to fiber
loss on the preceding fiber segment. The amplifier spontaneous emission factor is 1.5.
Bandwidths of any optical filters in the link are large compared to the bandwidth (5 GHz)
of the low-pass electrical filter preceding the electrical detector. Assume that thermal
noise in the receiver is negligible and that the signal-ASE beat noise is dominant.
(a) What is the maximum distance a signal with transmitter power of −13 dBm and
wavelength of 1550 nm can propagate before the electrical SNR drops below 20 dB?
(b) Repeat the calculation if the EDFAs are spaced every 30 km and their gain is adjusted
to compensate for the new fiber loss. Assume the same spontaneous emission factor.
Comment on the results.
6 (Upper bound on EDFA gain) An optical signal, with average power Psin and wave-
length λs and an optical pump with average power Ppin and wavelength λ p , enter an
erbium-doped fiber. From energy conservation, show that an upper bound on the EDF
gain is
λ p Ppin
G ≤1+ .
λs Psin
7 (APD SNR) A receiver is composed of an avalanche photodiode and a high-
impedance front end. The APD is illuminated with constant optical power, P. The
average photocurrent is enhanced by an average multiplication factor M̄ compared to
the PIN photodiode case.
The shot noise generated in the photodetection process can be approximated as white
Gaussian noise, with a power-spectral density of q M̄ 2+x R P, where q is the electron
charge, R is the PIN photodiode responsivity given by Equation (4.36), and x is an
excess-noise factor depending on the material (e.g., x = 0.3 to 0.5 for silicon APDs).
Both the shot and thermal noise are filtered by the low-pass filter composed of the
parasitic capacitance of the APD and the load resistance.
 Enabling Technology 313

Derive the expression for the signal-to-noise ratio at the input of the electronic ampli-
fier and show that there is an optimum value of the average multiplication factor M̄ for
which the SNR is maximized. Also show that at the shot-noise limit, the PIN photodiode
provides better SNR performance than the APD.
8 Suppose that the n ports of the device shown in Figure 4.54 are grouped into n 1
input ports and n 2 output ports, where n 1 + n 2 = n. Assume that when optical signals
are applied to the input ports only, there are no reflected waves at those ports.
(a) Derive the relation between the n 2 × n 1 power matrix, A [Equation (2.4)], and the
scattering matrix.
(b) Now suppose that the input and output ports are interchanged and that when optical
signals are applied to only the (new) input ports, there are no reflected waves at
those ports. Again, derive the relation between the new power transfer matrix, A′ ,
and the scattering matrix.
(c) If the device is reciprocal, show that A′ = At .
9 Consider a passive reciprocal element with n + 1 ports. Suppose that when the first
port is used as an input and the others are used as outputs, it acts as a 1:n power splitter.
Show that, operated in the opposite direction, it is an n:1 combiner with a combining
loss of 1/n.
10 Redo the previous problem in the case in which the element is a 1:n demultiplexer.
Show that it is a multiplexer when operated in the opposite direction.
11 (Filter cascading) In large networks, an end-to-end path often traverses many optical
filters contained in WSXCs and other components along the path. In this problem, we
study the bandwidth-narrowing effects of cascading for two types: FP and MI filters.
The (low-pass equivalent) amplitude transfer function of an FP filter can be modeled
using a first-order Butterworth function, whereas an MI filter can be modeled using a
third-order Butterworth. The general nth order Butterworth function is defined as
1
|H ( f )| =   2n ,
f
1+ fc

where f c is the (3 dB) cutoff frequency.

(a) Assuming cascades of 10, 20, 30, and 50 filters, plot the overall power transfer
function of the cascades in dB for each filter type.
(b) Compare the narrowing effects for the two filter types. Which seems more suitable
for large-scale networks?
(c) What is the cutoff frequency of a chain of MI filters with individual cutoff frequency
of 100 GHz?
(d) Comment on the effects of random filter center-frequency misalignments on the
effective bandwidth of a cascade of filters. Which of the two types do you think
would exhibit the best performance as a result of filter misalignments?
(e) Comment on how bandwidth narrowing and filter misalignment affect tolerances to
laser-tuning errors.
314 Multiwavelength Optical Networks

Bibliography

[Agrawal95] G. P. Agrawal. Nonlinear Fiber-Optics Communication Systems, Second Edition.

New York: Academic Press 1995.
[Agrawal97] G. P. Agrawal. Fiber-Optic Communication Systems: Second Edition (Wiley Series
in Microwave and Optical Engineering). New York: John Wiley & Sons 1997.
[Agrawal02] G. P. Agrawal. Fiber-Optic Communication Systems, Third Edition. New York:
John Wiley & Sons 2002.
[Akiyama+01] T. Akiyama, H. Kuwatsuka, T. Simoyama, Y. Nakata, K. Mukai, M. Sugawara,
and O. Wada. Nonlinear gain dynamics in quantum-dot optical amplifiers and its application to
optical communication devices. IEEE J. Quant. Electron., 37(8):1059–1065, 2001.
[Akiyama+06] T. Akiyama, K. Kawaguchi, M. Sugawara, M. Ekawa, H. Ebe, and Y. Arakawa.
Quantum-dot semiconductor optical amplifiers. IEEE LEOS Newslett., pp. 11–13, February
2006.
[Amann94] M. C. Amann. Wide tunable lasers for WDM applications. In Fifth Optoelectronics
Conf. Tech. Digest, pp. 208–209, Makuhari Messe, Japan, July 1994.
[ANSI88] American National Standards Institute. Safe use of optical fiber communication sys-
tems utilizing laser diodes and LED sources. ANSI Standard Z136.2, 1988.
[Antoniades+97] N. Antoniades, I. Roudas, R. E. Wagner, and S. F. Habiby. Simulation of ASE
noise accumulation in a wavelength add–drop multiplexer cascade. IEEE Photon. Technol. Lett.,
9(9):1274–1276, 1997.
[Antoniades+02] N. Antoniades, A. Boskovic, I. Tomkos, N. Madamopoulos, M. Lee, I. Roudas,
D. Pastel, M. Sharma, and M. Yadlowsky. Performance engineering and topological design of
metro WDM optical networks. IEEE J. Select. Areas Commun., 20(1):149–165, 2002.
[Antoniades+04a] N. Antoniades, I. Roudas, G. Ellinas, and J. Amin. Transport metropolitan op-
tical networking: Evolving trends in the architecture design and computer modeling. IEEE/OSA
J. Lightwave Technol., 22(11):2653–2670, 2004.
[Antoniades+04b] N. Antoniades, N. Menon, and N. Chbouki. Techno-economic value analysis of
the role of banded amplification in evolving WDM metro network architectures. In Proceedings
of the Laser and Electro-Optics Society Annual Meeting, Rio Grande, Puerto Rico, November
2004.
[Antoniades+06] N. Antoniades, K. C. Reichmann, P. P. Iannone, N. J. Frigo, A. M. Levine, and
I. Roudas. The impact of polarization-dependent gain on the design of cascaded semiconductor
optical amplifier CWDM systems. IEEE Photon. Technol. Lett., 18(20):2099–2101, 2006.
[Badding+06] J. V. Badding, V. Gopalan, and P. J. A. Sazio. Building semiconductor structures
in optical fiber. Photon. Spectra Mag., pp. 80–88, August 2006.
[Bao+96] Y. Bao, A. Sneh, K. Hsu, K. Johnson, J.-Y. Liu, C. M. Miller, Y. Monta, and
M. B. McClain. High-speed liquid crystal fiber Fabry–Perot tunable filter. IEEE Photon. Tech-
nol. Lett., 8(9):1190–1192, 1996.
[Bernstein+04] J. J. Bernstein, W. P. Taylor, J. D. Brazzle, C. J. Corcoran, G. Kirkos, J. E. Odhner,
A. Pareek, M. Waelti, and M. Zai. Electromagnetically actuated mirror arrays for use in 3-D
optical switching applications. J. Microelectromechan. Syst., 13(3):526–535, 2004.
[Boskovic+02] A. Boskovic, M. Sharma, N. Antoniades, and M. Lee. Broadcast and select OADM
nodes: Application and performance tradeoffs. In Proceedings of the IEEE/OSA Optical Fiber
Commun. Conf., Paper TuX2, pp. 158–159, Anaheim, CA, March 2002.
[Brener+00] I. Brenner, B. Mikelsen, G. Raybon, R. Harel, K. Parameswaran, J. R. Kurz, and
M. M. Fejer. 160 Gbit/s wavelength shifting and phase conjugation using periodically poled
LiNbO3 waveguide parametric converter. Electron. Lett., 36(21):1789–1790, 2000.
 Enabling Technology 315

[Bromage04] J. Bromage. Raman amplification for fiber communication systems. IEEE/OSA J.

Lightwave Technol., 22(1):79–93, 2004.
[Charlet+05] G. Charlet, E. Corbel, J. Lazaro, A. Klekamp, R. Dischler, P. Tran, W. Idler,
H. Mardoyan, A. Konczykowska, F. Jorge, and S. Bigo. WDM transmission at 6 Tbits/s capacity
over transatlantic distance, using 42.7-Gb/s differential phase-shift keying without pulse carver.
IEEE/OSA J. Lightwave Technol., 23(1):104–107, 2005.
[Chayet+04] H. Chayet, S. Ben Ezra, S. Tzadok, S. Tsadka, and J. Leuthold. Regenerative all-
optical wavelength converter based on semiconductor optical amplifier and sharp frequency
response filter. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper ThS3,
Los Angeles, CA, March 2004.
[Cheng+02] K. S. Cheng and J. Conradi. Reduction of pulse-to-pulse interaction using al-
ternative RZ formats in 40-Gb/s Systems. IEEE Photon. Technol. Lett., 14(1):98–100,
2002.
[Chow+07] A. Chow, K. Kikuchi, T. Nagashima, T. Hasagawa, S. Ohara, and N. Sugimoto.
Widely tunable wavelength conversion by four-wave mixing in 1-m dispersion-shifted bismuth-
oxide photonic crystal fiber. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf.,
Paper OTuI2, Anaheim, CA, March 2007.
[Chowdhury+99] D. Q. Chowdhurry, L. Bhagavatula, L. Button, P. Wigley, and C. Sutton.
PDL-induced gain ripple statistics in EDFA. In Proceedings of the European Conference on
Optical Communications, pp. 3.81–3.82, Nice, France, September 1999.
[Chowdhury+04] D. Chowdhury, M. Mlejnek, and Y. Mauro. Numerical modeling of PMD. J. of
Opt. Fiber Commun. Reports, Springer Science and Business Media Inc., pp. 141–149, October
2004.
[Chowdhury+06] A. Chowdhury, J. Yu, L. Xu, T. Wang, and G.-K. Chang. Experimental demon-
stration of 40 Gbps per channel transmission over 50 GHz DWDM systems for low-cost metro
and access networks. In Proceedings of the European Conference on Optical Communications,
Paper Mo4.5.5, Cannes, France, September 2006.
[Chraplyvy90] A. R. Chraplyvy. Limitations on lightwave communications imposed by optical
fiber nonlinearities. IEEE/OSA J. Lightwave Technol., 8(10):1548–1557, 1990.
[Chraplyvy+95] A. R. Chraplyvy, A. H. Gnauck, R. W. Tkach, R. M. Derosier, C. R. Giles,
B. M. Nyman, G. A. Ferguson, J. W. Sulhoff, and J. L. Zyskind. One-third terabit/s transmis-
sion through 150 km of dispersion-managed fiber. IEEE Photon. Technol. Lett., 7(1):98–107,
1995.
[Chu+02] P. B. Chu, S.-S. Lee, and S. Park. MEMS: The path to large optical crossconnects.
IEEE Communications Mag., 40(3):80–87, 2002.
[Clesca+94] B. Clesca, D. Ronarch, D. Bayart, Y. Sorel, L. Hamon, M. Guibert, J. L. Beylat,
J. F. Kerdiles, and M. Semenkoff. Gain flatness comparison between erbium-doped flouride
and silica fiber amplifiers with wavelength-multiplexed signals. IEEE Photon. Technol. Lett.,
6(4):509–512, 1994.
[Crossland+00] W. A. Crossland, I. G. Manolis, M. M. Redmond, K. L. Tan, T. D. Wilkinson,
M. J. Holmes, T. R. Parker, H. H. Chu, J. Croucher, V. A. Handerek, S. T. Warr, B. Robertson,
I. G. Bonas, R. Franklin, C. Stace, H. J. White, R. A. Woolley, and G. Henshall. Holographic
optical switching: The “ROSES” demonstrator. IEEE/OSA J. Lightwave Technol., 18(12):1845–
1854, 2000.
[Delorme+98] F. Delorme, G. Alibert, C. Ougier, S. Slempkes, and H. Nakajima. Sampled-
grating DBR lasers with 101 wavelengths over 44 nm and optimized power variation for WDM
applications. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper FB4,
pp. 379–381, San Jose, CA, February 1998.
316 Multiwavelength Optical Networks

[Desurvire+87] E. Desurvire, J. R. Simpson, and P. C. Becker. High-gain erbium-doped traveling-

wave fibre amplifier. Optics Lett., 12(11):888, 1987.
[Desurvire94] E. Desurvire. Erbium-Doped Fiber Amplifiers: Principles and Applications.
New York: John Wiley & Sons 1994.
[Doerr+03] C. Doerr, L. Stulz, D. Levy, L. Gomez, M. Capuzzo, J. Bailey, R. Long, A. Wong-Foy,
E. Laskowski, E. Chen, S. Patel, and T. Murphy. Eight wavelength add-drop filter with true
reconfigurability. IEEE Photon. Technol. Lett., 15(1):138–140, 2003.
[Domash+04] L. Domash, M. Wu, N. Nemchuk, and E. Ma. Tunable and switchable multi-cavity
thin film filters. IEEE/OSA J. Lightwave Technol., 22(1):126–135, 2004.
[Downie+03] J. Downie and A. Ruffin. Analysis of signal distortion and crosstalk penalties
induced by optical filters in optical networks. IEEE/OSA J. Lightwave Technol., 21(9):1876–
1886, 2003.
[Dragone+91] C. Dragone, C. A. Edwards, and R. C. Kestler. Integrated optics n × n multiplexer
on silicon. IEEE Photon. Technol. Lett., 3(10):896–899, 1991.
[Eiselt+95] M. Eiselt, W. Pieper, and H. G. Weber. SLALOM: Semiconductor laser amplifier in
a loop mirror. IEEE/OSA J. Lightwave Technol., 13(10):2099–2112, 1995.
[Elby02] S. Elby. Public network reliability and resiliency. Presentation at Columbia University,
http://www2.cvn.columbia.edu/course/elene6901/session/S.ElbyNetworkReliabilityTalkPT1.
ppt, October 30, 2002.
[Ellinas+04] G. Ellinas, J.-F. Labourdette, J. A. Walker, S. Chaudhuri, L. Lin, E. Goldstein, and
K. Bala. Network control and management challenges in opaque networks utilizing transparent
optical switches. IEEE Communications Mag., 42(2):S16–S24, 2004.
[Elrefaie+88] A. F. Elrefaie, R. E. Wagner, D. Atlas, and D. G. Daut. Chromatic dispersion limita-
tions in coherent lightwave transmission systems. IEEE/OSA J. Lightwave Technol., 6(5):704–
709, 1988.
[Feuer+03] M. Feuer, S. Huang, S. Woodward, O. Coskun, and M. Boroditsky. Electronic disper-
sion compensation for a 10-Gb/s link using a directly modulated laser. IEEE Photon. Technol.
Lett., 15(12):1788–1790, 2003.
[Filios+03] A. Filios, R. Gutierrez-Castrejon, I. Tomkos, B. Hallock, R. Vodhanel, A. Coombe, W.
Yuen, R. Moreland, B. Garrett, C. Duvall, and C. Chang-Hasnain. Transmission performance of
a 1.5-µm 2.5-Gb/s directly modulated tunable VCSEL. IEEE Photon. Technol. Lett., 15(4):599–
601, 2003.
[Ford+98] J. E. Ford and J. A. Walker. Dynamic spectral power equalization using micro-opto-
mechanics. IEEE Photon. Technol. Lett., 10(10):1440–1442, 1998.
[Ford+99] J. E. Ford, V. A. Aksyuk, D. J. Bishop, and J. A. Walker. Wavelength add-drop switching
using tilting micromirrors. IEEE/OSA J. Lightwave Technol., 17(5):904–911, 1999.
[Fracasso+03] B. Fracasso, J. L. de Bougrenet de la Tocnaye, M. Razzak, and C. Uche. Design
and performance of a versatile holographic liquid-crystal wavelengh-selective optical switch.
IEEE/OSA J. Lightwave Technol., 21(10):2405–2411, 2003.
[Francis+01] D. A. Francis, S. P. Dijaili, and J. D. Walker. A single chip linear optical amplifier.
In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper PD13, Anaheim, CA,
February 2001.
[Garrett+98] L. D. Garrett and A. H. Gnauck. 8 × 20 Gb/s–315 km, 8 × 10 Gb/s-480 km WDM
transmission over conventional fiber using multiple broadband fiber gratings. In Proceedings
of the IEEE/OSA Optical Fiber Commun. Conf., Paper PD18, San Jose, CA, February 1998.
[Gimlett+89] J. L. Gimlett and N. K. Cheung. Effects of phase-to-intensity noise conversion
by multiple reflections on gigabit-per-second DFB laser transmission systems. IEEE/OSA J.
Lightwave Technol., 7(6):888–895, 1989.
 Enabling Technology 317

[Gnauk+97] A. H. Gnauk and R. M. Jopson. Dispersion compensation for optical fiber systems.
In Optical Fiber Telecommunications IIIA, I. Kaminow and T. L. Koch (Editors), New York:
Academic Press 1997.
[Gnauk+02] A. H. Gnauk and P. J. Winzer. Optical phase-shift-keyed transmission. IEEE/OSA J.
Lightwave Technol., 23(1):115–130, 2002.
[Goh+99] T. Goh, A. Himeno, M. Okuno, H. Takahashi, and K. Hattori. High extinction ratio
and low-loss silica-based 8 × 8 strictly nonblocking thermooptic matrix switch. IEEE/OSA J.
Lightwave Technol., 17(7):1192–1199, 1999.
[Goldstein+94] E. L. Goldstein, L. Eskildsen, and A. F. Elrefaie. Performance implications of
component cross-talk in transparent lightwave networks. IEEE Photon. Technol. Lett., 6(5):657–
670, 1994.
[Greywall+03] D. S. Greywall, P. A. Busch, F. Pardo, D. W. Carr, G. Bogart, and H. T. Soh.
Crystalline silicon tilting mirrors for optical cross-connect switches. J. Microelectromechan.
Syst., 12(5):708–712, 2003.
[Gruner-Nielson+01] L. Gruner-Nielson, Q. Yujun, B. Palsdottir, P. B. Gaarde, S. Dyrbol, and
T. Veng. Module for simultaneous C+ L-band dispersion compensation and Raman amplifica-
tion. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper TuJ6, pp. 65–66,
Anaheim, CA, March 2002.
[Hamming62] R. W. Hamming. Numerical Analysis for Scientists and Engineers. New York:
McGraw-Hill 1962.
[Harumoto+02] M. Harumoto, M. Shigehara, and H. Suganuma. Gain-flattening fiber using
long-period fiber gratings. IEEE/OSA J. Lightwave Technol., 20(6):1027–1033, 2002.
[Hasnain00] C. J. Chang-Hasnain, Tunable VCSEL. IEEE J. Select. Top. Quant. Electron.,
6(6):978–987, 2000.
[Hayee+99] M. I. Hayee and A. E. Willner. NRZ versus RZ in 10-40 Gb/s dispersion-managed
WDM transmission systems. IEEE Photon. Technol. Lett., 11(8):991–993, 1999.
[Hengstler+03] S. Hengstler, J. J. Uebbing, and P. McGuire. Laser-activated optical bubble switch
element. In Proceedings of the IEEE/LEOS Int’l. Conf. on Optical MEMS, pp. 117–118,
Waikoloa, HI, August 2003.
[Henry+88] P. S. Henry, R. A. Linke, and A. H. Gnauk. Introduction to lightwave systems.
In Optical Fiber Telecommunications II, S. E. Miller and I. Kaminow (Editors), New York:
Academic Press 1988.
[Henry+89] C. H. Henry, G. E. Blonder, and R. F. Kazarinov. Glass waveguides on silicon for
optical packaging. IEEE/OSA J. Lightwave Technol., 7(10):1530–1539, 1989.
[Hibino00] Y. Hibino. An array of photonic filtering advantages: Arrayed-waveguide-grating
multi/demultiplexers for photonic networks. IEEE Circuits Devices Mag., 16(6):21–27, 2000.
[Hinton93] H. S. Hinton. An Introduction to Photonic Switching Fabrics. New York: Plenum
Press 1993.
[Ho05] K. P. Ho. Generation of arbitrary quadrature signals using one dual drive modulator.
IEEE/OSA J. Lightwave Technol., 23(2):764–770, 2005.
[Iannone+98] P. P. Iannone, K. C. Reichmann, and N. J. Frigo. Wavelength-dependent modulation
effects of light-emitting diodes in multiple-subband passive optical network. In Proceedings of
the IEEE/OSA Optical Fiber Commun. Conf., Paper FD4, pp. 401–403, San Jose, CA, February
1998.
[Iannone+06] P. P. Iannone, K. C. Reichmann, X. Zhu, and S. Gray. A discrete SOA-Raman
hybrid amplifier with 80-nm bandwidth. In Proceedings of the European Conference on Optical
Communications, Paper Th 2.3.7, Cannes, France, 2006.
318 Multiwavelength Optical Networks

[Igarashi+04] K. Igarashi, H. Tobioka, S. Takasaka, R. Miyabe, J. Hiroishi, T. Yagi, O. Aso, and

S. Namiki. Wavelength conversion characteristics of highly efficient fiberoptic four-wave mixer
for 160 GHz picosecond pulse train. In Proceedings of the IEEE/OSA Optical Fiber Commun.
Conf., Paper TuC6, Los Angeles, CA, March 2004.
[Inoue+93] K. Inoue, T. Hasegawa, K. Oda, and H. Toba. Multichannel frequency conversion
experiment using fiber four-wave mixing. Electron. Lett., 29(19):1708–1710, 1993.
[Iocco+99] A. Iocco, H. G. Limberger, R. P Salathe, L. A. Everall, K. E, Chishom,
J. A. R. Williams, and I. Bennion. Bragg grating fast tunable filter for wavelength division
multiplexing. IEEE/OSA J. Lightwave Technol., 17(7):1217–1221, 1999.
[Irmer+03] S. Irmer, J. Daleiden, V. Rangelov, C. Prott, F. Romer, M. Strassner, A. Tarraf, and
H. Hillmer. Ultra low biased widely continuously tunable Fabry–Perot filter. IEEE Photon.
Technol. Lett., 15(3):434–436, 2003.
[Islam02] M. Islam. Raman amplifiers for telecommunications. IEEE J. Select. Top. Quant. Elec-
tron., 8(3):548–559, 2002.
[Jaggi+98] P. Jaggi and H. Onaka. Optical ADM and cross-connects: Recent technical advance-
ments and future optimal architectures. In Proceedings of the National Fiber Optics Engineers
Conf. (NFOEC), pp. 311–321, Orlando, FL, September 1998.
[Jeruchim+92] M. C. Jeruchim, P. Balaban, and K. S. Shanmugan. Simulation of Communication
Systems. New York: Plenum Press 1992.
[Jopson+92] R. M. Jopson and A. A. M. Saleh. Modeling of gain and noise in erbium-doped
fiber amplifiers. SPIE, Fiber Laser Sources and Amplifiers III, 1581:114–119, 1992.
[Kamalakis+03] T. Kamalakis, T. Sphicopoulos, and M. Sagriotis. Accurate estimation of the
error probability in the presence of in-band crosstalk noise in WDM networks. IEEE/OSA J.
Lightwave Technol., 21(10):2172–2181, 2003.
[Kaminow+87] I. P. Kaminow, P. P. Iannone, J. Stone, and L. W. Stulz. FDMA-FSK star network
with a tunable optical filter demultiplexer. Electron. Lett., 23(21):1102–1103, 1987.
[Kaminow+97] I. P. Kaminow and T. L. Koch (Editors). Optical Fiber Telecommunications IIIB.
New York: Academic Press 1997.
[Kaminow+02a] I. Kaminow and T. Li (Editors). Optical Fiber Telecommunications IVA: Com-
ponents. New York: Academic Press 2002.
[Kaminow+02b] I. Kaminow and T. Li (Editors). Optical Fiber Telecommunications IVB: Systems
and Impairments. New York: Academic Press 2002.
[KaminowChap16+02] I. Kaminow and T. Li (Editors). Optical Fiber Telecommunications IVB,
Chapter 16. New York: Academic Press 2002.
[KaminowChap17+02] I. Kaminow and T. Li (Editors). Optical Fiber Telecommunications IVB,
Chapter 17. New York: Academic Press 2002.
[Kasahara+02] R. Kasahara, M. Yanagisawa, T. Goh, A. Sugita, A. Himeno, M. Yasu, and
S. Matsui. New structure of silica-based planar lightwave circuits for low-power thermooptic
switch and its application to 8 × 8 optical matrix switch. IEEE/OSA J. Lightwave Technol.,
20(6):993–1000, 2002.
[Kato+98] M. Kato, K. Kurokawa, and Y. Miyajima. A new design for dispersion-shifted fiber
with an effective core area larger than 100 pm2 and good bending characteristics. In Proceedings
of the IEEE/OSA Optical Fiber Commun. Conf., pp. 301–302, San Jose, CA, February 1998.
[Kazovsky+96] L. Kazovsky, S. Benedetto, and A. Willner. Optical Fiber Communication Sys-
tems. Norwood, MA: Artech House 1996.
[Keil+94] N. Keil, H. H. Yao, C. Zawadzki, and B. Strebel. 4 × 4 polymer thermo-optic directional
coupler switch at 1.55 µm. Electron. Lett., 30(8):639–640, 1994.
 Enabling Technology 319

[Koch+84] T. L. Koch and J. E. Bowers. Nature of wavelength chirping in directly modulated

semiconductor lasers. Electron. Lett., 20(12):1038–1039, 1984.
[Koch+86] L. Koch and R. A. Linke. Effect of nonlinear gain reduction on semiconductor laser
wavelength chirping. Applied. Phys. Lett., 48(3):613–615, 1986.
[Korotky+88] S. K. Korotky and R. C. Alferness. Waveguide electrooptic devices for optical
fiber communications. In Optical Fiber Telecommunications II. S. E. Miller and I. Kaminow
(Editors), New York: Academic Press 1988.
[Lee+94] E. A. Lee and D. G. Messerschmitt. Digital Communication, Second Edition. Norwell,
MA: Kluwer Academic 1994.
[Leuthold+04] J. Leuthold, J. Jaques, and S. Cabot. All optical wavelength conversion and re-
generation. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper WN1, Los
Angeles, CA, March 2004.
[Li+97] Y. P. Li and C. H. Henry. Silicon optical bench waveguide technology. In Optical Fiber
Telecommunications IIIA, I. P. Kaminow and T. L. Koch (Editors), Chapter 8, New York:
Academic Press 1997.
[Libori+01] S. B. Libori, J. Broeng, E. Knudsen, A. Bjarklev, and H. R. Simonsen. High-
birefringent photonic crystal fiber. In Proceedings of the IEEE/OSA Optical Fiber Commun.
Conf., Paper TuM2, Anaheim, CA, March 2001.
[Lin98] L. Y. Lin. Micromachined free-space matrix switches with submillisecond switching time
for large-scale optical cross-connect. In Proceedings of the IEEE/OSA Optical Fiber Commun.
Conf., Paper WH5, pp. 147–148, San Jose, CA, February 1998.
[Lu+04] H.-H. Lu, W.-J. Wang, and W.-S. Tsai. CSO/CTB performances improvement in a bi-
directional DWDM CATV system. IEEE Trans. Broadcast., (50)4:377–381, 2004.
[Madsen+99] C. K. Madsen, J. A. Walker, J. E. Ford, K. W. Goosen, and G. Lenz. A tunable
dispersion compensating MARS all-pass filter. In Proceedings of the European Conference on
Optical Communications, Paper WeB1.3, Nice, France, September 1999.
[Manolis+02] I. G. Manolis, T. D. Wilkinson, M. M. Redmond, and W. A. Crossland. Reconfig-
urable multilevel phase holograms for optical switches. IEEE Photon. Technol. Lett., 6(6):801–
803, 2002.
[Marom+02] D. M. Marom, D. Nelson, D. Greywall, N. Basavanhally, P. Kolodner, Y. Low, Pardo,
C. Bolle, S. Chandrasekhar, L. Buhl, and C. Giles. Wavelength-selective 1 × 4 switch for 128
WDM channels at 50 GHZ spacing. In Proceedings of the IEEE/OSA Optical Fiber Commun.
Conf., pp. FB7-1–FB7-3, Anaheim, CA, March 2002.
[Matsumoto+02] M. Matsumoto and O. Leclerc. Analysis of 2R optical regenerator utilizing
self-phase modulation in highly nonlinear fiber. Electron. Lett., 38(12):576–577, 2002.
[Mears+87] R. J. Mears, L. Reekie, I. M. Jauncey, and D. N. Payne. Low-noise erbium-doped
fibre amplifier operating at 1.54 µm. Electron. Lett., 23(19):1026, 1987.
[Mecozzi+04] A. Mecozzi and M. Shtaif. Signal-to-noise-ratio degradation caused by
polarization-dependent loss and the effect of dynamic gain equalization. IEEE/OSA J. Lightwave
Technol., 22(8):1856–1871, 2004.
[Meslener84] G. J. Meslener. Chromatic dispersion-induced distortion of modulated monochro-
matic light employing direct detection. IEEE J. Quant. Electron., 20(10):1208–1216, 1984.
[Mollenauer+92] L. F. Mollenauer, E. Lichtman, G. T. Harvey, M. J. Neubelt, and B. M. Nyman.
Demonstration of error-free soliton transmission over more than 15,000 km at 5 Gbit/s, single
channel and over 11,000 km at 10 Gbit/s in a two-channel WDM. In Proceedings of the
IEEE/OSA Optical Fiber Commun. Conf., Paper PD10, p. 351, Washington, DC, February
1992.
320 Multiwavelength Optical Networks

[Moodie+03] D. Moodie, A. Ellis, X. Chen, F. Garcia, D. Rogers, S. Perrin, P. Cannard,

R. McLaughlin, M. Robertson, S. Amos, S. Cole, C. Ford, and I. Reid. Applications of
electroabsorption modulators in high-bit rate extended reach transmission systems. In Pro-
ceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper TuP1, Atlanta, GA, March
2003.
[Mori+01] K. Mori, H. Takara, and S. Kawanishi. Analysis and design of supercontinuum pulse
generation in a single-mode optical fiber. J. Opt. Soc. Am. B, Opt. Phys., 18(12):1780–1792,
2001.
[Mori+05] K. Mori and K. Sato. Supercontinuum lightwave generation employing a mode-locked
laser diode with injection locking for a highly coherent optical multicarrier source. IEEE Photon.
Technol. Lett., 17(2):480–482, 2005.
[Murashige+98] K. Murashige, T. Miyakawa, H. Taga, S. Nakagawa, H. Tanaka, K. Goto,
M. Suzuki, and S. Yamamoto. Sixty 5.3 Gbit/s 1650 km straightline WDM transmission.
In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., pp. 48–49, San Jose, CA,
February 1998.
[Murphy+97] E. J. Murphy, T. O. Murphy, R. W. Irvin, R. Grencavich, G. W. Davis, and
G. W. Richards. Enhanced performance switch arrays for optical switching networks. In Pro-
ceedings of the Eighth Euro. Conf. on Integrated Optics, Paper EFD5, Stockholm, Sweden,
April 1997.
[Noguchi98] K. Noguchi. Optical free-space multichannel switches composed of liquid-crystal
light-modulator arrays and birefringent crystals. IEEE/OSA J. Lightwave Technol., 16(8):1473–
1481, 1998.
[Oh+98] M.-C. Oh, H.-J. Lee, M.-H. Lee, J.-H. Ahn, and S. G. Han. Asymmetric X-junction
thermooptic switches based on fluorinated polymer waveguides. IEEE Photon. Technol. Lett.,
10(6):813–815, 1998.
[Okamoto+96] K. Okamoto, M. Okuno, A. Himeno, and Y. Ohmori. 16-channel optical add/drop
multiplexer consisting of arranged-waveguide gratings and double-gate switches. Electron.
Lett., 32(16):1471–1472, 1996.
[Otani+97] T. Otani, T. Kawazawa, K. Goto, N. Takeda, and S. Akiba. 16 channel 2.5 Gbit/s
WDM transmission experiment over 9000 km by using gain equalised amplifier repeaters.
Electron. Lett., 33(4):309–310, 1997.
[Park+02] H. S. Park, K. Y. Song, S. H. Yun, and B. Y. Kim. All-fiber wavelength-tunable
acoustooptic switches based on intermodal coupling in fibers. IEEE/OSA J. Lightwave Technol.,
20(10):1864–1868, 2002.
[Park+05] S. Park, R. Leavitt, R. Enck, V. Luciani, Y. Hu, P. J. S. Heim, D. Bowler, and
M. Dagenais. Semiconductor optical amplifier for CWDM operating over 1540-1620 nm.
IEEE Photon. Technol. Lett., 17(5):980–982, 2005.
[Patel+95] J. S. Patel and Y. Silberberg. Liquid crystal and grating-based multiple-wavelength
cross-connect switch. IEEE Photon. Technol. Lett., 7(5):514–516, 1995.
[Petropoulos+03] P. Petropoulos, T. M. Monro, H. Ebendorff-Heidepriem, K. Frampton,
R. C. Moore, H. N. Rutt, and D. J. Richardson. Soliton-self-frequency-shift effects and pulse
compression in an anomalously dispersive high nonlinearity lead silicate holey fiber. In Pro-
ceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper FH3, Atlanta, GA, March
2003.
[Philipsen+00] J. L. Philipsen, D. Barbier, A. Kevorkian, C. Cassagnettes, N. Krebs, and P. Bruno.
Compact gain-block consisting of an Er-doped waveguide amplifier (EDWA) and a pump/signal
multiplexer, realized by ion exchange. In Proceedings of the OSA Optical Amplifier Applications
Conf., pp. 151–153, Quebec City, Canada, July 2000.
 Enabling Technology 321

[Pohlmann+91] T. Pohlmann, A. Neyer, and E. Voges. Polarization independent Ti:LiNbO3

switches and filters. IEEE J. Quant. Electron., 27(3):602–607, 1991.
[Ramo+94] S. Ramo, J. R. Whinnery, and T. Van Duzer. Fields and Waves in Communication
Electronics, Third Edition. New York: John Wiley & Sons 1994.
[Ranalli+99] A. R. Ranalli, B. A. Scott, and J. P. Kondis. Liquid crystal-based wavelength se-
lectable cross-connect. In Proceedings of the Euro. Conf. on Opt. Commun., Vol. I, pp. 68–69,
Nice, France, September 1999.
[Reichmann+01] K. Reichmann, P. Iannone, M. Birk, N. Frigo, D. Barbier, C. Cassagnettes,
T. Garret, A. Verlucco, S. Perrier, and J. Philipsen. An eight-wavelength 160-km transpar-
ent metro WDM ring network featuring cascaded erbium-doped waveguide amplifiers. IEEE
Photon. Technol. Lett., 13(10):1130–1132, 2001.
[Reichmann+06] K. C. Reichmann, P. P. Iannone, X. Zhu, N. J. Frigo, and B. R. Hemenway.
240 km CWDM transmission using cascaded SOA Raman hybrid amplifiers with 70-nm band-
width. IEEE Photon. Technol. Lett., 18(2):328–330, 2006.
[Renaud+96] M. Renaud, M. Bachmann, and M. Erman. Semiconductor optical space switches.
IEEE/OSA J. Select. Top. Quant. Electron., 2(2):277–288, 1996.
[Rotolo+95] S. Rotolo, S. Brunazzi, R. Cadeddu, E. Casaccia, and C. Cavazzoni. Strategy for a
computer-aided analysis of all-optical multiwavelength transparent networks. In Proceedings
of the IEEE/OSA Optical Fiber Commun. Conf., Paper ThB2, pp. 205–206, San Diego, CA,
February 1995.
[Roudas+97] I. Roudas, N. Antoniades, R. E. Wagner, and L. D. Garrett. Influence of realistic
optical filter characteristics on the performance of multiwavelength optical networks. In Pro-
ceedings of the IEEE/LEOS Annual Mtg., pp. 542–543, San Francisco, CA, November 1997.
[Roudas+99] I. Roudas, D. H. Richards, N. Antoniades, J. L. Jackel, and R. E. Wagner. An
efficient simulation model of the erbium-doped fiber for the study of multiwavelength optical
networks. Opt. Fiber Technol., 5(4):363–389, 1999.
[Roudas+00] I. Roudas, N. Antoniades, D. H. Richards, R. E. Wagner, J. L. Jackel, S. F. Habiby,
T. E. Stern, and A. F. Elrefaie. Wavelength-domain simulation of multiwavelength optical
networks. IEEE/OSA J. Select. Top. Quant. Electron., 6(2):348–362, 2000.
[Roudas+01] I. Roudas, M. N. Lascar, A. E. Kruse, B. Hallock, D. Chowdhury, R. S. Vodhanel,
N. Antoniades, I. Tomkos, and M. Sharma. 10 Gb/s uncompensated transmission in transparent
optical metropolitan area networks using electroabsorption modulators over negative dispersion
fiber. In Proceedings of the Conf. on Lasers and Electro-Optics (CLEO’01), Paper CFA3,
Baltimore, MD, May 2001.
[Roudas+02] I. Roudas, N. Antoniades, T. Otani, T. E. Stern, R. E. Wagner, and D. Q. Chowdhury.
Accurate modeling of optical multiplexers/demultiplexers concatenation in multiwavelength
optical networks. IEEE/OSA J. Lightwave Technol., 20(6):921–936, 2002.
[Roudas+07] I. Roudas and N. Antoniades. Performance outages in CWDM optical networks due
to the polarization-dependent gain of semiconductor optical amplifiers. IEEE Photon. Technol.
Lett., 19(1):48–50, 2007.
[Saleh+90] A. A. M. Saleh, R. M. Jopson, J. D. Evankow, and J. Aspell. Modeling of gain in
erbium-doped fiber amplifiers. IEEE Photon. Technol. Lett., 2(10):714–717, 1990.
[Saleh+91] B. E. A. Saleh and M. C. Teich. Fundamentals of Photonics. New York: John Wiley
& Sons 1991.
[Saleh+99] A. M. Saleh and J. M. Simmons. Architectural principles of optical regional and
metropolitan access networks. IEEE/OSA J. Lightwave Technol., 17(12):2431–2448, 1999.
[Sapriel+02] J. Sapriel, D. Charissoux, V. Voloshinov, and V. Molchanov. Tunable acoustooptic
filters and equalizers for WDM applications. IEEE/OSA J. Lightwave Technol., 20(5):892–899,
2002.
322 Multiwavelength Optical Networks

[Satorius+02] D. A. Satorius, T. E. Dimmick, and G. L. Burdge. Double-pass acoustooptic

tunable bandpass filter with zero frequency shift and reduced polarization sensitivity. IEEE
Photon. Technol. Lett., 14(9):1324–1326, 2002.
[Scobey+96] M. A. Scobey and D. E. Spock. Passive DWDM components using microplasma
optical interference filters. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf.,
Paper ThK1, pp. 242–243, San Jose, CA, February 1996.
[Sharony+92] J. Sharony, K. W. Cheung, and T. E. Stern. Wavelength dialated switches (WDS) –
A new class of high-density, suppressed cross-talk, dynamic wavelength routing cross-connects.
IEEE Photon. Technol. Lett., 4(8):933–935, 1992.
[Shen+03] L.-P. Shen, W.-P. Huang, and S.-S. Jian. Design of photonic crystal fibers for
dispersion-related applications. IEEE/OSA J. Lightwave Technol., 21(7):1644–1651, 2003.
[Shimoe+87] T. Shimoe, K. Hajikamo, and K. Murikami. A path-independent optical space
switching network. In Proceedings of the International Switching Symposium (ISS), vol. 4,
Paper C12.2, pp. 999–1003, Phoenix, AZ, March 1987.
[Smith+90] D. A. Smith, J. E. Baran, J. J. Johnson, and K. W. Cheung. Integrated-optic acoustically
tunable filters for WDM networks. IEEE J. Select. Areas Commun., 8(6):1151–1159, 1990.
[Soole+94] J. B. D. Soole and C.-E. Zah. Multiwavelength semiconductor lasers advance rapidly.
Laser Focus World, 30:S9–S15, June 1994.
[Spanke87] R. A. Spanke. Architectures for guided-wave optical space switching systems. IEEE
Communications Mag., 25(5):42–48, 1987.
[Spiekman03] L. Spiekman. Optical amplification for metro: EDFA/EDWA amplets & semicon-
ductor technologies. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., tutorial
ThC4-2, Atlanta, GA, February 2003.
[Srivastava+98] A. K. Srivastava, Y. Sun, J. W. Sulhoff, C. Wolf, M. Zirngibl, R. Monnard,
A. R. Chraplyvy, A. A. Abramov, R. P. Espindola. T. A. Strasser, J. R. Pedrazzani, A. M.
Vengsarkar, J. L. Zyskind, J. Zhou, D. A. Ferrand, P. F. Wysocki, J. B. Judkins and Y. P. Li. 1 Tb/s
transmission of 100 WDM 10 Gb/s channels over 400 km of TrueWave fiber. In Proceedings
of the IEEE/OSA Optical Fiber Commun. Conf., Paper PD10, San Jose, CA, February 1998.
[Stapleton+07] A. Stapleton, N. S. Cockerham, M. Bagheri, S. Farrell, and J. O’Brien.
Silicon photonic crystal directional couplers for power splitting, wavelength filtering, and op-
tical switching. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper OMI4,
Anaheim, CA, March 2007.
[Stern60] T. E. Stern. Some quantum effects in information channels. IRE Trans. on Information
Theory, IT-6:435–440, 1960.
[Stolen+73] R. H. Stolen and E. P. Ippen. Raman gain in glass optical waveguides. Appl. Phys.
Lett., 22(6):276–278, 1973.
[Sugimoto+03] Y. Sugimoto, Y. Tanaka, N. Ikeda, T. Yang, H. Nakamura, K. Inoue, K. Miyashita,
T. Maruyama, K. Ishida, Y. Watanabe, and K. Asakawa. Design and characterization of a
coupling-strength-controlled directional coupler based on photonic crystal waveguides. In Pro-
ceedings of the 5th Pacific Rim Conf. on Lasers and Electro-optics (CLEO/Pacific Rim 2003),
Taipei, Taiwan, December 2003.
[Suzuki+98] S. Suzuki, A. Himeno, and M. Ishii. Integrated multichannel optical wavelength
selective switches incorporating an arrayed-waveguide grating multiplexer and thermooptic
switches. IEEE/OSA J. Lightwave Technol., 16(4):650–655, 1998.
[Takiguchi+03] K. Takiguchi, K. Okamoto, T. Goh, and M. Itoh. Integrated-optic dispersion slope
equalizer for N × several tens of Gb/s WDM transmission. IEEE/OSA J. Lightwave Technol.,
21(11):2463–2469, 2003.
 Enabling Technology 323

[Tomkos+01a] I. Tomkos, R. Hesse, N. Antoniades, and A. Boskovic. Impact of filter concatena-

tion effects on the performance of metropolitan area optical networks utilizing directly modu-
lated lasers. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., pp. BB4.1–BB4.3,
Anaheim, CA, March 2001.
[Tomkos+01b] I. Tomkos, D. Q. Chowdhury, D. J. Conradi, D. Culverhouse, K. Ennser, C. Giroux,
B. Hallock, T. Kennedy, A. Kruse, S. Kumar, N. Lascar, I. Roudas, M. Sharma, R. S. Vodhanel,
and C.-C. Wang. Demonstration of negative dispersion fibers for DWDM metropolitan area
networks. IEEE J. Select. Top. Quant. Electron., 7(3):439–460, 2001.
[Tsang87] W. T. Tsang. Quantum confinement heterostructure semiconductor lasers. In Semicon-
ductors and Semimetals. New York: Academic Press 1987.
[van Berlo+95] W. van Berlo, M. Janson, L. Lundgren, A.-C. Morner, J. Terlecki, M. Gustavsson,
P. Granestrand, and P. Svensson. Polarization-insensitive, monolithic 4 × 4 InGaAsP-InP laser
amplifier gate switch matrix. IEEE Photon. Technol. Lett., 7(11):1291–1293, 1995.
[Vengsarkar+96] A. M. Vengsarkar, P. J. Lemaire, J. B. Judkins, V. Bhatia, T. Erdogan, and
J. E. Sipe. Long-period gratings as band-rejection filters. IEEE/OSA J. Lightwave Technol.,
14(1):58–64, 1996.
[Wagner+96] R. E. Wagner, R. C. Alferness, A. A. M. Saleh, and M. S. Goodman. MONET:
Multiwavelength optical networking. IEEE/OSA J. Lightwave Technol., 14(6):1349–1355, 1996.
[Watson+90] J. E. Watson, M. A. Milbrodt, K. Bahadori, M. F. Dautartas, C. T. Kemmerer,
D. T. Moser, A. W. Schelling, T. O., Murphy, J. J. Veselka, and D. A. Herr. A low-voltage
8 × 8 Ti:LiNbO3 switch with a dilated-Benes architecture. IEEE/OSA J. Lightwave Technol.,
8(6):794–801, 1990.
[West+01] J. A. West, N. Venkataramam, C. M. Smith, and M. T. Gallagher. Photonic crystal
fibers. In Proceedings of the European Conference on Optical Communications, Amsterdam,
the Netherlands, October 2001.
[Westbrook+05] P. S. Westbrook, J. W. Nicholson, K. S. Feder, and A. D. Yablon. Improved
supercontinuum generation through UV processing of highly nonlinear fibers. IEEE/OSA J.
Lightwave Technol., 23(1):13–18, 2005.
[Winzer+05] P. J. Winzer, F. Fidler, M. J. Matthews, L. E. Nelson, H. J. Thiele, J. H. Sinsky,
S. Chandrasekhar, M. Winter, D. Castagnozzi, L. W. Stulz, and L. L. Buhl. 10-Gb/s upgrade of
bi-directional CWDM systems using electronic equalization and FEC. IEEE/OSA J. Lightwave
Technol., 23(1):203–210, 2005.
[Woodward+03] S. Woodward, M. Feuer, S. Huang, and M. Boroditsky. Demonstration of an
electronic dispersion compensator in a 100 km, 10 Gb/s ring network. IEEE Photon. Technol.
Lett., 5(6):867–869, 2003.
[Yoo+95] S. J. B. Yoo, C. Caneau, R. Bhat, and M. A. Koza. Wavelength conversion by quasi-
phase-matched difference frequency generation in AlGaAs waveguides. In Proceedings of the
IEEE/OSA Optical Fiber Commun. Conf., Paper PD14-2, San Diego, CA, February 1995.
[Yoo96] S. J. B. Yoo. Wavelength conversion technologies for WDM network applications.
IEEE/OSA J. Lightwave Technol., 14(6):955–966, 1996.
[Yoo+96] S. J. B. Yoo, C. Caneau, R. Bhat, M. A. Koza, A. Rajhel, and N. Antoniades. Wavelength
conversion by difference frequency generation in AlGaAs waveguides with periodic domain
inversion achieved by wafer bonding. Appl. Phys. Lett., 68(19):2609–2611, 1996.
[Yoshikuni02] Y. Yoshikuni. Semiconductor arrayed waveguide gratings for photonic integrated
devices. IEEE J. Select. Top. Quant. Electron., 8(6):1102–1114, 2002.
[Zimmerman+04] D. Zimmerman and L. Spiekman. Amplifiers for the masses: EDFA, EDWA,
and SOA amplets for metro and access applications. IEEE/OSA J. Lightwave Technol., 22(1):63–
70, 2004.
5 Static Multipoint Networks

In static networks essentially all functionality resides in the network access stations
(NASs). The performance of the network is therefore determined by how the NASs
provide logical connectivity and throughput to satisfy the network’s traffic requirements.
This chapter explores the performance issues in static networks, viewing them all as
special cases of shared media, as described in Section 5.1. Existing and potential uses of
shared media abound, the most important of these being to provide efficient local access
for end users to a larger optical network. The multiplexing and multiple-access tech-
niques required to achieve multipoint logical connectivity in these networks are treated
in Section 5.2. Sections 5.3 through 5.6 deal with capacity allocation and control to serve
prescribed traffic requirements. We first point out some general flow conservation con-
straints that must be satisfied in any shared-channel system. Then the problems of traffic
scheduling and control are discussed in settings with increasing degrees of complexity:
dedicated connections (Section 5.4), demand-assigned connections (Section 5.5), and
packet switching in the optical layer (Section 5.6). Section 5.7 discusses network access
applications of static multipoint architectures. These include broadcast star-based and
wavelength-router-based passive optical networks (PONs) that provide the foundation
of fiber to the home/premises. In these applications the static network is the link between
the end user and an optical core or metropolitan area network.

5.1 Shared Media: The Broadcast Star

The simplest form of a transparent optical network, the static network, was defined in
Chapter 3 as a collection of fixed (passive) splitting/combining nodes without wavelength
selectivity, interconnected by fibers that provide full or partial connectivity among a set
of NASs. Of the static networks, the most elementary form is the broadcast star, in which
all signals transmitted from the NASs are combined at a star coupler and broadcast to
all receivers. We have seen, however, that many seemingly more complex topologies
are functionally equivalent to the broadcast star, including folded buses, trees (see
Figure 2.7), and embedded rooted trees (see Section 3.4.2). Thus, we can consider the
broadcast star as a fairly general representation of a large family of static networks.
Its essential feature is that it acts as a shared broadcast medium. This chapter explores
in detail the properties of shared-medium networks operated in a broadcast-and-select
mode, using the star as a prototypical example.
 Static Multipoint Networks 325

1
M
1
1

C
C

M N
2
Directed
Undirected
(a) Two stars

Network 1

Network 2

Subnet 1

Subnet 2 Combine

NAS
=
C Star coupler

Electronic concentrator
(b) Star-based access
Passive splitter/combiner Split
Sp

Figure 5.1 Star networks.

Before adopting abstract models for these networks, let us first examine a few ways in
which a star network might be used in practice. Figure 5.1 shows examples of networks
based on broadcast stars. Figure 5.1(a) shows the undirected and directed versions of
the star. Each NAS may serve several colocated end systems or may be a part of another
network. (These are identical to Figures 2.7[a] and 2.11, except that a general M × N
star is shown in the directed case.) In the undirected case, as illustrated in Figure 5.1(b),
we show bidirectional transmission on each link, and we assume here that a bidirectional
326 Multiwavelength Optical Networks

link is realized using a pair of unidirectional fibers and a combiner and splitter.1 The
figure shows how a broadcast star may be used as an access network connecting end
systems to two larger networks.
When the end systems are spread over a fairly wide area (say, tens of kilometers), it
may not be cost-effective to run a fiber link directly from the star coupler to each end
system. Instead, a more economical approach is to divide the overall broadcast network
into a central part (the star coupler and its access links) and several lower capacity local
access subnets. This would be especially suitable for a network serving several separate
but densely populated clusters of end users; for example, groups of users in different
buildings in the same city. The figure shows an example of two subnets connected
to an undirected star network. Each node of a subnet, combines traffic from two or
more sources onto a link carrying it upstream, so all traffic generated in the subnet is
concentrated onto a single fiber accessing the star node C. Similarly, outbound fibers
on the star carry traffic downstream to one of the local access subnets, which in turn
distributes it to the various destination points. A subnet can be configured in several
different ways.
An electronic approach is shown in subnet 1 of Figure 5.1(b). It connects a set of end
systems via electronic concentrating nodes to an NAS linked to the star node. The con-
centrators perform an electronic multiplexing function upstream and a demultiplexing
function downstream. Because demultiplexing is performed at the concentrators, each
end system receives only the information destined for it.
An optical approach is shown in subnet 2. Comparing this to subnet 1, we see that
the electronic concentrators and the NAS accessing the star node are replaced by optical
network nodes, which are composed of pairs of fixed (passive) optical splitters and
combiners. The end systems are now interfaced to the local access subnet through
NASs. Signals from the end systems are multiplexed and converted to optical form
in their NAS, and the optical signals from the NASs are combined in the tree so that
all transmitted traffic from an access subnet is concentrated on a single upstream fiber
accessing the star coupler. Conversely, optical signals carrying traffic on a downstream
fiber from the star coupler are distributed to all stations in the access subnet connected to
that fiber using fixed splitters at each node of the tree.2 Although the subnets aggregate
traffic from the end users and distribute traffic to the end users, they do not provide
connectivity among the end users. If that connectivity is desired, it must be implemented
at the level of the star coupler or beyond.
To complete the picture in Figure 5.1(b), two of the NASs are shown belonging to larger
networks. In this case the star coupler is acting as a bridge joining a large collection of end
users to the two larger networks. The NASs within the larger networks might represent
communication service providers (for example, Internet Service Providers [ISPs]) and

1
Later in this chapter networks using single bidirectional fibers will be discussed (see Section 5.7).
2
One technical problem associated with the optical approach is the fact that the signal powers from each
transmitting NAS may not be equal after they are combined on the upstream access fiber, because of
combining losses at the access network nodes and fiber attenuation. In principle, it is possible to adjust
the combining ratios at each node and/or insert optical amplifiers to ensure power equality and maintain
satisfactory power levels. This problem occurs in all passive tree networks.
 Static Multipoint Networks 327

the star network together with its subnets provides access for the end systems to the
larger networks.
It is important to observe the differences between the subnet configurations as far
as network operation is concerned. In the case of an electronic access subnet, the star
coupler sees one NAS, ignoring the fact that there is a subnet behind it.
Things are very different when the local access network is purely optical. Now the star
coupler sees all NASs in the access subnet as if they were directly connected to it. The
result is an equivalent broadcast star with a much larger number of NASs. The optical
multiplexing and multiple-access techniques used on the star must now be extended back
to the stations on all the leaves of the access trees. This arrangement would normally
be more costly than the electronic one, assuming that NASs are more expensive than
electronic concentrators. However, it would have more capacity and flexibility because
there are no electronic bottlenecks in the access networks.
The above discussion focused on an application where the star network is used as an
adjunct to one or more larger networks. However, as mentioned above, a star network
might be embedded on a larger optical infrastructure, as illustrated in Figure 3.27. In
this case, the star coupler together with associated optical access nodes are realized as
a rooted tree using linear divider-combiners (LDCs). Using waveband-selective LDCs,
the embedded star could be arranged to operate on a prescribed portion of the available
optical spectrum (one or more wavebands), with the spectrum allocation chosen to
suit the traffic requirements. In this application the embedded star might be part of a
multistar-based hypernet as described in Section 3.4.3.4. Although shared media exist
in many different physical incarnations, the basic multiplexing, multiple-access, and
control techniques discussed in this chapter are valid for any shared medium network.

5.2 Representative Multiplexing and Multiple-Access Schemes

Multiplexing and multiple access play an essential role in a shared-medium network

because this is the only way multiple logical connections (LCs) can be supported by
the network. Therefore, as a prelude to a more general discussion of shared-medium
networks, we first discuss some typical multiplexing and multiple-access techniques in
an optical broadcast-and-select environment.
Although multiplexing and multiple access have a long history, it is important to
review their essential characteristics in the context of optical networks. Because the
constraints of optical and electronic technologies are quite different, the advantages and
disadvantages of various techniques in the optical domain are quite different than in the
traditional electronic domain. Time and wavelength division multiplexing and multiple
access were introduced in Section 3.2.1, where it was pointed out that wavelength division
techniques alone cannot provide the degree of logical connectivity that is required in
a large network. Of the many possibilities for creating high logical connectivity, we
consider four representative schemes here: TDM/TDMA, TDM/T-WDMA, subcarrier
multiplexing and multiple access (SCM/SCMA), and code division multiple access
(CDMA).
328 Multiwavelength Optical Networks

The first two are treated together because TDM/TDMA is a special case of TDM/
T-WDMA operating on one λ-channel. All of these except CDMA are orthogonal
techniques, characterized by the fact that signals accessing a common medium are
nonoverlapping in time and/or optical frequency. CDMA is a quasiorthogonal technique,
requiring considerably more signal processing and transmission bandwidth than the other
three techniques, but which is sometimes advantageous in situations involving large
numbers of low-throughput users (typical of optical packet-switching applications).
Our discussion focuses on dedicated connections carrying stream-type traffic, with
some digressions for packet-switched traffic when appropriate.

5.2.1 Time-Wavelength-Division Multiplexing/Multiple Access

In Section 3.2.1, we showed how several LCs can be maintained simultaneously on a
3 × 3 broadcast star by time-sharing one or more λ-channels. To illustrate the funda-
mental features of TDM/T-WDMA, we expand on that example now.
Let us recall from Section 3.2.1 that the four ingredients necessary for realizing
multipoint logical connectivity in a TDM/T-WDMA system operating on a shared broad-
cast medium are

r Transmission channel multiplexing in the transmitting stations using TDM

r Optical broadcast on the shared medium
r Optical multiple access using T-WDMA
r Transmission channel demultiplexing in the receiving stations

Suppose the shared medium is an M × N directed star, and we have C λ-channels

available to carry the connections. For simplicity, we focus on the case in which all
optical transmitters (OTs) operate at a common bit rate R t bits per second, and each LC
is point-to-point. This still allows for one-to-many and many-to-one logical connectivity.
This means that the information flows on each LC are distinct, which rules out logical
multicast connections. Each connection carries stream traffic at a prescribed bit rate, and
we wish to provide dedicated connections by allocating fixed transmission capacities
that match the flow requirements. In this case, the traffic may be characterized by a set of
M N prescribed flows, described in terms of an M × N traffic matrix: Ŵ = [γi j ]. Each
entry γi j represents the flow requirement on an LC between the source i and destination
j in bits per second. (The matrix Ŵ alone is not sufficient to characterize logical multicast
connections [see Section 5.4.1.4].) The total traffic represented by Ŵ is

γ̄ ≡ γi j . (5.1)
ij

The connections are carried on transmission channels multiplexed according to a

fixed-frame schedule with L equal-size time slots. The channel–slot allocations can
be expressed as a C × L channel allocation schedule (CAS). If ci j is the number of
channel–slots allocated to connection [i, j], then this represents a capacity allocation of
R t ci j /L bits per second, which is the effective bit rate of that connection. Because ci j is
 Static Multipoint Networks 329

an integer, the basic quantum of capacity available in a fixed-frame system with frame
length L is R t /L bits per second.
To avoid information loss, the CAS must be collisionless (no more than one transmitter
may use the same channel in the same time slot) and conflict free. For stations equipped
with single receivers, conflicts occur if more than one station transmits to the same
receiver in the same time slot. (For a destination station with β receivers, as many as β
transmissions are allowed to that station in the same time slot, but they must all be on
distinct channels.)
The parameters R t and C are limited by technological constraints. Thus, there is
a limit on transmitter bit rate: R t ≤ Rmax , set by the speed of the electronics, the
modulation bandwidth of the laser transmitters, and the capacity of the λ-channels.
Similarly, the number of channels, C, is limited by the available optical spectrum and
the feasible channel-packing density. Finally, the functionality of the NASs (number of
transmitters/receivers and their tunability) determines the scheduling options within a
frame. The totality of these constraints determines whether it is possible to find a CAS
that satisfies a given traffic requirement, Ŵ. To illustrate this point, we return to the
undirected 3 × 3 network of Figure 3.6, redrawn for convenience in Figure 5.2.
Let the traffic matrix be expressed as Ŵ = R 0 T , where Ŵ has been decomposed into
a normalized (dimensionless and integer-valued) traffic matrix T multiplied by a scalar
constant R 0 (bits per second), called the basic bit rate. We denote the total normalized
traffic as

T̄ ≡ ti j . (5.2)
ij

Three cases of normalized traffic matrices and their corresponding CASs are provided
in Figure 5.3. In each case, the number of channel–slots ci j allocated to an LC is equal
to the normalized traffic requirement ti j . Now, equating the capacity allocation R t ci j /L
to the flow requirement R 0 ti j , and using the fact that ci j = ti j , we find that to provide
the required capacity, each transmitter must operate at a rate

Rt = R0L. (5.3)

Equation (5.3) represents an acceleration of R t relative to R 0 . This results in a “speed-

up” factor, R t /R 0 , which is equal to the frame length L for the channel–slot allocation;
in other words, the longer the frame the higher the required transmitter speed. Let us see
how this works out for the cases presented in Figure 5.3.
In Figure 5.3(a), the normalized traffic specified in the matrix T is to be carried
using TDM/TDMA; that is, using a single λ-channel. A suitable CAS is shown in the
figure. (Note that the traffic distribution and slot schedule corresponds to that shown
in Figure 3.8.) Because this CAS requires a frame of length L = 4, each transmitter
must operate at a bit rate of R t = 4R 0 . If this value exceeds Rmax , the traffic cannot be
supported by the network.
The remaining two cases in Figure 5.3 represent TDM/T-WDMA networks with
different traffic requirements, and with two channels operating on wavelengths λ1 and
330 Multiwavelength Optical Networks

NAS 1

[1, 2] [2, 1]
TP OT OR RP
[1, 3] λ1,2 λ1 [3, 1]

[2, 1] [1, 2]
TP OT C OR RP
[2, 3] λ1,2 λ1 [3, 2]

[3, 1] [1, 3]
TP OT OR RP
[3, 2] λ2 λ2 [2, 3]

(a) TT-FR System

NAS 1

[1, 2] [2, 1]
TP OT OR RP
[1, 3] λ1 λ2 [3, 1]

[2, 1] [1, 2]
TP OT C OR RP
[2, 3] λ2 λ1 [3, 2]

[3, 1] [1, 3]
TP OT OR RP
[3, 2] λ1,2 λ 1,2 [2, 3]

(b) TT-TR System

Figure 5.2 A 3 × 3 example.

λ2 . Stations are considered with varying degrees of tunability, and it is assumed in all
cases that each station has a single transceiver.
A nonuniform normalized traffic matrix is shown in Figure 5.3(b). The objective now
is to focus on the tunability options when C = 2 channels are available. The figure shows
a possible CAS for a TT-FR system in which receivers 1 and 2 are both permanently
tuned to channel 1, and receiver 3 is tuned to channel 2. Transmitting stations 1 and
2 tune over both channels, and transmitter 3 is fixed-tuned to channel 2. Figure 5.2(a)
shows the system, indicating the tuning of the various OTs and ORs.
Note that the schedule has a length L = 7 and that two channel–slots are empty.
Because they are both tuned to the same channel, the ORs in stations 1 and 2 capture all
data on channel 1. Thus, the information destined for each station must be selected at
the electronic level using synchronized gates in their reception processors (RPs). This
 Static Multipoint Networks 331

0 2 1 [1, 2] [1, 2] [1, 3] [2, 3]

T= 0 0 1
0 0 0 Slot 1 2 3 4
t
(a) TDM/TDMA System

0 1 3 Channel 1 [2, 1] [2, 1] [3, 1] [3, 1] [1, 2] [3, 2] [3, 2]

T= 2 0 2
2 2 0 Channel 2 [1, 3] [1, 3] [1, 3] [2, 3] [2, 3]

Slot 1 2 3 4 5 6 7

(b) TT-FR System

Channel 1 [1, 2] [3, 2] [1, 3] [1, 3] [3, 2] [1, 3]

Channel 2 [2, 3] [2, 1] [2, 1] [3, 1] [2, 3] [3, 1]

Slot 1 2 3 4 5 6

(c) TT-TR System

Figure 5.3 TDM/T-WDMA channel allocation schedules.

schedule requires a transmitter bit rate of R t = 7R 0 . For example, if the basic bit rate is
R 0 = 500 Mbps, then each transmitter must run at a rate R t = 3.5 Gbps.
For this example, no TT-FR schedule can produce a frame length shorter than seven,
and no FT-TR schedule can produce a frame length shorter than eight. Why? To obtain a
more efficient schedule, we must move to TT-TR. In this case, a possible six-slot schedule
for the normalized traffic of Figure 5.3(b) is shown in Figure 5.3(c). Only the transceiver
in station 3 is required to be tunable; transmitters 1 and 2, are tuned permanently to
channels 1 and 2, respectively. Similarly, receivers 1 and 2 are tuned permanently to
channels 2 and 1, respectively (see Figure 5.2[b]). The required transmitter bit rate of
R t = 6R 0 is an improvement over the TT-FR case, allowing the prescribed traffic to
be carried with slower transmitters. Alternatively, the throughput can be “scaled up” by
maintaining the same transmitter speed and increasing R 0 by a factor of 7/6.
Whenever tunability is required, wavelength-selectable arrays of transmitters or re-
ceivers can be used instead of tunable devices. (These are arrays with only one transmitter
or receiver enabled at a time.) Thus, in the TT-TR case, the same CAS can be realized
with arrays of two transmitters and two receivers in station 3, and single fixed transceivers
in the other stations.
In the previous examples, C was taken to be fixed. However, it is often of interest to con-
sider trade-offs between the number of channels used and the resultant traffic-handling
capacity of the network. For illustration, consider an undirected M × M broadcast star
network, where M = 81. We require full logical connectivity (without loopback connec-
tions) with uniform traffic; in other words, the traffic matrix is Ŵ = R 0 T , where ti j = 1
for i = j and is zero otherwise. Using an FT-TR system with C = 81, a CAS can be ob-
tained that is a direct generalization of the 3 × 3 case of Figure 3.9(b) (see Figure 5.4[a]).
In this case the traffic can be scheduled in a frame of length L = (M − 1) = 80.
332 Multiwavelength Optical Networks

Channel 1 [1, 2] [1, 3]

Channel 2 [2, 3] [2, 4]

Channel 3 [3, 4] [3, 5]

Channel 4 [4, 5] [4, 6]

Slot 1 2 3
t
(a)

Channel 1 [1, 2] [1, 3] [1, 4] [4, 1] [4, 2] [4, 3] [7, 1]

Channel 2 [2, 3] [2, 4] [2, 5] [5, 2] [5, 3] [5, 4] [8, 2]
Channel 3 [3, 4] [3, 5] [3, 6] [6, 3] [6, 4] [6, 5] [9, 3]

Slot 1 2 3 81 82 83 161

(b)

Figure 5.4 Illustrating channel reuse in an FT-TR system.

Now, suppose only three channels are available. Continuing in the FT-TR mode, let
us partition the set of stations into three subsets containing 27 stations each, and as-
sign one channel to all transmitting stations in the same subset. For example, assign
channel k to station j, where k = 1 + [( j − 1) mod 3]. With these assignments, each
transmitting station owns a channel, but there is an r -fold reuse of the channels, where
r = M/C. It is still possible to schedule the traffic, provided that we observe the con-
straint that all stations transmitting in the same time slot must use distinct channels.
A possible CAS for this case is shown in Figure 5.4(b), where r = 27 and a frame of
length L = r (M − 1) = 2160 is now required. Because each station is active for only
1/r of the time, the distinct channel constraint, which is required to avoid collisions,
translates into an r -fold increase in frame length. This results in a reduction of aggre-
gate throughput by a factor of r unless transmission speed R t is increased by the same
factor.
These examples were designed to show how traffic can be scheduled in a TDM/
T-WDMA system while accounting for the various system constraints. A general treat-
ment of scheduling in shared-channel systems is presented in Section 5.4.

5.2.1.1 Scheduling Efficiency

The previous section illustrates certain relations among traffic requirements, net-
work resources (available channels), and station functionality (maximum bit rates
and transceiver tunability) in shared-channel systems. There is an obvious trade-off
between bit rates and the number of available channels. More channels means lower
required bit rates, with the pure TDM/TDMA (single-channel) system requiring the
highest bit rate. However, the required value of R t also depends on how efficiently
the traffic is scheduled. To quantify this idea, let us define scheduling efficiency ηs as
 Static Multipoint Networks 333

the ratio of the total traffic requirement to the total capacity used to carry it:
ηs ≡ γ̄ /R t C. (5.4)
Using Equations (5.2) and (5.3), we find that Equation (5.4) can also be expressed as
ηs = T̄ /LC. (5.5)
It can be seen from these expressions that ηs = 1 when the available channels are
fully occupied, that is, when there are no idle slots in the frame. All schedules presented
earlier had ηs = 1, except for the TT-FR network of Figure 5.3(b), which had ηs = 6/7.
A well-balanced traffic matrix and/or a high degree of transceiver tunability permit CASs
with high scheduling efficiency. Maximizing ηs minimizes the frame length L and hence
minimizes the transmitter bit rate required to support the given aggregate traffic. The
interrelations among traffic requirements, system resources, and scheduling efficiency
are discussed more generally in Section 5.4.1.

5.2.1.2 Transmission Constraints: Optical Spectral Efficiency

We have seen that the total capacity of a shared-channel system is given by R t C (in bits
per second). Assuming that this capacity is realized within a limited optical bandwidth,
a high capacity requires high-speed channels and a high channel-packing density. Both
of these parameters are limited by many physical constraints. They include the following
(see Chapter 4):
r Laser characteristics: modulation bandwidth, linewidth, chirp, and wavelength
stability
r Transmission impairments: fiber dispersion, nonlinearities, and attenuation
r Receiver characteristics: optical filter imperfections and receiver noise and detection
method (direct or coherent)
r Signal processing: electronic speed limitations and modulation techniques used in the
TPs/RPs
Even though these represent a complex set of interrelated constraints, their overall
effect at the network level can be subsumed into a single parameter: the optical spectral
efficiency, which we define as
ηop ≡ R t C/Bop (bps/Hz), (5.6)
where Bop is the total optical bandwidth occupied by the shared-channel system. The
parameter ηop represents the throughput of the shared-channel system per unit of opti-
cal bandwidth, assuming that all channels are utilized fully (i.e., ηs = 1). Combining
scheduling efficiency and optical spectral efficiency, we have the following relation
between total throughput and system bandwidth:
γ̄ = ηs ηop Bop . (5.7)
To understand the factors influencing optical spectral efficiency in the context of
TDM/T-WDMA, we must examine the spectral characteristics of the signals at several
points in the system. Suppose a set of access stations shares a set of C λ-channels
334 Multiwavelength Optical Networks

[j, 1]
[j, 2] TP s(t)
OT
Bit rate Rt a
[j, k]

(a)

Ss(f )
Bb

f
0 1
2T
(b)

Sm(ν)
Bm

ν
0 νn
(c)

Bop

Bm

ν
0 ν1 ν2 νC
∆ν
(d )

Figure 5.5 Illustrating optical spectral efficiency.

running at optical frequencies ν1 , ν2 , . . . , νC . Figure 5.5(a) shows the structure of station

j, where a number of LCs, [ j, 1], [ j, 2], . . . , [ j, k], . . . are time division multiplexed
onto a common transmission signal at point a in the figure. The multiplexed signal is
used to modulate the laser transmitter O T. Any digital modulation format can be used
as long as it is compatible with the characteristics of the laser modulator and the ORs.
Consider a single LC [ j, k], and assume that it occupies one or more slots on a fixed
λ-channel at optical frequency νn . Also assume for simplicity that the data symbols are
binary, so each transmitted symbol carries 1 bit of information in a symbol interval of
T seconds for a symbol rate of 1/T = R t . Then the baseband power spectral density
(PSD) Ss ( f ) of a signal s(t) carrying one of the multiplexed connections appears as
shown in Figure 5.5(b). Note that it occupies an electrical bandwidth Bb > 1/2T , where
1/2T is the Nyquist frequency, which determines the minimum possible spectral width
for this signal. The PSD Sm (ν) of the modulated optical signal at the output of the laser
transmitter is shown in Figure 5.5(c).
The shape of Sm (ν) and its bandwidth Bm varies considerably, depending on effects
at the electrical (transmission channel) level and the optical level. At the electrical level,
 Static Multipoint Networks 335

the bandwidth of the signal modulating the laser depends on the modulation format
used for the original data. Some formats (e.g., OOK and phase-shift keying [PSK])
are bandwidth conserving; others (e.g., frequency-shift keying [FSK] and CDMA) are
bandwidth expanding. Furthermore, if this signal is used to intensity modulate the
optical carrier, there will be a bandwidth expansion introduced by the nonlinearity of
the modulation process (see Section 5.2.2.2). Additional bandwidth expansion factors
are introduced by two laser characteristics: (1) the linewidth of the unmodulated laser
and (2) the chirp (optical frequency modulation) introduced when the laser is modulated
directly (see Section 4.5.3.1).
Given R t and Bm , the spectral efficiency of a digital modulation scheme is defined as
ηm ≡ R t /Bm (bps/Hz). (5.8)
In “ideal” cases, in which bandwidth-conserving modulation methods are used, laser
linewidth is negligible, external modulation is employed, and reception is via coherent
detection, it is possible to obtain values of ηm on the order of 1 bps/Hz. However, typical
implementations using intensity modulation and direct detection3 have efficiencies that
are small fractions of the ideal.
When multiple stations are active using WDMA, the complete system PSD appears
as in Figure 5.5(d). Note that C separate spectral components are shown centered at the
available optical frequencies, which are spaced equally at intervals of ν Hz. In the case
of fixed-tuned transmitters with distinct frequencies assigned to each transmitter, these
would correspond to the spectra of individual transmitters. In other cases (tunable trans-
mitters and/or fixed-tuned transmitters time sharing the same channels), each separate
component might represent a superposition of signals from more than one station. The
value of ν is chosen to ensure that the individual spectral components do not overlap
(ν > Bm ) and to allow for a guard band, which is essential for providing tolerance for
laser frequency drift and imperfections in receiver tuning and filtering. (In practice, the
guard band might be one or more orders of magnitude greater than the signal spectral
width to allow for component manufacturing tolerances and parameter drift over time.)
In this system the total occupied optical bandwidth is Bop = Cν, so we have an
optical spectral efficiency of
ηop = R t /ν < ηm . (5.9)
Clearly, highest performance is achieved when both optical spectral efficiency and
scheduling efficiency are high. However, these come at a price. The former requires
high-quality optical transceivers, and the latter requires rapid transceiver tunability or
the use of laser and/or detector arrays, all of which are costly.
Although TDM/T-WDMA has been discussed as a specific multiplexing and multiple-
access technique using the λ-channel as the basic information-carrying medium, the
basic concepts used in this section apply to any multichannel scheme. The subcarrier
and CDMA techniques discussed next provide alternative ways of “carving” channels
out of the optical spectrum.
3
Intensity modulation precludes the use of optical phase to carry information, which prevents the use of
certain bandwidth-conserving modulation schemes such as PSK (see Section 4.7.1).
336 Multiwavelength Optical Networks

5.2.2 Subcarriers
Combined TDM and T-WDMA techniques are effective ways of achieving a high degree
of multipoint logical connectivity while maintaining high aggregate throughput. How-
ever, their station complexity (and cost) tends to increase rapidly with the number of
required connections. High connectivity requires many (and therefore densely packed)
λ-channels and rapidly tunable, highly stable transceivers (or large transceiver arrays)
and/or very-high-speed channels. Each of these requirements is costly. Therefore, it is
important to seek ways of supporting multipoint LCs without expensive or highly repli-
cated optical or electro-optic components. An interesting approach is SCM and SCMA.
Because an additional level of modulation/demodulation is introduced electronically in
these systems, operations that are costly at the optical level (e.g., rapid tunability) can
be replaced by electronic operations at the transmission channel level. Replacement of
optics by electronics is one of the main advantages of the subcarrier approach in shared
media systems, especially when cost is the primary issue.
Subcarriers were used for point-to-point transmission links in some of the earliest
fiber-optic systems: most notably in distribution of frequency division multiplexed cable
television (CATV) signals over hybrid fiber-coaxial (HFC) networks.4 This is an example
of an application in which the information to be transported by the network is already
modulated onto electrical carriers, so it is natural, efficient, and cost-effective to keep it
in that form when transporting it on the fiber medium. Other similar situations exist in
optical wireless access networks for mobile communication systems (see Section 5.7.4).
Subcarrier techniques can be used by themselves or in combination with other
multiplexing/multiple-access methods. In this section, we discuss several configurations
in which subcarriers are used as part of a shared-channel multiplexed/multiple-access
system to support multipoint logical connectivity. Three possibilities are described:
TDM/T-SCMA, SCM/SCMA, and SCM/WDMA/SCMA. We begin with a description
and analysis of SCMA.

5.2.2.1 Principles of SCMA

Subcarrier multiple access provides the means for supporting many-to-one logical con-
nectivity using only electronic signal processing in the NASs. For illustration, consider a
directed 4 × 4 broadcast star SCMA network in which all stations are configured to sup-
port potential full bipartite logical connectivity. Figure 5.6 shows an example in which
each transmitting station operates on an optical frequency ν j and a subcarrier frequency
f j . (The required values of the ν j s are discussed in Section 5.2.2.2.) As in any broadcast
network, one-to-many connectivity is achievable because of the signal splitting at the
star coupler. For many-to-one connectivity, the ORs must be wideband so that all optical
signals are picked up by the receivers, and thus all subcarrier frequencies are present in
the photocurrent at the output of the OR. Any desired LC can be created at a receiving
station by selecting the corresponding subcarriers, an operation that is executed in the
station’s RP.
4
One of the reasons for the appeal of subcarrier techniques is that the supporting microwave technology is
mature and inexpensive because it has been used widely in CATV systems for many years.
 Static Multipoint Networks 337

f1 ν1 f1 f4
TP OT OR RP

f2 ν2 f1 f4
TP OT OR RP

f3 ν3 f1 f4
TP OT OR RP

f4 ν4 f1 f4
TP OT OR RP

Figure 5.6 SCMA example.

Figure 5.7 shows the structure of the transmitting and receiving stations in more
detail for a 4 × 4 network. Each transmitting station, see Figure 5.7(a), has a fixed-tuned
OT, with station j operating on a λ-channel at optical frequency ν j . If transmitting
station j is supporting full connectivity to all four receiving stations, then four LCs –
[ j, 1], [ j, 2], [ j, 3], and [ j, 4] – destined for the corresponding receiving stations are
multiplexed onto a common data signal at point a ′ in Figure 5.7(a). (For the moment,
the type of multiplexing used is irrelevant.)
The information carried may be digital, analog, or a combination of both in the case of
multiple logical channels. Now, instead of using the data to modulate the laser directly,
in SCMA it is first translated from baseband to a microwave subcarrier frequency by an
electronic modulator (labeled SCMOD in the figure), which we consider to be part of the
TP. (As indicated above, in some cases the data may already be in the form of a modulated
subcarrier.) The modulated subcarrier signal s j (t) appears at point a in Figure 5.7(a).

fj
[j, 1]
[j, 2] sj(t) vj
MUX OT
[j, 3] a′ a b
[j, 4] SCMOD
TP
(a) Transmitting Station j

LO
fIF
[1, k]
i(t) h
OR BPF ED DMUX
e e′ [4, k]
SCDMOD

RP
(b) Receiving Station k

Figure 5.7 Transmitting and receiving stations equipped for SCMA.

338 Multiwavelength Optical Networks

A wide variety of modulation formats can be used in the modulator, employing standard
electronic techniques. For analog signals, single-sideband AM might be used to conserve
microwave bandwidth, whereas FM would be used to reduce the effects of noise. For
digital signals, the options range from bandwidth-conserving techniques such as PSK
to bandwidth-expanding methods, including FSK and CDMA (see Section 5.2.3).
In typical SCMA systems, the signal s j (t) intensity modulates the laser, which is
operating at a fixed optical frequency ν j . (Direct modulation is currently most common
because it is less costly, but external modulation is also used.) To fix ideas, suppose that
digital data are impressed on the subcarrier using quadrature phase-shift keying (QPSK),
so that the modulated subcarrier signal is of the form
∞
s j (t) = u(t − lT ) cos(2π f j t + φl ), (5.10)
l =−∞

where

1, −T /2 < t ≤ T /2
u(t) = (5.11)
0, otherwise.
In Equation (5.10), T is the symbol interval, u(t) is a rectangular pulse,5 and the phase
φl is drawn from the set {0, π/2, −π/2, π } and carries 2 bits of data. The subcarrier
frequency f j (in hertz) is assumed to be fixed at this point, so the transmitted signal has
a pair of carrier frequencies: f j and ν j . (Tunable subcarrier frequencies appear in some
of the multiplexing and multiple-access methods described in Section 5.2.2.3.)
On the receiving side (illustrated by receiving station k in Figure 5.7[b]), an optical
signal consisting of a superposition of all transmitted signals is received and converted
to electrical form in a direct-detection OR using a wideband photodetector capable
of detecting all transmitted λ-channels. (Because the photodetector is wideband, OR
tunability is not an issue.) It is then the function of the subcarrier demodulator and
demultiplexers in the RP to select the data signals destined for the receiver. The first
stage of the selection process is subcarrier demodulation, translating the modulated
subcarrier signal back to baseband at point e′ in Figure 5.7(b). The photocurrent i(t) is
first downconverted by mixing it with a signal generated by a tunable local oscillator so
the signal carried on a selected subcarrier frequency is translated to a fixed intermediate
frequency f IF . The desired signal is selected at point h using a bandpass filter centered
at f IF and is then converted to baseband by an electronic detector. The combination
of local oscillator, mixer, bandpass filter, and electronic detector constitute the tunable
subcarrier demodulator. Depending on the details of the multiplexing/multiple-access
method being used, additional processing may occur in the demultiplexer shown in
Figure 5.7(b). Note that this is the classical heterodyne configuration, similar to that
used in both radio receivers and coherent optical receivers.
The modulated subcarrier signal s j (t) is a bandpass signal that occupies an electrical
bandwidth B j . Because the signals from all transmitting stations must be distinguishable
5
This is a slight simplification. In subcarrier systems, it is important to conserve electrical bandwidth. But a
rectangular pulse has infinite bandwidth, so in real systems u(t) would be “shaped” to produce a signal that
is appropriately bandlimited.
 Static Multipoint Networks 339

in the photocurrent at the receiving station (point e in Figure 5.7[b]), some sort of
orthogonality (or quasiorthogonality) must be maintained at the electrical level; i.e., the
subcarrier level. (Orthogonality on the electrical level is required because the wideband
photodetector cannot separate the superimposed signals at the optical level.) Typically,
distinct subcarrier frequencies are used by all transmitting stations, with the frequencies
spaced apart far enough so that the signal spectra from all transmitters are nonoverlapping
in electrical frequency.6
Assuming that the electrical frequency bands do not overlap to ensure orthogonal-
ity, the aggregate bandwidth occupied by all active subcarrier signals combined is

BT ≥ j B j – see Figure 5.8(a), which shows a case where B j = B for all j. The
value of BT cannot exceed the maximum modulation frequency of the laser (assuming
direct modulation). The highest subcarrier frequency is also limited by fiber disper-
sion. Dispersion causes the relative phases of the sidebands in the optical signal to be
modified as the signal propagates along the fiber. As a result, the subcarrier intensity
modulation is attenuated and replaced by phase modulation, which cannot be observed
in a direct-detection receiver. Unless measures are taken to mitigate this effect, it can
significantly limit transmission distances and/or maximum subcarrier frequencies (see
Section 4.3.2.3).
Each transmitting station can operate at a different bit rate and use a different sub-
carrier modulation format as long as the spectra of all modulated subcarriers fit into the
total bandwidth BT without overlapping. This is a fundamental limitation of SCMA,
and it means that the aggregate bandwidth of the signals from all transmitting stations
is limited by the maximum allowable modulation frequency of the transmitter. Thus, for
subcarrier applications using direct modulation, lasers with especially large modulation
bandwidths are required. (Typical values are 20 GHz, corresponding to aggregate net-
work throughputs on the order of 20 Gbps.) This situation is analogous to TDMA, in
which the aggregate throughput is limited by the maximum transmitter bit rate Rmax ,
which is in turn limited by the maximum allowable modulation frequency of the laser
transmitter. But SCMA contrasts sharply with WDMA, in which the laser modulation
bandwidth limits only the bandwidth of its own transmitted data signal, so lasers with
modulation bandwidths on the order of 1 GHz are sufficient.

5.2.2.2 Spectra
A quantitative performance analysis of subcarrier systems requires a closer look at
their spectral characteristics. Let us continue with the 4 × 4 example, assuming that
each modulated subcarrier s j (t) has bandwidth B. If all subcarrier frequencies are
spaced equally, the spectral densities Ss( j) ( f ) of the modulated subcarriers appear as in
Figure 5.8(a). Each signal s j (t) intensity modulates its laser to give the combined optical
PSD So (ν) shown in Figure 5.8(b). The figure shows typical double sideband (DSB)
spectra. However, other optical modulation formats are possible.
Although the optical carrier frequencies ν j should be distinct, they play no role in the
SCMA process, so their precise values are unimportant. In fact, when very wideband
6
Orthogonality could be maintained in other ways. For example, the electrical spectrum could be shared
among the different transmitting stations using TDMA, CDMA, or other techniques.
340 Multiwavelength Optical Networks

BT
(1) (4)
Ss (f ) Ss (f )
B

f
0 f1 f2 f3 f4
(a) Subcarrier Spectra

So(ν)
Bop

ν
ν1 ν2 ν3 ν4
(b) Optical Spectrum

Si (f )
BT

OBI

f
0 f1 f2 f3 f4 (ν4 – ν3 )
(c) Photocurrent Spectrum

Figure 5.8 Subcarrier spectra.

sources are used, spectra from different users may overlap without serious performance
degradation. This is a significant advantage of subcarrier techniques, because require-
ments for optical wavelength stability are not as stringent as in WDMA.
Although precise control of optical frequencies is unimportant, there is an advantage
in wide spacing of optical carriers, because of optical beat interference (OBI) appearing
at the output of the photodetector. Because the direct-detection receiver is a wideband
square-law device, it accepts signals from many different sources, and pairs of optical
signals from different lasers “beat” against each other in the detection process, producing
“noiselike” OBI signals. The spectrum of each OBI signal is centered at an electrical
frequency that is the difference between the optical frequencies of the pair of beating
optical signals. If the transmitted signals are spaced too closely in optical frequency,
some of the OBI power may fall in the subcarrier band, possibly overwhelming the
desired signals. As is shown later, this effect may place a lower limit on λ-channel
spacing in an SCMA system.
To study the OBI effects in the general case, consider an M × M system with aggregate
subcarrier bandwidth BT = M B. On the optical level, we assume that all λ-channels are
separated by equal nominal spacings, ν, fitting into a total optical spectrum of width
Bop = Mν – see Figure 5.8(a) and 5.8(b) for the 4 × 4 case. Of particular interest is
 Static Multipoint Networks 341

the minimum value of spacing and hence the minimum optical bandwidth required to
achieve acceptable system performance. Let

E(t) = Ek (t) (5.12)
k
be the optical field present at the input of a receiver, where Ek is the contribution from
the kth transmitter. Then, using complex envelope representation, E takes the form
 
E(t) = k Re E k (t)e j2πν0 t ,

(5.13)
where
 
Ek (t) = Re E k (t)e j2πν0 t . (5.14)
In Equation (5.13), ν0 is a nominal optical frequency for all transmissions. The kth
complex envelope E k (t) is then defined as
√
E k (t) = 2Ik (t)e jk (t) (5.15)
with
E 02
Ik (t) = 2
[1 + m k sk (t)] (5.16)
and
k (t) = 2π (νk − ν0 )t + φk (t). (5.17)
Recall that intensity modulation (as opposed to amplitude modulation) is a non-
linear process due to the square root in Equation (5.15). This causes an irregular shape
and spreading of the spectral components in Figure 5.8(b). This effect becomes more
pronounced with increasing modulation index m k .
In Equation (5.16), Ik (t) is the intensity of the kth optical field, with a true optical
frequency νk that appears in the phase function k (t) of Equation (5.17). Now from
Equations (2.17), (5.13), (5.15), and (5.17), the detected photocurrent is
2
R   
i(t) = Ek =R Ik (t) + R Ik In e j2π(νk −νn )t + φk (t)−φn (t) (5.18)
2 k k k=n

where R is the responsivity of the photodetector.

The first summation in Equation (5.18) contains the desired detected intensities
from the various transmitters, and the second summation contains OBI terms centered
around difference frequencies (νk − νn ).7 Thus in a system with M transmitters there are
M(M − 1)/2 OBI terms.
Returning to our 4 × 4 example, the electrical spectrum Si ( f ) of the photocurrent at
a receiver [see Figure 5.8(c)] contains the desired signal spectra Ss( j) ( f ) together with
OBI. One OBI term from Equation (5.18) is shown, due to transmitters 3 and 4.
7
Note that when only two signals are present, this two-signal mixing process looks just like that in a
heterodyne receiver. As in the heterodyne case, the intensities of the difference frequency terms depend
on the relative polarizations of the contributing pairs of signals. In the heterodyne receiver, polarization
alignment is necessary to produce the desired signal, whereas in the subcarrier case polarization alignment
is detrimental, producing OBI. The OBI terms in Equation (5.18) represent the worst case, corresponding
to perfectly aligned polarizations of the two beating signals.
342 Multiwavelength Optical Networks

f0
a
1
a
BT
B
f
0 fk f0 = (ν4 ν3)

Figure 5.9 Effect of OBI.

A more detailed picture of the two-sided electrical PSD at the input of the demodulator
is shown in Figure 5.9. Again, only one OBI term is shown, centered at frequency
f 0 = (ν4 − ν3 ) = ν. Note that there is a “tail” of the OBI spectrum that extends
into the subcarrier band BT , overlapping the signal spectra. This overlapping OBI
acts as noise, which degrades the signal at the input to the electrical detector (point
h in Figure 5.7[b]). Normally, the intermediate frequency filter has a bandwidth, B,
equal to that of the modulated signal to eliminate as much noise and interference as
possible. Thus, the interference affecting a particular received signal is the part of the
total OBI spectrum falling within that signal’s bandwidth (shown as the shaded area in
Figure 5.9 for subcarrier f k ). In an M × M system with equally spaced optical carriers,
the M − 1 OBI terms due to beat notes between adjacent carriers, each centered at
f 0 = ν = (ν j+1 − ν j ), produce the dominant noise in the subcarrier band. To maintain
a satisfactory BER, this adjacent-channel beat noise must be kept small, which in turn
implies making the λ-channel spacing ν sufficiently large.
The minimum required optical spacing νmin for satisfactory performance depends
on the spectral characteristics of the OBI terms and the vulnerability of the subcarrier-
modulated signals to OBI noise. Each of these factors depends in a complex way on many
physical parameters, including the channel bit rate, modulation format and bandwidth,
the modulation index m, the acceptable BER, other sources of noise, the number of active
OTs, the unmodulated laser linewidth, the type of laser modulation (direct or external),
the subcarrier frequency, and so forth. These issues have been studied thoroughly in
the literature (see, for example, [Darcie87, Olshanksy+89, Saleh89]). In view of the
complexity of the problem, we make a number of simplifying approximations here
to obtain a closed-form relation between νmin and the other system parameters. The
objective is to obtain some insight into the role of the various parameters affecting
system performance. The development follows the approach of [Desem90]. The two
basic assumptions are the following: (1) a suitable performance measure for OBI is
the carrier-to-interference ratio (CIR) at the input to the electrical detector and (2) for
CIR calculations, the spectrum of each of the OBI components can be approximated as
Lorentzian.
By focusing on CIR, we avoid getting involved in the details of the modulation format
and electrical detection process. A CIR of approximately 20 dB at the input of the detector
is sufficient to produce a BER superior to 10−9 for binary digital modulation formats
while providing a reasonable margin for other sources of noise (see Section 4.6.3). For
example, in the QPSK illustration of Equation (5.10), the signal is a sinusoid of unit
amplitude and varying phase, so that the carrier power equals the signal power, and CIR
 Static Multipoint Networks 343

is the same as SNR. Assuming that the OBI can be treated as independent Gaussian
noise, the SNR sufficient to achieve a BER of 10−9 using QPSK is approximately 23 dB.
(QPSK requires an increase of 3 dB in SNR over binary PSK for the same BER.)
Assuming sk (t) in Equation (5.16) is a sinusoid of unit amplitude, the carrier power
for the kth signal present in i(t) is R 2 E 04 m 2k /8. The total power in the OBI because of the
beating of signals k and n is R 2 E 04 /8. To determine how much of this falls into a given
signal band, we need an appropriate model for the PSD of the OBI. Given a signal s(t)
with modulation index m, the Lorentzian spectrum shown in Figure 5.9 (characteristic
of a second-order resonant system) has been found empirically to model OBI well for
small m (see [Antoniades+95]). For large m a more exact approach is required. The
two-sided Lorentzian PSD of a signal centered at frequency f 0 and normalized to unity
total power is given by
a a
g( f ) = + , (5.19)
2π [a 2 + ( f − f 0 )2 ] 2π [a 2 + ( f + f 0 )2 ]
where the spectral width parameter a is one half the full width at half maximum (FWHM)
of the OBI spectrum. Because each OBI term in the electrical spectrum is the result of
convolving optical spectra from two signal sources, its width is the sum of the widths
of the two convolved spectra. The width of each optical signal spectral component
increases with laser linewidth and chirp (in the case of direct modulation of the laser).
Furthermore, because of the nonlinearity of the modulation process, the spectral width
is an increasing function of modulation index m and subcarrier frequency. However, the
latter effects only become important for large modulation indices.
To compute an approximation for CIR, we assume a system of M transmitters with
equal modulation indices m, equal signal bandwidths B, and with λ-channels at equal
frequency spacings ν. In computing the OBI power falling in the signal band, it is also
assumed for simplicity that ν ≫ BT , so the integral of g( f ) over the signal band can
be approximated as 2Bg(0). (This inequality is a fair assumption in systems in which
some effort is made to maintain optical wavelength separation.)
Using Equations (5.16) and (5.18), the total OBI power for this system is
M −1
 (M − k)R 2 E 04 B a
POBI = . (5.20)
k =1
π (a 2 + k 2 ν 2 )

The summation in Equation (5.20) represents the effects of all superimposed OBI
terms, with values of k ≥ 2 corresponding to OBI from nonadjacent optical carriers.
It can be seen that POBI tends toward zero for large a or large ν. In the former case,
the OBI spectrum is so spread out that little OBI power falls in the signal bandwidth, and
the spacing ν becomes unimportant. This suggests that when optical carrier spacing
is very small, or when it cannot be controlled, the best line of defense against OBI is to
increase a.8
In the case of large ν the center of the OBI spectrum is so far from the subcarrier band
that the tail that overlaps the signal bandwidth has negligible power. The cases of small
8
The fact that a large value of a reduces OBI has prompted a number of system designs in which the laser
spectrum is widened intentionally to reduce OBI.
344 Multiwavelength Optical Networks

and large a/ν are both of practical interest. The former case would arise in systems us-
ing narrow-linewidth, temperature-controlled lasers. These would typically be employed
in long distance and high bit-rate applications where it pays to use expensive transmit-
ters. The latter case represents wide spectrum light sources such as LEDs or multimode
lasers; e.g., the Fabry–Perot laser of Section 4.5.1. These might be used in low-cost,
short-distance and low-bit-rate applications such as LANs and access networks.
An approximate expression for CIR for the case of small a/ν is
m 2 π (ν 2 + a 2 )
CIR = , (5.21)
8(M − 1)Ba
where only the first (adjacent-channel) term in Equation (5.20) has been retained.9 For
large a M/ν we have
m 2 πa
CIR = . (5.22)
4M(M − 1)B
The appearance of M in the denominator of Equations (5.21) and (5.22) shows that
OBI imposes a serious limit on the number of simultaneously transmitting users in an
SCMA system. This may or may not translate into a connectivity limitation depending
on what multiplexing schemes are used (see Section 5.2.2.3).
An experiment that provides some verification of the previous development concern-
ing OBI is reported in [Antoniades+95], where the performance of an M × M SCMA
system is studied starting with empirical data for the case M = 2 and extrapolating to
larger sizes.
In a system with f 1 = 600 MHz, f 2 = 850 MHz, and m = 0.23 for each trans-
mitter, it is found empirically that the OBI spectrum is approximately Lorentzian,
with a = 100 MHz. Assuming a modulation bandwidth B = 100 MHz and an opti-
cal carrier spacing of ν = 7.6 GHz, this gives a CIR of approximately 20.8 dB us-
ing Equation (5.21), compared with the measured value of approximately 21 dB in
[Antoniades+95].
The laser can be modulated at frequencies as high as 1 GHz, which suggests that
10 users can be fit into the modulation bandwidth, with subcarriers spaced 100 MHz
apart, and each transmitting at R t = 100 Mbps. (This implies a spectral efficiency at
the subcarrier level of ηm = 1, which is accomplished easily using QPSK to modulate
the subcarrier.) But for M = 10, we find from Equation (5.21) that the optical carrier
spacing must be increased to ν = 22.8 GHz to maintain the same CIR as in the case
of two users. This gives a total throughput of 1 Gbps using an optical spectrum of
approximately 228 GHz, resulting in optical spectral efficiency ηop = 0.004.
Now suppose the same throughput is divided among 20 users each occupying a
bandwidth of B = 50 MHz while maintaining the same aggregate bit rate. In this case
we find that ν and BT must be increased to 33.1 and 662 GHz, respectively, reducing
the spectral efficiency to ηop = 0.0015.

9
For small values of a/ν, the error incurred by dropping OBI terms generated by nonadjacent optical
carriers increases with M, so the OBI is underestimated by approximately 12.5, 25, and 56% for M = 3,
10, and 100, respectively.
 Static Multipoint Networks 345

The spectral efficiencies in these examples were relatively low because of the small
value of m. The basic problem with a small modulation index is that the total trans-
mitted optical powers interact to produce OBI, whereas only a small fraction of those
powers represent useful signals. Significantly better efficiencies have been demonstrated
experimentally in systems using larger m.
The advantage of large m was demonstrated in [Wood+93]. Experiments were con-
ducted in a 4 × 4 star coupled system with values of m up to 1.8. This corresponds to
overmodulation of the laser, producing clipping of the subcarrier. In this application,
clipping has little detrimental effect on signal quality provided that the subcarrier har-
monics generated by clipping do not interfere with any useful signals. That is the case
here, where digital data at 155 Mbps was modulated onto subcarriers using FSK, and
the effects of modulation depth on BER were determined under varying conditions of
OBI and modulation depth. It was found that when two of the four lasers were tuned
to optical frequencies producing maximum OBI, the system had an error floor at 10−2
BER with m = 0.8. However, when the modulation index was increased to 1.8 only a
minor power penalty (1.5 dB) due to the OBI was experienced at 10−9 BER. Another
useful result of this work was the determination of a relation between CIR due to beat
interference, and the resultant BER. It was found that for purposes of affecting BER, the
CIR could be treated like the SNR in a system with additive Gaussian noise, which gives
some support to our assumption in this section that the CIR can be taken as a simplified
performance measure in studying systems with OBI.
In systems where transmitter wavelengths cannot be controlled, and which are operat-
ing in the large a M/ν regime (Equation [5.22]), increasing the modulation index has an
additional advantage: it spreads the laser’s optical spectrum, thereby increasing the OBI
spectral width parameter and improving CIR. An example of an unusual and effective
way of implementing this spectral spreading is reported in [Woodward+96]. In a three-
laser subcarrier system using uncooled multimode Fabry–Perot lasers, large-amplitude
sinusoidal “clipping tones” are added to the subcarrier signals modulating each laser.
The result is a broadening of the spectrum of the OBI. Experiments were conducted
in essentially worst-case conditions with identical optical carriers and aligned polariza-
tions. They showed that the clipping tones significantly improved BER performance in
the face of OBI.

5.2.2.3 TDM/T-SCMA
In a TDM/T-SCMA system, the multiplexing function needed for one-to-many logical
connectivity is realized using time division methods, whereas the multiple-access func-
tion (for many-to-one connectivity) uses a combination of time division and subcarrier
multiple access. For illustration we return to the 4 × 4 system of Figure 5.6, shown in
more detail in Figure 5.10. Transmitting station j transmits on a fixed and distinct optical
frequency ν j , with the optical frequency separation among the stations sufficient to make
OBI negligible. (This condition is relaxed later.) To provide full logical connectivity, four
logical channels are first time division multiplexed to produce a single baseband data
signal, which is then translated to a subcarrier frequency (in the boxes labeled MUX and
SCMOD, respectively, in Figure 5.10[a]). In the multiplexed data signal, the LCs are
346 Multiwavelength Optical Networks

f1 f1 f4

[1, 1] [1, 1]
[1, 2] 1
e e′ [2, 1]
MUX SCMOD OT OR SCDMOD DMUX
[1, 3] [3, 1]
[1, 4] [4, 1]
2

3
f4 f1 f4

[4, 1] [1, 4]
[4, 2] e e′ [2, 4]
MUX SCMOD OT OR SCDMOD DMUX
[4, 3] 4 [3, 4]
[4, 4] [4, 4]

(a) FT-TR System

f1

SCDMOD DMUX [1, k]

OR

f4

SCDMOD DMUX [4, k]

(b) Arrayed Receiver (Station k)

Figure 5.10 TDM/T-SCMA.

assigned to designated channel–slots (assuming a fixed-frame system), where channel

now means subcarrier channel.
In an FT-TR subcarrier system, source station j is assigned a fixed subcarrier fre-
quency f j , and each destination station has a wideband OR together with a rapidly
tunable subcarrier demodulator, as in Figure 5.10(a). The photocurrent at point e of
each destination station contains all of the subcarriers, and hence all of the LCs are
accessible at this point. Receiving station k selects information destined for it by tuning
its demodulator to each transmitter’s subcarrier frequency exactly at the times when that
information is present. This is completely analogous to what takes place at the optical
level in an FT-TR TDM/T-WDMA system. As in TDM/T-WDMA, the subcarrier system
may also operate in a TT-FR or TT-TR mode. The demodulated information streams,
at point e′ in the kth receiving station in Figure 5.10(a), are processed by an array of
synchronized gates in the demultiplexer box to extract the various LCs, [∗, k], and to
route them to the proper output ports.
Like all shared-channel networks, subcarrier systems may operate using either fixed
or dynamic capacity allocation. In the former case, channel–slots are allocated in a
 Static Multipoint Networks 347

fixed periodic frame just as in TDM/T-WDMA, except that optical frequency tuning
is replaced by subcarrier tuning. Similarly, in packet-switched systems, channel–slots
are allocated dynamically, and contention for the shared channels must be resolved
using an appropriate MAC protocol. All of the issues of synchronization, transmitter
collisions, and receiver conflicts that are present in TDM/T-WDMA are also present here.
Thus, for example, certain performance improvements and system simplifications can be
achieved by replacing the tunable subcarrier demodulator by an array of demodulators,
one tuned to each transmitted subcarrier frequency as shown in Figure 5.10(b). This
allows a destination station to see all transmissions concurrently, eliminating receiver
conflicts. Although this arrayed configuration is analogous to an arrayed OR, there is an
important practical difference: It is currently far less costly to replicate electronics than
optics.
An important special case of TDM/T-SCMA is its potential application in access
networks such as the PONs described in Section 5.7. In a network where M end users
access a single upstream network node, the upstream (downstream) traffic typically
uses an M × 1 (1 × M) directed star. The downstream link in this case just uses TDM,
in which case the full bandwidth of the transmitter, BT , is allocated to each end user
on a slot-by-slot basis. The upstream traffic uses T-SCMA as in Figure 5.10. Assum-
ing that OBI limits the number of users to a value N < M, a T-SCMA schedule can
be arranged where no more than N users are active in any one time slot. Maximum
upstream throughput is achieved if the receiver has an array of N SCDMODs, as in
Figure 5.10(b).
To share the bandwidth of the channel fairly among all users, they can be accom-
modated in assigned time slots according to their traffic needs. The simplest subcarrier
frequency assignment would allocate a unique subcarrier frequency to each user, in which
case M frequencies are required, and the upstream bandwidth available to each user is
B = BT /M. Also, the upstream receiver must have either a fixed-tuned SCDMOD array
of size M or a rapidly tunable array of size N . Alternatively, a total of N subcarrier
frequencies, reused among the M users, will suffice to permit N simultaneous upstream
transmissions. In this case the upstream receiver needs only a fixed-tuned SCDMOD
array of size N . This allows for more upstream throughput, because the bandwidth
available for each user is now B = BT /N . When M frequencies are used, time slot
assignments are not constrained by the user’s frequencies. With less than M frequencies,
receiver conflicts may occur at the subcarrier level where more than one user transmitting
in the same slot uses the same subcarrier frequency. This can be avoided by properly
scheduling the slot transmissions. If the upstream SCMODs are fixed tuned, the users
must be scheduled taking into account their assigned frequencies. If the SCMODs are
tunable, their frequencies can be tuned on a slot-by-slot basis to avoid conflicts.

5.2.2.4 Packet Switching Using SCMA

Subcarriers can also be used to implement packet switching in the optical layer. As
mentioned in earlier chapters, packet-switched networks typically have large popula-
tions of bursty users with low average throughput per user. The fact that the pop-
ulation is large suggests that OBI may be a serious problem. However, the traffic
348 Multiwavelength Optical Networks

characteristics of packet-switched networks suggest a different way of dealing with

it: Just ignore it!
Consider an SCMA system operating in a packet-switched mode in which each station
is active only when transmitting a packet. Now OBI occurs only when two or more trans-
mitters operating on closely spaced optical frequencies are active simultaneously. But
this event can be made reasonably rare by proper system design. This is the basis of the
system proposed in [Shankaranarayanan+91]. The system uses a random access MAC
protocol (see Section 5.6.1). M stations are assigned distinct subcarrier frequencies,
but their optical frequencies are random and unknown. Stations are active only when
they transmit a packet, and packet transmission times are random and uncoordinated.
Thus, there is a nonzero probability that transmissions from two or more stations will
overlap in time. Given a time overlap, there is also a nonzero probability of outage
due to OBI. Outage occurs if the OBI power generated by the overlapping signals is
large enough in the subcarrier band to mutilate the desired signals. Because the MAC
protocol has a mechanism for retransmission of mutilated packets, this type of outage is
not catastrophic provided that it occurs with sufficiently low probability.
The main advantage of this approach to SCMA is that it requires no special assignment
or control of optical frequencies. The main disadvantage is the probability of outage
due to OBI. This probability increases with increasing value of K , the number of
simultaneously colliding packets. It also increases with the BER caused by a pairwise
collision. In a system with uncontrolled optical frequencies, we have seen that the effect
of pairwise OBI can be controlled by using a high modulation index m and large spectral
width parameter a. (Forward error correction might also be an option.) By reducing the
pairwise OBI, more simultaneous packet collisions can be tolerated, as indicated by the
role of the parameter M in Equations (5.21) and (5.22). Finally, even in a completely
uncontrolled random access system the probability of K simultaneous collisions can be
controlled by limiting the traffic intensity (see Section 5.6.1). This has the consequence
of reducing the effective value of M in Equations (5.21) and (5.22).

5.2.2.5 SCM/SCMA
An alternative to TDM in providing one-to-many connectivity in subcarrier systems
is SCM. Figure 5.11 shows a transmitting and receiving station supporting full logical
connectivity in a directed 4 × 4 SCM/SCMA system. Transmitting station j, using an
array of four subcarrier modulators, generates a subchannel at carrier frequency f jk for
LC [ j, k] for k = 1, 2, 3, and 4. These are superimposed at point a in Figure 5.11(a) to
modulate the OT. Thus, the signal at point a is of the form

s(t) = m k s jk (t). (5.23)
k

(If a single multicast LC – [ j, {1, 2, 3, 4}] – is desired, only a single SCMOD and
single carrier frequency are needed.) Receiver k uses an array of four SCDMODs, as
shown in Figure 5.11(b), with demodulating frequencies f jk , j = 1, 2, 3, and 4. Note
that M N distinct subcarrier frequencies are needed in an M × N system, but no time
division multiplexing/demultiplexing equipment is required. This represents a simple
 Static Multipoint Networks 349

fj1 f1k

[j, 1] SCMOD SCDMOD [1, k]

a
Σ OT OR

fj4 f4k

[j, 4] SCMOD SCDMOD [4, k]

(a) Transmitter j (b) Receiver k

Figure 5.11 SCM/SCMA.

trade from time-division to frequency-division multiplexing techniques at the subcarrier

level.
In some applications (e.g., multichannel video distribution in HFC systems), the
frequency division approach is the natural one because the basic information is delivered
to the end user on sinusoidal carriers. Even when the information is digital, there may
be a cost advantage for SCM/SCMA over TDM/T-SCMA, in that relatively low-speed
electronics can be used at the baseband level and no rapid tuning or slot synchronization
is required at any level.
There are important differences in using subcarriers for multiplexing as opposed to
multiple access. In SCM, several subcarriers are superimposed on the same optical
signal. Because there is only one optical signal, there is no OBI. However, impressing
several subcarriers on one laser introduces cross-talk, clipping, and interference caused
by the inevitable nonlinearities in the modulation process and the various impairments
in the fiber.
Conversely, in SCMA, most of these effects are absent because each laser is modulated
by only one subcarrier. However, OBI occurs unless precautions are taken as described
in Sections 5.2.2.2 and 5.2.2.3. The SCM/SCMA system just described is subject to both
multiple subcarrier effects and OBI.

5.2.2.6 SCM/WDMA/SCMA
The constraint on throughput imposed by laser modulation bandwidth and the connectiv-
ity limit imposed by OBI in SCMA can be relaxed by using subcarriers in combination
with optically tunable receivers to yield an SCM/WDMA/SCMA network. Basically,
this is a generalization of SCM/SCMA, wherein optical tuning is used to enhance the
spectral efficiency of the system at both the subcarrier and optical levels. The approach,
first proposed by [Liew+89], is based on a two-level system implemented as shown
in Figure 5.12. The underlying network is a directed broadcast star. The system is de-
signed to operate in a circuit-switched mode, creating multiple LCs on demand. As in
SCM/SCMA, no rapid tuning either at the subcarrier or the optical levels is required
350 Multiwavelength Optical Networks

SCMOD

fm

[k, j] SCMOD OT νk

SCMOD

(a) Transmitting Station k

SCDMOD

fm
νk
b
OF OR SCDMOD [k, j]

SCDMOD

(b) Receiving Station j

Figure 5.12 SCM/WDMA/SCMA.

for one-to-many and many-to-one logical connectivity. (Tuning only changes during
connection establishment so that millisecond or slower tunability is acceptable.)
As shown in Figure 5.12(a), each transmitting station is equipped with an array of
tunable subcarrier modulators and a fixed-tuned OT, whereas each receiving station (see
Figure 5.12[b]) has a wideband OR preceded by a tunable optical filter, followed by an
array of demodulators (point b in Figure 5.12[b]). The filter is capable of independently
selecting several (possibly nonadjacent) λ-channels, so it has the functionality of the
multiwavelength switch described in Section 2.3.2.4.
Suppose that a distinct optical frequency νk is assigned permanently to transmitting
station k. Then, whenever an LC of the form [k, ∗] is needed, it is realized by assigning
some subcarrier frequency f m from a pool { f 1 , f 2 , . . . , f C } to that connection. The
number C of subcarrier channels available in the pool depends on the channel bandwidth
B and the usable subcarrier spectrum bandwidth BT . Any station can receive an LC from
station k by tuning its optical filter to select the optical carrier frequency νk and tuning
one of its subcarrier demodulators to receive the assigned subcarrier frequency f m . If
there are M transmitting stations, each with D modulators, the total number of possible
 Static Multipoint Networks 351

Table 5.1 SCM/WDMA/SCMA example.

Receiver Optical Subcarrier

number tuning Accessible LCs tuning Accepted LCs

2 ν1 [1, 2]1 , [1, 3]2 f1 [1, 2]

3 ν1 , ν2 [1, 2]1 , [1, 3]2 , [2, 3]3 , [2, 4]1 f2 , f3 [1, 3], [2, 3]
4 ν2 [2, 3]3 , [2, 4]1 f1 [2, 4]

simultaneous LCs is MD, a number that may be considerably greater than C. The
interesting aspect of the two-level selectivity available in this system is that the active
subcarrier frequencies do not have to be distinct, creating the possibility of reuse of
the subcarrier spectrum. All that is required is that the set of subcarriers assigned to each
transmitting station be distinct and that the set demodulated by each receiving station be
distinct.
Of course, adding tunable optical filters to a subcarrier system tends to nullify the
cost advantages of the subcarrier approach. However, the fact that only slow tunability
is required makes this a less costly approach than, say, TDM/T-WDMA.
To illustrate, suppose the connections [1, 2], [1, 3], [2, 3], and [2, 4] are to be set up
in a 4 × 4 system. Let the subcarrier assignments be [1, 2]1 , [1, 3]2 , [2, 3]3 , and [2, 4]1 ,
where the subscript j on an LC indicates that subcarrier frequency f j is assigned to that
connection. Table 5.1 indicates the required receiver tuning. Note that the accessible
LCs are those that appear (sometimes unintentionally) at the output of the ORs (point
b in Figure 5.12[b]), whereas the accepted LCs are those selected after subcarrier
demodulation. Figure 5.13 shows a system that is set up to carry the connections indicated
in Table 5.1. The tuning of the optical filters is indicated by the notation ν j associated
with the optical filter, and the tuning of the demodulators is indicated by the notation f j
on the RPs. In this case f 1 is being reused and, in fact, appears twice (unintentionally)
at receiver 3. However, this causes no interference because the only subcarriers being

[1, 2] [1, 2]1 [1, 3]2 ν1 ν1 f1

TP1 OT OR [1, 2]
[1, 3] OF RP2

[2, 3] [2, 3]3 [2, 4]1 ν2 f2, f3 [1, 3]

TP2 ν1, ν2
OT C OR
[2, 4] OF RP3 [2, 3]

ν2 f1
OR [2, 4]
OF RP4

Figure 5.13 SCM/WDMA/SCMA example.

352 Multiwavelength Optical Networks

demodulated at that receiver are f 2 and f 3 , carrying connections [1, 3] and [2, 3],
respectively.

5.2.2.7 Summary
The following sums up the advantages and disadvantages of using subcarriers. The
advantages include
r They make it possible to create multipoint logical connectivity using only fixed-
tuned (or slowly tunable) optical transceivers, together with fairly inexpensive tunable
microwave modulators and demodulators.
r More rapid and precise tuning is possible in the electronic domain than in the optical
domain.
r The current subcarrier technology is more mature and less expensive than the corre-
sponding WDM technology.

However, the disadvantages are

r SCMA makes inefficient use of the optical spectrum because of OBI.
r If direct modulation is used, lasers with wide modulation bandwidths are required to
accommodate high subcarrier frequencies and to achieve a high aggregate throughput
for the system.
r Fiber dispersion severely limits the reach of subcarrier systems.

5.2.3 Code Division Multiple Access

All multiple-access techniques discussed thus far have used orthogonality to achieve
channel distinguishability at the receivers. In contrast, code division multiple access
(CDMA) is only quasiorthogonal, in the sense that (using linear processing) different
channels are only “approximately” separable at the receivers, with the interference from
unwanted channels appearing as background noise. CDMA is an example of a spread-
spectrum technique, in which a modulation process is used that spreads the encoded
signal spectrum over a much wider bandwidth than that of the original data signal. This
spectral spread has many advantages in contexts other than optical networks, in particular,
immunity to noise and interference. In shared-channel optical network applications, the
most important advantage is that many simultaneous low-throughput logical channels
can be accommodated on a shared medium with a minimum of coordination. (No
time synchronization or precise optical tuning is required.) This makes it an attractive
approach for packet switching, using either electronic or purely optical signal processing.
The basic idea of optical CDMA (OCDMA) is to encode each bit in a data stream into
a waveform that is unique for each LC. The unique code is analogous to a unique wave-
length in a WDMA system or a distinct time slot in TDMA. Thus, an LC is established
between a source and a destination by “tuning” the transmitter and receiver to a common
code. The transmitter generates bits with a unique waveform, and the receiver recog-
nizes these bits only, filtering out all other waveforms. Quasiorthogonality means that
the filtering is not perfect. As the number of superimposed CDMA channels is increased,
 Static Multipoint Networks 353

the interference from the unwanted channels, called multiple-access interference (MAI),
tends eventually to degrade the desired signal. However, the degradation in CDMA is
“graceful,” with the level of interference proportional to the number of superimposed
channels and their traffic intensities.
In an FT-TR CDMA system using distinct assignments, each transmitter is assigned
a distinct code, and receivers connect to a desired source by configuring themselves to
capture only those bits with the assigned code. The usual problem of receiver conflicts
arises when more than one source transmits to the same receiver. In a TT-FR system,
the receivers are assigned distinct codes, and it is the source that must choose a code
matched to the desired receiver. In this case, collisions and bit errors occur if more than
one source transmits to the same receiver at the same time. Note that the CDMA codes
act like source (destination) addresses in the case of FT-TR (TT-FR).
There are many possibilities for encoding the CDMA bits. Two well-known methods
use frequency hopping and pseudorandom pulse sequences. In the case of frequency
hopping, the waveform is generated as an optical carrier with a frequency (wavelength)
that is varied (“hopped”) in a pseudorandom manner, where the timescale of the variation
changes from one scheme to another. Hopping may be on a per-bit basis, a per-packet
basis (in packet-switched systems), or within one bit. In the latter case, it is called
fast frequency hopping, FFH-CDMA. The wider the range of frequency variation and
the faster the variation, the more the spectrum is spread. In the case of pseudorandom
pulse sequences, also known as direct sequence (DS-CDMA), the waveform of each
bit consists of a pseudorandom sequence of binary pulses, known as chips, with the
chip duration Tc = T /L being much shorter than the bit duration T . The integer L is
known as the code length. Because bits of duration T in the data signal are replaced
by pulse trains with pulse widths T /L, this amounts to a bandwidth expansion of the
data spectrum by a factor of L. By combining frequency hopping with direct sequence
techniques, two-dimensional time-frequency (T-F-CDMA) codes can be constructed. In
the 2D case, each chip may contain optical power at one or more wavelengths, which
change from one chip to the next. As we shall see, other more elaborate codes are possible
as well, which involve using phase as well as time and frequency. One characteristic that
all encoded CDMA signals have in common is that they are “noiselike”; simple data
pulses are converted to complex waveforms usually spread in both time and frequency
for transmission on the channel. Then at the receiver they are restored to the original data
pulse form again by the CDMA decoder. The noiselike waveform is important for the
quasiorthogonality that is desired in a CDMA system and it also imparts some security
against eavesdropping.
The performance of a CDMA system can be characterized in terms of the number of
users that can be accommodated while maintaining a satisfactory BER. Clearly, there is
a trade-off; the more users, the more interference between the users, and the higher the
BER. One way of improving both the size of the system and its error performance is to
increase the dimensionality of the signal space. In the codes described above this can
be done either by increasing the number of chips per bit or the number of wavelengths
in a frequency-hopping system or both in a T-F system. However, this requires either
increasing the required transmission bandwidth (which generally augments the effects
354 Multiwavelength Optical Networks

of transmission impairments) or decreasing the bit rate. All of these considerations

represent the set of constraints that must be considered in the design of a CDMA system,
and they will come to the fore in different ways as we examine specific systems.
The signal processing for encoding and decoding the optical CDMA waveform may
be performed either electronically or optically. In direct sequence CDMA, intensity
modulation is the encoding technique, which makes it adaptable to either electronic
or optical processing. When frequency, phase, or polarization is a parameter in the
modulation format, some optical processing (i.e., filtering) is required at the receiver.
Highest performance is potentially achieved when all information in the pulse waveform
(amplitude, frequency, phase, and polarization) is used to decode at the receiver. This
can be done only by using full optical processing. CDMA systems that use this complete
information in optical processing are called coherent. Otherwise they are noncoherent.

5.2.3.1 Electronic Processing: Intensity-Modulated Direct-Detection Systems

In the intensity-modulated direct-detection case the CDMA signal intensity modulates
the optical carrier, which is combined with others in a shared-channel system, and the
combined signals are broadcast to all receivers. Each receiver’s photodetector converts
the combined signals to a photocurrent, which is then processed by a CDMA decoder to
extract the desired bit stream. This is a noncoherent system.
The block diagram of a typical FT-TR intensity-modulated DS-CDMA system is
shown in Figure 5.14. (Station j uses channel j.) The following notation will be used:

r x ( j) = Data signal at source station j

r h ( j) = Impulse response of the CDMA encoder for the jth channel
r s ( j) = CDMA encoded signal at source station j
r E j = Optical field at source station j
r E = Optical field at a receiver
r g (k) = Impulse response of the CDMA decoder for the kth channel
r y (k) = Decoded signal at a receiver tuned to the kth channel

Note that the CDMA encoder is shown conceptually as a linear filter operating on the
data signal. (This may not correspond to the actual implementation.)

(j)
h
(j) (j)
(j) Pulse x (t) s (t)
al
Generator
Encoder OT εj (t)
(a) Transmitting Station j

g(k)
(k) (k)
i(t) y (t) a
Threshold l
ε(t) OR Decoder
Detector
(b) Receiving Station Tuned to Code k

Figure 5.14 Block diagram of a direct-detection CDMA system.

 Static Multipoint Networks 355

Now suppose that the jth data signal is presented to the encoder in the form of
rectangular pulses u(t) of width Tc , spaced at the symbol interval T and carrying a
( j)
binary data sequence {al } (1s or 0s). Then we have
 ( j)
x ( j) (t) = al u(t − lT ) (5.24)
l

and
 ( j)
s ( j) (t) = al v ( j) (t − lT ), (5.25)
l

where

( j)
v (t) = u(τ )h ( j) (t − τ ) dτ (5.26)

represents a single encoded pulse. To generate the pseudorandom chip sequence, we use
an encoder with impulse response
L−1

h ( j) (t) = cn( j) δ(t − tn ), (5.27)
n=0

where tn = nTc and the cn ’s are 0s or 1s chosen to make h ( j) (t) look like a random
sequence of impulses. Equation (5.26) then becomes
L−1

( j)
v (t) = cn( j) u(t − tn ). (5.28)
n=0

The effect of encoding on one data pulse is shown in Figure 5.15, in which the
pulse u(t) in Figure 5.15(a) is convolved with the CDMA impulse response h ( j) (t) in
Figure 5.15(b) to produce the encoded waveform v ( j) (t) in Figure 5.15(c). The pulse
positions in v ( j) (t) are determined by the nonzero chip coefficients cn . (CDMA codes
with chip coefficients taking on only the values 1 and 0 are called optical codes.) The
complete encoded signal s ( j) (t) now has the appearance of a random sequence of narrow
pulses.
The encoded signal can now be used to intensity modulate the laser using OOK,
giving an optical field
 ( j)

E j (t) = Re E j (t)e j2πν0 t , (5.29)

where

E j (t) = 2I j (t) (5.30)

and
E 02 ( j)
I j (t) = 2
[s (t)]. (5.31)

The optical field at the input to the photodetector is


E(t) = E j (t) (5.32)
j
356 Multiwavelength Optical Networks

u(t)

Tc
t
T
(a)

h(j)(t)

t
(b)

v (j)(t)

t
(c)

Figure 5.15 Waveforms for a direct-detection CDMA system.

so the photocurrent i(t) is proportional to j s ( j) (t).10 The receiver in Figure 5.14(b) is



tuned to select the kth CDMA channel by making the impulse response of its decoder
equal to that of the encoder for the kth channel, reversed in time (and delayed one symbol
interval); that is,

g (k) (t) = h (k) (T − t). (5.33)

This form of decoder is called a matched filter, and its effect is to correlate the
incoming signal with a copy of the bit waveform for the kth channel. The signal at the
output of the CDMA decoder is
  (k)
(k)
y (t) = s ( j) (τ )g (k) (t − τ ) dτ = y j (t). (5.34)
j j

(k)
In Equation (5.34), y j (t) is the component of the output due to the signal from the
jth transmitter and is given by
(k)  ( j)
y j (t) = l al w (k, j) [t − (l + 1)T ], (5.35)

where

w(k, j) (t) = wn(k, j) u(t − tn )


n (5.36)

10
This assumes that there is no noise or distortion in the channel and no OBI in the signal band.
 Static Multipoint Networks 357

and
( j)
wn(k, j) = cm(k) cm+n .

m (5.37)

The sequence wn(k, j) in Equation (5.37) is the correlation sequence between the codes
for channels k and j, which becomes an autocorrelation sequence when k = j. In an
(unrealizable) ideal case, we would like to have

wn(k, j) = 0, for all n when k = j (5.38)

wn(k,k) = 0, for n = 0 (5.39)
(k,k)
w0 = K, (5.40)

where K in Equation (5.40) is the number of nonzero chips in the code (called the code
weight). In the ideal case, all interfering terms [wn(k, j) for k = j] in Equation (5.34) drop
out and the output of the CDMA receiver becomes
 (k)
y (k) (t) = K al u[t − (l + 1)T ] (5.41)
l

so the desired data sequence is recovered perfectly (with a delay of one symbol inter-
val). These ideal correlation conditions are characteristic of very-wide band random
signals; i.e., white noise, which is why the optical pulse sequences are designed to be
pseudorandom.
In practice, in intensity-modulated optical systems it is impossible to achieve the
ideal conditions for perfect recovery. However, a class of codes, called orthogonal
optical codes (OOCs), comes close. The OOCs obey the conditions

wn(k, j) ≤ 1, for all n when k = j (5.42)

wn(k,k) ≤ 1, for n = 0 (5.43)
(k,k)
w0 = K. (5.44)

An example of a pair of OOCs with L = 32 and K = 4 is shown in Figure 5.16. These

codes produce the lowest possible interference.

t
0 9 12 27

t
0 4 11 30

Figure 5.16 Orthogonal optical codes.

358 Multiwavelength Optical Networks

Equation (5.41) shows that by setting the threshold at the output of the decoder at a
value slightly lower than K , it is possible to detect 1s while minimizing errors due to
interference. (0s are detected as the absence of 1s in a synchronous bitstream.) In fact,
using OOCs, a bit error can occur only (in the absence of random noise) if more than
K − 1 stations interfere simultaneously with the desired station.
Unfortunately, only a relatively small number of OOCs exist for given values of L and
K . It has been shown [Salehi89] that the maximum number of distinct OOCs of length
L and weight K (and hence the maximum number of stations in an OOC-based system)
is bounded by
 
L −1
M≤ , (5.45)
K (K − 1)
where ⌈x⌉ is the smallest integer greater than or equal to x.
Thus, for example, the codes of Figure 5.16 are the only OOCs for L = 32 and
K = 4. Equation (5.45) poses a dilemma: To accommodate a large number of stations,
K should be made small. However, a small K makes the decoder vulnerable to noise
and interference. A way out is to increase L, but then the point is soon reached when
ultrashort pulses are required. This not only consumes optical bandwidth but also makes
the system highly vulnerable to fiber dispersion. Furthermore, implementation becomes
difficult and expensive.
There are a few ways of improving the performance of the CDMA system just
described. First, we note that MAI is worst when all stations are active simultaneously.
As in SCMA the interference problem is much less severe if the system is used for
packet switching in the optical layer. In that case, the stations are active only when
transmitting packets, so interference can be kept under control by keeping the aggregate
packet throughput sufficiently low.
Another way of reducing interference is by mutual synchronization of all stations to
align their symbol intervals. In that case, the conditions for perfect detection reduce to
(k, j)
w0 = 0, when k = j (5.46)
(k,k)
w0 = K. (5.47)

Not only are these conditions realizable, but also they can be achieved with a much
larger number of codes (and thus with systems of larger size) than in the case of OOCs.
Of course, synchronization presents its own difficulties, especially in cases of ultrashort
pulses.
Finally, to improve discrimination between 1s and 0s, a pair of codes can be used for
each CDMA channel. In this case, the transceiver structure takes the parallel form shown
( j) ( j)
in Figure 5.17, in which the jth channel 0s (1s) are encoded using the code h 0 (h 1 ).
(k) (k)
In the receiver, the photocurrent is decoded simultaneously using decoders g0 and g1 ,
matched to the encoders for the kth channel, with a comparator detecting a 0 or a 1,
depending on which decoder produces the largest output. This is a classic transceiver
structure and generalizes in an obvious way to q-ary data transmission: q > 2. The
downside is that more codes are required, reducing the maximum number of stations
that can be accommodated.
 Static Multipoint Networks 359

Encoders
(j )
al = 0 (j )
h0

(j)
Pulse s (t)
Σ OT
Generator

(j)
(j) h1
al = 1
(a) Transmitting Station j

Decoders
(k)
g0

i (t ) (k)
ε(t ) OR Comparator al

(k)
g1

(b) Receiving Station Tuned to Code k

Figure 5.17 Parallel CDMA transceiver structure.

5.2.3.2 CDMA with Optical Processing

There are several reasons for implementing CDMA using all-optical processing. First,
the processing speeds required in systems accommodating more than a few users at
reasonably high data rates are difficult to achieve electronically. Second, CDMA systems
based on frequency and/or phase modulation require at least some optical processing in
front of the photodetector. Finally, coherent CDMA, which can be realized only in the
optical domain, increases the dimensionality of the signal space and therefore reduces
MAI, allowing the system to accommodate more users at lower BER.
A conceptual block diagram of a CDMA system using all-optical processing is shown
in Figure 5.18. Note that it is just the system of Figure 5.14 with the encoder and OT
interchanged and the decoder and OR interchanged. As shown in Figure 5.18(a), the

(j)
H (ν)
(j )
(j ) Pulse x (t) ε′j (t)
al Generator
OT Encoder εj(t)
(a) Transmitting Station j

(k)
G (ν)
ε(k)(t) (k)
y (t ) Threshold (k)
ε(t) Decoder OR
Detector
al

(b) Receiving Station Tuned to Code k

Figure 5.18 CDMA with all-optical processing.

360 Multiwavelength Optical Networks

signal x ( j) (t) carrying the data at the jth transmitter is first converted to a series of short
optical pulses, constituting the optical field E ′j (t). The pulses are then encoded optically
to produce the transmitted field E j (t), in which each narrow data pulse is typically spread
out in time to produce a noiselike waveform. In a coherent system the encoding is most
easily understood in the frequency domain.
Let E′ j (ν) and E j (ν) be the Fourier transforms of E ′j (t) and E j (t), respectively, and
let H( j) (ν) be the transfer function of the encoder for the jth channel (an optical filter
operating on E ′j ). We then have E j (ν) = H( j) (ν)E′ j (ν). At each receiving station, the
combined field E is processed by a decoding filter G(ν), where the frequency-domain
condition for a filter matched to channel k is G(k) (ν) = H∗ (k) (ν); that is, the decoding filter
is the conjugate of the coding filter. In the time domain conjugate filtering corresponds
to convolving the encoded signal with itself, which produces an autocorrelation function
analogous to the autocorrelation sequence described in Section 5.2.3.1 for DS-CDMA
codes. In general, if we start at the transmitter with a narrow pulse representing a 1 and
then encode it to spread it over time, the effect of the conjugate decoding filter H∗ (k) (ν)
at the receiver is to “compress” the spread waveform back to a narrow pulse, which is
then converted to electronic form in the photodetector and then thresholded to determine
the received bit. The use of the conjugate filter produces true coherent detection, but to
realize conjugate filters in practice, phase and possibly polarization relations must be
correctly matched at the decoder and encoder, which requires a source of coherent light
at the transmitter as well as carefully matched encoders and decoders. Many proposed
systems using optical processing are noncoherent, with encoder/decoder structures that
are “almost matched” but do not preserve phase relations. Nevertheless, matched linear
filtering is a convenient way to conceptually represent a wide range of CDMA encoding
and decoding operations, even if they are not actually implemented as coherent systems.

5.2.3.3 Optical Processing in Noncoherent Systems

Optical processing has been demonstrated in noncoherent systems employing a variety
of codes. Essentially all of them are forms of DS-CDMA or generalizations of DS-
CDMA to additional dimensions in the signal space, typically referred to as 2D or
3D codes. Each divides the bit period into chips and assigns selected values of signal
parameters (amplitude, frequency, phase, and polarization) to each chip. Figure 5.19
shows two examples of 2D codes. The first represents fast frequency hopping (FFH-
CDMA), where the signal frequency, indicated as wavelength in the figure, is hopped
from one value to another in each chip period. The second shows a similar system that
may use more than one signal frequency in each chip. A complete 3D system including
the polarization state as a third variable is shown in Figure 5.21. These multidimensional
systems are generally implemented with incoherent sources; for example, a laser array
or spectrally sliced ASE noise.

Intensity-Modulated Systems
One of the earliest laboratory demonstrations of CDMA in an optical network using
optical processing was reported in [Prucnal+86]. In this case DS-CDMA was used
with intensity modulation, based on codes like those shown in Figure 5.16. It was
 Static Multipoint Networks 361

n n
t1 t2 t3 t4 t5 t1 t2 t3 t4 t5

λ5 λ5

λ4 λ4
m

m
λ3 λ3

λ2 λ2

λ1 λ1

SSP1 SSP3
(a) (b)

Figure 5.19 Multidimensional codes. (Adapted from [Chen01, Figure 1]. Copyright
c 2001 IEEE.
Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

implemented as a synchronized system in which ultrashort optical pulses were generated

by a mode-locked laser and were distributed to all transmitting and receiving stations.
The pulses acted as a system clock, and after gating by the data sequence, they provided
the signals x ( j) (t) in Equation (5.24). Because the signals were already in the optical
domain, the encoding was performed optically through a parallel arrangement of fiber
delay lines, each producing a delay of an integral number of chip times Tc to realize the
impulse response h ( j) (t). Different codes were programmed by selecting the appropriate
combination of fiber delays. At the receiver, the matched decoder was realized in a similar
fashion, with the output sampled at times synchronized with the symbol intervals. The
samples were then converted to electrical form in a photodetector, with the photocurrent
fed to a threshold detector. In this case the encoder and decoder are not true conjugate
filters, because the fiber delay lines do not preserve the necessary phase relations in the
filtering operations. Thus, this is a noncoherent system.

Fast Frequency Hopping CDMA Using FBGs

An implementation of FFH-CDMA has been proposed and experimentally demon-
strated using tunable FBGs [BenJaafar+01, Fathallah+99, Wei+03]. The FBG
encoder/decoder is shown in Figure 5.20. A circulator directs data pulses to the FBG

Time

Broadband G1 G2 G3 G4 G5 G6 G7
Encoded data

data

Figure 5.20 FBG encoder for FFH-CDMA (From [Wei+03, Figure 1]. Copyright

c 2003 IEEE.
Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)
362 Multiwavelength Optical Networks

encoder, which encodes each data pulse into a regularly spaced sequence of delayed
pulses, each at a different optical frequency determined by the reflective impulse re-
sponse of the gratings. A prescribed impulse response over L chips is obtained by
writing L individual FBGs into a fiber at regular spacings corresponding to the chip
durations. The pitch of each grating is chosen to reflect the selected chip frequency. To
facilitate assigning pseudorandom hopping sequences, all gratings are strain tunable,
with the same unstrained pitch (see Section 4.10.4.3). They can then be individually
tuned to different frequencies to provide unique frequency hopping codes for each user.
The reflected waveform is directed to the outbound fiber by the circulator. The receiver
performs the decoding operation by passing the received signal through a matched FBG
decoder, which is just a spatially reversed version of the encoder.
In [BenJaafar+01] experiments are reported using this system over 80 km of SMF
with dispersion compensation, involving simultaneous transmission from 16 differently
encoded interfering sources. In each transmitter, an eight-chip code was generated
over 30 frequencies by driving the encoder with broadband data pulses. The pulses
consisted of EDFA ASE noise, so that each individual grating reflected one slice of
the noise spectrum to produce the desired frequency hopping. Although processing
is all optical, this is a noncoherent system. Experiments were conducted at data rates
of 1.25–2.5 Gbps, achieving successful transmission in spite of the large number of
interferers.
The BER performance of the same architecture is reported in [Wei+03]. Measured
bit error rates were studied and compared with BER predicted theoretically due to MAI
and to OBI. This is one of the few experimental BER investigations of a fairly large
CDMA system, and it turned up some interesting conclusions. In a system running at
625 Mbps, the total number of users had to be limited to four to achieve 10−9 BER,
which is not surprising. Most other systems have about the same behavior. However,
the experimentally observed BER is much higher than that predicted theoretically due
to MAI alone, and much lower than that predicted theoretically due to OBI alone. The
theoretical overestimation of OBI effects may be explained by divergences between
the experimental conditions and the mathematical model used for OBI calculations.
Theoretical calculations of OBI assume coherence and polarization alignment for the
beating sources, which is a worst-case scenario that tends to overestimate the BER
under experimental conditions (see Section 5.2.2.2 and [Tancevski+00]). However,
the theoretical underestimation of MAI is not so easily explained and needs further
study.
A generalization of the FBG-based system is proposed in [Chen01], wherein several
parallel fibers are used in the encoder/decoder. Using this arrangement more complex
T-F codes are possible, including codes with more than one frequency in a given chip
period, and codes using the same frequency in two different chips.11 Again, an incoherent
source is proposed, but no experimental results are available.

11
A single-fiber FBG system cannot use the same frequency in two different chip periods, because this
requires using two gratings at the same pitch in two different positions in the fiber. This would result in
multiple reflections of the signal component on that frequency in the fiber.
 Static Multipoint Networks 363

Sample time

3D OCDMA
encoded data bit Star coupler Threshold
t1 ………………………..…t11 (to all users)
Input data Recovered
λ1
data out
OCDMA λ4 Matched
encoder λ1 decoder

λ4
t1 ………………………..…t11 Threshold
Decoded bit
(stacked chips)

Figure 5.21 A 3D CDMA system. (From [McGeehan+04, Figure 1]. Copyright
c 2004 IEEE.
Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

A 3D CDMA System
In [McGeehan+04], polarization is added to time and frequency to produce 3D codes.
This can more than double the number of users for the same level of MAI. The complete
CDMA system appears in Figure 5.21. In the transmitter, a laser array generates a set of si-
multaneous pulses on distinct wavelengths, which are optically encoded to form a data bit.
First, the pulses are split into two orthogonal polarized sets and demultiplexed and each
wavelength pulse is delayed to fall into a prescribed chip time slot. The delayed pulses are
then multiplexed and the two polarized streams are recombined to constitute the encoded
bit. The reverse operations are performed at the decoder: separation of the incoming
signal into two polarized signals, demultiplexing, delaying the demultiplexed pulses to
reverse the delays in the encoder and remultiplexing. In the matched decoder, these oper-
ations “stack” the wavelength pulses in the same time slot, so their intensities add at the
photodetectors (one for each polarization). The outputs of the photodetectors are thresh-
olded and a bit is declared to be a 1 if the decisions on both polarization paths are 1. The
system has been demonstrated in the laboratory, but no performance results are avail-
able for a multiuser environment. Although the technique may be applicable to limited
distance LAN usage, PMD, and chromatic dispersion can be a problem in larger systems.

5.2.3.4 Optical Processing in Coherent Systems

Typical coherent CDMA systems generate an encoded signal by starting with a very nar-
row (and therefore wideband) data pulse and spreading it in time by either (1) separating
its frequency components and modulating their phases or (2) effecting phase/amplitude
modulation of the frequencies in the pulse in one device without separating the compo-
nents. We now give examples of (1) using phase masks and (2) using FBGs.

Optical CDMA Using Phase Masks

An early proposal for OCDMA using coherent detection appeared in [Salehi+90]. A
conceptual block diagram is shown in Figure 5.22. The data stream in the form of
ultrashort (picosecond) optical pulses is first created by gating a periodic source pulse
364 Multiwavelength Optical Networks

Data
Stream Threshold
Device
<1 psec

c
pse
<1
Subpicosecond Data
100 psec M×M
Source Modulator
Coupler

Grating
Grating

Multielement Phase
Phase Modulator Conjugate Mask
Grating Other
Subscribers Grating

Transmitter Receiver

Figure 5.22 Realization of coherent optical CDMA. (From [Salehi+90, Figure 2]. Copyright

c 1990 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers,
Inc.)

source. The encoder consists of a grating demultiplexer that spatially separates the
frequency components of the pulses, followed by a multielement phase modulator or
phase mask that shifts the phases of the different frequencies pseudorandomly. The
frequency components are then remultiplexed into a signal whose pulses are now spread
out in time. As shown in Figure 5.22, the decoder has the same structure as the encoder,
using a conjugate phase mask to compress the desired pulses, whereas the pulses from
other channels pass through the decoder as low-level noise. The conjugate phase masks
produce a pair of true optical matched filters as used in the system of Figure 5.18.
Another system using the phase mask architecture of Figure 5.22 was reported in
[Sardesai+98]. It uses a mode-locked erbium fiber ring laser to produce short (275-fs)
pulses. These are spread into pseudonoise waveforms about 20 ps in duration using a
liquid crystal SLM as a programmable phase mask. In a multiuser system, the SLMs are
programmed to produce minimally interfering phase-coded waveforms. A single-user
system was demonstrated that proved the concept, with the autocorrelated femtosecond
pulse shape reproduced at the receiver after the encoding and decoding operations. These
are very broadband systems (the pulse has a bandwidth of several terahertz), so they
are very sensitive to fiber dispersion. Tests were performed over 2.5 km of SMF using a
dispersion-compensating module together with some “precompensation” realized in the
encoding SLM. The result was an essentially perfect decoded pulse.
A different approach to phase coding is presented in [Etemad+04], with the objective
of implementing coherent CDMA within a smaller (80-GHz) bandwidth. If a mode-
locked laser generates a periodic pulse stream, its frequency spectrum is composed of
discrete lines separated by the pulse repetition rate. In the proposed system, periodic
pulses with power confined to an 80-GHz bandwidth are generated by tuning the laser to
operate in a desired 80-GHz segment of the optical spectrum. In the case at hand, a 5-GHz
 Static Multipoint Networks 365

repetition rate is used, which generates phase-locked frequency components at a 5-GHz

separation, giving 16 frequencies in a selected 80-GHz bandwidth. By tuning to different
portions of the spectrum (using a bandpass filter in the laser cavity), different bands can be
assigned to different users, resulting in a relatively narrowband tunable coherent source.
The periodic pulse stream can then be modulated to produce a stream of data pulses.
Phase coding of the data pulses is implemented in an optical system that demultiplexes
the 16 frequency components, impresses binary phase shifts (0 or π ) on them according to
codes selected for desired CDMA performance, and remultiplexes them for transmission.
A matched decoder employing a conjugate phase mask at the receiver completes the
system. An additional consideration in this system is the width of the spectral lines being
phase encoded. When the periodic pulse stream is modulated, it is transformed into a
random data bit stream. Because it is no longer periodic, the spectral lines widen into
bands. The widths of these bands must be controlled for proper operation of the phase
modulator, and this is accomplished by using a data-modulation format that induces
correlation among the data pulses, thereby narrowing the bands. Single-user operation
was demonstrated experimentally with satisfactory results, with simulations used to
study MAI.

Optical CDMA Using Superstructure Fiber Gratings

A superstructure fiber Bragg grating (SSFBG) is a standard FBG with rapidly varying
spatial refractive index of uniform amplitude and pitch, on which is superimposed a
slow spatial refractive index modulation. A property of a weakly reflecting SSFBG is
that a prescribed optical impulse response can be written into it by the superimposed
slow spatial modulation [Eggleton+94]. As a result, when excited by an optical pulse,
the reflective output of the SSFBG is the convolution of the pulse waveform and the
SSFBG impulse response. This is a true coherent CDMA encoder, and the correspond-
ing decoder is just the spatially reversed image of the encoder, so they are conjugate
filters. This corresponds to the coherent system described in Figure 5.18. In [Lee+02,
The+01], a DS-CDMA system is described based on SSFBGs. The basic transmitter
and receiver structures use circulators and reflecting FBGs as in the noncoherent systems
(see Figure 5.20). In this case, the desired impulse response consists of a series of chips
for a DS-CDMA code, each one containing a spread and phase-shifted version of the
narrow data pulse.
When multiple users are active in any CDMA system, the output of the decoder
is a relatively sharp pulse that must be distinguished from a background noise and
interference level consisting of MAI from other users, as well as many other effects.
These include imperfections in the encoder/decoder, fiber impairments, intersymbol
interference, OBI, and random noise. The result is a reduction in the contrast between
the decoded pulse and the aggregate background effects. A novel nonlinear processing
approach is used in the proposed system to improve contrast. It consists of a NOLM
acting as a saturable absorber, inserted in front of the photodetector to sharpen the
decoded pulse before detection and thresholding (see Section 4.11.2).
Experiments at data rates up to 2.5 Gbps over 25 km of dispersion compensated fiber
were run with narrow (2.5 ps) data pulses and a 63-chip bipolar code. The encoder
366 Multiwavelength Optical Networks

spread the data pulse width to chip pulse widths of 6.4 ps and impressed a binary
DS code onto the chips using phase reversals; i.e., a binary DPSK modulation format.
BER measurements were taken with a single user and with one interferer. The NOLM
significantly improved system performance, permitting higher bit rates and reducing
power penalties as compared to linear processing alone.
More ambitious experiments were reported involving a 511-chip SSFBG using DPSK
modulation [Wang+06]. Acceptable BER12 was reported with MAI from eight interfer-
ers, and its performance was shown to be superior to optical CDMA using OOK in terms
of tolerance to MAI and OBI, as well as in simplifying threshold settings. A variant of
the basic scheme [Hamanaka+06] was designed to operate with compound data rates
running from 622 Mbps to 2.488 Gbps for demand provisioning of multiple services.

5.2.3.5 Concluding Remarks on CDMA

A feature of CDMA that distinguishes it from other multiple-access schemes is its
inherent ability to function in a shared medium in a self-contained uncoordinated manner.
All addressing is encoded into the bits, and no synchronization or time slot assignment
is required. As suggested at the beginning of Section 5.2.3, these qualities imply that
the best application for CDMA is as a packet-switched transport technique in the optical
layer. A basic property of packet switching is that it can support large numbers of low
traffic users. How does OCDMA perform in this respect? Typical performance results,
both theoretical and experimental, show that if low BER is to be maintained in OCDMA
systems then MAI tends to severely limit the number of allowable interfering users. This
would seem to nullify the value of CDMA in a packet-switched network. Yet, things are
not as bad as they seem, because theoretical and experimental BER results in CDMA
are normally based on continuously interfering bit streams rather than packet operation.
Using arguments similar to those for packet-switched SCMA (Section 5.2.2.4) or for
random access packet-switched systems in general (Section 5.6.1), it can be seen that the
key issue as far as MAI is concerned is the probability of several potentially interfering
packets overlapping in time. This probability is a function of the overall system traffic
intensity. For example, if no more than four interfering users can be tolerated to maintain
satisfactory BER, then many times that number can be accommodated provided that they
operate at a sufficiently low aggregate traffic intensity.
Large numbers of users do, however, require a large set of quasiorthogonal CDMA
codes, meaning that long and/or multidimensional codes are needed for large user
populations. These two requirements, low traffic intensity and large code space, do not
appear to be insurmountable obstacles in deploying optical CDMA. What the network
gains in return is simplicity of operation, automatic bandwidth allocation, and some
degree of privacy and fairness.
The reader has probably noticed that in discussing CDMA we frequently resort to
words like proposed, experimental, and demonstrated rather than commercially available
or deployed in the field. At this writing, CDMA is essentially an experimental technique,
limited to laboratory demonstrations and perhaps special niche applications. However,
12
Acceptable in this case is 6 × 10−5 , anticipating the use of FEC to improve the quality of the raw data
stream.
 Static Multipoint Networks 367

technology has evolved considerably since the first systems were proposed and tested.
Complex encoders and decoders can now be fabricated as integrated devices. More
importantly, packet switching in the optical layer is becoming increasingly competitive
as a transport medium in parts of the network close to the end user: the access and LAN
environments. These all suggest that CDMA has a place in future optical networks.

5.3 Traffic Constraints in Shared-Channel Networks

Section 5.2 provided a detailed picture of a range of multiplexing and multiple-access

techniques suitable for shared-channel networks. We now step back from the details of
data transmission to focus on the larger question of accommodating prescribed traffic
requirements in general shared-channel networks. The objective is to derive some basic
relations that determine whether a network with a limited pool of resources is capable
of carrying a specified traffic load. Of particular interest is the issue of traffic imbalance
and how it affects network performance.
Figure 5.23 shows a model of a directed M × N shared medium. (If M is equal to
N , and transmitting station i is colocated with receiving station i, then we have the
undirected case.) As in Section 5.2.1, the network traffic is characterized by a set of
MN prescribed flows, expressed as an M × N traffic matrix, Ŵ = [γi j ]. Each entry γi j
represents the flow on a point-to-point LC between the source i and destination j in bits
per second (see Section 5.4.1.4 for the case of multicast connections).
The actual traffic can be modeled in various ways:
r Synchronous bit streams requiring dedicated LCs
r Synchronous bit streams requiring demand-assigned (circuit-switched) connections
r Asynchronous (packet) traffic

In the case of synchronous traffic on dedicated connections, a “fluid-flow” model

is appropriate, in which the network capacity required to carry a specified traffic flow
γi j corresponds exactly to the quantity of traffic being carried. In the other cases, the
offered traffic is generally specified in terms of statistical averages. In these cases, a

Transmitting Receiving
Station Station
α1 β1
1 1
Transmitters Receivers

αM βN
M Transmitters N
Receivers

C Channels

Figure 5.23 Shared-channel broadcast medium.

368 Multiwavelength Optical Networks

certain amount of “excess” capacity is required to accommodate random fluctuations

exceeding the average. In this section, we focus on the fluid-flow traffic model. The
effects of random fluctuations in traffic flow are considered in Sections 5.5 and 5.6.
The network resources include the available channels and the station transmitters or
receivers. Let C = the number of available channels, αi = the number of transmitters in
station i, and β j = the number of receivers in station j.
The ith transmitting station can transmit simultaneously on αi distinct channels, and
the jth receiving station can receive simultaneously on β j distinct channels.
The channels are drawn from a pool containing a total of C channels. The precise
meaning of channel may vary from one system to another. For example, in a WDMA
system it would be a λ-channel, in an SCMA system it would be a microwave subcarrier,
and in CDMA it would correspond to a distinct bit code assignment.
The transmitters and/or receivers may be rapidly tunable (on a submicrosecond
timescale), slowly tunable (milliseconds), or not tunable at all. With rapid tuning, the
pool of channels can be time-shared to create multipoint LCs according to some fixed
or dynamic time schedule, as in the examples of TDM/T-WDMA and TDM/T-SCMA
given in Sections 5.2.1 and 5.2.2.3, respectively. With slow tuning, LCs can be created
on a demand-assigned (circuit-switched) basis, but each channel is dedicated to a single
point-to-point LC for the duration of that connection.
The results of this section are derived from flow conservation conditions only and
are therefore independent of the multiplexing/multiple access methods used. However,
to fix ideas, it is useful to think of the channels as λ-channels and to assume that the
transceivers are rapidly tunable so that multipoint logical connectivity can be achieved
using TDM/T-WDMA techniques.
Clearly, the larger C is, the larger the overall network throughput can be. Similarly, a
large αi (β j ) suggests that station i ( j) is able to transmit (receive) a large aggregate data
flow. But how are all of these quantities interrelated, and how do they limit the permissible
traffic distributions in the network? This section addresses these questions in a general
context, without requiring any detailed description of the channels, the transceivers, or
the traffic. We assume only that the amount of network traffic is specified for all LCs
linking the various source and destination stations.
As in Section 5.2.1, let us assume that all transmitters operate at a common, fixed bit
rate, R t ≤ Rmax .13 Each channel can accommodate the flow from one transmitter, up to
the maximum rate Rmax . One question that is addressed is whether a given traffic matrix
can be supported using a bit rate not exceeding Rmax .
We begin with the following flow conservation inequalities:
Ti ≤ R t αi , i = 1, 2, . . . , M
Rj ≤ Rtβj, j = 1, 2, . . . , N (5.48)
γ̄ ≤ R t C,
13
Among the factors limiting the channel bit rate are the speed of the electronics in the TPs and RPs, the
modulation bandwidth of the laser, the characteristics of the λ-channel and optical path it is using (band-
width, dispersion, noise, interference, attenuation, nonlinearities), and the bandwidth and noise properties
of the ORs. See Chapter 4.
 Static Multipoint Networks 369

where

Ti = j γi j = aggregate traffic from station i

Rj = i γi j = aggregate traffic to station j (5.49)

γ̄ = ij γi j = total traffic.

The first inequality in Equation (5.48) expresses the limitation imposed by each
station’s transmitter capacity, and the second is the limit due to each station’s receiving
capacity. The third is the fundamental limitation imposed by the network capacity,
relating the aggregate traffic to the aggregate transmission capacity available to carry it.
Combining the inequalities in Equation (5.48) we find a lower bound on the bit rate,
R t , needed to support a given traffic requirement:
Ti R j γ̄


Rmin = max , , . (5.50)
ij αi β j C
Equation (5.50) shows that a bit rate of at least γ̄ /C is always necessary because
of the aggregate traffic requirements, but this may not be sufficient. If one of the first
two inequalities in Equation (5.48) is violated for some rate that satisfies the third, this
signifies that there is a traffic imbalance with respect to the available station resources.
This phenomenon can be expressed in terms of a traffic balance factor:
γ̄ /C
η≡  Ti R j γ̄  . (5.51)
maxi j αi , β j , C

The parameter η attains a maximum value of unity when the traffic distribution is
matched to the station resources in the sense that the flow is limited only by the available
capacity of the shared medium. The traffic is unbalanced (η < 1) when station transceiver
limitations are the binding quantities. Another way of interpreting η is as the throughput
of the network, γ̄ , normalized with respect to the total capacity Rmin C needed to carry
it. The two parameters γ̄ and η subsume the key characteristics of the matrix Ŵ relative
to the network resources. The lower bound on bit rate can be expressed in terms of these
parameters as

Rmin = γ̄ /ηC. (5.52)

The point to note in Equation (5.52) is that poorly balanced traffic necessitates an
increased bit rate to support a given aggregate traffic requirement γ̄ . This in turn leads
to inefficient use of the communication channels, as shown in subsequent examples. It is
not hard to conjure up extreme cases. For example, in a directed M × M shared-channel
network with C equal to M channels, and with a single transmitter or receiver per
station, the balance factor is 1/M if all traffic originates from one transmitting station
or is destined to one receiving station. (This simply states the fact that M − 1 channels
are superfluous in this case.)
A special case of Equation (5.51) arises when C = 1 and αi = β j = 1 for all i, j,
which corresponds to a TDM/TDMA network. In this case, η = 1 independent of the
matrix Ŵ (because row and column sums of Ŵ are never greater than γ̄ ).
370 Multiwavelength Optical Networks

1 1 1 1 0 2 2 0
1 1 1 1 0 0 1 3
T=
1 1 1 1 2 0 1 1
1 1 1 1 2 2 0 0
(a) (b)

0 2 1 0 0 2 2 0
0 0 2 3 0 0 3 1
2 0 1 1 2 0 1 1
2 2 0 0 2 2 0 0
(c) (d)

Figure 5.24 Normalized traffic matrices.

To illustrate the significance of the traffic balance factor, we now examine η and
Rmin in a variety of situations. In each case the specified flows are expressed as integer
multiples of a basic bit rate R 0 ; that is, in the form Ŵ = R 0 T , where T is the normalized
traffic matrix first introduced in Section 5.2.1. Figure 5.24 gives some examples of
normalized traffic distributions in directed 4 × 4 systems. The total traffic in each case
is γ̄ = 16R 0 . We shall determine η and the value of Rmin necessary to support these
traffic distributions given several different sets of network resources.

5.3.1 Balanced Traffic

Figure 5.24(a) shows the case of uniform traffic. In this case, η = 1 for any C ≤ 4, so
that Rmin = γ̄ /C. (This is true for any number of transmitters and receivers per station.)
Thus, for a single-channel TDM/TDMA system, Rmin = 16R 0 , and for a four-channel
system, which can be operated in either an FT-TR or a TT-FR mode, Rmin = 4R 0 .
The traffic matrix of Figure 5.24(b) represents nonuniform and incomplete multipoint
connectivity. Nevertheless, η = 1 for all C ≤ 4 and any number of transmitters and
receivers per station. Thus, once again, Rmin = γ̄ /C. The observation here is that traffic
balance does not require uniformity of the traffic matrix.

5.3.2 Unbalanced Traffic

The next two examples, shown in Figures 5.24(c) and 5.24(d), illustrate traffic imbal-
ances. In each case we start with the assumption that C = 4, with a single transmitter or
receiver in each station. The normalized traffic of Figure 5.24(c) is a minor variation of
the case of Figure 5.24(b), in which the entries t13 and t23 have been interchanged. As
a result, we now have T2 = 5R 0 , which means that the balance factor is now η = 4/5.
Because of the traffic imbalance, Rmin = 5R 0 , an increase of 25% over the balanced
case. However, if station 2 had two transmitters (α2 = 2), the balance factor would again
be unity, returning Rmin to 4R 0 .
Another unbalanced case is shown in Figure 5.24(d), which is obtained from
Figure 5.24(b) by permuting entries t23 and t24 , with the result that R3 = 6R 0 . Now
 Static Multipoint Networks 371

when C = 4, we have η = 2/3 and Rmin = 6R 0 . Next, suppose that we reduce the num-
ber of channels to C = 3. Referring to Equation (5.52), we find that the balance factor
increases to 8/9, but Rmin remains at 6R 0 . If there are two receivers in station 3 (β3 = 2)
and C = 3, the balance factor returns to unity, reducing Rmin to 16R 0 /3.

5.4 Capacity Allocation for Dedicated Connections

Each of the examples in the previous section determines a minimum value of the
transmitter bitrate Rmin that is necessary to support a specified traffic distribution. It has
little practical meaning unless we can show that this minimum rate (or something close
to it) is also sufficient. In this section we show that using fixed-capacity allocations, a bit
rate close to Rmin can indeed support the specified traffic; that is, the lower bound on R t is
tight. A scheduling method is described for satisfying arbitrary traffic requirements in a
variety of settings for both stream and packet traffic. The underlying optical infrastructure
remains the M × N shared-channel broadcast medium of Figure 5.23.

5.4.1 Fixed-Frame Scheduling for Stream Traffic

The basic features of fixed-frame schedules for dedicated LCs in TDM/T-WDMA net-
works were presented in Section 5.2.1. A periodic time frame with L equal-size slots was
used to multiplex the traffic onto C channels. The approach there was to present the ideas
through some simple examples, limited to cases with a single transmitter or receiver per
station. It remains to present a general method of determining fixed-frame schedules that
are optimal for specified traffic flows and network resources and to show that these sched-
ules achieve a performance close to the best possible as indicated by the flow constraints
of Section 5.3. As in the previous section, our results apply to shared-channel broad-
cast media based on a wide range of multiplexing/multiple-access techniques, of which
TDM/T-WDMA is but one concrete example. We consider networks with multiple trans-
mitters or receivers per station and with all transmitters running at a common bit rate R t .
As in Section 5.3, the capacity of the broadcast medium is to be shared among a set
of up to M N point-to-point LCs carrying stream traffic at bit rates specified by a traffic
matrix Ŵ. The algorithm used for optimal scheduling is a generalization by [Gopal82]
of a technique by [Inukai79]. For illustration, CASs are exhibited for the examples in
Section 5.3.
In the scheduling problem, there are three interrelated parameters of particular interest:
1. L min = the minimum number of slots required to schedule the traffic
′
2. Rmin = the minimum transmitter bit rate that accommodates the required traffic
3. ηs = the scheduling efficiency of the CAS [defined in Equation (5.5)]
The frame time F and slot time F/L (in seconds) do not enter into the analysis. Once a
frame schedule is determined, any value of F consistent with other system requirements
can be used.
Clearly, a schedule with a minimum value of R t is desirable to accommodate as
much traffic as possible without exceeding Rmax . The purpose of keeping L small is that
372 Multiwavelength Optical Networks

there is a fixed overhead (the guard time) associated with each time slot. (In the case of
multichannel systems involving rapid transceiver tuning, the guard time must include
an allowance for transmitter and/or receiver tuning time.) Increasing the number of slots
in a frame while keeping the frame time small (to limit latency) means that the slot
size diminishes relative to the overhead, eventually producing a prohibitive reduction in
transmission efficiency.
Following Section 5.2.1, to construct a traffic schedule we first decompose the traffic
matrix into the product Ŵ = R 0 T , where the elements ti j of the normalized traffic
matrix T are relatively prime integers. A capacity of ci j = ti j channel–slots is allocated
to LC [i, j] so that each channel–slot contains one unit of normalized traffic, which
corresponds to R t F/L bits. (From now on we assume that the overhead due to guard
times is negligible.)
Building on the terminology of Section 5.3, let us define the following quantities:
T i′ = j ti j = Ti /R 0 = aggregate normalized traffic from station i


R′j = i ti j = R j /R 0 = aggregate normalized traffic to station j


(5.53)

T̄ = i j ti j = γ̄ /R 0 = total normalized traffic.

Using these definitions, it can be shown (see [Gopal82]) that the minimum number
of time slots necessary to create a collisionless conflict-free schedule accommodating a
normalized traffic matrix T , with one unit of normalized traffic per channel–slot, is
   ′   ′ 
T̄ T i Rj
L min = max , , . (5.54)
ij C αi βj

This number is also sufficient, as demonstrated in Appendix B, where an optimal

scheduling algorithm is presented. The schedule is found by decomposing the normalized
(s) (s)
traffic matrix into L min M × N matrices, C (s) = [ci j ], s = 1, 2, . . . , L min , where ci j is
a nonnegative integer representing the number of channels assigned to LC [i, j] in time
 (s)
slot s, and ci j = s ci j .
To match the traffic requirements, the C (s) ’s must satisfy
L min

C (s) = T. (5.55)
s =1

Each C (s) must also satisfy the system constraints

 (s)
ci j ≤ αi , i = 1, 2, . . . , M (5.56)
j
 (s)
ci j ≤ β j , j = 1, 2, . . . , N (5.57)
i
 (s)
ci j ≤ C. (5.58)
ij

Note that the inequality in Equation (5.58) ensures that the allocation in each slot
does not exceed the number of available channels, Equation (5.56) ensures that there are
 Static Multipoint Networks 373

enough transmitters available for the allocation, and Equation (5.57) ensures that there
are no receiver conflicts.
From flow conservation, we find that the traffic specified by a matrix Ŵ can be
transmitted in a schedule with L min slots using a value of transmitter bit rate:
′
Rmin = L min R 0 . (5.59)

Equation (5.59) gives the minimum possible bit rate sufficient to carry the traffic Ŵ.
The efficiency of the schedule with L min slots can be expressed as
γ̄ T̄
ηs = ′ C
= . (5.60)
Rmin L min C
Because there are L min C channel–slots in the frame, the value of ηs is unity when all
channel–slots are fully utilized. Otherwise, ηs < 1, indicating that there are some idle
channel–slots.
These scheduling results parallel closely and complement the development of
Section 5.3, which led to an expression for Rmin – the minimum bit rate necessary
to carry a specified traffic. In this case we have a method for scheduling a prescribed
′
traffic matrix, together with a bit rate Rmin sufficient for transmitting that schedule. Using
Equations (5.50), (5.51), (5.54), and (5.60), we find that
′ 
T̄ T i R j
 ′

Rmin ηs max i j ,
C αi
, βj
′
= =    ′   R′  ≤ 1 (5.61)
Rmin η T̄ T i
max ij , C
, j
αi βj

 T′ R′j
with equality if and only if maxi j CT̄ , αii , β j is an integer. The schedule of length L min

′
requires a bit rate Rmin ≥ Rmin , with equality when ηs = η. Otherwise a higher bit rate
is required to make up for the inefficiency caused by the fact that ηs < η.14
′
Noting that Rmin cannot exceed the maximum permissible transmitter bit rate Rmax ,
we find that a sufficient condition for a network to be able to carry a prescribed traffic
distribution is
γ̄
Rmax ≥ . (5.62)
ηs C

5.4.1.1 Channel Assignments: Tunability

Although the matrices C (s) defined earlier give a feasible set of channel–slot allocations
to match a given traffic matrix in a minimum-length schedule, only the number of
channels assigned to each LC in each slot is prescribed. The issue of specific channel
assignment (i.e., transceiver tuning) is not addressed. In general, once the C (s) ’s are
determined, any arbitrary assignment of channels yields an acceptable CAS, provided
that all assignments in a given slot are distinct (to avoid collisions). This freedom of

14
In some cases, the difference between ηs and η may be large. If this is bothersome, and if one is not
concerned about keeping L small, then the matrix T can always be “scaled up” through multiplication
by an integer (with a corresponding scaling down of R 0 ) until ηs ≈ η, and a schedule can be found with
′
Rmin ≈ Rmin . Thus, the scaling of T is a compromise between keeping L small (to reduce overhead and
latency) and keeping scheduling efficiency high. (Another aspect of scaling appears in Section 5.4.1.3.)
374 Multiwavelength Optical Networks

choice in making channel assignments can be used to the network designer’s advantage
in satisfying other system limitations: in particular, constraints on transceiver tunability.
An arbitrary set of channel assignments generally produces a TT-TR CAS. Because
tunability is costly, a natural question arises. Is it always possible to find an FT-TR or a
TT-FR CAS with L min slots for a prescribed traffic matrix? The answer is no, and this
leads to a number of related questions that we address now.

In the case where C ≥ i αi , enough channels are available to dedicate a distinct
channel to each transmitter so that an FT-TR CAS is possible. (For the case where all
source stations have a single transmitter, this corresponds to a system with C ≥ M.)

Similarly, if C ≥ j β j , enough channels are available to dedicate a distinct channel
to each receiver so that a TT-FR CAS is possible. (For the case where all destination
stations have a single receiver, this corresponds to a system with C ≥ N .) We refer to
these cases as systems with a “full complement” of channels. For smaller numbers of
channels, as might be expected, a TT-TR system is usually required to produce a CAS
with a minimum number of slots.
The next question is as follows: what is the minimum frame length for an FT-TR or
TT-FR system with a prescribed traffic matrix, and with less than a full complement
of channels? We answer this now for a restricted class of systems: those with a single
transmitter or receiver per station.
Consider an M × N shared-channel system with a single transmitter per station,
and with C = M̂ < M. We seek a minimum-length FT-TR CAS for this system for a
normalized traffic matrix T . Because there is less than a full complement of channels
for an FT-TR system, a channel assignment must be chosen wherein some transmitters
are necessarily tuned to the same channel. Let P = {S1 , S2 , . . . , S M̂ } denote a partition
of the set of source stations into M̂ subsets, where each Si is a set of source station
numbers. In our FT-TR CAS, all stations in the same subset share a common dedicated
channel. Let T̂ = [t̂ i j ] be a normalized M̂ × N traffic matrix derived from T , where

t̂ i j = tk j . (5.63)
kǫSi

The matrix T̂ defines normalized traffic for a fictitious system with M̂ source sta-
tions, where the ith fictitious source station has normalized traffic corresponding to the
aggregate traffic from all (real) stations in the set Si . Each source station in the fictitious
system is assumed to have a single transmitter, and each destination station has the same
number of receivers as it has in the original system. Under these assumptions, we find
that the minimum number of slots required to schedule this traffic in the fictitious system
is
 ′ 
Rj
 
T̄
L̂ min = maxi j , T̂ i′ , , (5.64)
C βj

where

′
T̂ i = t̂i j . (5.65)
j
 Static Multipoint Networks 375

By comparing Equation (5.64) with the value of L min for the original system in
Equation (5.54), we find that L̂ min ≥ L min with equality if and only if T̂ i′ ≤ L min for
i = 1, 2, . . . , M̂.
Once a partition has been chosen, the scheduling algorithm of Appendix B can be used
to find a channel–slot schedule for the fictitious system in the form of a set of M̂ × N
matrices Ĉ (s) , s = 1, 2, . . . , L̂ min . Because only one transmitter per station was assumed
in this system, at the most one channel is allocated to fictitious source station i in each
slot, with the total channel–slot allocation in the frame corresponding to the aggregate
traffic for the source set Si in the original system. Therefore, assigning a common
channel to all stations in the set, the total channel–slot allocation is sufficient to schedule
the aggregate traffic without collisions by distributing the channel–slot allocation among
all stations in Si . We have thus constructed an FT-TR CAS of length L̂ min corresponding
to the chosen partition of transmitting stations.
Clearly, L̂ min depends on the choice of partition. By computing the minimum value of
L̂ min over all partitions, we find the shortest possible frame length for an FT-TR system
with a given traffic matrix. In some cases (corresponding to traffic matrices with a high
degree of imbalance), it is possible to find partitions (and hence channel assignments)
with L̂ min = L min , so no penalty is incurred in using FT-TR.
A completely parallel development is possible for systems operating in a TT-FR mode.
In this case we assume that the destination stations have a single receiver per station
and C = N̂ < N . Now the set of all destination stations is partitioned into subsets S j ,
where all stations in a subset share a common dedicated channel. Letting T̂ = [t̂i j ] be a
normalized M × N̂ traffic matrix, where

t̂i j = tik , (5.66)
kǫS j

it can be shown that a TT-FR CAS exists with minimum slot length
   ′
T i

T̄ ′
L̂ min = max , , R̂ j , (5.67)
ij C αi
where
′ 
R̂ j = t̂i j . (5.68)
i

The details of the development leading up to this result are left as an exercise for the
reader. The next section provides some examples that apply these scheduling principles.

5.4.1.2 Examples of Optimal Schedules

All the examples in this section except the last one deal with 4 × 4 directed, shared-
channel networks. In each case, except Example 8, we use the normalized traffic matrices
introduced earlier in Section 5.3, wherein the total traffic is T̄ = 16.
The first four examples illustrate optimal scheduling for systems with a single trans-
mitter or receiver per station and a full complement of channels (C = 4). Although
optimal schedules can be found using the algorithm of Appendix B, these examples are
small enough to do the scheduling by inspection.
376 Multiwavelength Optical Networks

Channel 1 [1, 1] [2, 1] [3, 1] [4, 1]

Channel 2 [2, 2] [3, 2] [4, 2] [1, 2]

Channel 3 [3, 3] [4, 3] [1, 3] [2, 3]

Channel 4 [4, 4] [1, 4] [2, 4] [3, 4]

Slot 1 2 3 4

(a) TT-FR

Channel 1 [1, 2] [1, 2] [1, 3] [1, 3]

Channel 2 [2, 3] [2, 4] [2, 4] [2, 4]

Channel 3 [3, 4] [3, 3] [3, 1] [3, 1]

Channel 4 [4, 1] [4, 1] [4, 2] [4, 2]

Slot 1 2 3 4

(b) FT-TR

Channel 1 [1, 2] [1, 2] [1, 3]

Channel 2 [2, 3] [2, 3] [2, 4] [2, 4] [2, 4]

Channel 3 [3, 1] [3, 4] [3, 1] [3, 3]

Channel 4 [4, 1] [4, 1] [4, 2] [4, 2]

Slot 1 2 3 4 5

(c) FT-TR
Figure 5.25 CASs for systems with a full complement of channels.

Example 1 Figure 5.25(a) shows a TT-FR CAS for the uniform traffic matrix of
Figure 5.24(a), where receiver j uses channel j. It has a length L = L min = 4, so R t =
′
Rmin = 4R 0 . Reading down each column of the CAS we find the nonzero elements of the
(2) (2) (2) (2)
decomposition of T . Thus, for example, from column 2, c21 = c32 = c43 = c14 = 1.

Example 2 The traffic matrix of Figure 5.24(b) is nonuniform but balanced, and ηs = η.
Thus, a CAS can be found with L = L min = 4, as shown in Figure 5.25(b). The figure
shows an FT-TR schedule where transmitter j uses channel j.
 Static Multipoint Networks 377

Example 3 Now consider the traffic of Figure 5.24(c), where ηs = η = 4/5, indicating
′
a traffic imbalance. The minimum frame length is L min = 5, with Rmin = 5R 0 = Rmin .
An FT-TR CAS with minimum length is shown in Figure 5.25(c), in which the blanks
in certain entries of the schedule indicate that channels and station equipment are idle
in those channel–slots.

Example 4 Figure 5.24(d) gives another unbalanced normalized traffic matrix with
η = ηs = 2/3, so that this traffic can be scheduled in L min = 6 slots, as shown in
Figure 5.26(a).

We continue with three more examples based on the normalized traffic matrix of
Figure 5.24(d). The objective now is to illustrate how changing the system parameters
affects the feasible CASs.

Example 5 Suppose only three channels are available to schedule the traffic of
Figure 5.24(d), using a single transmitter or receiver per station. Despite this reduc-
tion of capacity, we find from Equation (5.54) that L min remains at 6 slots. However,
reducing C does pose a problem. Because we no longer have a full complement of chan-
nels, it is not clear whether a minimum-length FT-TR CAS can be found. To produce
such a schedule, we must assign the same channel to two transmitting stations with
aggregate traffic that does not exceed L min = 6. This is clearly impossible, and thus a
minimum-length FT-TR schedule does not exist. However, we note that the aggregate
traffic for destination stations 2 and 4 equals 6, so a minimum-length TT-FR schedule is
possible with those stations sharing a common channel. A feasible 6-slot CAS is shown
in Figure 5.26(b) with receiver 1 operating on channel 1, receivers 2 and 4 on channel
2, and receiver 3 on channel 3.

Example 6 If an FT-TR CAS is required for the previous example, the frame length
has to be increased. Because each source station has an aggregate traffic of 4, we
have L̂ min = 8, and ηs is reduced to 1/2. But by partitioning the source stations into
two subsets, {1, 2} and {3, 4}, we find that the same frame length can be achieved
with only two channels using FT-TR, with ηs = η = 1. A possible CAS is shown in
Figure 5.26(c) with transmitters 1 and 2 sharing channel 1 and transmitters 3 and 4 sharing
channel 2.
378 Multiwavelength Optical Networks

Channel 1 [1, 2] [1, 2] [1, 3] [1, 3]

Channel 2 [2, 3] [2, 3] [2, 4] [2, 3]

Channel 3 [3, 1] [3, 1] [3, 4] [3, 3]

Channel 4 [4, 1] [4, 1] [4, 2] [4, 2]

Slot 1 2 3 4 5 6

(a) FT-TR

Channel 1 [3, 1] [3, 1] [4, 1] [4, 1]

Channel 2 [1, 2] [1, 2] [2, 4] [3, 4] [4, 2] [4, 2]

Channel 3 [2, 3] [2, 3] [1, 3] [1, 3] [2, 3] [3, 3]

Slot 1 2 3 4 5 6

(b) TT-FR

Channel 1 [1, 2] [1, 2] [1, 3] [1, 3] [2, 3] [2, 3] [2, 3] [2, 4]

Channel 2 [3, 1] [3, 1] [3, 4] [4, 1] [4, 1] [4, 2] [4, 2] [3, 3]

1 2 3 4 5 6 7 8
Slot
(c) FT-TR

Channel 1 [1, 3] [1, 3] [1, 2] [1, 2]

Channel 2 [2, 3] [2, 3] [2, 4] [2, 3]

Channel 3 [3, 1] [3, 1] [3, 3] [3, 4]

Channel 4 [4, 2] [4, 2] [4, 1] [4, 1]

Slot 1 2 3 4

(d ) FT-TR (β3 = 2)

Figure 5.26 CASs for Examples 4, 5, 6, and 7.

Example 7 This example shows the effect of increasing the station equipment in the
case of the normalized traffic of Figure 5.24(d). Noting that it is receiving station 3 that
carries the largest load, we shall add a second receiver to that station (β3 = 2). Assuming
that C = 4, this improves the balance factor to unity, and thus we now have η = ηs = 1,
and L min = 4. Figure 5.26(d) shows an FT-TR CAS with transmitter j using channel j,
in which destination 3 is now allowed to appear as much as twice in each column of the
(1) (1) (2) (2)
schedule: c13 = c23 = 1 and c13 = c23 = 1.
 Static Multipoint Networks 379

1 1 1 1

2 0

0 1 0

0 1 0

1
T
1

( a)

Channel 1 [1, 1] [1, 2] [1, 3] [1, 8] [1, 9] [1, 10]

Channel 2 [3, 2] [4, 3] [5, 4] [10, 9] [2, 1] [2, 1]

Slot 1 2 3 4 5 6 7 8 9 10

(b)

Figure 5.27 Heterogeneous traffic scheduling.

Example 8 In this example, we use partitioning and FT-TR to accommodate a hetero-

geneous traffic mix on a small number of channels. Figure 5.27(a) shows a normalized
traffic matrix for a 10 × 10 directed, shared-channel system in which source station
1 transmits 10 units of traffic while all others transmit 1 or 2 units. The total normal-
ized traffic is T̄ = 20. We assume that all stations have a single transmitter or receiver
and that C = 2, in which case the traffic is balanced: η = ηs = 1 and L min = 10. Not-
ing that the aggregate traffic from all the low-traffic stations is 10, a natural partition
of the source stations to produce an FT-TR system is to assign channel 1 to source 1 and
channel 2 to all others. A possible minimum-length CAS is shown in Figure 5.27(b).

This last example suggests that when there are many more users than available chan-
nels, an efficient way of assigning channels is to reserve one or more channels to be
shared by the low-traffic population either on an FT-TR or a TT-FR basis and to assign
the remaining channels to the large users as needed.

5.4.1.3 Traffic Matrix Scaling

The first step in determining CASs for fixed-frame systems involves the conversion of
the traffic matrix Ŵ to a normalized traffic matrix T . In theory, if the elements of Ŵ are
rational numbers, a unique decomposition of Ŵ in the form of a matrix T of relatively
380 Multiwavelength Optical Networks

prime integers multiplied by a scalar R 0 is always possible.15 In practice, however, the

decomposition involves approximations, and thus it is not unique. Furthermore, the way
it is done can influence strongly the efficiency of the traffic schedule.
To illustrate the importance of a “good” approximation, consider the following 2 × 2
irrational traffic matrix:
 
1 π
Ŵ= . (5.69)
1 1
Two possible approximations for γ12 are 10/3 and 3.1416 = 3927/1250. These are
both rational numbers that overestimate π and thus represent the traffic conservatively.
Using the first approximation, the decomposition is
 
1 3 10
Ŵ= . (5.70)
3 3 3
Using the second, it is
 
1 1250 3927
Ŵ= . (5.71)
1250 1250 1250
Now suppose this traffic is to be scheduled in a single-channel system (e.g.,
TDM/TDMA). In this case each of these normalized matrices can be scheduled with
ηs = 1, so we have L min = T̄ . Thus, the first approximation gives a frame length of
19, whereas the second gives a length of 7677! Clearly, the first approximation is much
better in terms of frame length, even though approximately 7% of the capacity allocated
to LC [1, 2] is wasted. What has happened here? In both cases we approximated Ŵ by
“rounding up” γ12 to a rational number. But once the elements of Ŵ are approximated
as rationals, the scalar constant R 0 in the decomposition becomes the reciprocal of the
least common denominator of the γi j s. A large denominator means a small R 0 , which
in turn implies a large T̄ and hence a large L min . In the case at hand, the more accurate
approximation of π produced a very long frame length. If we are willing to accept the
poorer approximation, a much shorter frame can be used. Thus, two criteria must be
considered in obtaining an “efficient” rational approximation of Ŵ: keeping the elements
of T small while at the same time avoiding excessive rounding errors.

5.4.1.4 Multicast Logical Connections

If some of the LCs carried by a shared-channel broadcast medium are multicast instead
of point to point, then the previous discussion of flow constraints and capacity allocation
in the last two sections must be revised. In this section, we present the necessary
modifications.
The traffic requirements used for scheduling point-to-point LCs were defined in terms
of a traffic matrix Ŵ. A property of any traffic matrix is the flow conservation condition:
  
γ̄ = Ti = Rj = γi j , (5.72)
i j ij
15
Because any real number can be approximated as closely as desired by a rational number, the decomposition
is possible for all practical purposes for any Ŵ.
 Static Multipoint Networks 381

which expresses the fact that the total traffic carried by the network equals the aggregate
traffic transmitted, which equals the aggregate traffic received. But in a multicast LC
the total traffic received exceeds the total traffic transmitted, because the traffic leaving
the source arrives at more than one destination. Thus, a matrix description of multicast
traffic is incomplete. Instead, the traffic parameters must be obtained directly from a
prescribed set of LCs.
To formalize this, let lk = [τk , ̺k ] represent the kth LC in our shared-channel system,
where τk is the transmitting station and ̺k is a set of receiving stations. As in the case
of point-to-point connections, we assume that the traffic flows are normalized to integer
values, so that if γk represents the flow requirement (in bits per second) on LC k, then

γ k = R 0 tk (5.73)

γ̄ = γk (5.74)
k

T̄ = tk , (5.75)
k

where tk is an integer, and γ̄ , T̄ , and R 0 represent the total transmitted traffic, the total
normalized transmitted traffic, and the basic bit rate used for normalization, respectively.
The aggregate normalized traffic T i′ , transmitted from station i, is found by summing
the normalized traffic over all LCs emanating from i:

T i′ = tk , (5.76)
{k|τk =i}

and the aggregate normalized traffic R′j , destined for station j, is found by summing
the normalized traffic over all LCs with a receiving set that contains j:

R′j = tk . (5.77)
{k| jǫ̺k }

The relations among these quantities are

 
R′j ≥ T ′
i = T̄ . (5.78)
j i

By substituting the values of T̄ , T i′ , and R′j given in Equations (5.75), (5.76), and
(5.77), respectively, into Equation (5.54), we find the minimum frame length for the
multicast case. Note that it is the aggregate normalized transmitted traffic that determines
the total traffic requirement T̄ in Equation (5.54), and this quantity does not include the
effect of multicast. However, the aggregate received traffic requirements R′j needed
for determining the minimum frame length are computed by taking into account the
multicast connectivity.
To illustrate the changes in scheduling for multicast connections, let us revisit the
schedule for the normalized (point-to-point) traffic matrix of Figure 5.24(c). Recall that
for a four-channel network with one transmitter or receiver per station, this traffic was
scheduled with L = L min = 5 slots yielding R t = Rmin = 5R 0 – see Figure 5.25(c) for
an FT-TR CAS.
382 Multiwavelength Optical Networks

Table 5.2 Multicast connections.

Connection
number τk ̺k tk

1 1 2 2
2 1 3 1
3 2 4 1
4 2 {3, 4} 2
5 3 1 2
6 3 3 1
7 3 4 1
8 4 {1, 2} 2

Now suppose some of the connections in that example are multicast, with the traffic
flow prescribed in Table 5.2. Note that the connections [4, 1] and [4, 2], both carrying
two units of traffic in the point-to-point system, have been replaced by the multicast
connection [4, {1, 2}], and a similar replacement was made for two units of traffic
originating at station 2. The eight LCs are visualized graphically in Figure 5.28(a) using
a directed logical connection hypergraph (LCH), with connection lk represented by
hyperarc k in the LCH.

Connection
lk
1

2
1 1
3

2 4 2
Transmitting Receiving
Stations Stations

3 5 3

4 4 Channel 1 [1, 2] [1, 2] [1, 3]

6
Channel 2 [2,{3, 4}] [2,{3, 4}] [2, 4]

7 Channel 3 [3, 1] [3, 1] [3, 4] [3, 3]

Channel 4 [4,{1, 2}] [4,{1, 2}]

8 Slot 1 2 3 4
t
(a) LCH (b) CAS

Figure 5.28 Logical multicast CAS.

 Static Multipoint Networks 383

The multicast modifications result in revised values of the arguments of Equa-

tion (5.54), with T̄ reduced from 16 to 12, and T ′2 reduced from 5 to 3. The binding
quantities in determining frame length are now T ′3 = R′1 = R′2 = R′3 = R′4 = 4, so
the minimum number of slots is reduced from 5 in the point-to-point case to 4 in the
multicast case. An FT-TR CAS with L = 4 slots is shown in Figure 5.28(b).
Multicast receiving sets appear in certain entries of the CAS, indicating that all
stations in the receiving set must tune simultaneously to the given channel. As before,
for conflict-free operation each receiving station must appear no more than once in each
column of the schedule. The reduction in frame length from the point-to-point to the
multicast case reflects the fact that less information is being transmitted. Because of this
reduction, the transmitter bit rates can be reduced by a factor of 4/5 to R t = 4R 0 .
It is interesting to compare three possible implementations of logical multicast in
optical networks. In purely optical networks operating with only point-to-point optical
connections (e.g., wavelength-routed point-to-point networks) a k-fold multicast con-
nection must be implemented on k individual LCs fanning out from the source to the
receiving set and all carrying the same information. This replication increases the traffic
load on the network k-fold without carrying any additional information.
In logically-routed networks (Section 3.5) an electronically switched logical topology
is overlaid on the physical topology. In this case, the multicast connection can be realized
at the logical level in the form of a multicast tree. Each logical switching node in the
tree replicates the information for transmission on logical paths to the receiving nodes.
This, however, requires conversion of information between the optical and electronic
domain at each node of the tree. Furthermore, this still imposes an additional load on
the physical layer, where the extent of the load augmentation depends on how the logical
topology is embedded on the physical topology.
Finally, in networks in which the underlying optical connections can be broadcast or
multicast (i.e., static broadcast networks or LLNs), a multicast LC can be implemented
without any additional load on the network by exploiting the broadcast/multicast con-
nectivity in the optical medium.16

5.4.2 Fixed-Frame Scheduling for Packet Traffic

In our discussion of fixed-capacity allocation, it was assumed that the traffic was in
the form of synchronous bit streams, so the information flow on each LC was constant
and known precisely in advance. In that case, all variables are deterministic, and fixed-
capacity allocations that are matched exactly to a prescribed traffic matrix produce
maximum efficiency. With packet traffic we have the option of using either fixed or
dynamic capacity allocations. In this section, the fixed allocation approach is presented.
Section 5.6, which is devoted to packet switching in the optical layer, explores the
dynamic approach. (A brief review of the elements of random processes and queues that
constitute the background for this and the next two sections is found in Appendix C.)
16
To be fair in comparing efficiency of resource utilization, it must be recalled that a static broadcast or multi-
cast medium always replicates each transmission optically on each link outbound from a broadcast/multicast
optical node. Thus, in effect, it wastes capacity when used for point-to-point LCs and is most efficiently
utilized for multicast LCs.
384 Multiwavelength Optical Networks

In scheduling packet traffic the slotted, fixed-frame structure used in the stream
traffic case can still be employed. However, certain modifications are required to adapt
the capacity allocations to random traffic fluctuations. Returning to the example of
Figure 3.6, let us assume that the traffic is now in packet form. Suppose that the LCs
[1, 2] and [1, 3] constitute separate packet streams, transferred from external equipment
to station 1 through two separate logical ports as shown in the figure. This implies that
the packets destined for stations 2 and 3 have already been sorted and separated in a
logical switch before being passed on to the NAS. In this case, the logical switch has
already performed a packet-switching/routing function and sees the NAS as providing
two separate point-to-point LCs – one to each destination. Referring back to the layered
architecture of Figure 2.1, the packet-switching function is performed in the logical
layer, which means that no additional packet encapsulation is required for addressing
and control.
The traffic on an LC is now in the form of a sequence of packets possibly of random
lengths, arriving at random points in time. Therefore, it must be characterized by statis-
tical parameters, the most important of these being  = the average packet arrival rate
(in packets per second) and m = the average packet length (in bits).
This forces us to deal with traffic flow requirements stated as statistical averages
(which are generally estimated imperfectly). In this case, how much capacity should
be allocated to accommodate them? An exact match of capacity to estimated average
traffic flows (which we considered to be ideal in the case of stream traffic) would be a
disaster for random traffic. Any momentary increase of packet rate or packet size above
the average value would cause an overload. The burstiness and unpredictability of packet
arrivals in these systems is accommodated in two ways:
1. Buffering the packets awaiting transmission17
2. Allocating capacity in excess of that required to accommodate the average flow
To understand how these techniques can be implemented in a shared-channel system,
let us focus on one LC. With static capacity allocation, the packet arrivals on the
connection “see” the equivalent of a fixed-rate “pipe”: the transmission channel carrying
that connection. The capacity of the channel is determined by its channel–slot allocation,
ci j , giving an effective bit rate of R t ci j /L bits per second.
We assume for simplicity that each channel–slot holds one data frame. This means that
if the packets are of constant length (e.g., ATM cells), each channel–slot accommodates
exactly one packet. If they have random lengths, the maximum packet length fits into
one slot.18 Because each packet now occupies one channel–slot regardless of its true
length, the actual (possibly random) value of the packet length must be replaced by a
fixed effective packet length m ′ , where m ′ ≥ m.
Using our slotted scheduling rule, the TP in each station places a packet belonging to
a given LC in each channel–slot allocated to that connection. To allow for fluctuations
17
In the case at hand, the physical location of the packet buffers could be either in the external equipment or
in the NASs.
18
A more efficient arrangement is for long packets to be segmented and short packets to be concatenated to
fit into the slots.
 Static Multipoint Networks 385

Arrivals
Queue Server
Departures
µ

Losses

(a)

N Delay (ms)

M|M|1

2 M|D|1
5
1

ρ ρ
0 .5 1 0 1
(b) (c)

Figure 5.29 Single-server queue.

in traffic flow, packets must be queued in buffers to wait their turn for transmission
in allocated channel–slots. If no packets are awaiting transmission at the beginning of
an assigned slot, the slot is left empty. This type of buffered transmission link can be
modeled as a single-server queue as shown in Figure 5.29(a), in which the server is the
slotted transmission channel. The behavior of the queue can be complex because of the
random time each packet spends in the system and the possibility of buffer overflow.
However, with a few reasonable assumptions, an adequate picture of its performance
can be obtained.
A queue is characterized by its arrival process, service process, queue discipline, and
buffer size. The buffer size is assumed to be infinite for the time being. We model the
arrival process as a Poisson process19 with arrival rate  and assume a first-come-first-
served (FCFS) queue discipline. The service process is determined by the channel–slot
assignments and frame time F. With ci j channel–slots allocated to connection [i, j] in
each frame, and with each channel–slot serving one packet, the effective service time
for the queue is for all practical purposes constant at µ−1 = F/ci j , with a service rate
of µ = ci j /F = ci j R t /Lm ′ packets per second.20
The fundamental quantity determining the behavior of the queue is its traffic inten-
sity, ρ = /µ, where 1 − ρ is the probability that the queue is empty. Assuming that
19
Although the Poisson model is often a good one for packet traffic, it does not capture accurately the effect
of packet segmentation. In that case each packet is converted into a “batch” of segments with random
batch size. Our simplifying assumptions, which excluded segmentation, were designed to avoid the extra
complications of batch arrivals.
20
Strictly speaking, the service time is constant only if the assigned time slots are spaced regularly in the
frame, but irregular spacing normally makes little difference in system performance.
386 Multiwavelength Optical Networks

the queue is operating in the statistical steady state, the main quantities of interest in
describing its behavior are the average queue length N̄ (which includes the packet in
service as well as the buffered packets) and average delay D̄ that a packet experiences
in the system. These can be expressed as
N̄ = ρ 1−ρ/2
1−ρ
(5.79)
  
ρ 1 1 F
D̄ = 2µ 1−ρ
+ 2µ
+ L
. (5.80)

Equation (5.79) is the formula for the average queue length of an M|D|1 queue (see
Appendix C for queueing terminology), which is a fairly good approximation of our
slotted transmission channel.
The first term in Equation (5.80) represents the average waiting time in an M|D|1
queue with service time µ−1 . The second term is a correction on the M|D|1 formula to
account for the average latency incurred while waiting for the arrival of the next assigned
time slot. The third term is the effective packet transmission time (one slot time).
The expression for average queue length is plotted as a function of traffic intensity
in Figure 5.29(b). The obvious observation is that queue length (and consequently
queueing delay) both become infinite as ρ → 1, at which point the queue becomes
unstable and a statistical steady state no longer exists. (Note that the queue is stable only
if ρ < 1, implying that it must be empty occasionally.) This behavior is characteristic
of virtually all types of queues with infinite buffers. In practice, the capacity allocation
(proportional to the transmission rate R t ) must be large enough so that the traffic intensity
ρ is significantly less than unity to maintain stable operation of the queue. This is one
source of inefficiency in systems operating with bursty traffic. Assigning capacity so
that ρ is very close to unity (say, 0.95) is risky. A temporary deviation in average traffic
of 5% from its prescribed value is enough to overload the system.
Also plotted in Figure 5.29(b) is the average queue length for a simpler Markov-
ian model: the M|M|1 queue. In the M|M|1 model, the packet length is assumed to
be random and distributed exponentially21 with mean m ′ , giving an average queue
length of
ρ
N̄ = . (5.81)
1−ρ
To make these ideas concrete, a numerical example is presented in Figure 5.29(c). The
figure shows the average time delay from source to destination for a packet on LC [i, j]
arriving to the modified M|D|1 queue, which has an average length that is plotted in
Figure 5.29(b). The total delay includes D̄ plus propagation delay P at the speed of light
in the fiber medium. Because all quantities in Equation (5.80) are inversely proportional
to the transmission rate R t , and because R t is normally high in optical networks, the
queueing delay is normally very small. (Consequently, the differences in computed delay
resulting from using different queueing models are minor.)
21
The random packet length assumption is actually a good approximation of situations in which slots are small
compared with actual packet lengths, and long packets are divided into batches of fixed-length slot-size
segments for transmission over several slots.
 Static Multipoint Networks 387

Throughput Maximum
(pps) No Throughput
Loss

Decreasing
Buffer Size
ρ = Λ ⁄µ
1

Figure 5.30 Throughput versus traffic intensity.

The parameter values used in this illustration are R t = 109 bps, F = 125 µs, L = 100
(so that m ′ = 1250), ci j = 1, and P = 5 ms, corresponding to transmission over a
distance on the order of 1000 km. This delay plot illustrates the relative lack of importance
of queueing delay in wide area optical networks.22 Propagation delays clearly dominate
all others as long as the queue is stable. For example, if we maintain ρ < 0.9, then D̄
remains less than 626 µs in this case.
In addition to queueing delays, a more important issue is buffer overflow. Because
buffer capacity is always finite in any real system, there is always a nonzero probability
that the buffer will be full when a packet arrives (even if ρ is small), in which case the
packet is lost. Packet loss due to buffer overflow becomes larger as ρ increases and as
the traffic becomes more bursty.
When there is loss one must distinguish between offered traffic  and carried traffic
or throughput, which is the offered traffic reduced by the packet loss. The value of the
throughput is thus (1 − PL ), where PL is the probability of packet loss. For values of
ρ > 1, flow conservation dictates that packet loss must always be large enough so that
the throughput is less than the service rate µ. For realistic buffer sizes, the probability of
packet loss generally becomes unacceptably high when ρ is close to unity. Thus whether
buffers are finite or not, performance deteriorates unacceptably as ρ → 1. A typical
family of curves of throughput versus traffic intensity is shown in Figure 5.30. Note how
throughput is reduced as buffer size is decreased.
To keep packet loss under control when buffer capacity is limited, enough capacity
must be allocated to reduce ρ to an acceptable value. How much of a margin of security
is needed in allocating capacity depends not just on packet arrival rate but on the detailed
statistical behavior of the source, especially the “burstiness” and statistical dependencies
of the arrival process. Burstiness is often expressed as the ratio of peak-to-average arrival
rate. In very bursty systems, the statistical fluctuations of packet arrivals may cause
extreme excursions of queue length and very high packet loss rates. In some cases, these
may cause system deadlocks and failures.
Statistical dependencies can cause similar effects. For example, a packet source
that generates packets in a statistically dependent manner is frequently modeled as

22
Queueing delays may still be important on lower speed access links funneling traffic into a wide area optical
backbone network.
388 Multiwavelength Optical Networks

a Markov-modulated process. The source may alternate randomly between two states –
active and idle – where the underlying two-state modulating process is a Markov chain
(see Appendix C). In the active state, it generates packets at some specified rate, and in
the idle state it generates no packets. The queue holding the packets is stable as long
as the average arrival rate is less than the service rate.23 However, even if the queue is
stable, extreme queue excursions may occur if the average duration of an active state is
very long. This leads to high packet loss rates and the other consequences mentioned
earlier. In these cases traffic intensities much less than unity may be required to avoid
these problems.
Having explored the performance of a single LC, let us now adapt the fixed-frame
scheduling results for an M × N shared-channel network to the case of buffered packet
traffic. An M × N framed system acts like a set of M N independent queues, each
associated with one LC. To maintain a fixed-capacity allocation for each connection,
we follow the methodology of Section 5.4.1. Point-to-point LCs are to be scheduled
with traffic requirements that are defined by an M × N traffic matrix  = [i j ] (in
packets per second), with effective packet length m ′ . We consider the general case
of a system with C channels and multiple, fully tunable transmitters or receivers per
station.
Following Equation (5.51), the traffic balance factor is now defined as
¯
/C
η= 
¯
, (5.82)
R
maxi j αTii , β jj , 
C

where

Ti = j i j = aggregate traffic from station i

R j = i i j = aggregate traffic to station j (5.83)
¯ =  i j = total traffic.
 ij

The traffic matrix can be expressed in the form  = 0 T , where a basic packet rate
0 (in packets per second) is chosen so that the elements of the normalized traffic matrix
T are relatively prime integers, representing the number of packets to be scheduled in
one frame. Then a suitable frame schedule over L min slots can be found, with L min
defined as in Equation (5.54), and with ci j = ti j .
Because the traffic intensity for each LC must be strictly less than unity, it is clear
that excess capacity must be allocated to each connection. This can be done by scaling
the underlying bit rate R t to allow for some desired value ρ < 1. (The development that
follows could easily be carried out using different values of traffic intensity ρi j for each
LC. However, for simplicity, ρ is assumed to be identical for all connections.)
If each point-to-point connection is to operate with traffic intensity ρ, then the fol-
lowing constraints must be respected:
ti j ρ R t
i j m ′ = , i = 1, 2, . . . , M, j = 1, 2, . . . , N . (5.84)
L min
23
In this case, the average arrival rate is equal to the fraction of time the source is active multiplied by the
arrival rate in the active state.
 Static Multipoint Networks 389

Summing Equation (5.84) over i and j and solving for R t , we find a minimum bit-rate
requirement of
′ 0 L min m ′
Rmin = , (5.85)
ρ
which gives a scheduling efficiency of
¯ ′
m T̄ ρ
ηs = ′ = ≤ ρ. (5.86)
Rmin C L min C
As in the case of stream traffic, the scheduling efficiency in the packet case can never
exceed the traffic balance factor. In defining ηs , it is always assumed that for the given
bit rate the carried traffic equals the offered traffic, defined by the matrix ; in other
words, there is no traffic loss, even though there may be delays. If ηs < ρ, this means
that there are some unscheduled channel–slots due to a balance factor η < 1. In this
case the empty slots can be allocated as desired to reduce traffic intensities. A scheduled
channel–slot is unused whenever the corresponding LC queue is empty, and this happens
with probability 1 − ρ. Because a common traffic intensity has been chosen for each
LC, the average queue length is the same for all connections. However, the delay for the
higher traffic connections is less than that for the lower traffic ones, due to the fact that
the former are allocated more capacity than the latter (see Equation [5.80]).
The point to keep in mind in this discussion is that when traffic is bursty, a fixed-
capacity allocation for each LC is satisfactory as long as ρ is kept sufficiently small. How-
ever, a conservative capacity allocation producing a small value of ρ reduces scheduling
efficiency. If high efficiency is important, traffic arrival rates and perhaps other charac-
teristics of the traffic must be controlled carefully for each logical channel to maintain
satisfactory performance. Traffic control has two aspects: (1) limitation of average of-
fered traffic on each LC through admission control and (2) “shaping” of packet flow
fluctuations using access and flow control.
Implementation of these functions resides in the higher layers of the network archi-
tecture and is beyond the scope of this text.24
When dealing with a large number of low-throughput bursty users, the most effective
way of dealing with packet traffic is through aggregation of traffic streams together
with dynamic capacity allocation. These functions require moving the packet-switching
function down into the optical layer. This is discussed in Section 5.6.

5.5 Demand-Assigned Connections

In Section 5.4 capacity allocation was treated as a deterministic problem, wherein

a prescribed set of dedicated LCs was scheduled into channel–slots within a periodic
frame. Fixed-capacity “pipes” were assigned to connections, with the capacity allocation
proportional to the number of assigned channel–slots. The fixed-pipe approach was used
regardless of whether the actual traffic was subject to random fluctuations. Thus, we used
24
The interested reader should see, for example, [Schwartz96].
390 Multiwavelength Optical Networks

the same basic procedure for asynchronous (packet) traffic in Section 5.4.2 as we did
for stream traffic, and the random fluctuations in the packet case were accommodated
by buffering and by assigning excess capacity.
In this section, we treat the case of capacity allocation for demand-assigned (circuit-
switched) LCs. In a circuit-switched environment, we deal typically with a large number
of stations, each of which is connected (active) only a portion of the time. Because
these connections are established and released in response to a random sequence of
connection establishment and termination requests, it is the statistical nature of the
connection process that is of prime concern here. Again, what happens “inside” the
connections is ignored, so regardless of whether a connection carries stream or packet
traffic it is treated as a fixed-bit-rate pipe, which requires a corresponding fixed-capacity
allocation. Depending on the state of the network (i.e., the number and type of active
connections and their channel assignments), a connection request may be accepted
or blocked. (Unless otherwise mentioned, it is assumed throughout this section that a
connection request is blocked only if it is impossible to establish the desired connection.)
The key performance parameter in this type of system is blocking probability – a quantity
that depends on the offered traffic, the network resources, and the connection control
algorithms.
The circuit-switching terminology used here is carried over from telephone traffic en-
gineering, where the channels are telephone trunks and blocking occurs when all trunks
are busy. This applies directly to a single optical link carrying multiple wavelengths
and/or multiple LCs on a single wavelength. It also applies to a multiple hop optical
path carrying several wavelengths or LCs. However, demand-assigned capacity alloca-
tion also has meaning in networks carrying cell or packet traffic, such as the optical
access networks discussed in Section 5.7. These access networks generally operate with
admission-control procedures that limit the number of active users in a network to en-
sure satisfactory quality of service. The admission-control procedures block connection
requests from users when the network has reached its capacity limit, behaving just as if
it were operating in a circuit-switched mode. However, the network capacity is shared
dynamically through packet switching in the optical layer (see Section 5.6).

5.5.1 Blocking Calculations in WDMA Networks

We now study the blocking performance of the shared-channel network of Figure 5.23
under demand-assigned traffic. Because analysis of random traffic is a far more complex
problem than the deterministic case, we shall only do a complete analysis of some special
cases. The main objective is to present, through a series of illustrations, a methodology
that can be extrapolated to more general situations.
Assume that each of the M source stations has a single transmitter and makes con-
nection requests at random. If a request is accepted, the connection is held for a random
length of time and is then released. To fix ideas, let us assume that each connection
occupies one full λ-channel for the duration of a call, in which case we are discussing
a WDMA network. (The results are equally applicable to any other type of channel,
such as subcarrier or CDMA.) On the destination side, we consider two possibilities:
 Static Multipoint Networks 391

Idle Active
µ

(a) Source Station

Mλ (M − c ) λ

0 1 c−1 c c+1 C

µ cµ
(b) Birth-Death Process

Figure 5.31 Markov chain model for demand-assigned traffic.

nonblocking stations and stations equipped with a single receiver. In the former case,
each destination station has C ORs, one tuned to each channel. This allows for many-
to-one logical connectivity and eliminates connection blocking due to a busy receiving
station. Although this may be unrealistic, we shall see that it is a good approximation to
many real situations, and it allows us to separate blocking effects due to busy channels
from those due to busy destinations. In the case of single-receiver stations, blocking may
occur either because the requested destination station is busy or because of limitations
in tunability. Varying degrees of tunability are considered for the single-receiver case:
TT-TR, TT-FR, and FT-TR.
We shall use Markovian traffic models for performance analysis. They are simple
and mathematically tractable and more often than not provide good approximations to
real-world traffic statistics. Each source station is modeled as a two-state, continuous-
time Markov chain, as shown in Figure 5.31(a). When the station is in the idle state, it
generates connection (call) requests at a rate of λ requests per second, and if a request is
accepted, it moves to the active state. If a request is blocked, it “forgets” that request and
remains in the idle state, continuing to generate requests at the rate λ. This corresponds
to a lost calls cleared (LCC) traffic model. (In the case of nonblocking receivers, a
call request is accepted whenever there is at least one idle channel.) An active station
releases calls at a rate µ releases per second, which implies that the call holding time is
distributed exponentially with a mean value µ−1 .
In circuit-switched systems, it is convenient to express traffic flow in terms of a
dimensionless unit, the Erlang, which is the product of call arrival rate and average
call holding time. In the absence of blocking, our two-state model generates ρ/(1 + ρ)
Erlangs of traffic, where ρ = λ/µ. Thus, the offered traffic in the network is G =
Mρ/(1 + ρ) Erlangs, which represents the average number of sources that would be
active at any one time in the absence of blocking. The carried traffic is S = (1 − PL )G,
where PL is the probability of call loss.
Assuming that M > C (otherwise there is no blocking), the complete Markov chain
model for this system is the birth-death process shown in Figure 5.31(b), in which the
state c represents the number of calls in progress (i.e., the number of occupied channels).
392 Multiwavelength Optical Networks

The birth rate (call arrival rate) from state c for this case is
λc = (M − c)λ, c = 0, 1, . . . , C − 1 (5.87)
and the death rate (call departure rate) from state c is
µc = cµ, c = 1, 2, . . . , C. (5.88)
The equilibrium probability distribution for the number of calls in progress in this
system, known as the Engset distribution, is easily determined using the methods of
Appendix C. This distribution can then be used to determine the key parameters of this
system, in particular, the average number of occupied channels c̄, and the probability
PB that an arriving call request is blocked, which is just the equilibrium probability πC
that the chain is in state C when the request arrives:
 −1
 C 
M  M
PB ≡ ENG(M, ρ, C) = ρC  ρ j . (5.89)
C j
j =0
25
The probability PL that a call is lost is given by
PL = ENG(M − 1, ρ, C). (5.90)
The Engset model simplifies further if we assume that M ≫ C. This might correspond
to a situation in which many end users are competing for the resources of an access
network like that of Figure 5.1(b), whose capacity is channelized into C units.
If the network is to operate with a reasonably low blocking probability, the rate λ
at which any one station issues connection requests must be small enough so that the
aggregate offered traffic G is equal to or less than the network capacity C. Under these
circumstances we can approximate the finite population (Engset) model of Figure 5.31(b)
by an infinite population (Erlang) model, by letting M → ∞ and λ → 0, with Mλ ≡ 
held constant. The result is a simpler birth-death Markovian model in which there is no
longer any state dependence on the call arrival rate. Without state dependence, PB = PL .
The aggregate call arrival process is now a Poisson process with rate . Departures occur
at a rate µc when the chain is in state c and the offered traffic is G = /µ.
This model leads to the well-known Erlang-B formula for blocking probability:26
 −1
C C
G  
PB = PL ≡ B(G, C) = G j /j! (5.91)
C! j=0

Figure 5.32 compares PL computed with the Engset and Erlang models as a function
of offered traffic G. The upper (lower) set of curves shows the convergence of the Engset
loss model to the Erlang model with increasing M/C for the case C = 5 (C = 10). Note
that, for the case M/C = 5, the Erlang-B formula is a fairly good approximation of the
finite-population case.
25
It turns out that in systems such as this one, with state-dependent call arrival rates, the probability that an
arriving request is blocked, PB , is not generally equal to the fraction of calls lost, PL (see [Schwartz87,
p. 520]).
26
For curves of the Erlang-B formula see, for example, [Briley83].
 Static Multipoint Networks 393

0
10

−1
10

−2 C =5
Loss Probability

10

−3
10

−4 C = 10
10

−5
10
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Offered Traffic G (Erlangs)
Erlang
Engset: M/C = 5
Engset: M/C = 2

Figure 5.32 Comparison of Engset and Erlang models.

Instead of using blocking probability, the performance of a circuit-switched system

can also be displayed as a curve of normalized throughput, c̄/C, versus normalized
offered traffic, G/C, or traffic intensity. The dashed curves in Figures 5.33 and 5.34,
respectively, show the probabilistic effects of call blocking (using the Erlang-B model)
for the cases M = 250, C = 50 and M = 50, C = 10. Thus in Figures 5.33 and 5.34
we compare large and small systems with the same ratio M/C. In each case M ≫ C,
so the Erlang-B approximation is a reasonable one. Loss probability can be deduced as
the difference between traffic intensity and normalized throughput.
Ideal performance, represented by the solid line in both figures, is the throughput that
would be obtained if the connection process was deterministic, with connection requests
arriving at regular intervals of exactly −1 seconds and connections held for exactly µ−1
seconds each. The deviation of actual performance from the ideal is due to the random
fluctuations of the connection process; as in the case of packet traffic, randomness
reduces traffic-carrying capacity. Note that the large system achieves performance that
is closer to the ideal than the small one, a consequence of the effect of the law of
large numbers in “smoothing out” the fluctuations in the connection process. This is a
394 Multiwavelength Optical Networks

C = 50, M = 250
0.9

0.8

0.7
Normalized Throughput, S/C

0.6

0.5

0.4

0.3

0.2

0.1

0
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Traffic Intensity, G/C
Ideal
Erlang
TT-TR
TT-FR, FT-TR

Figure 5.33 Normalized throughput versus traffic intensity.

well-known effect in all circuit-switched systems: Scaling up a system while maintaining

the same traffic intensity reduces the blocking probability significantly.
The previous results were based on the assumption of transmitter tunability and
nonblocking destination stations, which simplifies the analysis but may not always
be feasible. In [Ramaswami+90], the authors examine the influence of tunability on
performance. In their analysis, an M × M system is assumed, with one transmitter in
each source station and one receiver in each destination station and with C < M channels
shared among the M stations. Each idle station is assumed to generate call requests at
a rate λ, with uniformly distributed destinations. Three cases are considered: TT-TR,
TT-FR, and FT-TR. A modified Engset model is appropriate in this case, in which the
call arrival rates must include blocking effects dependent on the tuning rules.
In the TT-FR case, each channel is assumed to be assigned permanently to a set of
M/C destination stations. Therefore, a connection request is blocked if it is directed to
a receiving station tuned to an occupied channel. In the FT-TR case, each transmitter is
assigned a fixed channel using the same rule as in the TT-FR case. To avoid collisions on
 Static Multipoint Networks 395

C = 10, M = 50
0.9

0.8

0.7
Normalized Throughput, S/C

0.6

0.5

0.4

0.3

0.2

0.1

0
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Traffic Intensity, G/C
Ideal
Erlang
TT-TR
TT-FR, FT-TR

Figure 5.34 Normalized throughput versus traffic intensity.

occupied channels, a connection request generated at a transmitter tuned to an occupied

channel is blocked.
Normalized throughput curves for the TT-FR and FT-TR cases, which have identical
blocking probabilities, are shown as dot–dashed lines in Figures 5.33 and 5.34. Note that
normalized throughput is significantly less than that for the TT-TR case, shown dotted.

5.5.2 Blocking in Combined Time-Wavelength-Division Networks

By abandoning the channel–slot frame structure in Section 5.5.1, the additional chan-
nelization possibilities afforded by combined time-wavelength-division techniques were
sacrificed. We now examine the circuit-switched shared-channel network in the framed
context. As before, the channels can be taken as λ-channels, which would make this a
TDM/T-WDMA network.
Consider an M × N system with C channels, one tunable transmitter in each source
station, and one tunable receiver in each destination station. Each source station
396 Multiwavelength Optical Networks

generates connection requests at random, distributed uniformly to all destinations, and

all transceivers run at a fixed bit rate R t bits per second, where each active LC is assigned
one channel–slot in a frame of length L. As a result, each connection is allocated the
same bandwidth, running at an effective bit rate R 0 = R t /L bits per second. In this case,
one-to-many (many-to-one) logical connectivity is possible by assigning more than one
slot in a frame to a given source (destination) station, and multicast LCs can be created
if more than one receiver tunes to the same channel in the same time slot. Furthermore,
with the random connection request model, there are cases where more than one slot
is assigned to the same source-destination pair. This can be interpreted as assigning
additional bandwidth on demand to a given LC.
The framed TDM/T-WDMA system uses the time dimension to refine the “granular-
ity” of the shared-channel system. Thus, if L is large, a system is created in which the
effective bit rate R 0 assigned to each LC is much less than the transmitter bit rate R t .
This is often desirable because circuit-switched connections frequently require lower bit
rates than dedicated connections. The former would typically be used for individual user
applications running at relatively low speeds, whereas the latter would most likely be
used in a semipermanent optical infrastructure for provisioning large, high-throughput
logical networks. At the same time, by dividing each λ-channel into many small channel–
slots, the “size” of the system in terms of the number of elementary information-bearing
entities is made much larger.
To illustrate, suppose M ≤ N . Then with C = M, the system can be run in an FT-TR
mode using distinct channel assignments for each source station.27 In this case, a call
request generated at source station i for an LC [i, j] is accepted if there is at least one
slot in the frame that neither station i nor j is using. Otherwise it is blocked. This slot-
matching condition is illustrated in Figure 5.35, in which a connection between source
station 2 and destination 4 is requested in a frame of length L = 6. Figure 5.35 shows calls
in progress from source 2 to destinations 3, 1, and 2 and calls in progress from sources 1
and 3 to destination 4, where the connection [3, 4] uses two units of bandwidth. Slots 1
and 3 are available for the new connection, but all other slots would lead to mismatches.
In cases where every vacant slot would lead to a mismatch, rearrangement of existing
slot assignments might solve the problem. This may or may not be feasible depending
on how the system is managed.

Channel 1 [1, 4]

Channel 2 [2, 3] [2, 1] [2, 2]

Channel 3 [3, 4] [3, 4]

Channel 4

Slot 1 2 3 4 5 6

Figure 5.35 Matching time slots.

27
The development applies equally well to the case of systems with M ≥ N and C = N (using the TT-FR
mode).
 Static Multipoint Networks 397

PB L = 25 L = 50 L = 100
10−2

10−3

10−4

ρ/L (Erlangs)
0 0.3 0.6 0.9

Figure 5.36 Framed system blocking probabilities.

Figure 5.36 shows curves of blocking probability as a function of normalized offered

traffic ρ/L for M × M systems with several values of frame size L, based on an
approximate analytical model presented in [Humblet+93]. The normalization is obtained
by dividing the total offered traffic G = M ρ by the total channel–slots available to carry
it, M L, so that the system size M is eliminated. Two sets of curves are presented: (1) PB
including effects of slot mismatches (solid curves) and (2) PB ignoring slot mismatches
(dashed curves).
Note that slot mismatch has a fairly minor effect on blocking, reducing the allowable
offered traffic for a given loss probability by less than 10%.
As mentioned earlier, the framed system is “fine grained,” which means that the laws
of large numbers work in our favor. This effect is clearly observable in Figure 5.36, in
which the blocking probability is seen to decrease sharply with increasing L.

Admission Control
One way of eliminating mismatches is to constrain the permissible number of connections
in a frame. Continuing with our M × N FT-TR system, suppose an admission-control
rule is applied in which each source station is limited to N simultaneous connections and
each destination station is limited to M connections. (Repetitions are allowed but are
counted as distinct connections.) We shall refer to this as the limited-connection case.
Furthermore, set

L = M + N − 1. (5.92)

For this frame length, blocking of a connection request occurs only if either the
source or the destination is already at its connection limit.28 Constraining the number of
connections in this way limits the maximum carried traffic to approximately 50% of the
system capacity.

28
The condition in Equation (5.92) is identical to a well-known nonblocking result for traditional time-
multiplexed electronic switches [Schwartz87].
398 Multiwavelength Optical Networks

[1, 3] [1, 2] [1, 2] [1, 3] [1, 2] [1, 3] [1, 2] [1, 3]

[2, 2] [2, 4] [2, 2] [2, 2] [2, 4] [2, 2]

[3, 1] [3, 1] [3, 1] [3, 1] [3, 1] [3, 1] [3, 1] [3, 1]

[4, 4] [4, 3] [4, 4] [4, 4] [4, 3] [4, 4]

(a) Dedicated Assignments (b) Demand Assignments Rearrangement

Figure 5.37 Illustrating rearrangeability.

Nevertheless, the limited-connection system might be of interest for other reasons,

such as fairness: Limiting connections to and from each station tends to prevent stations
from “hogging” the shared-channel system.

Rearrangeability Revisited
Another way of eliminating mismatches is by channel–slot rearrangement. To illustrate,
consider a 4 × 4 directed star network with C = 4, operating in an FT-TR mode.
Suppose the following sequence of connection requests is to be scheduled: [1, 2],
[3, 1], [2, 2], [1, 3], [4, 4], [2, 4], [4, 3]. All LCs require one channel–slot except [3, 1],
which requires two; that is, we are allowing multiple effective bit rates. If this is con-
sidered as a dedicated connection assignment problem, it can be stated in terms of a
normalized traffic matrix:
 
0 1 1 0
0 1 0 1
T= 2 0 0 0.
 (5.93)
0 0 1 1

This is clearly a balanced traffic matrix, and because T̄ = 8, it can be scheduled in

a frame of length L = 2, as shown in Figure 5.37(a). However, if demand-assigned
scheduling is used, the first CAS in Figure 5.37(b) shows the schedule as it would look
after the first five connections are scheduled. The scheduling rule used to produce this
result is the following: Place a connection in the first time slot if possible; if not, use
the second slot. A slot mismatch problem now arises when the last two connection
requests arrive. Neither can be scheduled without a conflict at the receivers, so they must
be blocked. There is no way to avoid this situation without knowing the sequence of
requests in advance. Thus, just as the statistical fluctuations of the call request arrival
process reduce the traffic-carrying capacity of a shared-channel network, the random
order of these requests also reduces the carried traffic. However, if rearrangement is
permitted, this problem can be circumvented.
In this example, the conflict is eliminated by moving connection [4, 4] to the sec-
ond time slot, as shown in the second CAS in Figure 5.37(b). Thus, the issue of
rearrangeability, which appeared in permutation switches in Section 2.3.2.1 and in WRNs
in Section 3.3.1.1, reappears once more in time-multiplexed shared-channel systems. In
general, scheduling with slot rearrangeability is equivalent to dedicated connection
scheduling, so that blocking of a new connection request occurs only if it is impossible
 Static Multipoint Networks 399

to accommodate the complete list of active connections plus the new request using the
existing frame length. In this case, as long as the minimum frame length for the active
connections plus the new request (computed using Equation [5.54]) does not exceed
L = 2, all connections can be accommodated with rearrangement. As in the previous
discussions of rearrangeability, this option may or may not be a practical alternative.29
The framed TDM/T-WDMA system can be operated in more general circuit-switched
contexts as long as the usual constraints required to avoid collisions and conflicts are
observed. Some generalizations include the following:
r Multicast connections
r Multiple transmitters/receivers per station
r Effective bit rate selected on demand
r More elaborate admission-control rules

5.6 Packet Switching in the Optical Layer

So far we have considered the capacity allocation problem for shared-channel systems
using either dedicated or demand-assigned connections. Thus, the focus has been on
connection-oriented transmission, and the capacity allocation was treated as if it was
fixed for the duration of a connection. This type of service is appropriate for applications
requiring a guaranteed bandwidth, which includes all synchronous traffic applications
and some (typically real-time) asynchronous (packet) traffic applications. It is less
appropriate in an environment consisting of a large population of low-throughput bursty
users who require a high degree of logical connectivity. In this case, packet switching
is the natural way of carrying the traffic. The most efficient way of implementing the
packet-switching function in a purely optical network is to extend it downward to span
the logical and optical layers. The properties of the shared-channel broadcast medium
make it a particularly effective support for packet switching in the optical layer.
The general features of this approach were presented in Section 3.2.2. Recall that in
packet switching, as opposed to dedicated or demand-assigned (circuit-switched) ca-
pacity allocation, addressing, scheduling, and capacity allocation are implemented on a
packet-by-packet basis. This entails extensive communication, control, and processing
functions, which are absent in the dedicated or circuit-switched case. Nevertheless, a
number of benefits accrue from the additional communication and processing load, the
two most significant being: (1) high connectivity realized using packet header infor-
mation for routing and (2) high throughput and low delay realized through dynamic
capacity allocation.
Dynamic capacity allocation is an implicit by-product of the MAC protocols designed
for packet switching on shared-channel media. The protocols that are discussed here are,
in most cases, direct generalizations of those used in traditional LANs and MANs and

29
Rearrangement of time slots in a shared broadcast medium is normally possible with less disruption than,
say, reassigning wavelengths and optical paths in wavelength-routed networks because it is possible to
“swap” time slots with little or no information loss.
400 Multiwavelength Optical Networks

must address the same problems: contention resolution, packet loss, and retransmission
if required.
A large number of MAC protocols have been devised for optical broadcast-and-select
networks, representing various trade-offs among performance, overhead, hardware cost,
and processing complexity [Mukherjee92, Ramaswami93]. Practical M station systems
generally require anywhere from one to M separate control channels for dynamically
scheduling packets. This represents significant out-of-band communication overhead and
transceiver hardware. Furthermore, packet scheduling requires a considerable amount
of real-time processing. As the size of the system grows, both in number of stations
and geographical spread, channel collisions, receiver conflicts, signaling overhead, and
propagation delays, as well as protocol processing overhead can reduce the efficiency of
these systems substantially, increasing packet loss and/or delay and reducing throughput.
Thus, our objective here is to study a number of representative protocols, focusing on
how cost–performance trade-offs change with the structure of the system.
A good MAC protocol achieves the two objectives of high logical connectivity and
high performance, and it does so by aggregating traffic and allocating capacity dynami-
cally. Recall from Section 5.4.2 that when fixed allocations were used for packet traffic,
estimates of average traffic requirements on each LC were needed at connection estab-
lishment time to produce a fixed-frame schedule with the required capacity allocation.
The performance of a fixed-capacity system is vulnerable to unpredictable traffic fluctu-
ations as well as errors in the estimated requirements. In the dynamic case, we hope to
achieve better performance and efficiency by using knowledge of the actual (rather than
the estimated) traffic during the information transfer phase. Ideally, this knowledge can
be used to allocate channel–slots “instantaneously” while at the same time coordinating
transmissions among all stations to resolve contention for resources.
Thus, the fundamental difference between capacity allocation in the fixed and dynamic
cases is that scheduling is done once and for all at connection setup time in the former
case and during information transfer in the latter case. Clearly, the more knowledge each
station has of the instantaneous “global” state of the traffic in all other stations, the more
efficiently the traffic can be scheduled. But this information is distributed geographically
throughout the network, so to establish a dynamic schedule, a certain amount of control
information concerning packets awaiting transmission must be communicated among
the stations, and this information must be processed to produce scheduling decisions.
Various degrees of control are exercised in typical packet-switched shared-channel
protocols. We categorize these as30

r Uncontrolled scheduling (random access)

r Scheduling with loss
r Lossless scheduling (reservations)
r Perfect scheduling

30
In reality, an enormous number of protocols have been proposed that do not always fit neatly into these
categories (see, for example, [Mukherjee92]). It is beyond the scope of this text to cover all of them.
 Static Multipoint Networks 401

In the descriptions that follow, the optical infrastructure is taken as an M × M un-

directed broadcast star with each station having a single transceiver for data.31 (There may
be additional transceivers for control.) Let us assume that all stations are equidistant from
the star coupler, so the propagation time from source to destination (the time required
for a signal to propagate to the star coupler and back) is the same for all stations. The
systems all operate in a slotted mode, with transmitters picking up slot synchronization
as well as other control information by seeing all transmissions (including their own).
The transmitted data are assumed to be in the form of fixed-length packets, with the
packet length equal to the slot size. The traffic is assumed to be distributed uniformly,
and connection oriented, with all connections point-to-point. However, connectionless
transmission as well as multicast connections are handled easily in the systems treated
here.
As indicated in Section 3.2.2, most protocols for packet switching in the optical layer
involve some probability of lost packets. Thus, we must keep track of two types of traffic:
original and retransmitted. Let
S = the normalized throughput (average packets per channel–slot)
G = the total normalized traffic offered to the network (average packets per channel–
slot)
E = the average number of retransmissions of a packet until success
ps = the probability that a transmission is successful.
The total traffic G contains both originally generated packets (the normalized through-
put S) and retransmitted packets. We are particularly interested in Smax – the maximum
normalized throughput the system can support. No matter what the protocol, the loss
mechanism (e.g., collisions or conflicts), or the retransmission rule, the quantities defined
here are related by32
G = S(1 + E) (5.94)
S = ps G (5.95)
E = (1 − ps )/ ps . (5.96)

5.6.1 Uncontrolled Scheduling: Random Access

The simplest possible MAC protocol is random access, with the earliest implementation
being the ALOHA system [Abramson73]. Each station transmits its packets and stores
a copy of a transmitted packet in case a retransmission is needed. The RP in each station
reads all packet headers, performing a packet-filtering operation to select those destined
for it. (Multicast destination addresses can be used in the case of multicast LCs.) Channel
collisions occur when two or more packet transmissions overlap in time, in which case
all collided packets are lost. The collisions are detected by the RPs in the stations that
31
The undirected case is most appropriate for optical layer packet switching because of the ease of distributing
control information.
32
These relations apply only to systems in which lost packets are retransmitted. In some applications, lost
packets are simply dropped without attempts at retransmission (see Section 5.6.2).
402 Multiwavelength Optical Networks

transmitted the collided packets, and this information is used to reschedule those packets
at a later time.
This system has spawned many similar protocols, but by far the most important of
these is Ethernet, which is essentially the ALOHA protocol refined and adapted to wired
LANs.33 Although ALOHA is probably the oldest MAC protocol, and not a very efficient
one at that, it serves as a useful point of comparison for more sophisticated techniques.
Two performance parameters are of interest: the maximum normalized throughput,
S max , and the average delay until successful receipt of a packet, D̄. The key parameter in
computing these quantities is the probability of successful transmission ps . In general, an
exact value of ps is difficult to obtain because of statistical dependencies among original
and retransmitted packets. However, with a few simplifying assumptions, a fairly simple,
accurate, and tractable mathematical model can be obtained (see [Lam74]).
In the slotted version of ALOHA, based on a slot-synchronized, single-channel shared
broadcast medium, any station having a packet to transmit sends it in the next available
slot. In this model, we assume that each station transmits into each slot with probability
p, so that G = M p, and the transmission is successful with probability ps , where p
and ps for each slot and each station are independent of activity in previous slots. (One
implication of this model is that there is no packet queueing in the transmitting stations
and hence no queueing delay.) This state of affairs approximates quite well a system
with a large number of stations, in which collided packets are retransmitted after random
delays, averaging about three or more slots. In this case, we have

ps = (1 − p) M−1 ≈ e−G , for M ≫ 1. (5.97)

Then from Equations (5.95) and (5.97) we have

S = Ge−G , (5.98)

which is maximized for G = 1, giving S max = e−1 = 0.37 and E = 1.7. For all values of
throughput less than S max there are two possible values of G. However, the larger value
(G > 1) corresponds to a larger value of E and to unstable operation of the network.
This illustrates the well-known difficulty with all systems involving packet loss: The
retransmitted traffic puts an extra load on the network, which may reduce significantly
the maximum throughput. To avoid congestion and instabilities, offered traffic must be
kept well below the value that maximizes throughput, and special retransmission control
algorithms must be used to maintain stability. (Similar problems arise in CSMA/CD.)
The average delay until successful reception of a packet is expressed in units of time
slots as

D̄ = 1 + P + E(P + K̄ ), (5.99)

where P is the source-to-destination propagation time (in slots) and K̄ is the average
time before retransmission of a collided packet (in slots). In a typical optical network,

33
In optical networks, as opposed to wired LANs, collisions are normally detected long after the packet
transmission has been completed, so protocols such as Ethernet using carrier sense (e.g., CSMA/CD) are
unusable (see Section 3.2.2).
 Static Multipoint Networks 403

S D/P
0.37
3 3

2 2

1 1

G G
1 1
(a) Throughput (b) Delay

Figure 5.38 Slotted ALOHA.

P is the dominant quantity in Equation (5.99), so

D̄ ≈ (1 + E)P. (5.100)
Because E is 1.7 at maximum throughput in slotted ALOHA, the average packet
delay is approximately two to three propagation delays, corresponding to one to two
retransmissions. Throughput and delay curves for slotted ALOHA as a function of total
traffic G are shown in Figure 5.38. (The unstable region corresponding to G ≥ 1 is
omitted.) The delay is plotted in Figure 5.38(b) in units of propagation delay P. Note
that the delay is less than 3P throughout the stable region.
Random access has the virtue of simplicity, but suffers from low throughput and
relatively high delays. To improve throughput and delay, additional protocol complexity
must be introduced.

5.6.2 Scheduling with Loss

In this section, the performance improvement produced by the availability of certain
limited global state information is investigated. To fix ideas, consider an M × M system
with C = M channels and a single transceiver per station, operating in FT-TR mode with
a distinct channel assigned to each source station. New packets awaiting transmission
are queued in each source station together with copies of packets that were transmitted
previously but not received.
Now suppose that each station with a packet to transmit first communicates to all
receivers the destination of the packet and then sends the packet in the next time slot.34
This is known as a tell-and-go protocol. Because it takes one source-to-destination
propagation time P for the information about transmissions to reach the receivers, and
the data packet is transmitted before that time elapses, the transmitting stations have
no way of knowing whether others are also transmitting when they send their packets.
Furthermore, they have no confirmation that the intended destination station will be
ready to receive the packet when it arrives. (Waiting for a confirmation would delay
34
Of course, the mechanism by which this control information is communicated is an essential feature of the
protocol. It could be done through a dedicated control channel, or it could be piggybacked on data packets.
At this point it suffices to assume that the information is conveyed correctly to all stations “somehow”
before the arrival of the packet.
404 Multiwavelength Optical Networks

the transmission an additional propagation time.) This “haste” is the advantage (and
disadvantage) of tell-and-go. It keeps delay to a minimum, but it also leads to conflicts,
which arise when more than one station transmits a packet destined for the same receiver.
In that case, the receiver must tune to one of the active transmitters, causing losses of
the other packets.
In tell-and-go, these packet losses are inevitable. But assuming that all stations know
the rules by which the receiving stations choose among conflicting packets, the source
stations all know which packets were received correctly and which were dropped. In
this manner, each station knows after an interval P whether it must retransmit a packet.
Although there are conflicts with this system, the use of FT-TR with a full complement
of channels avoids collisions, and thus it is an improvement over TT-FR multichannel
random access.
As an illustration, consider the 4 × 4 example in Figure 5.39(a). Each station main-
tains a single queue. The figure shows a “snapshot” of the state of the queues in the four
transmitting stations at one instant in time. Each queue contains four packets awaiting

Source
Station

1 1, 3 2
1, 2 2
1, 3 1
1, 2 1

2 2, 4 3
2, 4 2
2, 3 1
2, 4 1

3 3, 3 1
3, 1 2
3, 4 1
3, 1 1

4 4, 3 1
4, 1 2
4, 2 1
4, 1 1

(a)

Channel 1 [1, 2]1 [1, 3]1 [1, 2]2 [1, 3]2

Channel 2 [2, 4]1 [2, 3]1 [2, 4]2 [2, 4]3

Channel 3 [3, 1]1 [3, 4]1 [3, 1]2 [3, 3]1

Channel 4 [4, 1]1 [4, 2]1 [4, 1]2 [4, 3]1

Slot 1 2 3 4

(b)

Figure 5.39 Tell-and-go protocol.

 Static Multipoint Networks 405

transmission, labeled in the figure with their source-destination addresses and LC se-
quence numbers (subscripts). The distribution of destination addresses has been chosen
to match the normalized traffic matrix in Figure 5.24(b). Recall that the fixed-frame
FT-TR CAS of Figure 5.25(b) accommodates these packets in four time slots. The sta-
tions normally schedule their transmissions FCFS, with each packet at the head of its
queue transmitted in the next available channel–slot. This is shown in the schedule of
Figure 5.39(b), in which slots are now allocated dynamically rather than in a fixed frame.
In the case of conflicts, the best the designated receiver can do is to tune to one of the
contending packets, losing the others. Figure 5.39(b) shows a case in which each receiver
tunes to the lowest numbered transmitter having information for it, resulting in the loss
of the five crossed-out packets.
The schedule is implemented in four time slots, which is sufficient for conflict-free
packet delivery in a fixed-frame system. However, the loss due to conflicts occurs because
of the order in which the packets happened to be queued (and served). Comparing
fixed scheduling to tell-and-go, we note that in both cases packets within each LC are
transmitted in the correct order. However, because of the pipelining that occurs in tell-
and-go, the loss of a packet is not likely to be detected until several packets behind
it have been transmitted. Thus, the MAC protocol does not guarantee reliable and
ordered packet delivery to the higher layers of the network. If loss-free, ordered delivery
is required in an application, it must be realized through appropriate retransmission
protocols (which include keeping track of sequence numbers) in the higher layers of the
network architecture: the logical link layer and higher.
For the random access system of Section 5.6.1, it was seen that collisions produce
nonmonotonic relations between throughput and total traffic, together with tendencies
toward instability. It turns out that conflicts in tell-and-go limit normalized throughput
(in packets per channel–slot) to Smax = 1 − e−1 = 0.63, and there is a monotonically
increasing relation between offered traffic and throughput, without any instability. How-
ever, this example assumed a full complement of channels. If C < M, then there will be
collisions as well as conflicts, and the system begins to resemble random access.
Although we have assumed that all lost packets are retransmitted, not all applications
require this. For example, in real-time applications such as packet video, the retrans-
mission delay for lost packets would probably be unacceptable, but a small fraction
of dropped packets may be permissible. By running the network at sufficiently low
throughput, acceptable loss probability for real-time applications can be achieved.

5.6.3 Lossless Scheduling: Reservations

The conflicts and packet reordering problems incurred using tell-and-go can be avoided,
but this requires extending the packet schedule over more than the minimum number of
slots. To construct a lossless schedule, global information concerning the packets at the
head of each station’s queue is required. This entails the distribution of additional control
information among all transmitting and receiving stations and longer packet delays than
tell-and-go. (A minimum of one propagation time P is required to communicate the
global state to all stations before transmission begins.)
406 Multiwavelength Optical Networks

Channel 1 [1, 2]1 [1, 3]1 [1, 2]2 [1, 3]2

Channel 2 [2, 4]1 [2, 3]1 [2, 4]2 [2, 4]3

Channel 3 [3, 1]1 [3, 4]1 [3, 1]2 [3, 3]1

Channel 4 [4, 1]1 [4, 2]1 [4, 1]2 [4, 3]1

Slot 1 2 3 4 5 6

Figure 5.40 Lossless scheduling.

In a conceptually simple arrangement, each transmitting station broadcasts in each

time slot (on a separate control channel) the destination address of the packet at the head
of its queue. All stations receive this information and execute a distributed algorithm
that determines which stations are allowed to transmit in the next available slot. (The
next available slot in this case is delayed somewhat more than P from the time the
address information is broadcast to leave sufficient time for the control information
to be received and processed.) In an FT-TR system with distinct channel assignments,
a packet selection rule must be followed to eliminate conflicts. It might be as simple
as selecting the conflicting packet with the lowest numbered source address, but a
better arrangement would be to select the packet at random or based on some other
fairness, priority, or quality-of-service (QoS) criteria. In systems with less than a full
complement of channels, the protocol must determine which stations should transmit
and what channel they should use to avoid both conflicts and collisions.
Choosing a limited number of packets for transmission in the next slot amounts to
reserving channels on a slot-by-slot basis – an example of dynamic scheduling using
reservations.35 In a slot-by-slot reservation system, packet loss is replaced by delay.
An extra delay of at least P is incurred by each packet. This is a penalty incurred for
ensuring lossless and ordered delivery, and it is not a serious problem in networks of
limited geographic size such as LANs or access networks.
Figure 5.40 shows how a lossless schedule for the packets queued in Figure 5.39(a)
can be achieved in six slots. Note the relation between tell-and-go and reservations. In the
tell-and-go case, whenever the packets at the head of the queues created conflicts, they
were transmitted and all but one were lost. In the reservation system, potential conflicts
were resolved by holding back transmissions, creating idle channel–slots. These idle
slots are caused by a phenomenon known as head-of-the-line (HOL) blocking, which
reduces throughput.36 In fact, it can be shown [Hui90], using arguments similar to those
we used for tell-and-go, that HOL blocking limits the maximum normalized throughput
35
More elaborate schemes have been proposed that involve reserving several channel–slots in advance, piggy-
backing control information on data packets, combining static and dynamic slot assignments, and so forth
(see [Mukherjee92]).
36
This type of blocking can be avoided by taking packets out of order. For example, the packet at the head of
the queue in station 4 could not be scheduled simultaneously with the others in the first time slot without
a conflict. However, referring to Figure 5.39(a), the packet from the same station with destination 3 could
have been transmitted without a conflict if it had been taken out of order. (Delivery within each logical
channel would still occur in correct order.) The other three packets ahead of it in the queue blocked this
possibility because of the FCFS queue discipline.
 Static Multipoint Networks 407

in these lossless systems to 0.63, the same value for tell-and-go! Thus, the use of extra
control information has not improved the throughput of the system, but it has solved the
problem of ensuring ordered delivery.37

5.6.4 Perfect Scheduling

In both tell-and-go and lossless scheduling, performance is less than optimal from the
point of view of loss and/or delay because of a lack of complete information concerning
the state of the source queues. Now suppose that complete information was available
to all source and destination stations. How should capacity be allocated, and how well
could the system perform?
In principle, if global queue information is available to all stations, and if packets are
permitted to be scheduled in other than FCFS order, then a “perfect” schedule can be
computed in which there is no packet loss, packets are delivered in correct order, and
the channels are used with maximum efficiency; that is, a minimum number of channel–
slots are idle. The operation of such a system is illustrated conceptually in Figure 5.41.
We continue with a 4 × 4 FT-TR system with distinct channels assigned to the source
stations. The figure shows a modification of Figure 5.39(a), in which the system is in the
same state as in Figure 5.39(a), but its waiting packets are now stored in four separate
queues per station.
Assuming that this global-state information is available to all stations, an algorithm
can be executed in each station that determines a conflict-free schedule for the waiting
packets in a minimum number of time slots. (The channel–slot allocation algorithm of

Destination
1

2
Source Station 1
3

Destination
1

2
Source Station 4
3

Figure 5.41 Queues for perfect scheduling.

37
The phenomenon of HOL blocking has been observed and analyzed in the context of electronic packet and
cell switch fabrics, including crossbars and multistage switches.
408 Multiwavelength Optical Networks

Appendix B can be used in which each waiting packet represents one unit of traffic in
a normalized traffic matrix T .) Assuming that the global queue information is updated
and distributed on a slot-by-slot basis, the schedule can be extended continuously in
time and modified if necessary as new packets arrive. The longer the queues, the longer
the schedule can be extended forward in time, and the more efficient it will be.
Of course, an extra delay P is incurred to distribute global information, as in the case
of reservations. More importantly, the processing time in the perfect scheduling case
can be considerable for large systems.
Depending on the queue states, it may not always be possible to fill all channel–
slots. For example, occasionally some or all queues in a station are empty. (This is
inevitable in a stable queueing system.) That station’s channel may then be idle in the
next slot. However, in the heavy-traffic case, when a queue is rarely empty, the channel–
slot occupancy approaches the scheduling efficiency ηs dictated by the average traffic
distribution. For balanced traffic ηs → 1 as ρ → 1.
Is perfect scheduling a practical objective in real networks? Clearly, it produces less
average packet delay and higher throughput than the lossless case of Section 5.6.3,
because it avoids HOL blocking. Furthermore, it ensures reliable and ordered deliv-
ery (assuming no channel errors) because packets in the same LC are transmitted
in order even though packets in different LCs may be served out of order. How-
ever, the extra communication overhead, more complex data structure (for multiple
queues), and additional processing must be weighed against potential performance
improvement.
Several protocols have been proposed in which some attempt is made to approach
perfect scheduling (see [Chen+91, Chipalkatti+92]). They all involve keeping track
of the states of all queues and computing either optimal or suboptimal transmission
schedules. For large M these techniques pose significant problems in terms of the
volume of control information, the required databases, and the processing required in
each station.

5.6.5 Dynamic versus Fixed Capacity Allocation

All of the procedures discussed previously for dynamic scheduling tend to adjust capac-
ity allocations to the instantaneous traffic demand. Ignoring the issue of communication
and processing overhead, the procedure with the highest performance is perfect schedul-
ing. How does this compare with fixed-capacity allocation? Because of the complex
interactions among all of the source queues, an exact analysis of performance of perfect
scheduling is very difficult. However, certain general statements can be made.
Looking at the system from the point of view of flow conservation, it should be
clear that the conditions relating traffic requirements and network resources for an
M × N system using fixed-capacity assignment apply to the dynamic case as well. Flow
conservation requires that certain aggregate average flow conditions are satisfied. Let
us assume a slotted system with fixed packet lengths of m ′ bits fitting into one slot
and all transmitters operating at R t bits per second. Furthermore, let the aggregate
traffic intensity at each source station be limited to ρ < 1. Then, using the definitions in
 Static Multipoint Networks 409

Equation (5.83), the traffic is constrained by

Ti m ′ ≤ ρ R t i = 1, 2, . . . , M (5.101)
′
R j m ≤ ρ R t j = 1, 2, . . . , N (5.102)
¯ ′ ≤ ρ R t C.
m (5.103)

Comparing Equations (5.101) and (5.102) to the constraints in Equation (5.84), we

see that dynamic scheduling of packets imposes only M + N constraints on source and
destination flows rather than M N constraints in the case of fixed-frame scheduling.
Furthermore, the constrained flows in the former case are aggregate traffic, all of which
make dynamic capacity allocation via packet switching a more flexible system with
better performance in the face of unpredictable variations in flow.

5.7 The Passive Optical Network

As pointed out in Chapter 1, the access network, called the “last mile” or the “first
mile”, depending on your point of view, remains the final bottleneck in bringing the vast
bandwidth of optical networks to the end user. The reason is cost. As the optical network
approaches the end user the number of access stations and end systems connected to it
grows exponentially. Thus, as we have already seen in the case of metro networks, the cost
per user becomes the overriding concern. Conversely, distances between network nodes
and end systems become smaller, and traffic loads to those end systems become lighter,
so lower performance components can be tolerated. Just where the optical infrastructure
should end and other media; e.g., copper or wireless, should begin continues to be a
subject of much debate. Since the trial deployment of fiber to the home (FTTH) by
Northern Telecom in 1981 the various telecommunications carriers have retreated to
access fiber terminations short of the end user: fiber to the curb (FTTC), fiber to the
building (FTTB), and fiber to the cabinet (FTTCab), with tentative forays closer to the
home as economic conditions and demand dictate.
Wherever the optical/electronic boundary is set, there is a need for a cost-effective
access network to bridge the gap between the core/metro network node and the end
system. This sets the stage for the passive optical network (PON). The term PON was
coined when these networks were first deployed for limited local distribution by British
Telecom [Hoppitt+89]. It is, perhaps, a sign of the times that “distribution,” a term
long used by the telecommunications carriers, has been replaced by “access,” a term
preferred by the computer networking community. The former suggests the point of
view of the network operator, whereas the latter evokes the user’s needs. Although it
was (and is) clear that PONs fit the requirements of optical access networks, the traffic
demand that makes optical access economically feasible is only now materializing,
largely driven by high bandwidth applications such as multimedia Internet services and
video distribution. Increases in demand, perceived economies from aggregating voice,
data, and multimedia services on a common access network, and, perhaps most important,
competition among diverse telecommunications service providers (e.g., CATV, satellite,
410 Multiwavelength Optical Networks

End systems

ONU11

Core/metro network,
ONU22
Service providers

1
OLT

ONU
ONU33

ONUNN

Figure 5.42 Passive optical network.

and wireless) have generated considerable recent activity in commercial access network
deployment, standardization, and research.
As its name implies, the PON is characterized by its passivity. For access networks
designed to be deployed in the field (as opposed to laboratory or in-house networks),
passivity implies a complete absence of direct or ancillary power: no amplification or
active signal processing (our usual definition of passivity), no switching or controllable
routing (a broader definition of passivity), and no environmental (e.g., temperature)
control, because all of these activities require external power sources. Thus, PONs are
unamplified static multipoint networks. The reason for this is again cost. Deployment
in an inhospitable physical environment demands rugged components not dependent on
any kind of power source.
A typical PON is shown in Figure 5.42, where the optical line terminal (OLT) is
at the edge of the core/metro network and the optical network units (ONUs) are at
the interface between the PON and the end systems; i.e., the (electronic) end user
or network equipment. An end system might be a LAN serving a small business, an
individual user (as in FTTH), or a wireless base station in an optical-wireless access
network. A fixed optical path connects the OLT over a directed broadcast tree, via
splitting/combining nodes, to N ONUs. (In the early PON architectures N was usually
limited to 32–64 ONUs because of splitting/combining losses, but that limitation is now
being overcome.) Over the same physical topology, a many-to-one optical path provides
connectivity from the ONUs back to the OLT. In most currently deployed PONs, a single
wavelength (λ-channel) is used on a downstream multicast optical connection (OC) from
the OLT to the ONUs, and similarly all ONUs share a single λ-channel upstream to the
OLT. Thus, this structure comprises a pair of superimposed 1 × N and N × 1 directed
trees. Generally the distances from the OLT to the ONUs are limited to 20–25 km (but
 Static Multipoint Networks 411

the distance barrier is easing), and the PON may be one component of a more diverse
access network ultimately reaching the end users. This might be composed of LANs,
DSL, wireless links, and perhaps other PONs.
The network described so far is essentially the same as the generic access network
shown shaded in Figure 5.1(b), except that the OLT in Figure 5.42 replaces the star
coupler in Figure 5.1(b). The network can be implemented using either two-fiber or sin-
gle fiber links. In the two-fiber case, there are two separate directed trees, downstream
and upstream, with each splitter/combiner node realized as a splitter node and a com-
biner node as shown in Figure 5.1(b). A more economical realization uses bidirectional
transmission on single fibers joining the splitting/combining nodes, the OLT, and the
ONUs, usually with different wavelengths used in each direction. In the bidirectional
case, each node operates as a splitter in one direction and a combiner in the other. In
either case, the transmission properties of the splitter/combiners define two separate
counter-propagating optical paths with no connectivity among the ONUs, so this is
client-server (ONU-OLT) rather than peer-to-peer architecture.38 It is important to note
the difference between this and the peer-to-peer broadcast star architecture that has been
used in most places throughout this book to describe shared-channel systems in static
multipoint networks. (See Section 3.2 and the preceding sections of this chapter.) An
extension of the directed tree architecture to include limited optical connectivity among
the ONUs will be discussed in Section 5.7.2 below.
The OLT transmits multiplexed data on the downstream tree to the ONUs. Based
on addressing information embedded in the bit stream, each ONU selects information
destined for it and ignores the rest. Thus in the downstream direction this is a typi-
cal broadcast-and-select network. All ONUs transmit “upstream,” sharing the optical
medium on the many-to-one optical path back to the OLT. With only one wavelength
used in each direction, this constitutes a TDM/TDMA system. Clearly, some sort of
fixed or demand-assigned scheduling is necessary to resolve contention for the shared
upstream medium, and the MAC protocols for implementing scheduling are among the
main features that differentiate the various PON architectures.
The OLT will normally reside in a telecommunications carrier’s Central Office and will
forward traffic from one or more service providers (e.g., ISPs, voice and video services)
to the ONUs. In a single-wavelength system, the downstream aggregated traffic from
the OLT is transmitted on a one-to-many optical connection on a wavelength λd , which
carries multiplexed point-to-point logical connections to each ONU. In cases where the
same information is destined for more than one ONU the logical connections might be
point-to-multipoint. In the upstream direction, all optical and logical connections are
point-to-point from an ONU to the OLT, usually on a common wavelength λu . Generally,
the downstream traffic is carried in the 1.5-µm band with upstream traffic in the 1.3-µm
band.
Although the details of scheduling and multiplexing the downstream and upstream
logical connections vary from one PON architecture to another, they are all examples
of reservation-type packet switching in the optical layer, as described in Section 5.6.3.
38
Complete connectivity would be possible if the splitter/combiners were replaced by star couplers, but
this is not usually wanted, not the least of the reasons being that the network operator loses control over
management of the access network.
412 Multiwavelength Optical Networks

The features of the various PON architectures were developed with a number of criteria
in mind, including high capacity utilization, low delay, and fairness. In addition service
transparency has been an important issue, meaning that the access network should be
able to adapt well to the higher layer transport services it will support. As will be seen,
each proposed PON architecture deals with these criteria in its own way.

5.7.1 ATM and Fixed-Frame PONs

Commercial deployment of PONs has been accelerated by standardization activities over
the past few years. An ATM-based PON known as APON was proposed in the 1990s and
evolved to the Broadband PON (BPON), which has been embodied in a series of ITU-T
standards since 1998 [ITU-T05]. These were developed by the Full Services Access
Network (FSAN) initiative, made up of an international group of telecommunications
carriers and equipment suppliers.
BPON is based on ATM cells, which are embedded in a fixed-frame-length format as
shown in Figure 5.43. The frames are 125 µs long and are transmitted periodically. In a
symmetric 155.52 Mbps version,39 the downstream frame contains a payload consisting
of 54 user cells destined for individual ONUs, together with two Physical Layer Op-
erations Administration and Maintenance (PLOAM) cells. The ONUs recognize cells
destined for them based on the standard ATM addressing information in the cell headers
(see Section 2.6.2). The PLOAM cells transmitted from the OLT contain grants that are
in effect reservations for individual ONUs to transmit cells in slots assigned in the next
upstream frame. The ONUs queue cells awaiting transmission, and the slot assignments
are made by the OLT based on queue length information transmitted back to it from the
ONUs. This information can be used for priority-based dynamic bandwidth allocation
based on QoS requirements.

56 cells

PLOAM ATM ATM PLOAM ATM ATM

Downstream
cell 1 cell 1 cell 27 cell 2 cell 28 cell 54

53 bytes
53 cells

Upstream ATM ATM ATM ATM

cell 1 cell 2 cell 3 cell 53

3-byte overhead
56 bytes (Guard time, preamble, and delimiter)

PLOAM: Physical layer operations administration and maintenance

Figure 5.43 BPON frame. (From [Ueda+01, Figure 2]. Copyright

c 2001 IEEE. Used by permis-
sion of The Institute of Electrical and Electronics Engineers, Inc.)

39
Options are standardized for asymmetric flows, higher speeds, and additional wavelengths.
 Static Multipoint Networks 413

Distance from OLT

ONU11
ONU1

ONU22
ONU2

ONU33 Upstream frame

ONU3

OLT
ONU33 ONU22 ONU11 ...
t

Figure 5.44 Transmission scenario in a BPON system.

The upstream frame contains only 53 cells with the space left over in the frame (3 bytes
per cell) used for guard times and physical layer overhead (including a preamble for
synchronization) in front of each cell. This provides sufficient time for the OLT receiver
to resynchronize to each transmission burst from an ONU. A timing diagram for the
frame exchange is shown in Figure 5.44, where ONU1 , ONU2 , and ONU3 are shown
transmitting bursts that fall into their assigned time slots in an upstream frame, and that
frame is shown as it arrives to the OLT. All ONU transmissions must be synchronized
to arrive in their assigned slots when the upstream frame is fully assembled at node 1 in
Figure 5.42. Because propagation delays from each ONU to that node differ due to the
different distances to the ONUs, these must be taken into account when transmissions
are initiated. This is shown in the figure, where the start of transmission from each ONU
is adjusted to compensate for its distance from the OLT. All TDMA systems have this
ranging problem. In this case it is resolved by the transmission of a ranging grant from
the OLT to an ONU, followed by the return transmission of a ranging cell from the ONU,
allowing the OLT to determine the round-trip delay between the two. That value, relayed
to the ONU, provides sufficient information for burst synchronization.
Because of the broadcast of multiplexed data streams from the OLT to all ONUs, there
is a risk of users eavesdropping on other users’ information in PONs. This problem has
been addressed in the BPON standard by using an encryption technique for ATM cells
known as churning. Although this offers some limited protection, its 8-bit key is not long
enough to provide a reasonable level of security [Thomas+02]. But the security issue
in broadcast networks is not easily solved simply by using a better encryption standard.
Part of the problem is that if encryption is done in the higher layers of the protocol
stack, only the payload information is encrypted, leaving addressing information in the
physical and data link layers in plain text and subject to manipulation by malicious users.
Conversely, encryption at the lower layers, including packet headers presents difficulties
in transmission and control. Thus security remains an open question in broadcast PONs.
The structure of the APON and BPON MAC protocols evolved from antecedents in
legacy systems: the 125-µs cell-based frame was chosen to adapt to SONET/SDH and
414 Multiwavelength Optical Networks

ATM transport in the rest of the network. However, as IP traffic increases in the core
network and Ethernet becomes the dominant format in systems accessing the core, a
need is developing for more protocol flexibility as well as higher speeds. An updated
higher speed version, Gigabit PON (GPON), addresses these needs by providing variable
length time slots and speeds up to 2.5 Gbps. It represents a continuation of the activities
of the FSAN initiative. (See [ITU-T03a, ITU-T03b, ITU-T04].) GPON has options
that allow it to adapt to a range of transport services from SONET/SDH and ATM to
Ethernet. It keeps the downstream fixed-frame-length format to adapt to SONET/SDH
but uses variable length slot allocations to accommodate other types of traffic including
(segmented) Ethernet frames.

5.7.2 Ethernet-Based PONs

In parallel with the development of GPON a different group working within the IEEE,
the Ethernet in the First Mile (EFM) Task Force, has developed an Ethernet-based PON
(EPON) (IEEE standard 802.3ah). It normally uses downstream and upstream transmis-
sion at 1550 and 1310 nm, respectively. Reflecting the seemingly eternal differences in
perspective between the telecommunications carriers and the computer networking/IP
community, EPON focuses on the prevalence of Ethernet at the end user side of the ac-
cess network rather than SONET/SDH and ATM transport on the carrier side. Using the
same underlying optical layer as in the other PONs, the OLT in EPON broadcasts a mul-
tiplexed stream of Ethernet frames to the ONUs, which recognize them through logical
link identifiers (LLIDs) assigned to them when they “register” in the EPON. Registration
and all other scheduling features of EPON are handled by a Multi-Point Control Protocol
(MPCP). An ONU registers with the network through a discovery process implemented
in the MPCP wherein the OLT broadcasts a discovery message inviting any new (either
just-connected or not yet initialized) ONU to register by sending a reply to the OLT,
containing its MAC address. Not only does this tell the OLT the identity of the ONU but
it enables the OLT to determine the round-trip delay to that ONU, thereby solving the
ranging problem.40 Note that the CSMA/CD MAC protocol described in Section 3.2.2,
which is the basic feature of wired Ethernets, is not applicable here, because the differ-
ent ONUs do not “see” each others’ transmissions. Thus, although EPON is structured
around other features of Ethernet (frame format, addressing, etc.), it requires a reserva-
tion procedure to replace CSMA/CD in resolving contention for the upstream channel.
This is accomplished through an exchange of requests for transmission windows (TWs)
from the ONUs, followed by TW allocations from the OLT. The allocations must take
into account the propagation time to the ONU as well as the length of the requested TW,
which may accommodate a burst of several Ethernet frames. This interplay of requests
and allocations (called, respectively, REPORTs and GATEs in MPCP) provides a frame-
work for participation of both the OLT and the ONUs in dynamic bandwidth allocation
and in controlling frame delays based on priority classes and QoS requirements.

40
Conceivably, more than one waiting ONU might transmit during this procedure, leading to a collision. Such
collisions are resolved in the usual manner by retransmissions during subsequent discovery opportunities.
 Static Multipoint Networks 415

There are many ways of implementing media access control in this type of shared-
channel system. Historically, one of the earliest MAC protocols was polling, which was
used in the first “multidrop” wired networks linking sets of terminals to a central com-
puter or communication controller [Schwartz87]. In roll-call polling, a central controller
(the OLT in our systems) sends a polling message downstream to each user (our ONU)
inviting that user to transmit, possibly with some restrictions on how much data it is
allowed to transmit. After completion of the upstream transmission, the OLT repeats
this two-way transaction with the next user, continuing this procedure in an order deter-
mined by its “polling table.” (The OLT may be simultaneously or intermittently sending
downstream data to the ONUs, depending on whether this is a full duplex or half duplex
system.) Bandwidth can be allocated as needed by polling some users more than others.
This principle is embodied in an EPON protocol called IPACT [Kramer+02], using the
request and allocation framework described above. To provide dynamic bandwidth allo-
cation, when the ONU is polled it is given a grant for a given TW, based on the ONU’s
queue state information, which is stored in the OLT’s polling table. Because propagation
delays in optical systems may be long compared to burst transmission times from the
ONUs, it is important to “pipeline” the grants without waiting for responses from each
ONU. Based on knowledge of round-trip transmission times to each ONU, grants for TWs
are issued so that bursts are interleaved efficiently on the upstream channel. The queue
state information that each ONU provides to the OLT is piggybacked onto its upstream
transmissions, and the grants to the ONUs may be piggybacked on data transmissions
from the OLT. Because this is a master-slave relationship, with the OLT controlling the
shared medium, it has the power to implement dynamic bandwidth allocation to meet
various criteria, including fairness and prior service level agreements (SLAs).
Decentralized control is another way of organizing medium access in PONs. The
idea is for the ONUs to decide among themselves how to schedule their upstream
transmissions, implementing a distributed scheduling algorithm based on their own QoS
requirements and queue states. All of this can be done independent of the OLT, which
is concerned only with transmitting downstream data. However, decentralized control
is difficult to implement because the ONUs cannot exchange control information in
the system of Figure 5.42. This is remedied in [Sherif+04] as shown in Figure 5.45.
The proposed scheme relies on redirecting a portion of the aggregated upstream signals
from all ONUs and broadcasting it back to these ONUs. The figure shows the same
sort of splitting/combining node as node 1 in Figure 5.42, but instead of a 1 × N
node it is 3 × N , with all ports supporting bidirectional transmission on single fibers.
(As in Figure 5.42, the ONUs could be located on a tree instead of a star.) A portion
of the upstream signal (at 1310 nm) appears on one of the additional ports and is
redirected to the second additional port via an isolator, which ensures unidirectional
transmission back to the ONUs. In addition to a receiver for the downstream traffic at
1550 nm, each ONU now has a second receiver tuned to 1310 nm to pick up control
information carried as overhead on the upstream signal. Other more recent approaches
have proposed a ring-based connection of ONUs that can provide for distributed dynamic
bandwidth allocation, as well as inter-LAN connectivity and enhanced survivability
features [Hossain+06, Hossain+07].
416 Multiwavelength Optical Networks

Control Plane:
1310 nm channel
Data Plane:
Upstream: 1310 nm channel ONU
Redirected
Downstream: 1550 nm channel
1310 nm signal

1550 nm

OLT Splitter
Combiner
1310 nm
ONU
3xN S/C

ONU

ONU

Figure 5.45 PON equipped for decentralized control. (From [Sherif+04, Figure 1]. Copyright
c
2004 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

The proposed decentralized control mechanism is based on a cycle consisting of

a control update period followed by a data transmission period. During the control
period each ONU sends a control message that is broadcast to all ONUs (and the OLT,
which may or may not ignore it). To avoid collisions, this is done using a fixed-frame
TDMA schedule with one slot for each ONU. These control messages are used as
the input to a distributed control algorithm executed simultaneously by all ONUs to
produce a schedule for data frame transmissions in the next data transmission period.
Because all ONUs have executed the same algorithm they know the schedule and the
data transmission can commence without any further exchange of information. This
technique allows for the support of differentiated services via priority queueing while
at the same time folding bandwidth management and fair scheduling into the mix.
Performance results in [Sherif+04] show its superiority over centralized methods in
satisfying QoS requirements.

5.7.3 WDM PONs

None of the PONs described thus far uses more than one wavelength in each direc-
tion, although there are cases where extra downstream wavelengths can be overlaid for
additional services such as video distribution. However, a true multiwavelength PON
 Static Multipoint Networks 417

Headend (HE) or
Central Office (CO)

MFL
10 Km

1.3/1.5 Remote Node (RN)

WDM
ONU
DEMUX

SCM or TDM
REC
LED
1.3/1.5
WDM
REC
1 x N waveguide grating
router (WGR)

Figure 5.46 LARNet. (From [Zirngibl+95, Figure 1]. Copyright

c 1995 IEEE. Used by permission
of The Institute of Electrical and Electronics Engineers, Inc.)

has a number of advantages, especially when it operates with dedicated λ-channels for
each ONU. These include better utilization of the fiber bandwidth, elimination of split-
ting/combining losses (in some cases), tailoring channel speeds and formats to individual
users’ needs, and ensuring privacy for downstream traffic. Several proposed WDM ar-
chitectures are based on 1 × N directed star topologies with a 1 × N wavelength router
at the star node.
Two of the first proposals for WDM-based PONs were LARNet [Zirngibl+95] and
RITE-Net [Frigo+94]. Each of these was based on AWG routing and used some in-
teresting technologies for cost savings. In LARNet, a unique wavelength is dedicated
to each ONU for downstream traffic, and the different wavelengths λ1 , λ2 , . . . , λ M in
the downstream 1.5-µm band are distributed to the ONUs by an AWG router in the
routing node (RN) (see Figure 5.46). Using bidirectional transmission on single fibers,
each ONU transmits its upstream signal back to a receiver in the OLT using the 1.3-µm
band. A set of wavelengths λ′1 , λ′2 , . . . , λ′M must be used to route the signals from each
ONU through the RN back to the OLT. A clever arrangement is used to achieve this
economically. LEDs are used in each ONU for upstream transmission. Because the
broad spectrum of the LED covers the total spectrum used for upstream transmission, a
“slice” of this spectrum is automatically selected through the filtering properties of the
RN and transmitted back to the OLT. (The portion of the power spectrum transmitted
by the LED that is not guided to the upstream port is dumped by the RN, resulting in
a power loss determined by the finesse of the AWG – about 13 dB in the experiments
reported in [Zirngibl+95].) The properties of the AWG dictate that the LED slice for
each ONU is on a different wavelength, so that a wavelength multiplexed signal with
distinct wavelengths from all ONUs reaches the receiver of the OLT. An additional
benefit of the spectrally sliced LED is that drifts in AWG filtering characteristics due
to temperature changes do not effect upstream transmission. However, LED signals do
not have the coherence properties of laser signals, which limits the transmissions to
relatively low-bit-rate-distance products.
418 Multiwavelength Optical Networks

Optical Network
Unit

Remote Node λ1

Central Office λ2 λ1
R
T MOD
Router λ2

R R
λ1 λ2
MOD

ONU

Figure 5.47 RITE-Net. (From [Frigo+94, Figure 1]. Copyright
1994

c IEEE. Used by permission
of The Institute of Electrical and Electronics Engineers, Inc.)

On the downstream side a single multiwavelength laser source is used, where each
wavelength must be externally modulated for downstream data transmission, and the
wavelengths must track drifts in the RN characteristics. A specially designed semi-
conductor laser integrated with an AWG is used as a multifrequency laser source (MFL)
transmitting on a grid matched to the RN. The MFL can be temperature controlled to
translate the complete wavelength grid to track any drifts in the RN characteristics.
Several possible upstream multiplexing techniques are proposed in LARNet. In one
configuration, an array of N tuned receivers allows for full utilization of an upstream
wavelength channel by each ONU. A less costly approach uses SCMA for modulating
the data onto the upstream channels (see Section 5.2.2.1). A different subcarrier is used
by each ONU, and a single broadband optical receiver converts the combined optical
signals to a subcarrier multiplexed electrical signal, which is demultiplexed to extract the
individual upstream channels. A third approach uses TDMA, with the usual requirements
for burst synchronization on the upstream link.
The RITE-Net architecture is similar to LARNet but uses unidirectional fiber pairs,
as shown in Figure 5.47, which means that the number of ports on the AWG must be
doubled. The original design uses a single rapidly tunable laser source for downstream
transmissions on N wavelengths assigned to the ONUs. Because this is a TT-FR tech-
nique using only a single transmitter, the downstream bit rate must accommodate the
aggregate traffic to all ONUs, as opposed to the LARNet approach with N dedicated
downstream channels. The cost-saving innovation in RITE-Net is the use of the same
signal for both downstream and upstream traffic, which entails the replacement of the
ONU transmitter by a modulator. However, this also requires unidirectional fiber pair
transmission to avoid interference between the two counter-propagating channels on
the same wavelength. As shown in the figure, the incoming downstream signal is split
between the ONU receiver and a modulator that “writes” the upstream data onto the
incoming signal and sends it back to the RN. The RN is designed to direct all upstream
wavelengths back to a single fiber carrying it to a single broadband receiver in the
OLT. Various upstream modulation formats are possible, but in all cases the downstream
optical carrier must have unused capacity available for carrying the upstream data. For
 Static Multipoint Networks 419

ONU1

λ1 λu ONU2

OLT λ2 λu
λ1, λ2, ... λN
OT
OT
1.3/1.5 µm Router/ λ3
WMUX combiner ONU3
OR
OR
λu
λu

λN
λυ

ONUN
ONUN

Figure 5.48 WDM PON.

example, this could be in the form of unused time intervals if TDMA is the upstream
data format, or it could be a portion of the modulating frequency spectrum for SCMA.
The experiments reported in [Frigo+94] used SCMA, which avoided the problems of
scheduling upstream transmissions to prevent collisions in a TDMA system. (If an ar-
ray of N tuned receivers is used at the OLT the scheduling problem is avoided, the
assignment of SCMA carriers is simplified, and more upstream capacity is available.)
In another WDM PON [Giles+96], costs are reduced by using single bidirectional
fibers together with a star node that combines wavelength routing for downstream
traffic and power combining for upstream traffic. As shown in Figure 5.48, downstream
traffic from the OLT is transmitted by an MFL, which generates signals at wavelengths
λ1 , λ2 , . . . , λ N in the 1550-nm band. These wavelengths are routed to individual ONUs
through the 1 × N star node, which acts as a wavelength-selective router for downstream
traffic at 1550 nm and a combiner for upstream traffic in the 1300-nm band. The star
node consists of a 1 × N splitter with fiber gratings in each of its N downstream fibers.
The gratings in the kth downstream fiber reflect all downstream wavelengths except λk ,
so ONUk sees that wavelength only. The upstream traffic is generated at each ONU by
uncooled FP lasers on a nominal wavelength λu in the 1300-nm band. These enter the
N downstream ports of the star node without any interaction with the fiber gratings and
are combined onto the upstream port and delivered to the receiver in the OLT. SCMA is
used for the upstream signals, which allows for demodulation and demultiplexing of the
combined upstream optical signal using a single receiver of the form of Figure 5.7(b).
A coarse WMUX/DMUX in the OLT directs the upstream traffic from the star node to
the OR and the downstream traffic from the OT to the star node.
Since these early experiments, the basic wavelength-routed architecture has been
demonstrated in a number of applications. In [Iannone+98], a variety of services were
combined on a WDM PON. These included conventional switched voice, data and video
420 Multiwavelength Optical Networks

services using TDM/TDMA at 50 Mbps in the 1.3-µm band, broadcast video in the
1.5-µm band, and high-speed (2.5 Gbps) service at 1.5-µm. A novel feature of this work
is that LEDs and spectrum slicing are used for downstream transmission of broadcast
video in two bands separated by 11 nm, which is the FSR of the AWG router. In this way
the router delivers the video to each ONU on two separate slices of the LED spectrum
determined by its port on the AWG.
In other experiments, the use of spectrally sliced LEDs was demonstrated in still more
applications. In a 15-ONU system using bidirectional transmission on single fibers,
broadcast video was combined with wavelength division-multiplexed data channels
operating at up to 2.5 Gbps downstream using directly modulated DFB laser transmitters.
The upstream channel used LEDs operating at 155 Mbps. The overlaid arrangement of
broadcast video and wavelength-routed data required a more complex routing node
consisting of a 16 × 16 AWG and a 1 × 16 splitter as described in [Son+03]. Another
system serving 40 ONUs was demonstrated with 622 Mbps downstream and 155 Mbps
upstream using LEDs in both directions [Han+04]. To achieve these results, electrical
precompensation on the downstream side together with FEC on the upstream side were
required.
The idea of the wavelength-routed PON can be extended in various ways. For example,
if additional ports are provided on the OLT side of the router then additional OLTs can be
added, which is a way of scaling the network upward in capacity and connectivity. With
M OLTs serving N ONUs via an RN, we have an M × N generalization of the 1 × N
case in Figure 5.46. Through combinations of time and wavelength division techniques
in each direction, there are many possibilities for connectivity between the OLTs and
the ONUs. For example, an ONU might have simultaneous bidirectional connections
with two OLTs on two different wavelengths or using two different sets of time slots
on the same wavelength. (Figure 5.1[b] illustrated this approach where a broadcast star
links a population of end users to two different networks.) An M × N arrangement of
this type might be used as the infrastructure of an open access network, as suggested
in [Banerjee+05], wherein different service providers operate through their own OLTs,
sharing the capacity of the underlying access network to provide a mix of services such as
voice over IP, video-on-demand, and Internet access to end users connected to the ONUs.

5.7.4 Optical-Wireless Access

With the convergence of wireless and optical networking, passive access networks in-
volving a merging of photonic and wireless technology – radio-over-fiber (ROF) – have
become an important area of activity. Typical applications are for fixed or mobile users
requiring broadband access for services such as HDTV and interactive multimedia. In
these applications, the ONUs are base stations (BSs) and the OLT is a central station
(CS) distributing signals over fiber to the BSs, where antennas are normally located.
The fiber distribution links might be single bidirectional fibers or fiber pairs. The special
feature of ROF systems is that signals are delivered to and received from the BSs in
the form of an information stream modulated onto a microwave or millimeter wave
electrical carrier. Given the large number of base stations being served, ROF is the most
efficient architecture for the fiber access network, because it delivers the radio frequency
 Static Multipoint Networks 421

Figure 5.49 Integrated system for dual services. (From [Jia+06b, Figure 1]. Copyright

c 2006
IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

signal directly to and from the antennas without requiring costly equipment in the base
stations. A number of different approaches have been proposed for ROF [Imai+06,
Jia+06a, Kuri+02, Yu+06]. We illustrate using a combination of a WDM PON with
ROF to produce a wired/wireless access network.
Figure 5.49 shows a network for delivering broadband services to the customer’s
premises on fiber or by wireless [Jia+06b]. A dedicated wavelength is used to supply
replicas of the information signal at baseband and on a millimeter wave subcarrier. (In
this case the antennas and wireless portion of the network are completely contained
in the customer’s premises rather than the base station.) Experiments were conducted
delivering 270 Mbps real-time uncompressed video together with 2.5-Gbps data to the
end user’s premises for access either directly from the fiber or from a short range (10.2 m)
wireless connection. As shown in the experiment of Figure 5.50, a dedicated wavelength
carries the baseband and subcarrier version of the signal to the customers’ premises,
where it is split with one part going to a high bandwidth photodetector (60 GHz PIN),

Figure 5.50 Dual services testbed. (From [Jia+06b, Figure 2]. Copyright

c 2006 IEEE. Used by
permission of The Institute of Electrical and Electronics Engineers, Inc.)
422 Multiwavelength Optical Networks

which converts it to a millimeter wave radio signal for broadcast at the destination.
(Optical and electrical amplification are included in the wireless path.) The other part
goes to a low bandwidth photodetector (APD) that eliminates the ROF part and detects
the baseband signal for delivery to the end user. The transmitter at the CS generates
a signal on each wavelength channel using the optical suppressed-carrier techniques
described in Section 4.7.1 [Chowdhury+06]. The wavelengths are multiplexed at the
CS for transmission on the access network and then demultiplexed at BSs for distribution
to the customers’ premises. The testbed showed that this approach could be successfully
implemented in an integrated access platform at the CS to provide simultaneous services
to fixed and mobile users.

5.7.5 Recent Trends

Over the past decade, the PON has evolved to the point where it is recognized as a
key building block of the optical network hierarchy. Most of the originally proposed
architectures described above remain the basis for current deployments and further
research. Because the technology has matured, the issues in this area have now shifted
to economics: extending the reach of PONs, increasing the number of customers being
served, increasing the user bit rates, and doing all of this while reducing the cost of the
hardware, especially in the ONUs.
The focus of recent attempts to achieve these results has been on a few main avenues
of development: removing the light source from the ONU and replacing it by modulation
or remodulation of a signal supplied by the OLT, using inexpensive light sources in both
the OLT and ONUs, and inserting fiber amplification in the overall system to extend
reach and increase the number of ONUs. Most of the important recent developments
have been on WDM PONs. We consider a few illustrative examples now.
The network of Figure 5.51 portrays an experimental demonstration of a method
of extending the size (in reach and in users) and reducing the cost of a PON using a
combination of DWDM and TDM [MacHale+06]. In many ways it is an extrapolation
of the basic configuration of LARNet (Figure 5.46), and it is interesting to compare the
two. The DWDM/TDM experiment simulates a network connecting a core exchange
(CE) through a local exchange (LE) to multiple PONs. This is a typical scenario for
recent PON proposals: a purely optical connection from the core to the ONUs, with a
relatively long “backhaul” from the CE to the LE, for a total of 116 km from core to
ONU. (The data paths are passive except for the EDFA located in the LE on the backhaul
link.) Each PON is associated with a downstream/upstream pair of wavelength channels
(all in the 1.5-µm band). The “blue” end of the spectrum is used downstream and the
“red” end upstream. Red/blue filters (RBFs) are used at various points to separate the
downstream and upstream spectra. Data are carried from the core using TDM on each
downstream wavelength channel to distribute information to all users on a given PON,
and TDMA is used upstream to share each PON’s wavelength among all its ONUs. The
data path between the core and the ONUs consists of a single bidirectional fiber, where
each PON has a bidirectional data path consisting of a feeder fiber, distribution fiber, and
drop fiber, with three splitting points along the way allowing for a total of 16N ONUs.
 Static Multipoint Networks 423

Backhaul fiber

Downstream Tx AWG Distribution Fiber Customer ONU

AWG DCF EDFA
circulator RBF
DFB +
EAM-SOA 1xN 1x8 Drop
Feeder Fiber split split Fiber
RBF
100 Km 2x2 SOA
SMF split 4.1 Km 0.4 Km
EDFA
SMF SMF
12.1 Km SMF SOA EAM
Rx SOA EAM
EDFA
Ch 24 DFB
25Km
AWG 1st TDMA-PON SMF
Ch 40 DFB
Core th
17 TDMA-PON Attenuator
CW Centralized
Exchange Rx
source
Local ~5 nm guard band
Exchange Downstream Upstream
channels channels

Ch1 Ch 17 Ch 24 Ch 40
1529 nm 1541.6 nm 1547.2 nm 1580.1 nm

Figure 5.51 DWDM/TDM PON. (From [McHale+06, Figure 1]. Copyright
c 2006 IEEE. Used
by permission of The Institute of Electrical and Electronics Engineers, Inc.)

A separate fiber feeds a carrier from a CW source in the LE to all ONUs in a specific
PON, and the ONUs use it to carry their upstream data. The ONU uses an SOA and
electro-absorption modulator (EAM) to impress the upstream data on the carrier.
The basic elements of LARNet are present here: bidirectional single-fiber data trans-
mission and WDM using AWGs. The add-ons are TDM/TDMA to accommodate mul-
tiple users on each wavelength, and a “reflective” ONU that uses a carrier supplied
from the CE to replace a local source. Centralized carrier distribution means that the
ONU can be “colorless”; that is, it has no wavelength-specific components in it, greatly
reducing its cost. However, the presence of the same wavelength in opposite directions
on the fibers in the PON makes data transmission vulnerable to Rayleigh back-scattering
(RBS). In RBS scattered light from the downstream signals interferes with the upstream
data.41 (The lower-right branch of the 2 × 2 splitter in Figure 5.51 simulates the RBS
contributed by the other paths in the PON.) Introducing the CW carriers into each PON
on separate fibers tends to reduce the RBS effect. The experiments demonstrated that an
aggregate bit rate per PON of 10 Gbps can be supported with 64 ONUs per PON (cor-
responding to N = 4), giving a system accommodating 1088 users on 34 wavelengths,
at 155 Mbps per user.
The objective of a colorless, reflecting ONU is achieved in different ways in a number
of other proposed networks. In [Payoux+06a], the same carrier is used for downstream
and upstream data in a network similar to that in [MacHale+06]. The upstream data
are impressed on the signal carrying downstream data by superimposing it at a higher
modulation index. Also, a reflective SOA (R-SOA) [Healey+01] is used as a modulator,
replacing the SOA-EAM combination. Other works also use a similar approach. For
example, [Arellano+06] superimposes upstream/downstream data on the same carrier

41
The effect of double RBS in Raman amplifiers is discussed in Section 4.4.2.
424 Multiwavelength Optical Networks

using various combinations of ASK, FSK, and SCM, employing R-SOAs for upstream
modulators, and using forward error correction to improve BER.
Other technologies have also been investigated at the OLT and ONU ends of WDM
PONs using colorless source free ONUs. In [Lee+06], ASE noise from an EDFA is
passed through a highly nonlinear fiber to produce a high-power supercontinuum source
at the OLT (see Section 4.5.1). The output of this source is transmitted downstream and
sliced by an AWG to create CW carriers for the ONUs. These are used as injection sources
for Fabry–Perot laser diode modulators that impress data on the upstream channels. In
[Zhang+06], a supercontinuum source produced by a mode-locked laser working into an
EDFA and a highly nonlinear photonic crystal fiber is used to generate the carriers that
are used in the ONUs. Finally, in [Payoux+06b], a sliced LED source in the OLT is used
to provide carriers for both the downstream and upstream data. The modulators at both
ends are R-SOAs and a red/blue wavelength allocation plan is used as in [MacHale+06,
Payoux+06a]. The proliferation of recent articles on this subject is a sure sign that the
last mile is the object of considerable current commercial interest.

5.8 Summary

This chapter provided an overall picture of how static multipoint networks are struc-
tured, what their constraints are, when they are useful, and when they are not. The
prototypical physical example used throughout was the broadcast star, operated as a
shared communication medium.
The basic constraints that must be considered when designing a static network are the
station functionality (at both the optical and electronic levels) and the available optical
spectrum. Within those constraints, we demonstrated how one can design and optimize
a network to satisfy given traffic requirements, taking into account various alternatives
for trading off cost versus performance. The main objective was to obtain maximum
connectivity and throughput at minimum cost and complexity.
It was seen that the highest efficiency is achieved when the design problem is deter-
ministic; that is, when dedicated connections are implemented for synchronous traffic.
Randomness generally reduces performance and increases cost: More sophisticated con-
trol is required and more resources (bandwidth in the fibers and processing power in the
stations) are needed to provide a given quality of service. One way of mitigating effects
of random fluctuations is by increasing system size and by aggregating traffic whenever
possible. This effect appeared several times in this chapter. In packet-switched systems,
packet loss is reduced by increasing buffer size, and vulnerability to traffic fluctuations is
reduced by dynamic capacity allocation. In systems with demand-assigned connections,
blocking probability at a given traffic intensity is reduced by scaling up the system.
Randomness is always present in the real world because of imprecision in predicting
demand, statistical fluctuations of demand-assigned connection requests, or burstiness
of packet traffic. Thus, the mathematical models introduced here to deal with random
traffic are particularly important in a wide range of network applications.
 Static Multipoint Networks 425

The most important current application of static networks is in the access network. The
broadcast star and its variants form the foundation for a number of passive access network
architectures including WDM PONs. Static networks are also used to support radio-over-
fiber for combined optical and wireless services. A few examples of recent developments
in this area were discussed in this chapter. Finally, as described in Chapter 3, larger
networks based on the LLN concept can be constructed using embedded broadcast
stars as subnets. For this reason, the generic techniques discussed in this chapter for
controlled sharing of the resources of a static multipoint network provide a foundation
for the design of more general WANs, which will be treated in the remainder of this
book.

5.9 Problems

1 A three-node directed star is configured with optical local access subnets. One node
consists of a concentration subnet for “servers” (e.g., video jukeboxes), and the other
two are distribution subnets for end systems.
(a) Show the fiber arrangements for these subnets.
(b) Because the transmission through the subnets is unidirectional from servers to end
systems, there will be a problem in network control (i.e., in implementing and
controlling requests from the end systems for server-to-end system connections).
Suggest a reasonable network control and management arrangement, including pro-
visions for signaling.
2 It is indicated in Section 3.2.1 that TDM/TDMA or TDM/T-WDMA requires guard
times between successive bursts from different transmitters.
(a) Explain the operations the receivers must perform during guard times and during
the initial portions of each transmitted burst.
(b) How is receiver synchronization affected when different transmitters use different
bit rates?
(c) How is NAS operation simplified when it has an array of ORs?
3 In the case of the normalized traffic matrix of Figure 5.3(b), why is it that no TT-FR
schedule can produce a frame length shorter than seven and no FT-TR schedule can
produce a frame length shorter than eight?
4 Revise the TDM/T-WDMA development of Section 5.2.1 for the case in which
different transmitters use different bit rates Rti . Apply this to the case of the 3 × 3
system of Figure 5.3(b), using TT-TR with two channels available. Assume that the
bit rate Rt1 for transmitter 1 is twice as fast as for the other two transmitters. Find the
required bit rates for the system as well as a feasible CAS.
5 Compare the SCMA receiver of Figure 5.7(b) with the coherent optical receiver of
Figure 2.33. Point out the similarities, differences, and technological problems of each.
Comment on relative costs.
426 Multiwavelength Optical Networks

6 For the case M = 50, determine the error incurred in neglecting the nonadjacent
channel OBI terms to give the approximation of CIR in Equation (5.21).
7 Derive Equation (5.22).
8 Redo the example at the end of Section 5.2.2.2 for M = 10 and 20 using a modu-
lation index of m = 1 and assuming that the Lorentzian shape of the OBI spectrum is
maintained with a width broadened to a = 200 MHz.
9 Give an example of how a packet-switched CDMA system could be implemented
at the subcarrier level, along the lines of Section 5.2.2.4. Compare this to an IM/DD
CDMA system without subcarriers. How does the subcarrier packet-switched approach
change the type of codes that are applicable, the effect of OBI, and the resultant overall
system performance? For example, how might the correlation properties of the codes
differ in the two cases? (Hint: In a subcarrier CDMA system based on pseudorandom
pulse sequences, the coefficients in a chip sequence can have negative as well as positive
signs or can even be complex.)
10 For the SCM/WDMA/SCMA system of Section 5.2.2.6, devise an algorithm for
“static” subcarrier frequency assignment. Assume a general M × M system with C
available subcarrier channels with fixed and distinct optical frequency assignments at the
transmitters. Based on a list of required LCs, the algorithm should determine whether
a feasible subcarrier frequency assignment can be found and should give a possible
assignment.
11 For the traffic matrix γ shown here, plot the balance factor, η, as a function of X ,
assuming that C = 3 and each station has a single transceiver.
 
0 1 1 0
0 1 1 0
γ =
3
.
0 0 0
0 0 0 X

12 Consider the following normalized traffic matrix:

 
1 0 2 0
0 4 0 3
T=
2

0 1 0
0 1 0 1

(a) For an FT-TR system with a full complement of channels and a single transmitter or
receiver per station, find L min and an optimal CAS.
(b) Repeat the previous part when α2 = 2.
(c) Repeat the first part for the case C = 2, finding an optimal schedule for an FT-TR,
TT-FR, and TT-TR system.
(d) Assume that the traffic from station 2 consists of a point-to-point connection, [2, 2],
carrying one unit of traffic, and a multicast connection, [2, {2, 4}], carrying three
units. Redo the first part for this case.
 Static Multipoint Networks 427

13 Program the scheduling algorithm of Appendix B. Try out your program on some
test cases.
14 In the example of Section 5.4.1.3, approximate π by 22/7 and determine the mini-
mum frame length L min . How much capacity is wasted in this case?
15 In the queueing delay example of Figure 5.29, let all parameters be the same as
before except the propagation delay. Take the speed of light in the fiber to be 2 × 108 m/s.
Determine the fiber distance at which the propagation delay in the fiber equals the average
system delay D̄ when ρ = 0.5. ( D̄ includes all delays except propagation delay.)
16 In the tell-and-go example of Figure 5.39, assume that all packets in transmitting
station 4 are destined for receiving station 3.
(a) Show how Figure 5.39(b) would look in this case, indicating conflicts and lost
packets.
(b) Show how the lossless schedule of Figure 5.40 would look in this case.
(c) Show a perfect schedule for this case.

Bibliography

[Abramson73] N. Abramson. The ALOHA system. In Computer Networks. Englewood Cliffs,

NJ: Prentice Hall 1973.
[Antoniades+95] N. Antoniades, W. Xin, B. Pathak, E. Yang, and T. E. Stern. Use of subcarrier
multiplexing/multiple access for multipoint connections in all-optical networks. In Proceedings
of the SPIE Conf. on All-Optical Commun. Syst.: Architectures, Philadelphia, PA, October 1995.
[Arellano+06] C. Arellano, C. Bock, J. Prat, K.-D. Langer. RSOA-based optical network units
for WDM-PON. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper OTuC1,
Anaheim, CA, March 2006.
[Banerjee+05] A. Banerjee, G. Kramer, Y. Ye, S. Dixit, and B. Mukherjee. Advances in passive
optical networks (PONs). In Emerging Optical Network Technologies, K. M. Sivalingam and S.
Subramaniam (Editors), New York: Springer-Verlag 2005.
[BenJaafar+01] H. Ben Jaafar, S. LaRochelle, P.-Y. Cortes, and H. Fathallah. 1.25 Gbit/s trans-
mission of optical FFH-OCDMA signals over 80 km with 16 users. In Proceedings of the
IEEE/OSA Optical Fiber Commun. Conf., pp. TuV3-1–TuV3-3, vol. 2, Anaheim, CA, March
2001.
[Briley83] B. E. Briley. Introduction to Telephone Switching. Reading, MA: Addison-Wesley
1983.
[Chen+91] M. Chen and T.-S. Yum. A conflict-free protocol for optical WDMA networks. In
Proceedings of the IEEE Global Telecommunications Conference (Globecom), pp. 1276–1281,
Phoenix, AZ, December 1991.
[Chen01] L. R. Chen. Flexible fiber Bragg grating encoder/decoder for hybrid wavelength-time
optical CDMA. IEEE Photon. Technol. Lett., 13(11):1233–1235, 2001.
[Chipalkatti+92] R. Chipalkatti, Z. Zhang, and A. S. Acampora. High-speed communication
protocols for optical star networks using WDM. In Proceedings of the IEEE Conference on
Computer Communications (Infocom), Florence, Italy, May 1992.
[Chowdhury+06] A. Chowdhury, J. Yu, L. Xu, T. Wang, and G.-K. Chang. Experimental demon-
stration of 40 Gbps per channel transmission over 50 GHz DWDM systems for low-cost metro
428 Multiwavelength Optical Networks

and access networks. In Proceedings of the European Conference on Optical Communications,

Paper Mo 4.5.5, Cannes, France, September 2006.
[Darcie87] T. E. Darcie. Subcarrier multiplexing for multiple-access lightwave networks.
IEEE/OSA J. Lightwave Technol., 5(8):1103–1110, 1987.
[Desem90] C. Desem. Optical interference in subcarrier multiplexed systems with multiple optical
carriers. IEEE J. Select. Areas Commun., 8(7):1290–1295, 1990.
[Eggleton+94] B. J. Eggleton, P. A. Krug, L. Poladian, and F. Ouellette. Long period superstruc-
ture Bragg gratings in optical fibers. Electron. Lett., 30(19):1620–1622, 1994.
[Etemad+04] S. Etemad, T. Banwell, S. Galli, J. Jackel, R. Menendez, P. Toliver, J. Young,
P. Delfyett, C. Price, and T. Turpin. Optical-CDMA incorporating phase coding of coherent
frequency bins: concept, simulation, experiment. In Proceedings of the IEEE/OSA Optical Fiber
Commun. Conf., Paper FG5, Los Angeles, CA, March 2004.
[Fathallah+99] H. Fathallah, L. A. Rusch, and S. LaRochelle. Passive optical fast frequency-hop
CDMA communications system. IEEE/OSA J. Lightwave Technol., 17(3):397–405, 1999.
[Frigo+94] N. J. Frigo, P. P. Iannone, P. D. Magill, T. E. Darcie, M. M. Downs, B. N. Desai,
U. Koren, T. L. Koch, C. Dragone, and H. M. Presby. A wavelength-division multiplexed passive
optical network with cost-shared components. IEEE Photon. Technol. Lett., 6(11):1365–1367,
1994.
[Giles+96] C. R. Giles, R. D. Feldman, and C. R. Doerr. Access PON using downstream 1550 nm
WDM routing and upstream 1300 nm SCMA combining through a fiber-grating router. IEEE
Photon. Technol. Lett., 8(11):1549–1551, 1996.
[Gopal82] I. S. Gopal. Scheduling Algorithms for Multibeam Communications Satellites. Ph.D.
thesis. Columbia University, New York, 1982.
[Hamanaka+06] T. Hamanaka, X. Wang, N. Wada, and K. Kitayama. 511-chip SSFBG and
DFG-based optical thresholder enabled compound data rate (10 Gbps-622 Mbps) OCDMA
experiment for multiple service provisioning platform. In Proceedings of the IEEE/OSA Optical
Fiber Commun. Conf., Paper OThT4, Anaheim, CA, March 2006.
[Han+04] K. H. Han, E. S. Son, K. W. Lim, H. Y. Choi, S. P. Jung, and Y. C. Chung. Bi-directional
WDM passive optical network using spectrum-sliced light-emitting diodes. In Proceedings of
the IEEE/OSA Optical Fiber Commun. Conf., Paper MF98, Los Angeles, CA, March 2004.
[Healey+01] P. Healey, P. Townsend, C. Ford, L. Johnston, P. Townley, I. Lealman, L. Rivers,
S. Perrin, and R. Moore. Spectral slicing WDM-PON using wavelength-seeded reflective SOAs.
Electron. Lett., 37(19):1181–1182, 2001.
[Hoppitt+89] C. E. Hoppitt and D. E. A. Clarke. The provision of telephony over passive optical
networks. Br. Telecommunications Technol. J., 7(2):100–113, 1989.
[Hossain+06] ASM D. Hossain, A. Shami, C. Assi, and M. Ali. Supporting private networking
capability in EPON. In Proceedings of the IEEE Int’l Conf. Commun. (ICC), Istanbul, Turkey,
June 2006.
[Hossain+07] ASM D. Hossain, R. Dorsinville, A. Hadjiantonis, G. Ellinas, and M. Ali. A
simple self-healing ring-based local access PON architecture for supporting private networking
capability. In Proceedings of the IEEE Global Telecommunications Conference (Globecom),
Washington DC, November 2007.
[Hui90] J. Y. Hui. Switching and Traffic Theory for Integrated Broadband Networks. Norwell,
MA: Kluwer Academic 1990.
[Humblet+93] P. Humblet, R. Ramaswami, and K. N. Sivarajan. An efficient communication
protocol for high-speed packet-switched multichannel networks. IEEE J. Select. Areas Commun.
11(4):568–578, 1993.
 Static Multipoint Networks 429

[Iannone+98] P. P. Iannone, K. C. Reichmann, and N. J. Frigo. High-speed point-to-point and

multiple broadcast services delivered over a WDM passive optical network. IEEE Photon.
Technol. Lett., 10(9):1328–1330, 1998.
[Imai+06] K. Imai, Y. Horiuchi, and N. Edagawa. Radio-over-fiber distribution system with OEO
repeater for 3G cellular systems. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf.,
Paper OW G2, Anaheim, CA, March 2006.
[Inukai79] T. Inukai. An efficient SS/TDMA time slot assignment algorithm. IEEE Trans. Com-
mun., 27(10):1449–1455, 1979.
[ITU-T03a] ITU-T. Gigabit-capable passive optical networks (GPON): General characteristics.
ITU-T Recommendation G.984.1, 2003.
[ITU-T03b] ITU-T. Gigabit-capable passive optical networks (GPON): Physical Media Depen-
dent (PMD) layer specification. ITU-T Recommendation G.984.2, 2003.
[ITU-T04] ITU-T. Gigabit-capable passive optical networks (G-PON): Transmission convergence
layer specification. ITU-T Recommendation G.984.3, 2004.
[ITU-T05] ITU-T. Broadband optical access systems based on passive optical networks (PON).
ITU-T Recommendation G.983.1, 2005.
[Jia+06a] Z. Jia, G.-K. Chang, and J. Yu. A full-duplex radio-over-fiber system with 2.5 Gbit/s
data symmetric delivery over 40 km SMF-28. In Proceedings of the European Conference on
Optical Communications, Paper Tu1.6.3, Nice, France, September 2006.
[Jia+06b] Z. Jia, J. Yu, Y.-K. Yeo, T. Wang, and G.-K. Chang. Design and implementation of a
low cost, integrated platform for delivering super-broadband dual services simultaneously. In
Proceedings of the European Conference on Optical Communications, Paper Tu1.6.6, Nice,
France, September 2006.
[Kramer+02] G. Kramer and B. Mukherjee. IPACT: A dynamic protocol for an Ethernet PON
(EPON). IEEE Communications Mag., 40(2):74–80, 2002.
[Kuri+02] T. Kuri and K. Kitayama. Novel photonic downconversion technique with optical
frequency shift for millimeter-wave-band radio-on-fiber systems. IEEE Photon. Technol. Lett.,
14(8):1163–1165, 2002.
[Lam74] S. S. Lam. Packet Switching in a Multi-access Broadcast Channel. Ph.D. thesis. Depart-
ment of Computer Science, UCLA, Los Angeles, 1974.
[Lee+02] J. H. Lee, P. C. The, P. Petropoulos, M. Ibsen, and D. J. Richardson. A grating-based
OCDMA coding-decoding system incorporating a nonlinear optical loop mirror for improved
code recognition and noise reduction. IEEE/OSA J. Lightwave Technol., 20(1):36–46, 2002.
[Lee+06] J. H. Lee, C. H. Kim, Y.-G. Han, and S. B. Lee. 1.25 Gbit/s WDM PON upstream trans-
mission using Fabry–Perot laser diodes injected by depolarised CW supercontinuum source.
In Proceedings of the European Conference on Optical Communications, Paper Tu3.5.5, Nice,
France, September 2006.
[Liew+89] S. C. Liew and K.-W. Cheung. A broadband optical network based on hierarchical
multiplexing of wavelengths and RF subcarriers. IEEE/OSA J. Lightwave Technol., 7(11):1825–
1838, 1989.
[MacHale+06] E. K. MacHale, G. Talli, and P. D. Townsend. 10 Gb/s bidirectional transmission
in a 116 km reach hybrid DWDM-TDM PON. In Proceedings of the IEEE/OSA Optical Fiber
Commun. Conf., Paper OFE1, Anaheim, CA, March 2006.
[McGeehan+04] J. E. McGeehan, S. M. R. M. Nezam, P. Saghari, T. H. Izadpanah, A. E. Willner,
R. Omrani, and P. V. Kumar. 3D time-wavelength-polarization OCDMA coding for increasing
the number of users in OCDMA LANs. In Proceedings of the IEEE/OSA Optical Fiber Commun.
Conf., Paper FE5, Los Angeles, CA, March 2004.
430 Multiwavelength Optical Networks

[Mukherjee92] B. Mukherjee. WDM-based local lightwave networks – Part I: Single-hop systems.

IEEE Network, 6(3):12–27, 1992.
[Olshanksy+89] R. Olshanksy, V. A. Lanzisera, and P. M. Hill. Subcarrier multiplexed lightwave
systems for broadband distribution. IEEE/OSA J. Lightwave Technol., 7(9):1329–1342, 1989.
[Payoux+06a] F. Payoux, P. Chanclou, T. Soret, N. Genay, and R. Brenot. Demonstration of a
RSOA-based wavelength remodulation scheme in 1.25 Gbit/s bidirectional hybrid WDM-TDM
PON. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper OTuC4, Anaheim,
CA, March 2006.
[Payoux+06b] F. Payoux, P. Chanclou, and R. Brenot. WDM PON with a single SLED seeding
colorless RSOA-based OLT and ONUs. In Proceedings of the European Conference on Optical
Communications, Paper Tu4.5.1, Nice, France, September 2006.
[Prucnal+86] P. Prucnal, M. A. Santoro, and T. R. Fan. Spread spectrum fiberoptic local area
network using optical processing. IEEE/OSA J. Lightwave Technol., 4(5):547–554, 1986.
[Ramaswami93] R. Ramaswami. Multi-wavelength lightwave networks for computer communi-
cation. IEEE Communications Mag., 31(2):78–88, 1993.
[Ramaswami+90] R. Ramaswami and R. Pankaj. Tunability Needed in Multichannel Networks:
Transmitters, Receivers or Both? IBM Technical Report No. RC-16237, 1990.
[Saleh89] A. A. M. Saleh. Fundamental limit on number of channels in subcarrier-multiplexed
lightwave CATV system. Electron. Lett., 25(12):776–777, 1989.
[Salehi89] J. A. Salehi. Code division multiple-access techniques in optical fiber network –
Part I: Fundamental principles. IEEE Trans. Commun., 37(8):824–833, 1989.
[Salehi+90] J. A. Salehi, A. M. Weiner, and J. P. Heritage. Coherent ultrashort light pulse code-
division multiple access communication systems. IEEE/OSA J. Lightwave Technol., 8(3):478–
491, 1990.
[Sardesai+98] H. P. Sardesai, C.-C. Chang, and A. M. Weiner. A femtosecond code-
division multiple-access communication system test bed. IEEE/OSA J. Lightwave Technol.,
16(11):1953–1964, 1998.
[Schwartz87] M. Schwartz. Telecommunication Networks: Protocols, Modeling, and Analysis.
Reading, MA: Addison-Wesley 1987.
[Schwartz96] M. Schwartz. Broadband Integrated Networks. Englewood Cliffs, NJ: Prentice Hall
1996.
[Shankaranarayanan+91] N. K. Shankaranarayanan, S. D. Elby, and K.-Y. Lau.
WDMA/subcarrier-FDMA lightwave networks: Limitations due to optical beat interference.
IEEE/OSA J. Lightwave Technol., 9(7):931–943, 1991.
[Sherif+04] S. R. Sherif, A. Hadjiantonis, G. Ellinas, C. Assi, and M. A. Ali. A novel decentralized
Ethernet-based PON access architecture for provisioning differentiated QoS. IEEE/OSA J.
Lightwave Technol., 22(11):2483–2497, 2004.
[Son+03] E. S. Son, K. H. Han, J. K. Kim, and Y. C. Chung. Bidirectional WDM passive optical
network for simultaneous transmission of data and digital broadcast video service. IEEE/OSA
J. Lightwave Technol., 21(8):1723–1727, 2003.
[Tancevski+00] L. Tancevski and L. A. Rusch. Impact of the beat noise on the performance of
2-D optical CDMA systems. IEEE Commun. Lett., 4(8):264–266, 2000.
[The+01] P. C. The, P. Petropoulos, M. Ibsen, and D. J. Richardson. A comparative study of the
performance of seven- and 63-chip optical code-division multiple-access encoders and decoders
based on superstructured fiber Bragg gratings. IEEE/OSA J. Lightwave Technol., 19(9):1352–
1365, 2001.
 Static Multipoint Networks 431

[Thomas+02] S. Thomas and D. Wagner. Insecurity in ATM-based passive optical networks. In

Proceedings of the IEEE Int’l Conf. Commun. (ICC), New York, NY, April/May 2002.
[Ueda+01] H. Ueda, K. Okada, B. Ford, G. Mahony, S. Hornung, D. Faulkner, J. Abiven, S. Durel,
R. Ballart, and J. Erickson. Deployment status and common technical specifications for a B-PON
System. IEEE Communications Mag., 39(12):134–141, 2001.
[Wang+06] X. Wang, N. Wada, T. Miyazaki, and K. Kitayama. Demonstration of DPSK-OCDMA
with balanced detection to improve MAI and beat noise tolerance in OCDMA system. In
Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper JTh B57, Anaheim, CA,
March 2006.
[Wei+03] D. Wei, S. Ayotte, W. Mathlouthi, S. LaRochelle, and L. Rusch. BER performance of an
optical fast frequency-hopping CDMA system with multiple simultaneous users. In Proceedings
of the IEEE/OSA Optical Fiber Commun. Conf., Paper ThQ1, Atlanta, GA, March 2003.
[Wood+93] T. H. Wood and N. K. Shankaranarayanan. Operation of a passive optical network
with subcarrier multiplexing in the presence of optical beat interference. IEEE/OSA J. Lightwave
Technol., 11(10):1632–1640, 1993.
[Woodward+96] S. L. Woodward, X. Lu, T. E. Darcie, and G. E. Bodeep. Reduction of optical-
beat interference in subcarrier networks. IEEE Photon. Technol. Lett., 8(5):694–696, 1996.
[Yu+06] J. Yu, Z. Jia, L. Xu, T. Wang, and G. K. Chang. A novel scheme for reduction of phys-
ical layer interference in a DWDM radio-over-fiber network. In Proceedings of the European
Conference on Optical Communications, Nice, France, September 2006.
[Zhang+06] B. Zhang, C. Lin, L. Huo, Z. Wang, and C.-K. Chan. A simple high-speed WDM
PON utilizing a centralized supercontinuum broadband light source for colorless ONUs. In
Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper OTuC6, Anaheim, CA,
March 2006.
[Zirngibl+95] M. Zirngibl, C. H. Joyner, L. W. Stulz, C. Dragone, H. M. Presby, and I. P. Kaminow.
LARNet, a local access router network. IEEE Photon. Technol. Lett., 7(2):215–217, 1995.
6 Wavelength/Waveband-Routed
Networks

In Chapter 5 we discussed shared-channel networks, and the emphasis was on satisfy-

ing traffic requirements on a static, multipoint physical topology (a broadcast star or
its equivalent). The traffic requirements were expressed in terms of flows on logical
connections (LCs), and satisfaction of these requirements involved multiplexing and
multiple access to share the available channels efficiently. When combined time and
wavelength division techniques were employed, the optical connections supporting the
LCs were set up and time shared by rapidly tuning the transceivers over a given set of
wavelengths. Because all optical connections shared a common broadcast medium in
a static configuration, all optical paths supporting these connections were permanently
in place. We now move on to optical connection routing and wavelength/waveband as-
signment – issues that were absent in the static case. We treat both point-to-point and
point-to-multipoint (multicast) logical connections.

6.1 Introduction

In this chapter we focus on the optical layer of the architecture shown in Figure 2.1(a);
that is, we treat purely optical (transparent) networks with reconfigurable optical paths, in
which reconfiguration is achieved by space switching together with wavelength and/or
waveband routing. Unless otherwise stated, we assume that there is no wavelength
conversion in these networks, so the constraint of wavelength continuity is in force. The
earliest proposals for wavelength-routed networks (WRNs) appeared in [Brain+88] and
[Hill88].
In much of the subsequent work on these networks, a recurring issue has been to
determine the number of wavelengths required to achieve a desired degree of connectivity
as a function of network size and functionality of network nodes (e.g., static wavelength
routers, static wavelength interchangers, or WSXCs). This is a critical issue because the
available bandwidth of a fiber is not infinite, and the imperfections of the supporting
optical technology place a lower limit on the feasible wavelength spacings.
In some cases some very general results on wavelength requirements have been
obtained. For example, it has been shown (see [Barry93a, 93b; Barry+93]) that in a
static N -node network (basically a network with nodes√ that are limited to static routers)
with tunable transmitters and receivers, at least N /e wavelengths are required to
support all permutations of connections from inputs to outputs (i.e., full permutation
 Wavelength/Waveband-Routed Networks 433

connectivity). Wavelength and switch requirements in switched WRNs were presented

in [Barry+94].
As indicated in Section 3.5, when the connectivity limits of purely optical networks
are reached, it is necessary to turn to logically-routed networks (LRNs). In this case,
the purely optical network is responsible for the transport of bulk flows (e.g., gigabits
per second per λ-channel), whereas the logical switching node (LSN) does fine-grained
(e.g., IP packet) switching. This is the essence of the multihop concept, in which data
packets reenter the optical layer several times and are switched multiple times in the
LSNs before reaching their final destination [Acampora94]. Chapter 7 focuses on the
logical-layer properties of LRNs. But what is the wavelength requirement in the optical
layer supporting an LRN? Using a model similar to that used by [Barry93b], it has been
shown that for a cell-/packet-switched LRN [Bala+96], a fixed number of wavelengths
is sufficient to build a scalable network that supports full permutation connectivity on
a packet basis. This demonstrates the power of combining electronics with optics for
achieving both high connectivity and high throughput.
Bounds on wavelengths required for achieving full permutation connectivity in linear
lightwave networks (LLNs) were obtained by [Pankaj92] and [Pankaj+95]. In one result,
lower bounds on the number of wavelengths required for full permutation connectivity,
given a constraint on nodal degree, were found. Because they are based on “counting”
the network resources used in the connections, they are similar to the aggregate network
capacity bound of Section 6.3.1.1. In additional results, special network topologies
were sought that achieve full permutation connectivity with a minimum number of
wavelengths. The chosen topologies are similar in structure to the multistage switch
fabrics (e.g., Benes) discussed in Chapter 2 and the multihop networks (e.g., ShuffleNet)
discussed in Chapter 7.
Although the general approaches to bounding wavelengths are important for indicat-
ing basic relations between network structure and performance limits, they are often
asymptotic in nature and/or tied to unrealistic topologies. Thus, they are not very helpful
in guiding the network designer, who deals with control and performance optimization
in networks with realistic sizes and topologies.
This chapter focuses on these more realistic issues. We begin with a review of some
general properties of physical topologies, stressing their relation to network performance.
Then the problem of achieving a prescribed logical connectivity and throughput is studied
in both WRNs and LLNs. Various network configurations and various methods of control
are considered. If the traffic patterns are reasonably well-known in advance, the most
effective technique for assigning optical resources (paths, wavelengths/wavebands, and
channels) to LCs is by using static routing and channel assignment (RCA) rules. For
example, static routing is appropriate for provisioning a list of specified, dedicated
LCs; that is, for “semipermanently” embedding a specified logical topology onto a
given physical topology. (As we shall see, static routing rules may also be useful for
demand-assigned connections.) For more flexibility and efficiency in the face of random
traffic, dynamic assignments are preferable. In Section 6.3 we explore static RCA in
wavelength-routed networks. The dynamic case is treated in Section 6.4. The concluding
Sections 6.5 and 6.6 treat routing, waveband, and channel assignment for LLNs in the
434 Multiwavelength Optical Networks

static and dynamic cases, respectively. (In WRNs a channel is always a λ-channel,
whereas in LLNs channels may take many different forms.)

6.2 Physical Topologies

The topological features of the fiber interconnections play a fundamental role in deter-
mining the key performance indices of a network. They influence optical signal quality,
optical spectral efficiency, potential connectivity, maximum throughput, and survivabil-
ity. Thus we begin with a brief description of common network topologies and their
distinctive features.
Let G(V, E) be a graph representing the topology of a network (excluding its access
stations) with a set of vertices V representing the network nodes and a set of edges E
representing the network links. The degree  of a vertex (or node) is the number of
internodal links connected to the corresponding network node. Because the cost of a
node (e.g., a static router or optical cross-connect) grows rapidly with the number of
ports, it is desirable to keep  small.
The graph diameter D gives the longest distance between pairs of network nodes (in
optical hops). Because long optical paths lead to poor signal quality, network congestion,
and vulnerability to network faults, it is important to use topologies that have small
diameters. Thus, good topologies are as “dense” as possible in the sense that they have
a high order (number of nodes) for a small diameter. To be cost-effective, this high
order should be achieved with reasonably small node degrees. The well-known Moore
bound gives the maximum order NMoore (, D) of a graph of maximum degree  and
diameter D:
 D−1
NMoore (, D) = 1 +  i=0 ( − 1)i
( − 1) D − 2 (6.1)
= ,  > 2.
−2
Graphs that achieve this bound are known as Moore graphs. There are an infinite
number of Moore graphs of diameter 1 (the fully connected graphs of any degree).
Also, Moore graphs of degree 2 and any diameter exist (all rings with an odd number
of vertices). However, Moore graphs of small diameter grow scarce as  increases. For
D = 2, there is only one Moore graph of degree 3 (the Petersen graph, with N = 10,
Figure 6.41) and only one of degree 7 [Hoffman+60]. Moore graphs of diameter 2 and
3 do not exist for any other degrees except possibly  = 57, for which the jury is still
out.
Of course, graphs that come close to the Moore bound are also good candidates as
physical topologies of optical networks. Figure 6.1 gives the number of vertices in known
graphs of maximal order for degrees and diameters up to 8.1 Some of these are “almost”

1
With the exception of the numbers in parentheses in Figure 6.1, these are the graphs that were known as of
1982 [Bermond+82]. The numbers in parentheses correspond to Cayley graphs, discussed by [Arden+91,
Chudnovsky+88].
 Wavelength/Waveband-Routed Networks 435

D
2 3 4 5 6 7 8

3 10 20 38 70 128 180 280

4 15 40 95 364 731 (1081) (2943)

5 24 60 174 532 2734 2988 7600

6 32 105 317 820 7817 10,920 16,385

7 50 122 420 1550 8998 31,248 54,168

8 57 200 807 2550 39,223 40,593 154,800

Figure 6.1 Number of vertices in known maximal graphs.

Moore graphs. For example, the maximal graph with 38 vertices shown in Figure 6.2
has  = 3 and D = 4, so its order is within 8 of the Moore bound of 46.
For purposes of physical topology design and performance evaluation, it is generally
easier to work with topologies that have a high degree of symmetry. For example,
Figures 6.3(a)–6.3(c) show the three regular tessellations of the plane: the hexagonal,
square, and triangular grids of degrees 3, 4, and 6, respectively. Each of these is made
into a finite regular graph by closing it back on itself on two pairs of opposite sides,

Figure 6.2 Thirty-eight-vertex graph. (From [Ghafoor85, Figure 2.5]. Copyright
1985,
c A.
Ghafoor, Reprinted by permission.)
436 Multiwavelength Optical Networks

(a) (b)

(c) (d)

Figure 6.3 Tessellations of the plane.

producing a torus as in Figure 6.3(d) for the square grid. It is then completely symmetric
(i.e., it looks the same when viewed from any vertex).
The more common regular networks typically fall far short of the Moore bound. Thus,
the n-cube has degree  = D = n and order Nn -cube = 2 D , whereas the square grid (as a
torus) has  = 4 and Ntorus = n 2 for D = n − 1, where n is odd.
Two families of networks that are considerably more dense than the grids and the
hypercubes are the deBruijn and Kautz networks.2 Their directed versions (in the form
of regular digraphs) are useful candidates for topologies of LRNs, where the values of
 and D can be specified arbitrarily.3 The undirected versions (which are not regular)
provide a wide range of possibilities for physical topologies of purely optical networks.
Figure 6.4 shows the undirected deBruijn and Kautz graphs with maximum degree
 = 4 and diameter D = 3.
The orders of undirected deBruijn and Kautz graphs of diameter D and maximum
degree  are given, respectively, by
 D

NdeBruijn (, D) = (6.2)
2
 D  D−1
 
NKautz (, D) = + . (6.3)
2 2
A comparison of the orders of the n-cube, and deBruijn and Kautz physical topologies,
is presented in Table 6.1. The known maximal graphs are also shown in the table together

2
An extensive discussion of these networks and their properties appears in [Bermond+89].
3
Construction procedures for deBruijn and Kautz graphs and their generalizations are presented in
Section 7.2.2.
 Wavelength/Waveband-Routed Networks 437

deBruijn Kautz

Figure 6.4 Undirected deBruijn and Kautz graphs.

with the Moore bound.4 Note that the deBruijn and Kautz graphs are far more dense
than the n-cube, and they come within an order of magnitude of the sizes of the known
maximal graphs.
The diameter of a network only gives the “worst-case” internodal distance. It is more
important, from the point of view of network performance, to have some measure of
¯ For a regular N -node network of degree , it can be
the average internodal distance d.
shown (see [Cerf+74]) that a lower bound on d¯ is
1
d¯min (, N ) =  d (, N ), (6.4)
(N − 1) min
where
 
D
 D

d
min (, N ) =  j( − 1) j−1 − D 1 +  ( − 1) j−1 − N  . (6.5)
j=1 j=1

d
In Equation (6.5), min is a lower bound on the sum of the distances (shortest paths)
from any vertex of a regular graph of degree  to its N − 1 closest vertices. (D is
the smallest integer for which NMoore [, D] ≥ N .) An example of the construction
d
for min (3, 7) is shown in Figure 6.5. It is based on the assumption of a Moore graph
d
topology. The figure shows a tree of depth D = 2, where min = 9. Note that the same tree,
d
completed with the dashed lines, can be used to find min (3, 10) = 15. The Petersen graph
shown in Figure 6.41 later in this chapter satisfies this bound exactly because it is a Moore
graph. Although the bound in Equation (6.4) is not always very tight, it is a useful starting
point for determining other important network properties. (See the aggregate network

Table 6.1 Orders of some graphs.

D= n-cube deBruijn Kautz Maximal Moore

4 16 16 24 95 161
6 64 729 972 7,817 23,437
8 128 65,536 81,920 154,800 7,686,401

4
Only cases with D =  are listed so that comparisons can be made with the n-cube.
438 Multiwavelength Optical Networks

d
min(3, 7)
d
Figure 6.5 Construction for min .

d
capacity wavelength bound in Section 6.3.1.1.) Figure 6.6 is a plot of min as a function
of N , with  as a parameter.
Because physical topologies for real WANs usually have irregular and somewhat
random structures, it is of interest to seek estimates of internodal distances for random
graphs. The distance properties of “connected, semirandom” directed networks have
been studied by [Rose92]. These are networks of N nodes and L directed links, in
which every node is reachable from every other on a directed path, each node has p
links emanating from it (i.e., its out-degree is p, its average in-degree is p, and
L = N p), and the topology is completely random otherwise. Figure 6.7 shows plots of

d
min

400 3

4
Sum of Shortest Paths

300
5

6
200

100

0
0 20 40 60 80 100
Number of Nodes N
d
Figure 6.6 Plot of min as a function of N .
 Wavelength/Waveband-Routed Networks 439

1000

900

800
Semirandom
100 Trials
700
Nodes Reached in k Hops

600

500

400
ShuffleNet

300

200

100

0 k
0 5 10 15 20 25
N 4608; p 2

Figure 6.7 Internodal distances in random networks. (From [Rose92, Figure 11a]. Copyright
1992
c
IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

the number of nodes reached from a single node in k hops for a random ensemble of 100
such networks, with N = 4608 and p = 2.5 This curve can be viewed as a probability
distribution for internodal distances, in which distance is now measured in optical hops
on directed paths. Note the striking resemblance to a Gaussian distribution. The mean
of this distribution corresponds to d¯ for the ensemble. In this case, the curves show a
mean of d¯ ≈ 11 and standard deviation of σd ≈ 2. (The standard deviation decreases
with increasing p, remaining relatively independent of the order of the network.) The
narrow Gaussian shape in Figure 6.7 suggests that individual node-pair distances in large
random networks deviate very little from d¯ (i.e., most node pairs are about the same
distance apart!).
For comparison, the distribution of internodal distances for a k = 9 stage ShuffleNet
(with the same values of N and p) is also shown in Figure 6.7. The figure illustrates
the surprising fact that the regular and highly symmetric structure of the ShuffleNet
5
To ensure connectedness, these networks are constructed as random chordal rings (i.e., each network is a
directed ring containing all nodes, with chordal links added at random).
440 Multiwavelength Optical Networks

Figure 6.8 Recursive grid.

actually produces a wider distribution of internodal distances with a slightly larger mean
(d¯ ≈ 12).
Applying these results to undirected networks, we note that if link directions are ig-
nored, a random directed network with L links and with nodes of average out-degree
and in-degree p becomes a random undirected network with L links and average degree
 = 2 p. It should therefore be expected that the internodal distances for random directed
networks can be used as a guide to those in random undirected networks. Because there
are no constraints on directions in the latter case, the hop distances for the derived undi-
rected networks are bounded above by those for the corresponding directed networks.
For example, the curves of Figure 6.7 represent “conservative” estimates of internodal
distance distributions in an undirected network with 4608 nodes and an average node
degree  = 4.6
Often it is desirable, when planning a network, to start with a basic topology and then
“grow” it to larger sizes matched to the evolution of traffic demand. A simple example of
a hierarchical mesh/ring/access network structure was presented in Figure 1.5. Another
useful approach is recursive design, leading to structures in which the same basic unit
is repeated at different levels of the hierarchy. Two approaches to recursive design are
shown in Figures 6.8 and 6.9. The first begins with a square grid, which is refined
recursively in neighborhoods in which demand has increased [Brown+96]. The second
is a hierarchical design based on growing new Petersen networks recursively out of the
nodes of an existing network. In this design we begin with the 10-node Petersen network
of Figure 6.41, containing the set of nodes
N = { A, B, . . . , J }. (6.6)
6
An indication of differences between the directed and undirected case can be obtained by studying families
of regular graphs. For example, for ShuffleNets ranging from N = 8 to 81, the undirected version has a d¯
approximately 18% less than the directed version, on the average.
 Wavelength/Waveband-Routed Networks 441

E GA
B
GB

DE I C

DEE
CA CB
DEA DA
DEB DB
DBA
DG DBH DBC

DC

Figure 6.9 Hierarchical Petersen graph.

(Because all subnets in the hierarchy have the same topology, it suffices to define the
network by its nodes alone.) The original network forms level 0 in the hierarchy. At each
of its nodes, a subnet of the same form can be grown, as exemplified by the subnets
NC , N D , and NG , which represent level 1. With these first-level subnets, the expanded
network has the set of nodes

N = { A, B, NC , N D , E, F, NG , H, I, J }, (6.7)

where

NC = {C A, C B, . . . , C J }
N D = {D A, NDB , DC, D D, NDE , D F, DG, D H, D I, D J } (6.8)
NG = {G A, G B, . . . , G J }.
442 Multiwavelength Optical Networks

Note from Figure 6.9 that the level 1 subnet N D in turn includes level 2 subnets:
NDB = {D B A, D B B, . . . , D B J }
(6.9)
NDE = {D E A, D E B, . . . , D E J }.

A node from which a higher level subnet is grown is called a gateway node. The
nongateway nodes are called end nodes. In the case shown, NASs are only connected to
end nodes, and at most one higher level subnet is grown from each gateway. The recursive
grid has good survivability (many alternative paths) but rather poor distance properties.
The Petersen design has good distance properties (because each Petersen graph has
diameter 2) but poor survivability (because a failure of a gateway node disconnects the
network).
A key feature of the hierarchical Petersen design is that all subnets at the same level
in the hierarchy (e.g., NG , NC , and N D or N D E and N D B ) are mutually disjoint. This is
an important property as far as wavelength/waveband assignment is concerned because
it implies that these subnets can reuse the same wavelengths/wavebands. Variations on
these themes can ameliorate their shortcomings (see Problems 1 and 2 at the end of this
chapter).
As we illustrate how to utilize network resources efficiently in this chapter and the next,
the examples used are based frequently on dense topologies of the type presented here.

6.3 Wavelength-Routed Networks: Static Routing and Channel Assignment

In this section, we consider static RCA in wavelength-routed networks. Unless stated

otherwise, the networks under discussion are assumed to operate without wavelength
interchange. (Thus the constraint of wavelength continuity must be observed.) Because
a WRN has a single λ-channel (wavelength) in each waveband, we shall use the term
wavelength to denote both wavelength and waveband. (In Section 6.5, which treats
waveband-routed networks, these two terms have a different significance.)
In static routing, the traffic requirements are given as a prescribed set of point-to-
point LCs between pairs of access stations. The set of connections defines a logical
topology; that is, a directed logical connection graph (LCG). These connections are
assumed to remain in place for relatively long periods of time, so it is worthwhile to
attempt to optimize the way in which they are embedded onto the physical topology,
even if optimization requires considerable computation.
The physical topology of the WRN may either be given or left free for the network
designer. In the former case, it most likely corresponds to the deployment of cables
in some existing fiber infrastructure. In the latter case, the physical topology may be
custom-tailored to a prescribed logical topology (see Section 6.3.6).
In exploring RCA, we begin with some general bounds on the number of wavelengths
required to satisfy prescribed traffic requirements, followed by a general formulation of
RCA as a graph coloring problem. We then treat some special physical topologies: rings
and combinations of rings in Sections 6.3.4 and 6.3.5 and multistars in Section 6.3.6.
 Wavelength/Waveband-Routed Networks 443

In Section 6.3.7 general mesh topologies are considered, using optimization techniques.
It turns out that the complexity of the optimization approach limits its applicability
to fairly small networks. For larger problems, computational efficiency is the most
important issue, suggesting the use of heuristic solution techniques for RCA. Two
particularly effective heuristics are presented in Section 6.3.8. They are quite simple
computationally and often achieve surprisingly good (and frequently optimal) results.
The heuristic algorithms used for the static case are actually closely related to the
dynamic rules discussed in Section 6.4.
The specification of the RCA problem begins with the traffic requirements, given
as a set of M point-to-point LCs {l1 , l2 , . . . , l M }, where li = [si , di ], and si and di are
the network nodes accessed by the source and destination stations, respectively.7 Unless
stated otherwise, we assume that each LC requires the full capacity of one λ-channel.
Therefore, each LC li is supported by a dedicated point-to-point optical connection
ci = (si , di )k , which is in turn supported by a dedicated point-to-point optical path
pi = si , di k . The subscript k indicates that ci and pi operate on wavelength k. For
purposes of RCA, logical and optical connections are equivalent, so we shall frequently
drop the adjectives, referring to both of them simply as connections.
Setting up an optical connection in a WRN requires two operations:

1. Creation of a source-destination optical path on a chosen wavelength (by setting the

optical cross-connects along its route)
2. Creation of an optical connection by tuning the source transmitter and destination
receiver to the chosen wavelength

An optical path p consists of a fiber path π ; that is, a sequence of fibers directed from
source to destination, and a wavelength assignment on that fiber path. Thus, several
optical paths may share the same fiber paths as long as they use different wavelengths
(to avoid violating the DCA condition). The term lightpath is often used in the literature
to denote various related entities that we refer to here as fiber paths, optical paths, λ-
channels, optical connections or logical connections. (Figure 2.1 defines the place of
each of these in our multiwavelength network architecture.)
In a WRN with a given physical topology, solving the RCA problem consists of
finding an appropriate optical path for each connection. This means finding a fiber path
(routing) together with a wavelength (channel or wavelength assignment). Static RCA
is equivalent to finding an embedding of a prescribed logical topology onto a given
physical topology while observing the constraints of the network.
In some cases the network resources may not be specified completely, so RCA
may include specification of certain design parameters. For example, the number of

7
There is no restriction on the LCs, so more than one can exist between the same source and destination
nodes. This could happen either because several stations are accessing the same source-destination node
pair or because several optical transceivers are being used in one source-destination station pair. When more
than one NAS is attached to a node, the source si and destination di must refer to a station rather than a node.
However, in most cases we assume that a single station is associated with each node, so node designations
are sufficient.
444 Multiwavelength Optical Networks

wavelengths may be left free, in which case a “spectrally efficient” solution might be
sought using a minimum number of wavelengths. In other cases we may be free to adjust
the number of fiber pairs on each link to match the expected traffic and hence reduce the
number of required wavelengths. In this case the problem consists of combined RCA
and link capacity (fiber) assignment. In still other cases, additional constraints might be
imposed, such as a limit on the optical path lengths. Also, an incomplete solution might
be acceptable, in which only most of the prescribed connections are established. All of
these represent variants of the basic problem that appear in subsequent sections.
The RCA problem often can be solved in an efficient manner by taking advantage
of special features of the network’s topology, such as symmetry. This is the case for
rings and other symmetric networks that are used here as illustrative examples. For
more general topologies, RCA can be cast as a general optimization problem and solved
using various combinatorial optimization techniques. However, due to the complexity
of the problem, optimal solutions can be found only for relatively small networks, and
suboptimal heuristics must be used in the larger cases.
Because of its complexity, it is usually necessary to separate the joint RCA problem
into its two components. Solving the routing problem separately from the channel
assignment problem frequently gives suboptimal results but is much easier to accomplish
and is generally satisfactory for engineering purposes.
Before going into the details of the RCA problem, it is helpful to obtain an overall
picture of the relations between logical layer requirements and physical layer resources.
From these relations we derive certain important connections between network “flows”
and optical spectrum (wavelength) requirements.

6.3.1 Flow Bounds: Matching the Physical and Logical Topologies

Suppose a wavelength-routed network has a given physical topology with N nodes and
L links, and each link has the same number of fiber pairs F. If the network operates
with an available spectrum of W wavelengths, the product WF represents the one-way
capacity of a link, expressed in units of λ-channels. By varying either F or W , we vary
the link capacity.8 The ability of the network to support a prescribed set of LCs depends
principally on the form of the logical topology, the form of the physical topology, and
the link capacities WF.
Assuming that all other network parameters are fixed, it is important to be able to
determine the number of wavelengths required to support a prescribed logical topology.
As we shall see, making an exact determination of this number by obtaining an optimal
solution to the RCA problem is generally extremely difficult. When suboptimal heuristics
are used, it is desirable to have some idea of how close the solution is to optimal. For
this reason, bounds on W can be very useful. We now present two simple lower bounds
based on network flows.

8
However, as pointed out in Section 3.3.1, F and W are not interchangeable quantities. Because of the
wavelength continuity constraint, capacity in the form of fibers is more valuable than capacity in the form
of wavelengths.
 Wavelength/Waveband-Routed Networks 445

6.3.1.1 Aggregate Network Capacity Bound

Consider a prescribed set of M logical/optical connections {ci }. Now if d(si , di ) is the
distance between the source and the destination nodes for connection ci (in optical hops),
M
we call  d = i=1 d(si , di ) the total internodal distance for the connection set. The
network capacity used in carrying this traffic must be at least  d “link-channels,” but
the aggregate capacity of L links in the network is 2LFW link-channels. Comparing
the two, we find that the number of wavelengths necessary to support all connections is
bounded below by
d
W ≥ . (6.10)
2L F
Equation (6.10) is satisfied with equality when all connections are routed on shortest
paths and all fibers are saturated; that is, when W wavelengths are used on every fiber. If
connection ci is routed on an optical path having H i hops, we define the total hop count
M
for the connection set as  H = i=1 H i , and the average hop count as H̄ =  H/M.
Also we denote the average node degree by  ¯ = 2L/N . Then a lower bound on W to
support all connections becomes9
H M H̄ d
W Netcap = = ≥ . (6.11)
2L F ¯ F N
N ¯ F
Suppose we apply the wavelength bound to a case in which the logical topology is
fully connected, with one connection per node pair; in other words, M = N (N − 1). If
the physical topology is regular of degree , Equation (6.11) gives
(N − 1) H̄  d (, N ) (N − 1)d¯min (, N )
WF ≥ ≥ min = , (6.12)
  
where d¯min and min
d
are defined in Equations (6.4) and (6.5), respectively.
Equation (6.12) was written with W F on the left side to express the bound in terms
of required link capacity in units of λ-channels. Because the bound was obtained using
flow conservation only, the wavelength continuity condition was not used. The bound
therefore makes no distinction between capacity in the form of fibers and capacity in
the form of wavelengths. In reality, as has been stated previously, a network in which
each link has a total capacity of k λ-channels generally performs better if that capacity
is provided by F = k fiber pairs per link and W = 1 wavelength than if the capacity is
provided by F = 1 fiber pair per link and W = k wavelengths.
These bounds clearly indicate the importance of packing the network nodes into a
dense (small-diameter) topology to reduce the required number of wavelengths. For ex-
ample, increasing  decreases the bound directly (in the denominator of Equation [6.12])
as well as indirectly by reducing d¯min . How closely the bounds can be approached de-
pends on (1) the effectiveness of the RCA algorithm and (2) the uniformity of the link
loading.

9
Equation (6.11) can be modified easily to apply to networks with different numbers of fiber pairs in different
links.
446 Multiwavelength Optical Networks

Uniformly distributed link loads can occur only if the prescribed traffic distribution
is “matched” correctly to the physical topology. To obtain a bound that reflects this
matching condition, we need to look more closely at the network flows.

6.3.1.2 Limiting Cut Bound

The limiting cut bound is based on the max flow–min cut theorem, which is well-known
in operations research (see Appendix A).
Consider a connected N -node network supporting a prescribed set of LCs. Suppose
the nodes are partitioned into two subsets containing K i and N − K i nodes. Let Ci be
the cut separating the network into the two disconnected subnets induced by this node
partition. Let Fi be the total number of fiber pairs in the links comprising the cut and
let M i be the number of LCs from the first to the second subnet. Then a lower bound on
the number of wavelengths W required to support these connections is
Mi
W ≥ . (6.13)
Fi
It follows that a lower bound on W to support all connections is
 
Mi
WLimcut = max , (6.14)
i Fi
where the maximum is taken over all cuts in the network. A cut that maximizes the right
side is called a limiting cut. A necessary condition for this bound to be achieved is that
traffic be distributed evenly among all fibers in each limiting cut.
In the case of a fully connected logical topology, Equation (6.14) becomes
 
K i (N − K i )
W Limcut = max . (6.15)
i Fi
Let us apply this bound to the N -node ring network of Figure 6.15. Taking the case N odd,
we cut the ring almost in half, so that K i and N − K i in Equation (6.13) are, respectively,
(N − 1)/2 and (N + 1)/2. With Fi = 2 (a single fiber pair bidirectional ring), the
2
bound becomes W ≥ N 8− 1 , verifying that the wavelength assignments described in
Section 6.3.4.2 are optimal. For the case N even, we have W ≥ N 2 /8, which is tight if
and only if N is divisible by 4 (see Equation [6.16]).
The limiting cut bound is very useful in detecting poor matches between logical
and physical topologies. In [Baroni+97], the authors use network cuts effectively for
this purpose. Figure 6.10 shows limiting cuts (indicated by dashed lines) in four real
or hypothetical WANs: ARPANet, UKNet, the European Optical Network (EON), and
NSFNet. For each of these networks, W Limcut and the ratio W Limcut /W Netcap are given,
together with various other network parameters. For example, for the ARPANet, the
limiting cut has three links, and for full connectivity a total of 11 × 9 = 99 connections
pass across this cut. Hence, assuming F = 1, the minimum number of wavelengths
required is 33 and is achieved when the routing algorithm divides the connections equally
among the three links. A routing algorithm that fails to distribute the connections evenly
requires more than the minimum number of wavelengths. Thus, the right choice of
routes determines the efficiency of the wavelength assignment process. Furthermore,
 Wavelength/Waveband-Routed Networks 447

Network N L H D ( min, max) W Limcut Nλ W Limcut /W Netcap

ARPANet 20 31 0.16 2.81 6 (2, 4) 33 33 1.92

UKNet 21 39 0.19 2.51 5 (2, 7) 19 22 1.41

EON 20 39 0.2 2.36 5 (2, 7) 18 18 1.57

NSFNet 14 21 0.23 2.14 3 (2, 4) 13 13 1.40

Figure 6.10 Limiting cuts for four networks. (From [Baroni+97, Table 1]. Copyright
1997
c IEEE.
Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

when routing is done independently of channel assignment (as is done when RCA is
separated into two subproblems), inefficiencies and increased wavelength requirements
may result.
The ratio of W Limcut to W Netcap indicates the degree of match between logical and
physical topologies. A high ratio, as in the ARPANet, indicates a relatively poor match,
suggesting a bottleneck at the limiting cut. (The fact that all ratios in these examples are
less than 2 indicates that the networks are fairly well designed.)
The limiting cut bound on wavelengths suggests a simple and cost-effective method of
link capacity assignment in wide area WRNs, to match a given logical topology. Starting
with a given fiber topology, and assuming a single fiber pair per link, the limiting cut
bound can be calculated based on the prescribed LCs. If the resultant W Limcut is too
large, a reduction can be achieved by augmenting the number of fiber pairs on the
links in the limiting cut(s). Because fiber cables normally contain many fiber pairs, this
is a fairly efficient way of conserving the optical spectrum. As the number of fiber
pairs in a limiting cut is increased, a point is reached when some other cut becomes
limiting. This procedure can then be continued on the new limiting cut until it is no
longer possible (or cost-effective) to improve the bound by increasing the fibers in the
448 Multiwavelength Optical Networks

links.10 Essentially, this amounts to a heuristic for assigning link capacities to match a
prescribed traffic distribution. Of course there is no guarantee that the actual wavelength
assignments in the network can achieve the bound. The actual number of wavelengths
used depends on the routing and channel assignments. Once RCA is complete, another
iteration of fiber assignment can be executed to reduce the number Nλ of wavelengths
actually used. This combination of RCA and fiber assignment on the “bottleneck” links
comprises a combined RCA and capacity assignment problem. Although an optimal
solution cannot be expected, empirical results (see [Baroni+97]) indicate that simple
RCA heuristics together with some judicious increase in link fibers at strategic locations
produce excellent results in minimizing Nλ . More details are provided in Section 6.3.8.
We have seen that the limiting cut bound can be an extremely useful tool for network
design, provisioning, and other related problems. However, the expression for W Limcut in
Equation (6.14) is deceptively simple. The computation of M i /Fi for each cut is trivial.
However, the enumeration of all cuts in a large network is out of the question, because
the total number of cuts to be examined in an N -node network is 2 N − 2; in other words,
it grows exponentially with the number of network nodes. Thus, it is important to be
able to find a good lower bound on W Limcut for large networks using a minimum of
computation. A computationally efficient heuristic that produces a tight lower bound,
W Limcut , on W Limcut is provided in Appendix D.11

6.3.2 Nonblocking Stations

As mentioned in Section 3.3.2, constraints on wavelength assignments are dependent on
the way an NAS is connected to its network node. Thus, to separate station and access
link constraints from network constraints, we reintroduce the concept of a nonblocking
access station, first proposed in Section 3.3.2. (It is assumed here that each access station
is attached to only one network node.)
We call an NAS with a single fiber pair access link, as shown in Figure 2.29, an ele-
mentary station. A station is called nonblocking if it can terminate all optical connections
arriving at or departing from its network node. Thus, in a network with single fiber pair
internodal links, a nonblocking NAS takes the form shown in Figure 3.17. It has as many
access fiber pairs as the degree  of its node, and thus it is equivalent to  elementary
stations in the same box.12 Each access fiber pair is attached to its own input/output
port on the network node so that it can be switched independently by the node. This
implies that nonblocking stations require  access ports on their attached network node,
compared with a single port for elementary stations. (In the case of networks with F
fiber pairs on each link, a nonblocking access station requires F access fiber pairs and
access ports.) Assuming that the network operates on W wavelengths, the nonblocking

10
The room for improvement of W Limcut can be ascertained by comparing new values of W Limcut and W Netcap
at each stage of the fiber augmentation process.
11
A tighter fluid flow bound on W using linear programming is formulated in Section 6.3.7.4. However, the
linear programming approach requires considerably more computation.
12
Recall that the degree of a node is defined here as the number of internodal links incident on the node. The
access links are not counted.
 Wavelength/Waveband-Routed Networks 449

A A

B C B C

(a) (b)

Figure 6.11 Three-node network.

station must contain an array of W transceivers connected to each access fiber pair, with
one transceiver in each array operating on each wavelength.
To remove the constraints of the access stations from the RCA discussion, we shall
frequently assume that one nonblocking station accesses each network node, in which
case the network can be represented by a model containing only the network nodes and
the internodal fiber links, suppressing the NASs.13 The difference between elementary
and nonblocking stations is illustrated in Figure 6.11(a). If the NAS at node A is an
elementary station, only one optical connection on each wavelength can originate at that
station. However, if it is nonblocking, two connections per wavelength can originate at
the station because node A has degree 2. With the assumption of nonblocking stations,
the network can be represented by a graph containing only the vertices A, B, and C as in
Figure 6.11(b). When the characteristics of the stations and access links are pertinent to
the RCA problem (as in the case of elementary stations), we include them as additional
vertices and edges in the network graph.

6.3.3 RCA as a Graph Coloring Problem

In this section we cast the RCA problem in terms of a path interference graph, G PI , incor-
porating all of the admissible RCA choices and the network constraints.14 Determination
of routing and channel assignments then reduces to the coloring of selected vertices
of G PI , subject to certain interference constraints. Section 6.3.7 provides alternatives
such as mixed integer program (MIP) or integer linear program (ILP) formulations.
All of these formulations are equivalent, but depending on the particular problem, one
formulation may be more convenient than another.
As before, we assume that each required logical/optical connection is supported by an
optical path. Let πk = a, b, . . . , z denote a fiber path using a sequence of fibers a, b . . . z,
and let p = A, B j,πk denote an optical path between source-destination pair ( A, B)

13
The nonblocking NAS is used here as a mathematical convenience. In practice, it would probably not be
cost-effective to equip each station with W F transceivers.
14
Interference graphs have been used by [Gopal82] for channel assignment in satellite systems and by
[Chlamtac+89] and [Bala+91b] for proving NP-completeness of the channel assignment problem in
WRNs and LLNs, respectively.
450 Multiwavelength Optical Networks

using a wavelength j on fiber path πk .15 Then, a feasible solution of the RCA problem
consists of a choice of an optical path (OP) (i.e., a fiber path and wavelength) for each
prescribed connection so that all network constraints are satisfied. An optimal solution
of the RCA problem consists of a feasible solution that minimizes some cost function;
for example, the number of wavelengths used.
Because we assume no wavelength interchange at this point, the channel assignments
are constrained by wavelength continuity (as well as by the DCA condition). Two optical
paths in a WRN violate DCA and thus interfere if they share a common fiber and
are assigned the same wavelength. In formulating the RCA problem, it is essential
to identify interfering optical paths. This is done by means of G PI , which exhibits
potentially interfering OPs. Wavelength continuity is included in the network model by
assigning the same wavelength to all fibers on a given OP.
For a given physical topology and prescribed connection set {ci }, a path interference
graph, G PI , is constructed by identifying each vertex of G PI with an admissible fiber
path. Assuming that there is at most one connection required for each source-destination

pair, and that there are K i admissible paths for connection ci , there will be i K i
vertices in G PI . Admissibility of paths is arbitrary; for example, the admissible paths
for a given source-destination pair might be all minimum-hop paths, all paths less than
a given physical length, or all possible paths. Two vertices of G PI are joined by an edge
(i.e., are adjacent) if their paths share a common fiber in the network. (Adjacent vertices
represent potentially interfering OPs.)
Having constructed G PI , solving the RCA problem consists of selecting one vertex
of the graph (a fiber path) for each prescribed connection and choosing a color for
each selected vertex (corresponding to a wavelength for the connection) so that adjacent
vertices are assigned different colors. If the assignments are made with a minimum
number of wavelengths, this is known as minimal vertex coloring of the subgraph
induced by the selected vertices. It is a classic and difficult (NP-complete) problem
in graph theory (see Appendix A). In cases in which several connections are required
between the same source-destination pair, they may be routed on different fiber paths
and/or on the same fiber path provided that connections on the same fiber path are
assigned different wavelengths. The latter case corresponds to a vertex coloring of G PI
using multiple colors per vertex.
For illustration, consider the network of Figure 6.12(a). We wish to determine a
routing and channel assignment for the connection set {(A, C), (A, F), ( A, E), (C, E)}.
To simplify matters, we assume that a single nonblocking station is attached to each
node so that multiple optical connections using the same wavelength can begin and end
on any node without risk of interference on the access fibers. We also assume that only
minimum-hop paths are admissible to reduce the number of possible routing alternatives.
All admissible fiber paths that support the required connections are as follows:
π1 = ab
π2 = ad f
15
When the network links carry more than one fiber pair, the specification of the path must distinguish a
particular fiber on each link.
 Wavelength/Waveband-Routed Networks 451

1 1
(A, C) 1 1

2 2
(A, F) 2 3 2
a b
A B C

1 1
(A, E) 4 5 4
c d

2 3
D E F (C, E) 6 6
e f
(a) (b) (c)

Figure 6.12 Illustrating RCA in a wavelength-routed network.

π3 = ce f
π4 = ce
π5 = ad
π6 = bd
Connection (A, C) is supported by π1 , connection ( A, F) is supported by π2 or π3 ,
connection (A, E) is supported by π4 or π5 , and connection (C, E) is supported by π6 .
These produce the path interference graph shown in Figure 6.12(b). The vertices of
the graph are arranged in the figure so that all fiber paths for a given connection appear
in the same row. To satisfy the connection requirements, one vertex must be chosen from
each row. Now suppose our objective is to minimize the number of wavelengths used
to carry all connections. In that case the vertices must be chosen so that a minimum
number of colors are used. The figure indicates a choice of paths π1 , π3 , π4 , and π6
(shaded), which are assigned colors (wavelengths) 1, 2, 1, 2, respectively, so a total of
W = 2 wavelengths are needed. Using the limiting cut bound of Section 6.3.1.2, it is
easy to show that this is the minimum possible number of wavelengths.
In this case, the joint routing and wavelength minimization problem was solved easily
by inspection. However, even in networks with small numbers of nodes, RCA is an
extraordinarily complex combinatorial problem. To obtain some idea of its complexity,
consider the following “brute force” approach.
1. Make an arbitrary choice of vertices of G PI (a routing assignment) to satisfy the
connection requirements. Form the subgraph of G PI induced by these vertices.
2. Determine the chromatic number of this subgraph.16
3. Repeat this process for all choices of vertices of G PI , and select a routing assignment
that produces a subgraph with a minimum chromatic number.
4. Do a minimal vertex coloring (wavelength assignment) of the selected subgraph.
16
The chromatic number of a graph is the number of colors required for a minimal vertex coloring of the
graph. It is equal to or greater than the size of the maximum clique of the graph (see Appendix A).
452 Multiwavelength Optical Networks

The complexity of this approach is mind-boggling. For a set of M connections, the

&M
number of subgraphs that must be examined is S = i=1 K i , and each subgraph has
M vertices. For full connectivity, M = N (N − 1). Determining the chromatic number
of each subgraph is itself an NP-complete problem (see [Garey+79]), and this must be
done S times. Finally, once a routing assignment is selected, a minimal vertex coloring
must be found – again, an NP-complete problem.
A particularly disagreeable feature of the joint RCA problem is that the number of
possible paths (and hence S) grows exponentially with the number of nodes and fibers
in the network. One way of keeping the paths under control is to place bounds on the
admissible path lengths, as was done in the previous example. This is important from
performance considerations as well, as pointed out in Section 6.2. Long paths not only
lead to poor signal quality because of accumulated noise, distortion, and cross-talk,
they also cause congestion in the network. Thus a path that requires two optical hops
consumes twice as much network capacity as one that requires only one hop.
In addition to reducing the size of the admissible path set, the RCA problem on G PI
can be simplified further by decoupling routing and channel assignment as follows:

1. Choose paths for the required connections according to some desired routing criterion
and form the subgraph of G PI induced by this choice.
2. Do a minimal vertex coloring of the chosen subgraph (a minimal wavelength
assignment).

This decomposition does not lead necessarily to a minimum-wavelength solution of

the RCA problem. For example, in the network of Figure 6.12 the routing criterion
might be to choose the lowest numbered path when more than one minimum-hop path
exists. This produces the subgraph shown in Figure 6.12(c). The chromatic number
of this subgraph is 3, with a possible minimal coloring shown in the figure. In sep-
arating the routing and wavelength assignment problem in this case, optimality was
sacrificed.
Fortunately, some good heuristics have been developed for solving the decoupled
RCA problem, so brute force techniques are not generally required. Many of these
are very simple, and computational experience has shown that they yield optimal or
close to optimal results in wide ranges of cases.17 Heuristic approaches are examined
in Section 6.3.8. The concept of the interference graph appears in a somewhat different
setting when we consider channel assignment for LLNs in Section 6.5.6.

6.3.4 Rings
As mentioned earlier, networks with some degree of symmetry make the RCA problem
simpler. Rings are an especially good choice for physical topologies because they are
symmetric and they are Moore graphs. Fiber networks in the form of SONET rings
have been deployed widely in the telecommunications industry for the past decade due
to their advanced protection and network management capabilities. As a consequence,

17
A good heuristic for minimal vertex coloring is provided in Appendix A.
 Wavelength/Waveband-Routed Networks 453

wavelength division-multiplexed (WDM) rings were the next step in the evolution of
WDM from its use in point-to-point transmission facilities to deployment in wavelength-
routed networks.
We therefore analyze static RCA in rings in this section before moving on to more
general topologies. The development is prefaced by a brief explanation of automatic
protection in rings. This digression is inserted because fault protection is such an impor-
tant issue in optical networks and because the method of protection influences the way
wavelength assignment must be performed in a WDM ring. The section ends with some
ideas on extrapolating the RCA methods for rings to more general mesh topologies,
using ring decomposition techniques.

6.3.4.1 Shared-Protection Rings

Shared-protection rings (SPRINGs), also known as MS-SPRINGs (multiplexed section
shared-protection rings), have been used for many years in traditional networks to
provide for rapid fault recovery using automatic protection switching (APS). At this
point we briefly describe a few basic features of a WDM SPRING as a prelude to
discussing the RCA problem. (Chapter 8 treats the issue of fault recovery in rings in more
depth.)
Figure 6.13 shows an example of a four-fiber WDM SPRING. It has a bidirectional
fiber pair in each link, used as working fibers, paralleled with another pair used for

WADM Optomechanical switch

Protection
Working
Working
Protection
Node 1

Node 5 Node 2

Node 4 Node 3

Figure 6.13 A four-fiber SPRING.

454 Multiwavelength Optical Networks

Optomechanical Switch

WADM Working 1

Working 2
Node 1

Node 5 Node 2

Node 4 Node 3

Figure 6.14 A two-fiber SPRING.

protection. Stations access the ring through WADMs at each node. Section 8.3.2.2
describes the protection properties of this ring. Under normal operation, using shortest
path routing, traffic from node 1 to node 2, for example, is carried in the outer (clockwise)
working fiber, and traffic from node 2 to node 1 is carried in the inner (counterclockwise)
working fiber.
To recover from link or node failures the optomechanical switches close, sealing off
the failed element, and reroute the affected connections the “long way” around the ring
using the protection fibers.18 Figure 6.14 shows a two-fiber WDM SPRING. Because
there are two working fibers and no protection fibers, each working fiber reserves half
of its capacity for failure recovery. This is done by using no more than half the number
of available wavelengths on each fiber. The unused wavelengths serve as protection
when traffic is looped back under fault conditions. To ensure that no distinct channel
assignment (DCA) violation occurs in case of failure, the working wavelengths in each
direction must be distinct.
In this section we focus on four-fiber SPRINGs, so the wavelength assignments on the
two counterrotating working fiber rings can be made independently, reusing the optical
spectrum on each ring. It is easy to show that a two-fiber SPRING requires exactly

18
Another protection technique, called path protection, maintains an additional connection on a different
path for each active connection (see Section 8.2.2.1).
 Wavelength/Waveband-Routed Networks 455

N
1
2

Figure 6.15 Bidirectional ring.

twice the number of wavelengths as a four-fiber SPRING to support the same logical
connectivity.

6.3.4.2 RCA in WDM Shared-Protection Rings

We examine the RCA problem in an N -node, four-fiber shared-protection ring; that is,
a network with working fibers that form a bidirectional ring, as shown in Figure 6.15.
The objective here is to determine routing and wavelength assignments over the ring to
provide prescribed logical connectivity among the N nodes using a minimum number
of wavelengths. A related objective is to do the assignments so that the ring can be
scaled up while trying to disturb the minimum number of existing assignments. It is
assumed that each LC is carried by a dedicated optical connection, so it is not necessary
to distinguish between logical and optical connections.
In most of this section, it is assumed that each node is associated with a nonblocking
NAS; that is, the station accesses its node through a two fiber pair access link.19 The
bidirectional ring was first encountered in Section 3.3.2. A typical node–station con-
figuration is shown in Figure 3.19, in which each access fiber pair connects to a 2 × 2
multiwavelength switch acting as a WADM. (There is no λ-interchange in the nodes.)
In effect, the bidirectional ring in Figure 3.19 is equivalent to two unidirectional rings,
each containing N WADMs accessed by elementary stations.
We present a routing and wavelength assignment algorithm that achieves full bidirec-
tional connectivity among all stations using a minimum number of wavelengths. The
algorithm is scalable in the sense that the size of the ring scales gracefully while try-
ing to disturb the minimum number of existing wavelength assignments [Ellinas+98b].
Routing is always on the shortest path.

19
We shall, however, introduce a single fiber pair access link at one point in the development to illustrate the
effect of elementary NASs.
456 Multiwavelength Optical Networks

Number of Wavelengths Required for Full Connectivity

In determining the minimum number of wavelengths required for full connectivity, we
distinguish three cases: N odd, N /2 odd, and N /2 even. It can be shown [Ellinas+98a]
that the number of wavelengths necessary and sufficient for full connectivity is
 2
N −1
, N odd
 8



2
W = N 8+4 , N /2 odd (6.16)


 N2

8
, N /2 even.
For example, the number of wavelengths required for the N (N − 1) connections on rings
of size N = 7, 8, 9, and 10 are, respectively, W = 6, 8, 10, and 13.
Recall that lower bounds on W for the cases N odd and N even were derived in
Section 6.3.1 using the limiting cut bound. Equation (6.16) indicates that those bounds are
tight for all cases except N /2 odd. If we replace the assumed nonblocking access stations
with elementary stations (with single fiber pair access links) and apply the limiting cut
bound to the access links, we find that at least W = N − 1 wavelengths are required for
full connectivity on any N -node network. This implies that for N < 6 a single-fiber-
pair access link causes a bottleneck, necessitating a larger number of wavelengths on
the ring than required with nonblocking stations. The difference between the two cases
was illustrated in Figures 3.18 and 3.19, in which RCA tables were given for a fully
connected five-node ring with single and double fiber pair access links, respectively.
These assignments are optimal, requiring four and three wavelengths, respectively.

Wavelength Assignment: Double Fiber Pair Access Case

Because the network is symmetric and shortest path routing is used, it is sufficient to focus
on connections using one direction of the ring. Half of the shortest paths between node
pairs use each direction. For N odd, shortest paths are unique, eliminating any routing
decisions. (The case of N even has a minor complication due to the fact that there are two
routing alternatives for the longest connections.) A matrix approach to RCA, valid for all
three cases and having desirable scalability properties, can be found in [Ellinas+98b].
Identical procedures are followed on both the clockwise and counterclockwise rings,
reusing the same wavelengths.

Wavelength Assignment: Single Fiber Pair Access Case

The procedure described in [Ellinas+98b] produces optimal wavelength assignments
for rings with nonblocking two fiber pair access stations. It is easy to verify that the
resultant assignments do not work for elementary single fiber pair access stations. Yet,
the limiting cut bound suggests that it might be possible to find feasible wavelength
assignments for the elementary access station case using no more wavelengths than
in the nonblocking case, provided that N ≥ 6. For N < 6 a bottleneck exists on the
access links, which in turn means that a larger number of wavelengths is required in the
single-fiber-pair access case than for the double fiber pair access case.
A general algorithm has been developed for determining routing and wave-
length assignments in rings for both odd and even numbers of stations and
 Wavelength/Waveband-Routed Networks 457

single fiber pair access links [Liu+99]. It requires the same number of wavelengths
as given in Equation (6.16) for the double fiber pair access case when N ≥ 6. The RCA
approach is based on the fact that a set of connections between access stations forming a
cycle on one ring can be assigned one wavelength that can be reused on all connections
in the set.
The algorithm for the single fiber pair access case for N even and (N ≥ 8)20 is as
follows:

1. List all possible cycles Ci, j 21 with i < j in the ring network. Route them on the
clockwise ring (CR) and assign to them wavelengths 1 to (N 2 /8 − N /4).
2. List all possible cycles Ci, j with i > j in the ring network. Route them on the coun-
terclockwise ring (CCR) and assign wavelengths as follows: λC C R( j, i) = λC R(i, j)
(where λC C R[ j, i] is defined as the wavelength in the counterclockwise ring for the
connection from node j to node i and λC R[i, j] is defined as the wavelength in the
clockwise ring for the connection from node i to node j). Two cycles Ci, j and Ck,l
routed on different directions on the ring may be assigned the same wavelength only
if i, j, k, and l are all distinct integers (this ensures that the access fiber constraint is
observed).
3. List all possible cycles Ci,i in increasing order of i. Route the first ⌈(M/2)⌉ on the
clockwise ring and assign wavelengths (N 2 /8 − N /4) + 1 to W. Assign the same set
of wavelengths to the latter part of the list in opposite order.

Arbitrary Connectivity on the Ring: One Wavelength Interchanger

The previous development assumed full logical connectivity on the ring, with a single
bidirectional connection between each node pair. Now suppose that an arbitrary LC
pattern is desired instead. Of course, if the desired connections are a subset of the full
connection pattern, the wavelength assignment rule presented for full connectivity can
be applied to the subset. But can we make do with fewer wavelengths in this case? It
turns out that this is often possible, provided that one wavelength-interchanging node is
included in the ring.
In general it can be shown that with one wavelength-interchanging cross-connect
(WIXC) in the ring, any set of connections can be set up using a number of wavelengths
determined only by the link capacity bounds (i.e., the DCA constraint). This constraint
requires only that enough wavelengths must be available to accommodate the maximum
number of connections carried on any fiber. Because this does not take into account
wavelength continuity, the implication is that a single wavelength interchanger in a ring
provides the same performance benefits achievable by putting WIXCs everywhere!

20
The case N = 6 uses the same approach but requires a special wavelength assignment. The case N odd is
treated by first deleting a node, solving the resultant problem for N even, reinserting the deleted node, and
then adding the necessary wavelength assignments for the added node.
21
A cycle Ci, j consists of four connections [(i, j), ( j, i ′ ), (i ′ , j ′ ), and ( j ′ , i)] in an even-node ring where
the N ring nodes (N = 2M) are labeled 0, 1, 2, . . . , (M − 1), 0′ , 1′ , 2′ , . . . , (M − 1)′ in clockwise order
around the ring and a pair of nodes i and i ′ constitute the end points of a ring diameter Di . When i = j, a
cycle Ci,i consists of only two connections: (i, i ′ ), (i ′ , i).
458 Multiwavelength Optical Networks

1 1

2 1 2

1 2 1

1
2 2

3 1

(a) (b)

Figure 6.16 Five-node WDM ring.

The reason why this is so is that a WIXC “breaks” the ring into an open linear chain
as far as the wavelength continuity condition is concerned. But for a prescribed set of
connections on an open-chain network there always exists a wavelength assignment that
satisfies the DCA and wavelength continuity conditions such that the required number
of wavelengths is equal to the number of connections on the maximally loaded link
[Bala+96]. We do not provide a proof here; we leave that as an exercise for the reader.
To illustrate, consider the five-node WDM ring shown in Figure 6.16(a). Five con-
nections are shown with a maximum link loading of two. Because of the wavelength
continuity constraint, the number of wavelengths required is 3. Now if a WIXC is
present at node 1, as shown in Figure 6.16(b), the same five connections require only
two wavelengths.
The economic implications of this result are substantial, suggesting that the benefits
of wavelength interchange technology can be obtained for a modest increase in overall
cost by installing this technology at only a few nodes (e.g., less than 10%) in a large
network. If the wavelength interchange results applied only to rings, they would be of
limited value. However, in the next section, we show that results for rings can sometimes
be generalized to much wider classes of networks.

6.3.5 Ring Decomposition of General Mesh Networks

We have seen that rings have a number of desirable properties that make them attrac-
tive in numerous applications. An application already mentioned is the use of a ring
as a basic topological structure for public carrier networks, as in the case of SONET
rings. Another is as a high-speed local or metropolitan area data network. Early ex-
amples of the latter are token rings and their higher speed counterparts: the fiber dis-
tributed data interface (FDDI) [Ross89] and the distributed queue dual bus (DQDB)
[IEEE91].22

22
DQDB was actually an open-bus network that could be configured physically as a ring to aid in fault
recovery.
 Wavelength/Waveband-Routed Networks 459

A G

E F

C D

(a) (b )

Figure 6.17 Ring decomposition.

Although the topology of a general mesh network is richer than a ring, there are
some advantages in decomposing a general network topology into a set of rings that
“cover” the network. Two examples are shown in Figure 6.17. Figure 6.17(a) shows a
network with N = 9 nodes and diameter D = 4. A bidirectional ring decomposition of
the network is shown as dashed lines in the figure. Figure 6.17(b) shows a unidirectional
ring decomposition of a hexagonal grid. The grid contains N = 24 nodes and has
a diameter of D = 7. Note that “bridge” nodes (indicated as shaded circles in each
network) are placed at strategic points in the network, linking two or more rings. The
placement of the bridge nodes is chosen so a path can be found through each network
between any node pair, which consists of a concatenation of segments confined to rings
(intraring segments). An end-to-end path between nodes on different rings hops from
ring to ring at the bridge nodes. As indicated later, the network’s performance can often
be improved by equipping bridge nodes with special functionality.
Ring decompositions of networks can be derived from the following general properties
of graphs.

Eulerian Networks
A large class of networks, the Eulerian networks, are amenable to decomposition into
bidirectional rings. This is done by finding a closed Eulerian trail through the network.23
(Algorithms for finding Eulerian trails can be found in [Fleischner90].) Because an
Eulerian trail usually passes through some nodes more than once, it can generally
be broken into sets of smaller cycles. By using both directions on each cycle (that
is, “double tracing” it), an Eulerian network with a single fiber pair in each link can
be decomposed into a set of bidirectional rings, intersecting at some of the network
nodes – see Figure 6.17(a) for an example.

23
An Eulerian network is a connected network, all nodes of which are of even degree. An Eulerian trail is
a trail that traverses every node in the network at least once and every link exactly once (see Appendix A
for properties of Eulerian graphs). The existence (or nonexistence) of Eulerian trails is at the heart of the
famous Konigsberg bridge problem, which marked the birth of graph theory [Harary72].
460 Multiwavelength Optical Networks

General Networks: Cycle Double Covers

The unidirectional ring decomposition of Figure 6.17(b) is derived from the concept of
an orientable cycle double cover. A well-known conjecture in graph theory is that every
bridgeless graph possesses an orientable cycle double cover: a set of cycles that “covers”
all edges, traversing each exactly once in each direction. (A more precise definition is
given in Appendix A.) The double-traced cycles in an Eulerian graph are a special case
of an orientable cycle double cover. The unidirectional rings in Figure 6.17(b) are a
subset of a cycle double cover of that graph. (Note that the graph in Figure 6.17[b] is
not Eulerian, so it cannot be covered using an Eulerian trail.)
In wavelength-routed networks, all nodes are WSXCs so that end-to-end connections
can, in principle, be made directly through the network without any additional equip-
ment. These wavelength-routed connections can often be made on paths shorter than
those confined to concatenated rings. (For example, a path from node A to node B
using the unidirectional ring decomposition in Figure 6.17[b] requires 11 hops, whereas
the shortest path between the nodes is only of length seven.) Nevertheless, the ring
decomposition of a general network is useful as a basic structure to support a number
of performance enhancements.

Sparse Wavelength Interchange

A network with bridges that are wavelength interchangers implements sparse wavelength
interchange throughout the network. If the bridges are located at well-chosen points,
they have the ability to nullify the wavelength continuity constraint and thereby reduce
the number of wavelengths required to support a given logical connectivity. Each bridge
effectively breaks all rings to which it belongs into open linear chains, as was explained in
Section 6.3.4.2. Under static routing rules, it is only the maximal link load that determines
the wavelength requirement on each “broken” ring. Wavelength interchange is also
important in reducing connection blocking under dynamic (demand-assigned) traffic
conditions (see Section 6.4). It has been verified by numerous simulation experiments that
networks with large diameters can benefit most from wavelength interchange, because
long paths tend to interfere with many other paths. In these cases it is reasonable
to expect that wavelength-interchanging bridges may be useful in breaking up long
paths into smaller segments to find wavelength assignments that satisfy DCA on each
fiber.
To illustrate, suppose the bridges in Figure 6.17(b) are WIXCs. Any connection
between nodes on different rings can be considered a concatenation of segments within
rings, joined at the bridges along the path. Note that it is possible to move among all
rings using the bridges. Because all rings are broken, wavelength continuity does not
constrain the wavelength assignments under static routing conditions.
Now consider a case of dynamic traffic in which the following optical connections are
in place using the wavelengths indicated by the subscripts: (C, D)1 , (D, E)2 , (E, F)3 ,
and (F, G)2 . Assume that wavelength rearrangement is not permitted, shortest path
routing is required (i.e., we ignore the orientations of the ring decomposition), and there
is a total of three wavelengths available. Then if a new connection (C, G) is requested,
it would have to be blocked if there were no wavelength interchangers. However, on the
 Wavelength/Waveband-Routed Networks 461

network with wavelength-interchanging bridges, the new connection can use wavelength
2 on the first link, and wavelength 1 on the remaining links.

Bridges as Overlay Processors

Another possible use for ring decomposition is in providing convenient regeneration
points for long paths. For example, in the network of Figure 6.17(b), it may be that
some of the paths in the network are too long for acceptable signal quality. In that case,
placement of regenerative repeaters at the bridge nodes may solve the problem (see
Section 2.5). Using the indicated unidirectional ring decomposition, the longest path
between bridges is 5 hops, as opposed to 7 hops for end-to-end connectivity on the
complete grid. Assuming that a 5-hop purely optical path is acceptable but 7 hops is
not, the longer paths can be realized as concatenations of segments on the rings, with
regeneration at the bridges ensuring end-to-end signal quality. (Shorter end-to-end paths
can bypass the repeaters using purely optical switching.) For example, a connection
(A, B) requires 7 hops without repeaters and 11 hops on four segments in the ring
decomposition, but no segment has more than four optical hops.

Bridges as Logical Switches

As noted, rings are a preferred topology in traditional voice and data networks. Thus,
to provide some “backward compatibility” in evolving wavelength-routed networks, it
may be worthwhile to set aside a portion of the optical spectrum to configure a ring
decomposition of a network as a logically switched overlay (see Section 2.6). This can
be particularly useful for exploiting the management capabilities that have been highly
refined for ring structures. In this case, the bridges are in the form of LSNs, in which
traffic hopping from one ring to another is processed electronically before being passed
to the next ring.
As an example, in the case of bidirectional rings, one wavelength, λd , in each direction
might be set aside to configure the overlay as a data network. In that case each overlay
station would access an ONN using λd in both directions on its ring, forming the
usual bidirectional access configuration. For traffic moving from one ring to another,
the bridges would access each ring to which they are connected, performing a routing
function as is done in typical LAN and MAN bridges or routers (see, for example,
[Halsall96]).
As shown in the example of Figure 6.18, traffic from ring 1 in the form of data frames
flows through the bridge, which “filters” the frames, reading destination addresses and
detecting those addressed to stations on ring 2. It copies those frames and transmits them
onto ring 2. The reverse process from ring 2 to ring 1 is also executed. Note that only
the clockwise direction has been shown on each ring, but both directions are present
using the same wavelength, so shortest path routing can be used. Each node performs
an add/drop operation on wavelength λd , so all optical connections are only one hop:
from each station to its neighbor on each side. (In general, a bridge can access more than
two rings, and the filtering and routing functions are basically the same.) One possible
application of the data ring overlay on a general optical network might be as a means of
distributing information for network control and management.
462 Multiwavelength Optical Networks

Bridge
1

Figure 6.18 Bridged ring overlay.

In a ring decomposition of a network, the optimal size of the rings is an open question.
If a network is decomposed into many small rings, there are many routing alternatives,
so routes limited to concatenated ring segments might approach shortest path routing.
However, a large number of rings means many intersections, which in turn requires many
bridges. Conversely, a few large rings imply long paths, high link loading, and a larger
number of required wavelengths. But fewer bridges are needed.

6.3.6 Multistar Wavelength-Routed Networks

The multistar topology illustrated in Figure 6.19 is a direct generalization of the star
configuration, discussed extensively as a shared broadcast medium in Chapter 5. As

T R
1 1
1
2 2
2
3 3
3
4 4
4
5 5

6 6

Figure 6.19 A multistar network.

 Wavelength/Waveband-Routed Networks 463

shown in the figure, each station transmits to several stars through separate outbound
access fibers and receives signals from several stars through a separate inbound access
fiber for each star. In the case shown, each subnet is a 3 × 3 star, and there are four
stars connecting six stations. Each station has two transceivers. The convention used
throughout this section is that like-numbered transmitters and receivers belong to the
same access station.
In a general (static) multistar WRN, we assume that the central nodes of the stars
are wavelength routers providing full bipartite connectivity between their transmitting
and receiving sets.24 That is, an optical connection is created from any transmitter
on a star to any receiver on the same star by tuning to a unique wavelength associ-
ated with that connection. This is known as permutation routing. If the stars provide
full fiber connectivity between all transmitting and receiving stations, then any sta-
tion can connect to any other station by finding a star to which they are both attached
and tuning to the wavelength supporting a connection between that source-destination
pair.
We first encountered permutation routers in Section 2.3.1.2, where an n × n router
was exhibited (Figure 2.10) with the following routing rule. A signal on input port j,
carried on wavelength λk , is routed to output port i, where k = (i − j) mod n. In the
permutation router, full connectivity among all input and output ports is achieved using
n wavelengths.
Now consider a general, regular multistar network containing M stars based on r × r
permutation routers operating on W = r wavelengths, where we call r the size of the
star. The network has N stations, each transmitting to and receiving from  stars,
where M = N /r . It turns out that a multistar configuration of this type, with full fiber
connectivity among N stations, using M stars of size r exists, provided that certain
relations hold among the parameters M, , r , and N . Systematic design procedures for
these networks are given in Sections 6.5.7 and 7.5.4.
Assuming that a full fiber connectivity pattern exists among the stations in a multistar
wavelength-router-based WRN, full optical connectivity is established by equipping each
station with  arrays of r fixed-tuned transceivers. Each transmitter array is connected
to one outbound access fiber, and each receiver array is connected to one inbound access
fiber (not necessarily on the same star). In this way each station acts as the source
and destination of r optical connections, so that full connectivity (including loopback
connections) is achieved.
To illustrate, the network of Figure 6.19 has N = 6, M = 4,  = 2, and r = 3. With
W = 3 wavelengths, each star provides nine connections between its transmitting and
receiving sets, which are enough to provide full bipartite connectivity between the sets.
All nine connections are necessary on stars 2 and 3 because the subnets are directed stars
with disjoint transmitting and receiving sets. However, the transmitting and receiving

24
In a simpler configuration, each central node in Figure 6.19 is a star coupler, so each transmitting station
multicasts to three receiving stations and shares each star with two other transmitting stations, creating a
hypernet logical topology embedded on an LLN. Multistar LLNs based on optical multicast are discussed
in Sections 6.5.7 and 7.5.4.
464 Multiwavelength Optical Networks

sets for stars 1 and 4 are identical, and thus three of the connections are loopback
connections. Excluding these, the total number of useful connections is 30, which is just
enough for full connectivity.
Multistar networks are very efficient in utilizing the optical spectrum (they provide M-
fold spectrum reuse) because their fiber topology is tailored to the desired connectivity
pattern. Furthermore, they offer high throughput, but at the (inevitable) expense of using
large numbers of optical transceivers. A negative feature of the multistar structure is that
it requires very long fibers, and the fibers are not arranged in a fashion that enhances
survivability. (There are no alternate paths between pairs of stations.) For example,
suppose we wish to connect N stations with geographic locations that are on a circle of
diameter D. Comparing the total fiber length for a large multistar, a single star, and a
ring, we find the approximate values, N D, N D, and 2π D, respectively.25 Clearly, the
multistar is the worst from the point of view of fiber length, but it has the best optical
efficiency. This suggests that multistars are best for use as LANs interconnecting high
throughput end systems over short distances; for example, as interconnection networks
for high-speed processors (see, for example, [Birk91]).

6.3.7 RCA as an Optimization Problem

In this section we consider the static RCA problem in its most general form. Any
network topology is allowed, and networks with and without wavelength interchange are
considered. Routing and channel assignment are considered together as an optimization
problem using MIP formulations. These turn out to have extremely large numbers
of variables and are intractable for large networks. For tractability the variables must
typically be “pruned” by limiting the admissible solutions. This can be done, for example,
by allowing only shortest path routing.
Joint routing and channel assignment can be cast as an optimization problem in
a number of different ways using various cost functions. Some possibilities are the
following:
r Establish all connections using a minimum number of wavelengths.
r Establish all connections using minimum path lengths.
r Maximize the number of connections established, subject to a constraint on the number
of wavelengths and/or the path lengths.

The following is an MIP formulation for solving the RCA problem. The network’s
physical topology is modeled as a set of nodes, N = {1, 2, . . . , N } and a set of links
L = {l x y }, where l x y denotes a unidirectional link from node x to node y. The elements
of N may be either the ONNs alone or the ONNs and the access stations. In these
two cases, the elements of L are, respectively, the internodal links alone or both the
internodal and access links. The available spectrum consists of a set of wavelengths
 = {1, 2, . . . , W }, where W = ||. Two versions of the problem are presented: for

25
This assumes that the stations on each star are positioned randomly around the circle, so that the best
placement of the central node is at the center.
 Wavelength/Waveband-Routed Networks 465

networks without and with wavelength interchange. We refer to these as RCA-I and
RCA-II, respectively.

6.3.7.1 No Wavelength Interchange: RCA-I

The prescribed traffic on the network is given in the form of a logical topology to be
realized by a set of optical connections C = {ck }, where ck = (sk , dk ) is a connection
with a fiber path and wavelength that are to be determined.26 In the case considered here,
the objective is to find an embedding of the prescribed connections, which minimizes a
linear combination of total wavelengths and total path lengths. Each ONN consists of a
WSXC, and there are no wavelength interchangers in the network. (In the case when the
model includes access stations, the stations are assumed to be equipped with arrays of
optical transceivers that cover the full set of available wavelengths.)
The wavelength continuity constraint is represented by replicating the network topol-
ogy into as many layers (copies of the network) as the number of available wavelengths.
To use wavelength w for a connection, the source and destination stations access the
wth copy of the network, as visualized in Figure 6.20. Note that in this representation
of the topology the number W of available wavelengths must be known in advance, so
the number of variables and equations in this formulation increases in proportion to W .
There is, therefore, a trade-off involved in the choice of W : Too large means too many
variables, but too small means that a solution may not exist.

WSXC

Wavelength 1
Access Fiber

λ-Interchange
NAS

Wavelength 2

Wavelength 3

Figure 6.20 Layered view of RCA.

26
Note that more than one ck may have the same source-destination pair, indicating that several parallel
connections are required for that pair.
466 Multiwavelength Optical Networks

In designating parameters and variables, subscripts are used to indicate physical

layer entities (e.g., nodes, links, and wavelengths), and superscripts are used to indicate
connections. The network parameters used in this formulation are
r f : The number of fibers on link l
xy xy
r d : A weight assigned to link l ; a positive constant “length” chosen for design
xy xy
purposes
r D k : The maximum permissible path length for connection c (a positive constant)
k

The variables used are

r k : A binary variable equal to 1 if connection c is carried on wavelength w and 0
w k
otherwise
r a k : A flow variable equal to 1 if connection c is carried on wavelength w on link
w,x y k
l x y and 0 otherwise
r ew : A binary variable equal to 1 if wavelength w is used for at least one connection
and 0 otherwise

Now RCA-I can be formulated as follows.

Minimize
 
k
m ew + dx y aw,x y (6.17)
w k w,x y

With
 k

 w , if x is the source of ck

 −k , if x is the destination of c

 k
 k w k
j = x aw,x j − j = x aw, jx = (6.18)


 0 otherwise

∀w ∈ , ck ∈ C, x ∈ N


k k

dx y aw,x
w,x,y y ≤ D , ∀ck ∈ C (6.19)
 k
ew = min 1, k w , ∀w ∈  (6.20)
 k
k aw,x y ≤ f x y , ∀w ∈ , x, y ∈ N (6.21)
k

w w = 1, ∀ck ∈ C (6.22)
k
aw,x y ∈ {0, 1}, ew ∈ {0, 1}, kw ∈ {0, 1}. (6.23)

In Equation (6.17), a large value of m is chosen to emphasize wavelength minimization,

and a small one is selected to emphasize total path length minimization.
Equation (6.18) is the flow conservation equation, which states that a connection
ck entering a node x on wavelength w must leave the node on the same wavelength,
thus ensuring wavelength continuity. If x is a source (destination) node for ck , the flow
conservation relation is completed with the additional term kw (−kw ) to account for
flow entering (leaving) the network.
Figure 6.21 illustrates how external traffic is modeled in the cases when access
stations are either excluded from or included in the model. In Figure 6.21(a), node x
 Wavelength/Waveband-Routed Networks 467

k k
Ωw x Ωw x j

(a ) (b)

Figure 6.21 External traffic in flow conservation equations.

is an ONN, and Equation (6.18) describes flow conservation including external flow. In
Figure 6.21(b), node x is an NAS accessing the network through node j. Now for node
x, Equation (6.18) reduces to the form:
 k

 w , if x is the source of ck

 −k , if x is the destination of c

k k w k
aw,x j − aw, j x = (6.24)


 0 otherwise

∀w ∈ , ck ∈ C.


Inclusion of access stations in the model is useful for dealing with the constraints of
elementary access stations, wherein the presence of an access link containing a single
fiber pair limits the number of connections that can be supported. Also, extra constraints,
such as limited tunability of transceivers, are accommodated easily when access stations
are included in the model. For example, the variables kw can be constrained to exclude
forbidden wavelengths.
The inequality in Equation (6.19) specifies a limit on the length of any path. (It can
be relaxed if there is no limit on the path length.) The limit can be used in a number
of ways. For example, if all link weights are set to 1 and D k = d(sk , dk ) – the optical
hop distance between source and destination – then the inequality specifies that only
minimum-hop paths can be used. In this case, the inclusion of the path length summation
in Equation (6.17) is redundant. However if D k > d(sk , dk ), some opportunity is left for
choices of longer paths, and inclusion of path lengths in the objective function may be
appropriate.
A very important role of these path length constraints is to reduce the complexity of
the problem by reducing the number of possible choices of fiber paths. For example,
if only shortest paths are allowed, a list of all shortest paths can be computed before
proceeding with the rest of the optimization problem. If the shortest paths happen to be
unique, the routing problem is solved, and what remains is channel assignment. Relating
this to the two-step formulation of the RCA problem described in Section 6.3.3, the
computation of the unique shortest paths completes the first step.
The second step is solved by coloring the resultant subgraph of G PI . If the shortest
paths are not unique, we still have a joint RCA optimization problem but of considerably
reduced dimensionality. In effect, the paths satisfying the inequality in Equation (6.19)
468 Multiwavelength Optical Networks

k
define a reduced set of alternatives for the flow variables aw,x y , forcing most of them
to 0. In this case it is generally more efficient to reformulate the problem using a set of
permissible paths rather than the flow variables.27 Optimization then reduces to choosing
a subset of the permissible paths and assigning wavelengths to them, as is done in the
path interference graph approach. The details are left as an exercise for the reader (see
Problem 11).
The significance of the remaining equations in the optimization formulation is as
follows:
r Equation (6.20) indicates whether a wavelength w is used for the embedding.
r Equation (6.21) is a constraint that limits the number of connections on a link and a
wavelength to the number of fibers on that link.
r Equation (6.22) requires that each connection be assigned to one and only one wave-
length.
When used with care, this formulation can be a useful optimization tool. However,
with no constraints to reduce its complexity, the number of variables and constraints can
become enormous, even for problems of relatively small size. Thus, for a network with
L bidirectional links, N nodes, and M connections, there are 2L M W flow variables
and N M W flow conservation equations. Another difficulty with this and most other
MIP formulations of the RCA problem is that solutions are not unique. Given an optimal
routing and channel assignment, another one can be found by relabeling the wavelengths.
Thus, there are at least W ! optimal solutions. Nonuniqueness of solutions can sometimes
wreak havoc on optimization algorithms.
To illustrate the importance of pruning the number of choices in an optimization
calculation, consider the problem of embedding optimally a fully connected 10-node
logical topology into a Petersen network (Figure 6.41). If elementary access stations
are used, the limiting cut bound indicates that this requires at least W = 9 wavelengths.
Modeling the physical topology to include the access stations, we have L = 25 and
M = 90. Taking the minimum possible value for W yields 40,500 flow variables in this
formulation, and it takes a considerable amount of time to obtain a feasible (but not
optimal) solution. (RCA assignments using nine wavelengths were found by computer,
but the path lengths were not minimum.) In contrast, an optimal solution to the problem
was found by hand very quickly using shortest paths and exploiting the symmetry of
the network (this is left as an exercise for the reader).28 This experience was, to say the
least, humbling for the computer but gratifying for the human!

6.3.7.2 Full λ-Interchange: RCA-II

Introduction of wavelength interchange at each node makes the WRN equivalent to
a traditional electronically switched telecommunication network in which each λ-
channel is equivalent to a “trunk.” When full λ-interchange is allowed, the RCA
27
A set of k shortest paths between all node pairs can be found in polynomial time.
28
Because the Petersen network has unique shortest paths between all node pairs, the RCA problem reduces
to channel assignment alone if a shortest path constraint is imposed.
 Wavelength/Waveband-Routed Networks 469

problem can still be visualized in terms of the layered picture of Figure 6.20. How-
ever, in this case, optical paths are allowed to move between the layers. The vertical
lines in the figure represent the links between different wavelength layers created by
λ-interchangers.
Because wavelength continuity is absent in this case, the separation of the wavelength
layers in the MIP formulation is no longer necessary. This means that the number of
wavelengths W need not be fixed in advance, nor do we need separate flow variables for
each wavelength. Therefore, the formulation of RCA-II simplifies to

Minimize

mW + dx y axk y (6.25)
k lx y

With


 1, if x is the source of ck


 −1, if x is the destination of ck
axk j − a kj x
 
j=x j=x = (6.26)


 0 otherwise

∀ck ∈ C, x ∈ N


dx y axk y ≤ Dk ,

lx y ∀ck ∈ C (6.27)
k

k ax y ≤ W f x y , ∀x, y ∈ N (6.28)
axk y ∈ {0, 1}, W integer. (6.29)
Comparing RCA-I with RCA-II, we see that the subscript w is missing from all
variables in RCA-II because there is no need to maintain counts of flows for each
wavelength. Instead, Equation (6.28) places a global limit on the number of connections
using each link. Another difference is that a feasible solution always exists for the problem
with wavelength interchange, provided that the path length constraints are sufficiently
loose. This is not the case in the formulation without wavelength interchange because
the choice of W may be too small to admit a feasible solution.

6.3.7.3 Sparse λ-Interchange

Although networks with full wavelength interchange may have better performance than
networks without wavelength interchange, the improvement may not be worth the price.
(WIXCs are considerably more complex and hence currently more costly than WSXCs.)
A middle ground was proposed by [Subramaniam+96b] and [Bala+97], who showed
that by placing WIXCs judiciously at a small number of sparsely dispersed locations
in the network, the full performance benefits of this technology can be obtained. An
example of this approach is the use of wavelength-interchanging bridges in the ring
decompositions discussed in Section 6.3.5. This suggests that with a small increase in
overall network cost, the performance benefits associated with wavelength interchange
can be obtained. Another variant of this approach is to install “limited” wavelength
interchange at all the nodes in the network [Gerstel+97, Yates+96]. In this case, the
470 Multiwavelength Optical Networks

wavelength interchange function is limited to a small set of permutations (i.e., it is not

nonblocking). Of course, the two concepts can be combined to provide sparse and limited
wavelength interchange in the network.
The RCA-I formulation for optimal routing and wavelength assignment can be adapted
to networks with sparse wavelength interchange. Referring back to Figure 6.20, the
inclusion of a WIXC at a given network node corresponds to the inclusion of the vertical
lines as sets of extra links, allowing paths to move from one layer to another. With
appropriate modifications of the flow conservation Equation (6.18) to take into account
the possibility of switching wavelengths at the WIXC nodes, the RCA-I formulation is
still valid. An alternative approach in this case is to include wavelength interchangers
everywhere but assign weights to the wavelength-interchanging links to penalize the
use of wavelength interchange. In this way, connection patterns are created that tend to
avoid wavelength interchange if possible, thereby allowing the elimination of many of
the original WIXCs.
The issue of wavelength interchange is revisited in the context of dynamic traffic in
Section 6.4.

6.3.7.4 The Fluid Flow Bound

One additional simplification is possible to make the optimization problem of RCA-
II much more tractable: removal of the integer constraint. So far, Equations (6.25)
through (6.29) constitute an MIP. Relaxation of the integer constraint in Equation (6.29)
and replacement by axk y ≥ 0 converts the MIP to an ordinary linear program, which is
much easier to solve. In this case the connections are treated as fluid flows, and thus
the optimal flow variables generally are not integers. Because each connection requires
one unit of capacity (one λ-channel), a fractional value of axk y indicates a bifurcated flow,
which is not physically realizable in a WRN. The value of W found in solving the linear
program represents the minimum capacity of each fiber (in units of λ-channels) required
to accommodate the aggregate flow in the network with optimized (bifurcated) routing.
This value generally is not an integer so it also is a nonphysical result.
The fluid flow formulation is an example of a “multicommodity” flow problem (see
Appendix D). In this version we are attempting to optimize routing to minimize required
capacity for prescribed flows, but various alternatives are possible. For example, we
might wish to maximize carried traffic subject to a bound on capacity or minimize the
average path length (by setting m = 0 in Equation [6.25]).
Although noninteger solutions of the fluid flow problem are nonphysical, they can
be useful in placing a lower bound on the wavelength requirement. Thus, if we let
W I∗ = w ew be the number of wavelengths found in solving RCA-I (no λ-interchange),


W I∗I be the number of wavelengths found in solving RCA-II for the same problem (with
full λ-interchange), and W L∗P be a solution of the fluid flow version of the same problem,
we then have
W I∗ ≥ W I∗I ≥ ⌈W L∗P ⌉ ≥ W Limcut ≥ W Limcut . (6.30)
These inequalities follow directly from the fact that in proceeding from left to right,
each value is the solution of a problem that is a relaxed version of the previous one.
 Wavelength/Waveband-Routed Networks 471

In the case of W Limcut , its value (and the value of its lower bound, W Limcut ) is obtained
by assuming that traffic (as a fluid flow) is distributed equally among the fibers in the
limiting cut. But this distribution is a “best case” and may not be possible given the flow
constraints within the partitioned network.
One way of interpreting the linear programming problem is as a limiting case of a
network in which the number of connections and wavelengths increases without bound,
whereas their granularity (required bit rate) becomes infinitely small. Thus, imagine that
each connection, ck in our original formulation, is replaced by n connections, with each
new connection operating at 1/n times the bit rate of the original connection. In this way
the total throughput remains constant in going from the original formulation to the new
one. Then if W I∗I is a solution of RCA-II, a feasible solution W (n) of the new MIP (with
an n-fold increase in required connections) exists with W (n) = nW I∗I , because for each
wavelength in the original problem there are now n wavelengths available to route the n
replicated connections in the new problem. However, for an optimal solution W (n)∗ of
the new problem we typically have W (n)∗ < nW I∗I , because the n replicated connections
can be spread over as many as n parallel optical paths, allowing for more routing
options.
So far, all we have done is increase the wavelength requirement. To complete the
fluid flow picture, it is conceptually reasonable to imagine that the n-fold increase
in connections and wavelengths is accompanied by an n-fold decrease in the optical
spectrum occupied by each wavelength because each λ-channel now requires only 1/n
times the capacity required in the original problem. This would be the case under ideal
conditions (no guard bands between λ-channels). Now, as n → ∞, the granularity of the
connections becomes infinitely fine and the connections resemble a fluid. The solution
W L∗P of the linear programming problem now represents the total optical spectrum
occupied by these infinitely fine channels.
Because of technological limitations (e.g., finite guard bands), this conceptual de-
crease in granularity cannot be pursued to any significant degree in WRNs. However,
in Section 6.5 it is shown that in LLNs it is possible to refine the granularity of LCs
to almost any degree through the use of multipoint optical connections. (Unfortunately,
granularity refinement in LLNs does not necessarily improve routing alternatives.)

6.3.7.5 Maximizing Carried Traffic

Another formulation of the RCA problem distinguishes between offered and car-
ried traffic, maximizing the carried traffic subject to network resource constraints
[Ramaswami+95]. We shall henceforth refer to this as the RS version of the RCA
optimization problem.
The network consists of L single fiber pair links and M source-destination (s–d)
pairs. It operates on W wavelengths. All prescribed connections are assumed to be full
duplex, with both directions of a connection routed on the same path using the same
wavelength. The topology of the network is described in terms of its (s–d) pairs, the set
of fiber paths between them, and the path–link incidence matrix B, which gives the links
on each path. The paths can be precomputed before the rest of the optimization process
is executed.
472 Multiwavelength Optical Networks

The parameters describing the network are

r P: The total number of available paths on which connections can be routed
r A = [ai j ]: The P × M path (s–d) pair incidence matrix, where

1, if path i is between source–destination pair j,
ai j = (6.31)
0, otherwise
r B = [bi j ]: The P × L path–link incidence matrix, where

1, if link j is on path i,
bi j = (6.32)
0, otherwise

The traffic requirements are specified as a total offered load ρ, with pi ρ being the
offered load between (s–d) pair i, i = 1, . . . , M. Thus the pi ’s define the distribution of
the offered traffic. In [Ramaswami+95], both deterministic and stochastic interpretations
of the offered load are considered. We focus on the deterministic (static) traffic case here,
and thus the offered load represents the number of connections available to be routed.
The objective of the RS formulation is to maximize the carried traffic for a given
value of ρ, with a constraint W on the number of wavelengths. This is the dual of RCA-I,
which (for large m) minimizes the number of wavelengths for a given carried traffic.
The variables used in RS are
r m : The number of connections carried between (s–d) pair i, i = 1, . . . , M
i
r C = [c ]: A P × W path–wavelength assignment matrix containing the RCA vari-
ij
ables, where

1, if wavelength j is assigned to path i
ci j = (6.33)
0, otherwise

The optimization problem is then stated as

Maximize
M

mi (6.34)
i =1

With
P

cki bk j ≤ 1, i = 1, . . . , W, j = 1, 2, . . . , L (6.35)
k =1
W 
 P
mi ≤ ck j aki , i = 1, . . . , M (6.36)
j =1 k =1
m i ≤ pi ρ, i = 1, . . . , M
m i ≥ 0, integer, i = 1, . . . , M (6.37)
ci j ≥ 0, integer, i = 1, . . . , P, j = 1, . . . , W
 Wavelength/Waveband-Routed Networks 473

Equation (6.35) expresses the fact that no wavelength can be used more than once on
any link, Equation (6.36) is a bound on the amount of carried traffic for (s–d) pair i, and
Equation (6.37) expresses the relation between carried and offered traffic. The optimal
values of the RCA variables ci j can be interpreted as the result of executing an RCA
algorithm that maximizes carried traffic.
Several other versions of the RS problem are given in [Ramaswami+95], including
a formulation with λ-interchange as well as a linear programming fluid flow bound
found by relaxing the integer constraints. They also present an alternative and equiv-
alent MIP formulation based on the path interference graph G PI , which we refer to
as RS ′ .
The first step in the RS′ formulation is to determine the set of all maximal independent
sets of G PI (see Appendix A). Wavelengths are then assigned to vertices in these sets,
with each assignment representing an optical path for one connection. Each wavelength
assigned to an independent set can be reused in that set to support as many connections
as there are vertices in the set. This assignment problem is equivalent to vertex coloring
a graph with multiple colors per vertex. In this case the objective is to maximize the
carried load, which corresponds to assigning as many colors (wavelengths) as possible
to the vertices subject to the constraints on offered traffic and available wavelengths.
Because this formulation is based on graph coloring, it becomes quickly intractable as
the number of vertices of G PI (i.e., the number of admissible paths) becomes large.
Thus, as in other formulations, either shortest path routing or similar path limitations
must be imposed to make the problem tractable.
The various RCA formulations in [Ramaswami+95] are used to obtain bounds on
carried traffic in a variety of scenarios, with and without wavelength interchange. The
reader is referred to the cited reference for more information.
As in the case of problems RCA-I and RCA-II, the usefulness of the RS formulations
depends on their dimensionality. Considering all possible source-destination paths is
out of the question because there are PW RCA variables in RS and the number of
paths P can be exponential in the number of network nodes or links. The situation in
RS′ is still worse, because the number of maximal independent sets can be exponential
in P. Because the path list is arbitrary, some subset of all paths can be used. This
would correspond to a predetermined static routing rule. For example, if single shortest
paths are selected, this is similar to our two-step separation of the RCA-I problem in
Section 6.3.3. Of course, this approach gives a suboptimal solution of the complete RCA
problem.
Although the RS formulation with a complete set of fiber paths generally has far more
variables than the RCA-I formulation, the number of variables can be reduced drastically
by using a reduced set of paths. Thus, the number of RCA variables is PW in RS, and the
number of flow variables is 2L M W in RCA-I. Now suppose k shortest paths are used
for each source-destination pair in the RS form. We then have KMW RCA variables in
the RS version compared with 2L M W flow variables in the RCA-I form. In this case
the RS form normally has fewer variables for reasonably small values of k. However,
as mentioned earlier, path pruning can also be done in RCA-I, reducing the number of
flow variables.
474 Multiwavelength Optical Networks

In making comparisons of RS and RCA-I, it must be understood that we are comparing

apples and oranges. The former formulation maximizes carried traffic using a fixed
number of wavelengths, whereas the latter minimizes the number of wavelengths used
for fixed offered (and carried) traffic. For example, in the versions of each involving
single shortest paths, both problems reduce to wavelength assignment only. In RS this
is equivalent to coloring a maximum number of vertices in G PI with a fixed number
of colors, but in RCA-I it is equivalent to coloring all vertices of G PI with a minimum
number of colors. The former will generally use fewer wavelengths than the latter if the
same number of connections are assigned in both cases. The difference is caused by
the fact that in attempting to maximize carried traffic, as in RS, one is free to drop the
“difficult” connections (i.e., those involving long paths) in favor of “easy” (short) ones.
This makes the problem easier but tends to produce unfairness, with discrimination
against the longer connections. In RCA-I, all traffic must be carried so that everyone is
treated equally.

6.3.8 Heuristics for Static RCA

A large number of heuristic RCA algorithms have been proposed for WRNs over
the years. Some early representative examples can be found in [Ayanoglu+96,
Bala+91a, Banerjee+96, Baroni+97, Chen+96, Chlamtac+89, Karasan+96,
Mokhtar+98, Mukherjee+96, Nagatsu+95, Ramaswami+94, Subramaniam+97,
Wauters+96]. Other more recent RCA heuristics are mostly variations of the earlier
proposed techniques and are not analyzed here. There is a large body of work in the
literature on RCAs that the reader is encouraged to explore.
Most of these attack the RCA problem by separating it into the two subproblems of
routing and channel assignment, resulting in an enormous simplification of the overall
problem. Some, such as that by [Banerjee+96], do some loosely coupled iterations of
both subproblems. In some cases (e.g., [Baroni+97]), the admissible paths are con-
strained, and in others (e.g., [Banerjee+96]), any paths are allowed. Although the latter
unconstrained approach might appear to be better, empirical results over a wide range
of examples show that this is not the case. Allowing paths of any length in heuristic
algorithms tends to congest the network, leading to an increase in the required number
of wavelengths. Thus, heuristics restricted to shortest paths or “almost-shortest paths”
tend to give better results in terms of spectrum usage, as is illustrated next.
Two heuristics are described in this section. They have been chosen because they
exhibit superior performance in a wide variety of test cases and at the same time are
computationally simple and efficient. The first is a minimum-hop algorithm presented
in [Baroni+97], and the second is a Monte Carlo approach [Bouillet98].

6.3.8.1 A Minimum-Hop Heuristic

The algorithm in [Baroni+97] separates RCA into its two subproblems and confines all
routes to shortest (minimum optical hop) paths. It has been shown to give optimal or
close to optimal results (in terms of wavelength usage) in fairly extensive experiments
with both actual and randomly generated network topologies.
 Wavelength/Waveband-Routed Networks 475

Given a list of M prescribed LCs, the routing portion of the algorithm attempts to
minimize fiber congestion (the number of paths using any one fiber), which tends to
minimize the number of required wavelengths. It operates as follows:

1. List all connections in random order and compute all minimum-hop paths for each
connection.
2. Assign any minimum-hop path to each connection.
3. For each connection on the list, substitute an alternative minimum-hop path for the
one previously assigned if and only if the number of connections (congestion) on
the most loaded link in the alternative path is lower than the congestion of the most
loaded link in the previously assigned path.
4. Repeat the previous step until no further substitutions are possible.

Once the paths are assigned, a wavelength assignment algorithm is executed, making
assignments to the longest paths first. The steps are as follows:

1. Group paths with same lengths into common sets and rank the sets in decreasing
order of length. Assign a numerical order to the wavelengths.
2. Select a path randomly from the first set.
3. Assign to the selected path the lowest numbered wavelength previously unused on
any link in the path.
4. Continue this wavelength assignment process for all paths in the first set. Then repeat
for the next and subsequent sets until all assignments have been made.

It should be noted that the total number of wavelengths used in this process is unknown
until the algorithm terminates. Thus, the RCA process may fail if it results in a wavelength
requirement greater than the total number of available wavelengths.
These algorithms are noteworthy for their simplicity. Comparing them with the brute
force techniques of Section 6.3.3, we see that the complexities of the routing and
wavelength assignment algorithms are O(M) rather than being exponential or worse in
the number of connections when graph coloring is used.
In [Baroni+97], the RCA heuristic is applied to a large number of networks, including
the four shown in Figure 6.10. Recall that these networks were used in Section 6.3.1.2
to illustrate an application of the limiting cut bound. For the networks shown, the figure
gives the number of wavelengths, Nλ , needed to support full connectivity using the
heuristic RCA algorithms stated here. For comparison, W Limcut and the ratio of the
two lower bounds on wavelengths W Limcut /W Netcap are also included. The topological
parameters listed in the figure are
r N = number of nodes
r L = number of links
r α = N (N2L−1) = physical connectivity
r H̄ = average path length
r D = network diameter
r δmin = minimum node degree
r δmax = maximum node degree
476 Multiwavelength Optical Networks

70
65 N ′λ 8

60
N ′λ 16
55
N ′λ 8
50
Wavelengths Saved (%)

45

40 N ′λ 22

35 N ′λ 16 N ′λ 8
30
25

20

15

10

5 N ′λ 16

0
0 10 20 30 40 50 60 70 80 90 100 110 120 130 140
Fiber Added (%)
ARPANet
UKNet
EON
NSFNet

Figure 6.22 Wavelength savings by increasing fibers. (From [Baroni+97, Figure 11]. Copyright

c 1997 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

Note that the heuristics either achieved or came very close to Nλ = W Limcut in all
cases. In addition to the fact that the algorithms are very effective, there are at least
two other explanations for this excellent performance. First, the networks are of fairly
small diameters, with H̄ < 3 in all cases so that long paths, where considerable path
interference might be expected, are rare. This means that sophisticated RCA algorithms
and/or wavelength interchange are not necessary to conserve wavelengths. Second,
W Limcut /W Netcap ranged from approximately 1.4 to 1.9 in these networks. A large ratio
suggests that the network contains bottlenecks, forcing an increase in Nλ to provide for
the connections on the bottleneck (limiting cut) links. But this increase in turn produces
idle channels on the nonbottleneck links: 40% to 90% idle λ-channels in this case. Thus,
if it is possible to optimize RCA on the bottleneck links, the RCA problem for the
remaining connections is simplified considerably. Hence, it should not be necessary to
exceed the W Limcut limit for these connections.
In Section 6.3.1.2 it was pointed out that one way to reduce the number of wavelengths
required to support a given traffic is to add capacity in the form of extra fibers to the
links in the limiting cut. This approach is used by [Baroni+97]. The results of selective
increase in fiber capacity for the four networks of Figure 6.10 are shown in Figure 6.22.
Because ARPANet has the highest W Limcut /W Netcap ratio, it is the best candidate for
 Wavelength/Waveband-Routed Networks 477

40

35 ARPANet ( )
Mean Number of Wavelengths (Nλ)

30

25
UKNet ( )

20 EON ( )

15 NSFNet ( )

10

0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Connectivity (α)
N= 5 N = 28
N= 6 N = 35
N= 14 N = 50
N= 20

Figure 6.23 Mean values of N λ versus α. (From [Baroni+97, Figure 5]. Copyright
1997
c IEEE.
Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

capacity augmentation. For example, the number of wavelengths used is reduced from
Nλ = 33 to Nλ′ = 22 by adding one additional fiber pair to six of the network links,
following the reasoning in Section 6.3.1.2. This results in a 33% reduction in required
wavelengths achieved by a 19% increase in fiber (and no change in the physical topology).
In addition to studying real network topologies, empirical results are given for Nλ
for randomly connected networks (RCNs). Families of RCNs are generated by picking
values of N and the physical connectivity parameter α and constructing random (but
connected) networks for these values.
The parameter α represents the ratio of L, the number of links in the network,
to L FC = N (N − 1)/2, the number of links in a fully connected network, with the
same number of nodes. Noting that the average node degree is  ¯ = 2L/N , we have
¯
 = α(N − 1). An intriguing aspect of the empirical RCA results is that α seems
to be a key parameter for large classes of networks, predicting quite accurately the
mean number of wavelengths N̄ λ and minimum number of wavelengths Nλmin needed
to support full connectivity over a wide range of network sizes.29 Figures 6.23 and 6.24
show the mean and minimum values of wavelengths, respectively, for a large set of

29
Nλmin is the minimum value of Nλ over the ensemble of RCNs tested.
478 Multiwavelength Optical Networks

40

)
min
Minimum Number of Wavelengths (Nλ 35

30

25

20

15

10

0
0.0 0.1 0.2 0.3 0.4 0.5
Connectivity (α)
N = 14 N = 35
N = 20 N = 50
N = 28

Figure 6.24 Minimum values of N λ versus α. (From [Baroni+97, Figure 8]. Copyright
1997
c
IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

networks as a function of α. The networks in the figures include the four real networks
of Figure 6.10 as well as RCNs for N = 5 through 50. Note that the curves for Nλmin
are of the approximate form Nλmin ≈ 2.4/α. Very similar results were obtained for Nλ
in regular networks – undirected ShuffleNets and deBruijn networks over the range
N = 8 to 125. This suggests that regularity and symmetry of the physical topology do
not necessarily improve performance. In fact, they inhibit the network designer because
they prevent modular network growth.
These empirical results can be related neatly to the wavelength bound W Netcap . Sub-
¯ = α(N − 1), M = N (N − 1), and F = 1 into Equation (6.11), we have
stituting 

W Netcap = H̄ /α. (6.38)

This suggests that spectrum usage for the best cases of RCNs is close to the bound
W Netcap and that over the range of sizes examined, the RCNs have H̄ ≈ 2.4 independent
of α and N . (The value of H̄ = 2.4 for the RCNs is consistent with the known value
H̄ = 2.33 for the regular network examples.)
Without looking carefully at these results, one might conclude that using a fixed
number of available wavelengths, WRNs can be “grown” to large sizes while maintaining
full connectivity. There is only one problem with this: α must be held fixed as N increases,
 Wavelength/Waveband-Routed Networks 479

which implies that ¯ must grow linearly with N . That kind of growth cannot be sustained
because the cost of each network node grows faster than . ¯ In fact, it is unreasonable to
expect to maintain full optical connectivity among all nodes in large networks because
that means that N − 1 simultaneous optical connections must be supported by each
station (assuming one station per node) no matter how large N is. This in turn means
that the number of optical transceivers in each station, and hence its cost, increases
linearly with the size of the network.
A more reasonable assumption is that each station supports some fixed number
M STA of optical connections independent of network size. In that case, we have from
Equation (6.11)
M STA H̄
W Netcap = , (6.39)
¯
so with a fixed average node degree the lower bound on wavelengths does not grow with
network size as long as the average internodal distance is kept under control.30 But with
fixed M STA , how do we maintain full connectivity among the stations? The only way to
do this is to provide a logically switched overlay on the wavelength-routed network; that
is, turn it into an LRN. This is a fundamental motivation for LRNs.
Our overall conclusion concerning RCA in wavelength-routed networks is, therefore,
wavelength-routed networks cannot by themselves achieve full connectivity among a
large number of nodes in a cost-effective manner!

6.3.8.2 A Monte Carlo Approach

The RCA algorithms and optimization procedures discussed earlier are all deterministic.
A fixed set of rules is applied to determine routes and channel assignments for a set
of prescribed connections, with the outcome determined uniquely by the procedures
used. However, with optimization and design problems as complex as these, experience
has shown that stochastic approaches using simulated annealing, genetic algorithms,
or neural networks often give better results than deterministic ones and require less
computation [Aarts+97, Rayward-Smith+96, Reeves93].
We now present a stochastic approach to the static RCA problem. Its objective is
to minimize the number of wavelengths required for a prescribed connection set while
observing a given set of constraints on allowable fiber paths. (The constraints can
be chosen to suit the needs of the designer.) In a large number of test problems, a
small sample of which appear in Table 6.2, the stochastic method performs better than
competing deterministic heuristics and does so with considerably less computation. It
is a Monte Carlo technique, wherein the static problem is converted to a dynamic one.
The prescribed set of LCs is replaced by a stochastic offered load on the network, where
a sequence of connection requests is drawn at random from the prescribed set.
An attempt is made to assign a fiber path and a λ-channel to each requested con-
nection using a simple shortest path RCA rule. The characteristics of the algorithm are
“adjustable” in the sense that path constraints and link weights (or “lengths”) may be
30
Of course, H̄ does grow, albeit slowly (logarithmically), with network size, so something has to give (either
¯ to keep the bound, and hence Nλ , from growing.
M STA or )
480 Multiwavelength Optical Networks

Table 6.2 Comparative performance of three RCA heuristics.

Case N L M D d¯ ¯
 σ2 W Limcut W mc WBar WBC

1 30 45 870 6 3.133 3.0 0.0 38 40 42 48

2 30 36 870 9 4.064 2.4 0.24 72 72 83 78
3 30 38 870 10 4.306 2.53 0.249 88 89 89 94
4 40 60 1560 6 3.546 3.0 0.0 57 63 70 74
5 40 61 1560 8 3.464 3.05 0.498 59 61 62 73
6 40 54 1560 9 4.038 2.7 0.46 91 93 143 111
7 51 62 2550 10 4.898 2.43 0.245 126 138 156 158
ARPANet 20 31 380 6 2.805 3.1 0.19 33 33 33 36
UKNet 21 39 420 5 2.505 3.71 1.823 19 19 22 25
EON 20 39 380 5 2.363 3.9 2.99 18 18 18 20
NSFNet 14 21 182 3 2.143 3 0.286 13 13 13 16
Atlantis 66 107 4290 14 5.135 3.24 1.94 173 175 212 250

assigned dynamically depending on link utilization and possibly other factors. Thus,
“shortest path routing” does not necessarily mean minimum hop. (General dynamic
RCA rules are discussed in Section 6.4.) If no feasible path exists given the constraints
of the algorithm, the request is blocked and tried again later. Requests that are accepted
are held for a random time and then released. Statistics on connection blocking proba-
bility are kept, and link weights are fine-tuned to select paths and to assign wavelengths
depending on blocking statistics. Lists of “good” paths for each connection are also
accumulated as the process proceeds so that the algorithm “learns” about the network
topology and the load distribution resulting from the prescribed connections. Connection
holding times are adjusted dynamically, increasing the holding times of the connections
having highest blocking probability. In this way the more “difficult” connections tend
to become “sticky,” remaining in place for longer periods of time than the “easy”
connections.
The algorithm starts with a light Erlang load and a minimum number of available
wavelengths W . As it proceeds, the load is increased and, if necessary to reduce blocking,
W is increased. At convergence all connections are accommodated and the algorithm
terminates. Because the performance of the algorithm depends on the specific RCA
rule and constraints that are applied, this is not one algorithm but rather a class of
algorithms.
The execution of the algorithm, in the form of an event-driven simulation, is described
in more detail in the flow chart of Figure 6.25. The notation used in the figure is as follows:

r li = an LC
r Pi = a set of paths used previously for li
r P ∗ = a subset of feasible fiber paths in Pi
i
 Wavelength/Waveband-Routed Networks 481

Pi {}
W WLimcut

Process Next
Event

Departure
Choose SP, Release Connection
Request or
P i* Departure? Update Fiber
A Occupancy
Assign Lowest Request li
Feasible to Attempt RCA
Find Set of Feasible
Accept Request li Paths P i* P i
Update Fiber
Occupancy in
Schedule Departure N
P i* Empty?
for l i

Y
N Find Shortest Block Request
All Connections
Feasible Path, π, for li
Are Active? Update bi
1 i 1 i c bi
Y
N
STOP π Exists?
N
Pb Too High?
Y
Pi Pi Y
W W 1

Figure 6.25 Flow chart of the Monte Carlo algorithm.

r W = the number of available wavelengths

r π = a feasible fiber path for li
r µi = the departure rate for connection li
r bi = the current blocking probability for li
r Pb = the current blocking probability for all connections

Referring to Figure 6.25, the simulation is initialized with all path sets, Pi , empty.
The initial number of available wavelengths, W , is set equal to W Limcut , which can be
calculated using an algorithm in Appendix D. Events (either connection requests or
releases) are scheduled and processed according to the following rules:

1. Connection requests are generated for all inactive connections, with exponentially
distributed interarrival times. The average interarrival times and holding times deter-
mine the Erlang load on the system. The simulation begins with a light load, which
482 Multiwavelength Optical Networks

is increased as it proceeds. A connection, li , is released after an exponentially dis-

tributed holding time with mean µi−1 . The simulation keeps track of the fiber path π
occupied by each active connection as well as the wavelength it uses.
2. If the next event is a departure, the fiber occupancy is updated by removing the
wavelength used by the released connection from each fiber on its path, and the
simulation processes the next event.
3. If the next event is a connection request, li , an attempt is made to assign an optical
path to it. First, the set of feasible fiber paths, Pi∗ ⊆ Pi , is examined. (A feasible
path is a fiber path in Pi for which a λ-channel is available that satisfies wavelength
continuity.)31 If Pi∗ is nonempty, a “shortest path” (SP) in Pi∗ is chosen based on
any appropriate link weight criterion. For example, link weights might be adjusted
dynamically for congestion, current blocking probability, or some other phenomenon.
Having chosen a fiber path, the algorithm proceeds to the channel assignment phase
(point A in Figure 6.25).
4. If a feasible path is not found in Pi , a shortest feasible fiber path for li is sought outside
of Pi , satisfying specified path constraints. The feasible paths might be limited to
minimum-hop paths or they might be constrained in any other way. If an acceptable
fiber path π is found, it is added to the path set Pi , and the algorithm again proceeds
to the channel assignment phase.
5. If no feasible fiber path can be found that satisfies the path constraints, the request
is blocked, its blocking probability bi is updated, and its average holding time µi−1
is incremented by an amount proportional to bi . The overall blocking probability
Pb is also incremented, and if it is deemed to be too high, the number of available
wavelengths is increased by 1. The next event is then processed.
6. Once a fiber path, π , has been chosen, the lowest numbered feasible wavelength is
assigned to the connection.32 The request is accepted, the fiber occupancy on path π
is updated, and the departure for that connection is scheduled. If all connections are
active, the algorithm terminates. Otherwise, the next event is processed.

Figure 6.26 shows a typical time trace of a Monte Carlo experiment. The fraction
of active connections is plotted as a function of the cumulative number of connection
requests. With the initial value of W , the fraction of accepted requests quickly reaches
approximately 0.94. Then a transient period ensues, with the value of W incremented
several times until convergence occurs, with all requests satisfied. In this example,
convergence required 25,000 connection requests.
Table 6.2 gives the results of 12 computational experiments with the Monte Carlo
algorithm, with comparisons with two other heuristics. Seven random networks of orders
ranging from N = 30 to N = 51 nodes are listed, together with the four networks of
Figure 6.10, and a 66-node version of Atlantis – a more detailed version of the network
in Figure 3.36(b). As indicated in Table 6.2, networks with various numbers of links L,
diameters D, average node-pair distances d, ¯ average degrees ,
¯ and degree variances σ2

31
We assume that there is no wavelength interchange in the network. However, a simple modification of the
algorithm at this point would allow for full or sparse wavelength interchange.
32
A feasible wavelength is a free wavelength that satisfies wavelength continuity.
 Wavelength/Waveband-Routed Networks 483

Convergence of Monte Carlo Algorithm

1

0.99
Active Connections/Total Connections

0.98

0.97

0.96

0.95

0.94

0.93

0.92

0.91

0.9
0 10,000 20,000 30,000
Number of Requests for Connections

Figure 6.26 Time trace of Monte Carlo algorithm.

were studied. In each case a routing and channel assignment was determined for a set of
M = N (N − 1) connections (full connectivity) using W mc wavelengths. The objective
of the experiments was to determine how close W mc was to optimum, considering a range
of network sizes and topologies. The optimality of W mc was checked by comparing it
with W Limcut .33 Because W Limcut ≤ W Limcut , and the latter is a lower bound on the
required number of wavelengths for the prescribed connection set, a small difference
between W mc and W Limcut indicates a good solution and perhaps an optimal one.
To compare this stochastic approach with some good deterministic algorithms, the
performance of the Monte Carlo algorithm was compared with the deterministic heuristic
of [Baroni+97] and with the algorithm of [Banerjee+96] as shown in the columns labeled
W Bar and W BC , respectively. Note that in all cases W mc is the lowest of the three, with
W Bar = W mc in only four of the cases and W BC always greater than W mc . In five of the
cases in Table 6.2, W mc = W Limcut , which implies that the value of W mc is optimal and
that the limiting cut heuristic of Appendix D computed the true value of the limiting cut
wavelength lower bound, W Limcut = W Limcut .
Because full connectivity was required (and achieved) in all experiments, the solutions
obtained for this static RCA problem can be applied directly to the dynamic case. All
that is required is to keep a table of the route and wavelength obtained for each (s–d) pair

33
The value of W mc could not be checked against the true optimum, or even against W Limcut , because neither
one of these quantities could be computed in a reasonable amount of time for problems of the size being
tested.
484 Multiwavelength Optical Networks

and to use table lookup to do routing and wavelength assignment on demand. However,
for networks operating with less than full connectivity, this will generally produce a less
efficient utilization of the optical spectrum than the dynamic algorithms described in the
next section.
There is an important observation that comes as a by-product of these experiments.
Because all results on the required number of wavelengths W either meet or come
close to W Limcut , and because W Limcut is a lower bound on W obtained ignoring the
constraint of wavelength continuity, the results demonstrate empirically that wavelength
interchange is unnecessary in the networks studied in these experiments.

6.4 Wavelength-Routed Networks: Dynamic Routing

and Channel Assignment

Under dynamic traffic conditions, a sequence of LC requests arrives to a network con-

troller in some random fashion, and these requests may or may not be accommodated,
depending on the current state of activity in the network. The network state consists of
all active connections together with their optical path (route and wavelength) assign-
ments. The state evolves randomly in time as new connections are admitted and active
connections are released. In performing routing and channel (wavelength) assignment
under dynamic traffic conditions, an algorithm must be executed in real time to accom-
modate each request if this is feasible. If a request cannot be accepted (either because
of physical constraints or admission control limitations), it is blocked. Because of the
real-time nature of the problem, algorithms for RCA in a dynamic traffic environment
must be very simple.
The performance of a dynamic RCA algorithm is generally measured in terms of
blocking probability. However, we shall see that blocking probability as a single param-
eter may hide some other very important aspects of network behavior – in particular,
fairness. Both blocking probability and fairness are influenced by the presence or ab-
sence of wavelength interchangers in the network. These effects are examined in the
context of some specific examples.
Because connections among the various node pairs all share the same sets of links,
there are statistical dependencies among all fiber and wavelength occupancies, and
these influence blocking strongly. These dependencies make it impossible to perform
an exact analysis of blocking probabilities and fairness in networks containing more
than a few nodes. Instead, we present some case studies of special topologies (rings
and interconnected rings) through simulation and through some approximate analytical
models. For more general topologies, some simple but highly approximate expressions
for blocking probabilities are presented.

6.4.1 Some Basic Routing and Channel Assignment Algorithms

Most dynamic RCA algorithms for wavelength-routed networks have the following
general form:
 Wavelength/Waveband-Routed Networks 485

r Arrange all admissible fiber paths for a given source-destination pair in some pre-
scribed order in a path list.
r Arrange all wavelengths in some prescribed order in a wavelength list.34
r Attempt to find a feasible route and wavelength for a requested connection starting
with the path and wavelength at the top of the list.

The specific nature of an RCA algorithm is determined by the constraints used in defining
admissible paths, the order in which paths and wavelengths are listed, and the order in
which the path and wavelength lists are accessed.

6.4.1.1 Characterization of Routing

In a static routing algorithm, the admissible paths are chosen and ordered independently
of the network state, whereas in an adaptive algorithm admissibility and ordering may
vary according to the current network state. A fixed routing algorithm is a static algorithm
in which every source-destination pair is assigned a unique path (i.e., the path list for each
node pair contains only one admissible path). In a fixed routing algorithm, a connection
is blocked if there is no wavelength available on the designated path, even if a different
route for the connection exists with a free wavelength at the time of the connection
request.35 In alternate routing each connection is assigned a set of admissible paths for
which the ordering of the path list, and hence the choice of paths, typically depends on
the state of the network: the adaptive routing case.36
In most practical cases, path admissibility and ordering is based on path length. The
paths typically are listed in increasing order of path lengths, and path length is normally
defined as the sum of link weights along the path. The link weights are typically chosen
using some desirable routing criterion, and because they can be assigned arbitrarily, they
offer a wide range of possibilities for selecting path priorities. Link weights may reflect
the load or “interference” on a link (i.e., the number of active connections sharing the
link [Ayanoglu+96, Bala92]) or they may reflect the interference lengths of the paths
sharing the link (i.e., the number of hops shared by pairs of interfering paths).37 Many
other weighting functions are possible. By assigning small weights to least loaded links,
for example, we place the paths with the maximum number of free channels on their
links at the top of the path list, producing a least loaded routing algorithm. Paths that are
congested become longer and are moved farther down on the list. This tends to avoid
heavily loaded bottleneck links.
For example, in a static routing algorithm, the path list might consist of all minimum-
hop paths, listed in order of increasing geographic length. In an adaptive routing scheme,
the path with the least interference might be placed at the top of the path list. Note that

34
This assumes no wavelength interchange.
35
A slight variant of fixed routing is a procedure in which only one shortest path is considered, but a random
choice is used when more than one shortest path exists [Chlamtac+89].
36
Alternate routing schemes can also be state independent (i.e., static). For example, if a path is selected
from the list by random choice, the alternate routing rule is state independent and therefore static (but
probabilistic).
37
The term interference is used here in the sense of potential interference. That is, the connections sharing a
link would interfere with each other, garbling their information if they were assigned the same wavelength.
486 Multiwavelength Optical Networks

with static algorithms, the path lists can be computed and ordered off-line. However, in
the adaptive case, real-time shortest path computation is required for each connection
request, because link weights change with the network state.
Most routing algorithms are constrained in the sense that the admissible paths are
selected as a predetermined subset of all possible paths. For example, the set of admissible
paths for a given connection might be limited to those having no more than d + 1 hops,
where d is the source-destination distance (in optical hops).
Among unconstrained algorithms, the simplest is shortest path routing. An example of
an unconstrained shortest path algorithm is one that chooses a minimum-hop path from
among all those that have at least one free channel available (see Section 6.4.1.4). More
general shortest path algorithms comprise two parts: (1) the assignment of appropriate
link weights (“lengths”) and (2) the computation of a shortest path based on link weights.
Although shortest path routing would appear to be a very simple procedure, compli-
cations may arise because
r The link weights may be state dependent
r The path lengths may be influenced by the presence or absence of wavelength inter-
changers
r The length of a path may not be equal to the sum of the link weights

The first item means that the routing algorithm is adaptive, which in turn implies
that global state information must be stored at the site (or sites) executing the routing
algorithm. If the algorithm is executed in a distributed fashion, this information must
be exchanged among the various network nodes performing the computation, creating
significant control overhead. The significance of the last two items is explained in the
least loaded routing algorithm, described in Section 6.4.1.5.

6.4.1.2 Characterization of Channel Assignment

The definition of a free channel depends on whether the network contains wavelength
interchangers. If there is no λ-interchange, a free channel (wavelength) exists on a path
only if at least one wavelength is unused on every fiber on the path (i.e., wavelength conti-
nuity and DCA must be observed throughout the path). If there is sparse λ-interchange,
the wavelength continuity condition is relaxed at each node containing a wavelength
interchanger so that different wavelengths may be assigned on different segments of a
path. With full wavelength interchange, a free channel exists as long as there is at least
one unused wavelength on each fiber on the path, but the wavelengths need not be the
same (i.e., only the DCA constraint must be observed). A path is called feasible if a free
channel exists on that path. (In the discussion that follows, wavelength interchange is
excluded unless indicated otherwise.)
The wavelength list is typically ordered either by wavelength number (the static case)
or by usage (the adaptive case). The usage of wavelength λi is defined here as the number
of active connections using λi . In the adaptive case, wavelengths might be listed in either
decreasing or increasing order of usage. We refer to the former case as the maximum
reuse method (Max Reuse) and the latter as the minimum reuse method (Min Reuse). In
 Wavelength/Waveband-Routed Networks 487

the wavelength assignment phase of an RCA algorithm, an attempt is made to assign a

wavelength to a chosen path, typically starting from the top of the wavelength list and
working down.38 Thus, in the Max Reuse case, the most used wavelengths are tried first,
the rationale being to reuse active wavelengths as much as possible before trying others.
This in turn makes it more likely that a channel that satisfies wavelength continuity is
available on a long path. In the Min Reuse case, the idea is to spread the load as equally
as possible over all available wavelengths. Experience with simulations has shown that
the Max Reuse method usually works best in WRNs (see, for example, [Bala+91a,
Ramaswami+95]).
There is no “universal” algorithm that performs best in all circumstances and under all
performance criteria. Thus we shall present a few alternative approaches in this section,
pointing out some of their pitfalls and advantages.

6.4.1.3 k Shortest Path

In the k shortest path (k-SP) algorithm [Bala+91a], the first k shortest paths for a
connection are considered admissible and placed on the path list in increasing order of
length. (For k = 1, this reduces to shortest path routing.) Assuming that path lengths are
state independent, k-SP is a static routing algorithm. Each path is checked in order, and
the first that is feasible is assigned the first free wavelength on the wavelength list.39 If
none of the k paths are feasible, the connection is blocked.
Note that if k-SP is used, there is no limit on the length of the admissible paths. If k
is large and if the network is congested, the first feasible path might exceed a reasonable
length limit. Furthermore, if k is small, there is a significant possibility that the path list
contains no feasible paths, resulting in a high blocking probability. For example, in the
case k = 1, a connection is blocked if the unique path on the list does not have a free
wavelength.

6.4.1.4 Shortest Path with Deletions

To reduce the risk of blocking using the k-SP algorithm, an adaptive version can be
used: shortest path with deletions (SPD). In SPD, shortest paths are computed taking
feasibility into account. A copy of the network topology is examined for each wavelength,
following the layered model of Figure 2.6(b). (The order in which the wavelengths are
selected is arbitrary.) All fibers carrying active connections on a particular wavelength
are deleted from the corresponding copy of the network, and a shortest path is computed
on the resultant “deleted topology.” (This is equivalent to finding a shortest path on
a given wavelength when the occupied fibers are assigned infinite weights.) At least
one path exists for a given connection request on a given wavelength unless the source
and destination nodes fall into two disconnected subgraphs for that wavelength. After
scanning all wavelengths, the lengths of all resultant paths are compared, and the shortest

38
Another technique is to pick wavelengths from the list at random [Ramaswami+95].
39
A modified version of k-SP adapted to LLNs is discussed in Section 6.6.1.2. It uses a minimum interference
criterion to choose a path from the admissible list.
488 Multiwavelength Optical Networks

(A, E) (A, C) (A, D)

Initial Path Found No Path No Path
State Links Deleted Found Found

A B A B A B A B

λ1 C C C C

D E D E D E D E

Path Found Path Found

Links Deleted Links Deleted

A B A B A B A B

λ2 C C C C

D E D E D E D E

Time

Figure 6.27 An example of SPD routing.

among them is selected using the corresponding wavelength. If a feasible path exists for
the given source-destination pair, at least one path and wavelength are found. Otherwise
the connection is blocked.
An example of the operation of SPD is shown in Figure 6.27. The network has five
nodes, it operates on two λ-channels, and all link weights are assumed to be 1. A sequence
of states (deleted topologies) of the network is shown in Figure 6.27, with the complete
topology (the network with all fibers idle) shown as the initial state on both wavelength
copies. A request for connection (A, E) arrives. It can be satisfied by two-hop paths on
both copies. To break the tie, the lower numbered wavelength, λ1 , is used, producing the
second state in the sequence. A request for connection ( A, C) arrives next and is placed
on the shortest path, A–C, using λ2 . Finally, a request for connection (A, D) arrives
and is placed on the only feasible path, A–B–C–E–D, on λ2 . If shortest path routing
had been used without deletions, the path A–C–E–D would have been chosen for the
connection (A, D), but the connection would have been blocked because the link A–C
is saturated, making the chosen path infeasible.
The SPD algorithm always finds the shortest feasible path for a connection, given the
current state of the network. It therefore results in a blocking probability that is as good
or better than k-SP for any given connection request. However, its overall performance
may be worse in some cases because it tends to produce longer average path lengths
than k-SP, thereby creating more network congestion.
 Wavelength/Waveband-Routed Networks 489

6.4.1.5 Least Loaded Routing

A constrained least loaded routing algorithm has been proposed by [Ayanoglu+96]. A set
of admissible paths is precomputed for each source-destination pair, and the admissible
path with the maximum free capacity (MFC) is chosen for a connection. The admissible
paths might be a set of k minimum-hop paths for the source-destination pair, which
can be computed off-line. Thus, this is a static constraint. Complications arise in this
algorithm due to the definition of free capacity. For a given connection, let
r  = the set of admissible paths
r π = a path in 
r M l = the number of fibers on link l
r A l j = the number of fibers on link l for which wavelength j is active

r U l = j A l j = total number of channels active on link l (the link utilization)
r  = {λ1 , λ2 , . . . , λW }, the set of wavelengths available in the network, where W = ||

The least loaded path is chosen as follows:

1. Without wavelength interchange, let


M FC = max max min(M l − A l j ) . (6.40)
π∈ λ j ∈ l∈π

The path and wavelength assigned are those that achieve the indicated maximizations.
2. With full wavelength interchange,


M FC = max min(W M l − U l ) . (6.41)
π∈ l∈π

The path assigned is one that achieves the indicated maximization.

In both cases the quantity in brackets is the free capacity on a path π . The link l, which
has the minimum free capacity over all the links on the given path π , is the “bottleneck”
link. The inverse of the free capacity of a path can be thought of as the effective path
length. In the case without wavelength interchange, the algorithm finds the path (and
wavelength) on which this free capacity is greatest (i.e., the path/wavelength with the
shortest effective length). In the wavelength interchange case, the wavelength is not
involved. In either case, a connection is blocked if MFC is zero. (This corresponds to
all admissible paths having infinite effective length.)
The fact that a small, precomputed set of admissible paths, , is used in the compu-
tation means that the maximization can be performed easily. In an unconstrained case,
the admissible set includes all possible paths, meaning that the maximization process
is considerably more complex. A simple shortest path algorithm such as Dijkstra’s (see
[Tarjan83]) does not suffice here because the effective path lengths are not sums of link
weights.

6.4.1.6 Fairness
In general, any network has the property that longer paths are likely to encounter more
blocking than shorter ones. Depending on the network topology and the routing rules,
this can result in unfair treatment of the connections between more distant node pairs.
490 Multiwavelength Optical Networks

0
10

1
10
W=1

2
10
Blocking Probability

3
10

4
10

5
10

W=4 W=8 W = 16 W = 32
6
10 2 1 0 1
10 10 10 10
Load per Node (Erlangs)
Asymptotic
Simulation

Figure 6.28 Blocking on an 11-node WDM ring.

Blocking of the longer connections leaves more network resources free for shorter paths,
so the paths allocated in the network tend to be short ones. Unfortunately, these shorter
paths “fragment” the wavelength mapping, making it less likely that the same wavelength
is available throughout the network for longer paths. It is useful to quantify these fairness
relations so the influence of effects such as path constraints and wavelength interchange
on fairness can be measured. To this end we define the unfairness factor, U f , as the ratio
of the blocking probability on the longest path to that on the shortest path under a given
RCA rule.
The problem of unfairness is more pronounced in networks without wavelength in-
terchange because finding long paths that satisfy the wavelength continuity constraint
is more difficult than without the constraint. To quantify the effect of wavelength inter-
change on unfairness, we define the fairness ratio as
U f (without λ-interchange)
Fr = . (6.42)
U f (with λ-interchange)
A related quantity, the average gain in blocking probability, G p , is defined in
[Ayanoglu+96] as the ratio of overall blocking probability without λ-interchange to
that with λ-interchange. In defining both Fr and G p , the Max Reuse wavelength assign-
ment rule is assumed here for the case without wavelength interchange.
 Wavelength/Waveband-Routed Networks 491

8 W = 32

7
Gain in Blocking Probability

3 W = 16

2 W=8

0
−1 0 1 2
10 10 10 10
Load per Node (Erlangs)

Figure 6.29 Gain in blocking; 11-node WDM ring, simulation.

The influence of various network parameters on fairness is examined through simula-

tion and asymptotic analysis in some case studies that follow. One important observation
is that networks with wavelength interchange sometimes exhibit orders of magnitude
improvement in fairness compared with networks without the wavelength interchange
capability. The improvement in overall blocking probability due to wavelength inter-
change is much less pronounced.

6.4.2 Case Study: Bidirectional Rings

Figures 6.28 through 6.31 show the results of a case study of a bidirectional ring under
dynamic channel assignment rules. Shortest path routing is used, and channels are
allocated using the Max Reuse rule.
Each node on the ring generates connection requests in a Poisson fashion, with an
average arrival rate α, and connections have exponentially distributed holding times
with mean 1/µ. Under uniform traffic, each request is directed equiprobably to all other
nodes. Under a worst-case traffic distribution, each source node generates connection
requests equiprobably to the two most distant destination nodes – (N − 1)/2 hops away
for N odd. Blocking and fairness are studied as a function of offered load per node
(ρ = α/µ), available wavelengths (W ), and wavelength interchange capability.
492 Multiwavelength Optical Networks

2
10
W = 32
Fairness Ratio

W = 16
101

W=8

0
10
−1 0 1 2
10 10 10 10
Load per Node (Erlangs)

Figure 6.30 Fairness ratio; 11-node WDM ring, simulation.

In the case W = 1, it can be shown (see Problem 18 at the end of the chapter) that for
rings of size N (odd), where N ≥ 5, the blocking probability under a worst-case traffic
scenario is given by
β(N − 2) + Nβ 2
Pb = , (6.43)
1 + Nβ + Nβ 2
where β = ρ/2. Blocking calculations under uniform traffic are far more difficult. Thus,
although the worst-case is somewhat artificial, it leads to some simplifications.
For larger values of W , the following asymptotic formula provides a simple and
fairly good “back-of-the-envelope” estimate of blocking probability on the ring in this
worst-case scenario:
αd W

αd
Pb ≈ , for < 1, (6.44)
µW µW
where d = (N − 1)/2 is the ring diameter. This approximation can be derived directly
from the Erlang-B formula, which gives low blocking for cases when total load offered
to a set of channels (in this case αd/µ) is less than the number of channels in the set (in
this case W ).
Simulation results, together with the asymptotic approximation, are shown in
Figure 6.28 for the case N = 11, using a worst-case traffic distribution. Considering
 Wavelength/Waveband-Routed Networks 493

102

Load = 15 Erlangs
Fairness Ratio

101

Load = 100 Erlangs

Load = 9 Erlangs

100
0 10 20 30 40 50 60 70 80 90 100
Wavelength Interchanger Density (%)

Figure 6.31 Fairness ratio improvement versus interchanger density; 11-node WDM ring with 32
wavelengths.

the simplicity of the asymptotic model, it works reasonably well, providing a simple way
to estimate the blocking probability for general WDM rings.

6.4.2.1 Rings with Full Wavelength Interchange

The effect of full wavelength interchange capability in an 11-node WDM ring is pre-
sented in Figures 6.29 and 6.30. These and subsequent results dealing with fairness and
wavelength interchange were obtained by simulation using uniform traffic distributions.
Figure 6.29 shows the average gain in blocking probability, G p , due to wavelength
interchange for the ring with 8, 16, and 32 wavelengths. It is clear that for 8 and 16
wavelengths there is limited improvement in blocking. However, with 32 wavelengths
the improvement reaches a peak of an order of magnitude at a fairly high load and then
drops.40
The case study of Section 6.4.4 shows that for ring interconnection with a large
number of wavelengths, the value of G p is even larger. In general for WDM rings (or

40
Intuition suggests that the improvement in blocking is small with very light and very heavy loads. In the
former case, wavelength interchange is not needed for good performance, and in the latter case, links tend
to saturate, and thus wavelength interchange does not help.
494 Multiwavelength Optical Networks

ring interconnections), G p increases with the number of wavelengths and the lengths of
the paths (i.e., the ring sizes).
The fairness ratio, Fr , obtained by simulation for the 11-node WDM ring is shown in
Figure 6.30. Observe that for 32 wavelengths, Fr reaches a peak of approximately 100.
As in the case of G p , the peak is attained at fairly high loads, after which Fr drops.
The significance of this fairness improvement becomes clearer when we examine
some of the numbers more closely. Consider the case with 32 wavelengths at a load of
12 Erlangs, giving an Fr of 80.
Without wavelength interchange the simulation yields blocking on the longest paths
of 4.5 × 10−2 , whereas it is 1.4 × 10−4 on the shortest paths. (The average blocking is
2 × 10−2 .) Thus, the unfairness factor, U f , is 320 at this load. However, with wavelength
interchange the blocking on the longest paths is 8 × 10−3 and the blocking on the shortest
paths is 2 × 10−3 , with an average blocking of 5 × 10−3 , giving a U f of 4. Clearly,
wavelength interchange results in a very large improvement in the blocking of long
paths in this case.

6.4.2.2 Rings with Sparse λ-Interchange

Given the cost of wavelength interchangers, it is important to determine whether it is
possible to achieve the benefits of wavelength interchange by introducing it sparsely, at
a few strategic places in the network [Bala+97, Subramaniam+96a].
In the case of a ring carrying uniform traffic, best results are obtained when the
interchangers are placed uniformly around the ring. Figure 6.31 shows the results of
simulations of the fairness ratio for approximate uniform placement on an 11-node ring,
with Erlang load per node as a parameter. The improvement is a roughly linear function
of the λ-interchange density (on the semilog scale). Improvement peaks at a fairly high
load. (Note that the 15-Erlang curve shows the highest fairness ratio.)

6.4.3 Performance of Dynamic Routing Rules on Meshes

Even in the case of simple topologies such as rings, using fixed routing rules, the
calculation of blocking probabilities in multiwavelength networks is extremely difficult.
In networks with arbitrary mesh topologies, and using alternate routing, the problem
is much more complex. The network typically has many alternate paths between a
source-destination pair, so a connection that is blocked on one path can be carried on
another. Once paths have been chosen, however (e.g., using a fixed routing rule), the
blocking problem is similar in structure to that in rings, although the dependencies
among interfering connections are generally more complicated.
In view of the complexity of the problem of predicting blocking in general mesh
networks, it is doubtful that mathematical models can be found that are accurate, general,
and tractable. Nevertheless, some models have been proposed. In [Subramaniam+96a],
a Markovian model is used to capture the effect of link dependency on connection
blocking. In [Birman95], a more accurate method for calculating the blocked traffic in
a mesh topology is proposed. However, this method is limited to fixed routing schemes
with routes of three hops or less.
 Wavelength/Waveband-Routed Networks 495

Probably the simplest and most intuitive approach is due to [Barry+95]. The model
they use is similar to that used for finding approximate blocking probabilities in tra-
ditional telephone networks and circuit switches. As in telephone network models, the
key assumption is that the channel (wavelength) occupancies on each link are mutually
independent random variables.
Consider a network with W available wavelengths, with a single fiber pair on each
link, and without wavelength interchangers. Suppose the probability ρ that a wavelength
is used on a link is known and equal for all links. Then, subject to the aforementioned
independence assumption, the probability that a connection is blocked on a specific H
hop path is

Pb = [1 − (1 − ρ) H ]W . (6.45)

The result is derived easily. The term (1 − ρ) in Equation (6.45) is the probability
that any given wavelength is free on any given link. Raised to the power H , it repre-
sents the probability that the wavelength is free on every link of the given path. Thus,
[1 − (1 − ρ) H ] is the probability that the wavelength is occupied on at least one link
on the path and therefore is unusable on the given path. Equation (6.45) therefore gives
the probability that all W wavelengths are unusable, so the connection is blocked on the
given path.
The independence assumption, on which the previous argument is based, is approx-
imately valid in cases when there is a high degree of “mixing” at each node (i.e., the
nodes are of high degree and/or W is large), so many connections enter and leave a
node in many different directions and on many different channels (wavelengths). This
is often the case in telephone applications. Unfortunately, the independence assumption
is less valid in networks with small nodal degree and relatively few wavelengths (less
mixing), which is more typical of current optical networks. (Even in DWDM the number
of wavelengths carried on a fiber is much less than the number of trunks carried on a
typical telephone link.) A ring is an extreme case, in which the nodal degree is only
2. (To check validity in the case of a ring, the reader should try to match some of the
simulation points in Figure 6.28 with Equation [6.45]. See Problem 19 at the end of the
chapter.)
How does Equation (6.45) change in the presence of wavelength interchangers? With
full wavelength interchange, a mesh optical network behaves like a traditional network,
with no wavelength continuity constraint. In this case, a connection request is blocked
on a given H hop path only if all wavelengths are used on one of the links. Under the
same assumptions as stated earlier, an expression for blocking probability in this case is
[Barry+95]:

Pb = 1 − (1 − ρ W ) H . (6.46)

6.4.4 Case Study: An Interconnected Ring

In the case of rings, various reasonable approximations of blocking probabilities are
possible. In the case of meshes, the simple formulas of Equations (6.45) and (6.46) are
useful in gaining some insight into how connection blocking is related to the various
496 Multiwavelength Optical Networks

100

10−1

W=1

10−2
Blocking Probability

10−3

10−4

10−5

W=8 W = 16 W = 32
10−6
10−3 10 −2 −1
10 100
Load per Node (Erlangs)
Asymptotic
Simulation

Figure 6.32 Simulation and asymptotic analysis; 195-node interconnected WDM rings.

network parameters. However, they are not very useful in obtaining quantitative results.
To obtain more accurate results in realistic networks, simulation is currently the only
viable approach.
Interconnected rings are examples of practical topologies going beyond rings, which
are frequently used by telecommunications carriers. Figures 6.32 through 6.34 show
simulation results for blocking and fairness ratios in an interconnected ring network
with 195 nodes. The results are qualitatively similar to those for the 11-node ring, with
more significant improvements in fairness.
In this example, the rings are interconnected pairwise, using WSXCs, some of which
are replaced by WIXCs when wavelength interchanging is introduced. The overall net-
work resembles a “necklace” with 15 “beads” (the rings). Each ring has 13 nodes. The
traffic distribution is uniform, and shortest path routing is used.
A good placement for sparse wavelength interchange in the network under study is
at the ring interconnection nodes, where the path interferences are the highest under
uniform traffic. As the density of wavelength interchangers increases, the interchangers
are placed approximately uniformly throughout the network. As shown in Figure 6.35,
most of the benefits from the use of wavelength interchange are achievable by installing
this capability at 10% to 20% of the nodes in the network.
 Wavelength/Waveband-Routed Networks 497

20

W = 32

15
Gain in Blocking Probability

10

W = 16
5

W=8

0
10−2 10−1 100 101
Load per Node (Erlangs)

Figure 6.33 Blocking improvement with wavelength interchange; 195-node interconnected WDM
rings.

6.4.5 Routing Multicast Connections in WRNs

Multicast, the ability to transmit a message from a single source node to multiple desti-
nation nodes, is a key networking capability underlying many widespread applications,
from video delivery to access network operation. As explained in Section 3.2.1, optical
networks are particularly well adapted to multicasting; it can be achieved by simply
including a light-splitting capability in the optical network nodes. Figure 6.36 shows
an example of an optical multicast connection, a light-tree, with source S and desti-
nations D1 , D2 , and D3 , in a transparent network utilizing multicast-capable optical
cross-connects. The splitting operation is inherent in some optical switches – such as
the generalized optical switches (waveband-selective linear divider-combiners [LDCs])
used in LLNs. Early work on optical multicasting in LLNs in the context of dynamic
traffic appears in [Bala+93]. Multicasting in LLNs is analyzed in Section 6.6.2. As
opposed to LLNs, WRNs operate with only one λ-channel per waveband (that is, they
switch wavelengths rather than wavebands) and use only permutation switches in their
OXCs. The permutation switch makes a WRN incapable of producing optical multi-
cast connections. However, by expanding the functionality of the OXCs to include an
optical signal splitting function (which is one half of the functionality of the LDC) the
498 Multiwavelength Optical Networks

104
W = 32

103
Fairness Ratio

102
W = 16

101 W=8

100 −2
10 10−1 100 101
Load per Node (Erlangs)

Figure 6.34 Fairness ratio improvement with wavelength interchange; 195-node interconnected
WDM rings.

network becomes a multicast-capable WRN (MC-WRN). Results on multicasting in

these networks are described in this section.
Figure 6.37 shows a possible realization of the basic building block of the multicast-
capable optical cross-connects: a single-wavelength split-and-deliver (SaD) switch. Fig-
ure 6.38 shows a wavelength-selective multicast-capable optical cross-connect based on
the general three-stage architecture of Figure 2.21. It uses SaD switches in the middle
stage to create a multicast-capable WSXC to accommodate multicast connections in a
transparent multiwavelength network [Ali+00a]. In Figure 6.39, a single power splitter
and optical amplifiers are shared by all input ports. This architecture introduces savings
in terms of cost, switch complexity, and power efficiency (as point-to-point connections
do not pass through the splitters now). However, it is limiting in the sense that only one
multicast connection can be accommodated at any time through this cross-connect.
Routing and channel assignment for a multicast connection in a WRN require opera-
tions similar to those for a point-to-point (unicast) connection:

1. Creation of a source-destination light-tree (a set of fibers, which constitute a multicast

tree) [Sahasrabuddhe+99] on a chosen wavelength (by setting the optical cross-
connects along the light-tree)
 Wavelength/Waveband-Routed Networks 499

104

Load = 2 Erlangs

103

Load = 4 Erlangs
Fairness Ratio

Load = 10 Erlangs
102

101

100
0 10 20 30 40 50 60 70 80 90 100
Wavelength Interchanger Density (%)

Figure 6.35 Fairness ratio improvement versus interchanger density; 195-node interconnected
WDM ring, 32 wavelengths.

2. Creation of an optical connection by tuning the source transmitter and destination

receivers of the light-tree to the chosen wavelength.

An optical light-tree T consists of a sequence of fibers directed from source to

destinations and a wavelength assignment on that fiber path.
In a WRN with a given physical topology, solving the RCA problem for a multicast
connection consists of finding an appropriate light-tree for each connection (routing)
together with a wavelength for that light-tree (channel assignment). Static multicast
RCA is equivalent to finding an embedding of a prescribed multicast logical topology
onto a given physical topology while observing the constraints of the network. Dynamic
multicast RCA can use algorithms of the same general form as dynamic unicast RCA.
We consider the dynamic case first. A general approach is the following:
r Arrange all admissible light-trees for a given source-destination set pair in some
prescribed order in a light-tree list.
r Arrange all wavelengths in some prescribed order in a wavelength list.41
r Attempt to find a feasible route and wavelength for a requested connection starting
with the light-tree and wavelength at the top of the list.

41
This assumes no wavelength interchange.
500 Multiwavelength Optical Networks

D1
D2

D3

Multicast-Capable OXC

Figure 6.36 Multicast connection in a transparent network.

As in the heuristics for unicast RCA described in Section 6.4.1, the characteristics of
a multicast RCA algorithm are determined by the constraints on admissibility and rules
for list ordering. Admissibility may be based on physical layer constraints (e.g., power,
OSNR, distance) or logical layer constraints (e.g., maximum number of hops, maximum
number of splits). The ordering of the light-tree list could be determined by the same or
similar constraints, and the ordering of the wavelength list could be based on specific
network requirements (such as Max Reuse, or Min Reuse for load balancing).
Before examining the details of the MC-WRN light-tree approach it is instructive to
consider how multicasting would be accomplished in a WRN without optical splitters.
In that case, every multicast request would be treated as a set of separate unicast
connections, each one on its own optical path, routed independently in the manner
prescribed in the previous sections. In this case information duplicated electronically
at the source is carried on each unicast path. This, in effect, relegates the multicasting
problem to the logical layer.42 A heuristic outlined in [Sun+01] finds multiple unicast
shortest path optical connections to solve the routing part of this problem. The first-
fit approach is then used to assign wavelengths for the optical paths. (In this scheme,
the first feasible choice on the wavelength list is assigned to the connection.) First-
fit, however, results in a large waste of network resources (link bandwidth, number of
receivers, etc.).

42
In fact, the more common approach to multicasting is to use a logical network overlay to create a multicasting
tree using combinations of logical connections, which involves duplicating the information electronically
at some of the intermediate nodes on the tree.
 Wavelength/Waveband-Routed Networks 501

Inputs
1 ...

2
...

. .
.
. .
.
. .
.

P
...

Outputs 1 2 P

Configurable Optical
amplifier 2 × 1 switch
splitter

Figure 6.37 A P × P split-and-deliver switch. (From [Rouskas03, Figure 2]. Copyright

c 2003
IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

The use of light-trees on a multicasting WRN produces more efficient solutions to the
multicasting problem. The routing subproblem, which entails finding appropriate light-
trees, is often cast as that of finding a Steiner tree, which is NP-complete [Hwang+92].43
Let G = (V, E, m) denote an undirected network graph where V is the set of vertices
in the network, E is the set of edges (links) in the network, and m is a positive cost function
on the set of edges in the network. A Steiner tree is defined as a tree T of G that includes

a given subset M of the vertices of G and minimizes e∈T m(e) where m(e) is the cost
of the edge e. (The set M comprises the multicast source and destinations.) A number of
heuristics have been proposed to find an approximate (suboptimal) solution to the prob-
lem of finding Steiner trees [Takahashi+80, Winter87], which is consequently a solution
to the multicast routing problem [Sahasrabuddhe+00, Sun+01]. Among these are:

1. Shortest path tree (SPT) heuristic: The SPT heuristic combines the shortest paths from
the source to each of the destinations to obtain the multicast tree. The time complexity
of the SPT algorithm is O(|V 2 ||E|). However, if all the shortest paths (from any source
to any destination) are calculated a priori, the time complexity to calculate a multicast
tree using the SPT algorithm reduces to O(|V ||E|) [Takahashi+80].
43
The Steiner tree problem is NP-complete except for the trivial cases of unicast (where it is a shortest path
problem), broadcast (where it is a minimum spanning tree problem), or where the network itself is a tree
[Sahasrabuddhe+00].
502 Multiwavelength Optical Networks

λ1
Input λ1 Output
Port 1 Port 1
SaD
λW λ1 Switch 1 λ1

Input λ1 Output
Port 2 Port 2

λW
. λW .
. .
λ1
. .
SaD λW
Switch W
Input Output
Port P λW λW Port P

Wavelength Wavelength
Demux Mux
W wavelengths

Figure 6.38 A P × P multicast-capable optical cross-connect based on a split-and-deliver switch.

(From [Rouskas03, Figure 3]. Copyright
c 2003 IEEE. Used by permission of The Institute of
Electrical and Electronics Engineers, Inc.)

2. Distance network heuristic (DNH): Find the shortest paths from the source to each
of the destinations and the shortest paths between each pair of destinations. Form
a distance network graph (a complete graph) whose vertices are the source and the
destinations, where the cost of each edge is the cost of the shortest path in the
original graph G between the vertices of that edge in the distance network graph.
Find the minimum spanning tree of the distance network, and convert the tree back
to the original graph to obtain the multicast tree. The time complexity of the DNH
algorithm is O(|V 2 ||E|). Again, however, if we calculate all the shortest paths (from
any source to any destination) a priori, the time complexity to calculate a multicast
tree using the DNH algorithm reduces to O(|V ||E|) [Kou+81].

The DNH algorithm gets closer to the optimal tree than the SPT algorithm. However,
the DNH algorithm is a centralized algorithm that requires recalculating the tree every
time a destination is added or removed. In contrast, the SPT algorithm can be imple-
mented in a distributed manner and does not require recalculating the tree every time a
destination is added or removed. We simply add or delete (prune) a branch of the tree.
Four additional heuristics for the Steiner tree problem: average distance heuristic
(ADH), contraction heuristic (CH), set covering heuristic (SCH), and dual ascent heuris-
tic (DAH) are described in detail in [Winter87].
Once the multicast tree is found, and assuming there are no wavelength converters,
the wavelength assignment subproblem becomes that of finding a single available wave-
length on all the multicast tree branches. A number of different schemes, such as random,
first-fit, or most-used, can then be employed for wavelength assignment.
 Wavelength/Waveband-Routed Networks 503

Point-to-point connections

Inputs Outputs
1 1

2 2

.
.
.
.
.
.

P P

Multicast connections

Configurable Optical
amplifier 2 × 1 switch
splitter

Figure 6.39 A P × P multicast-capable optical cross-connect based on splitter sharing. (From

[Rouskas03, Figure 4]. Copyright

c 2003 IEEE. Used by permission of The Institute of Electrical
and Electronics Engineers, Inc.)

In the static case, an ILP model can be utilized for the multicast RCA problem in the
MC-WRN. The network parameters and variables used in this formulation are [Ali+00a]
r M: Set of optical cross-connects in the network
r N : Set of stations in the network
r V : Set of nodes in the network. V = M ∪ N
r W : Set of wavelengths
r L: Set of links in the network
r G = (V, L): undirected graph representing network topology
r Q: Set of multicast sessions
r ψi : Multicast session number i
r K i : Set of alternate trees for multicast session ψi
r ei : = 1 if session ψi is established; = 0 otherwise
r λi, j,c : = 1 if session ψi is established using wavelength c and the jth tree; = 0
otherwise
r fl : Number of fiber pairs on logical link l
r Vl : Set of trees with link l in their link set
r Ti, j : The jth tree for session ψi
r σi : Profit of session ψi
504 Multiwavelength Optical Networks

Now the MC-RCA problem can be formulated as follows:

Maximize

Z= σi ei (6.47)
i:ψi ∈Q

With
 
λi, j,c ≤ fl , l ∈ L , c ∈ W (6.48)
i:ψi ∈Q j:Ti, j ∈K i ∩Vl

λi, j,c ∈ {0, 1} (6.49)

ei ∈ {0, 1} (6.50)

λi, j,c = ei , ψi ∈ Q. (6.51)
j∈K i c∈W

The goal of the ILP is to maximize the overall profit Z by choosing the appropriate
tree for each session ψi , from the set {Ti,0 , Ti,1 , . . . , Ti,|K i | }. Constraint (6.48) ensures
that a wavelength c ∈ W is used by at most fl trees. Constraints (6.49) and (6.50) force
the variables to be binary, and constraint (6.51) forces the use of at most one tree for each
session. In this formulation of the MC-RCA problem, profit is maximized by establishing
as many sessions as possible (keeping blocking low) while choosing the most profitable
ones.

Optical Multicasting with Optical Layer Constraints

Power loss is a consideration when optical multicast connections are accommodated in
the network, as the connections can potentially pass through a large number of optical
splitters. Power loss in these networks is due to signal attenuation along the fiber links and
signal splitting in the multicast-capable optical cross-connects. There is an upper bound
on the number of times a signal can be split and amplified, because optical amplifiers also
amplify noise levels. Unnecessary power losses can be eliminated by either modifying the
architecture of the multicast-capable optical cross-connects or implementing multicast
RCA algorithms that take into account the power budget constraints in the network.
A design that can potentially eliminate some of the unnecessary power losses in the
network was shown in Figure 6.39. Such a design would be useful in networks where
multicast traffic is only a small fraction of the overall traffic in the network. We focus
on power-budget-aware RCA algorithms that include two quality-of-service parameters:
source-destination loss and loss-variation among different source-destination pairs. The
second parameter is used to ensure that there is a fairness factor among all destinations
in a multicast connection.
There are two cases of interest for power-budget-aware RCA algorithms:

1. Splitting losses are ignored and only losses due to attenuation are taken into account.
This case is equivalent to finding delay-constrained Steiner trees, an NP-complete
 Wavelength/Waveband-Routed Networks 505

problem, for which there exist a number of heuristic algorithms [Wang+00]. For
example, Haberman’s algorithm constructs a reference tree with least-cost paths
from the source node to all destination nodes. For each destination node Di , an initial
tree is constructed that contains the path in the reference tree from the source node
to node Di . This initial tree is augmented by adding feasible paths from tree nodes to
destination nodes not on the tree, until all destination nodes are included in the tree.
Then the feasible tree with the least cost is selected (if more than one feasible tree is
found).
2. Attenuation losses are ignored and only losses due to splitting are taken into ac-
count. This is again an NP-complete problem [Xin+04]. A balanced light-tree (BLT)
constrained tree algorithm is introduced in [Xin+04]. It is shown that to guarantee
optical signal quality the light-trees obtained must be as balanced as possible, which
in addition ensures that there is a fairness factor among all different destinations in
a multicast connection. The BLT algorithm is as follows:
r Find an initial tree spanning all destination nodes.
r Find the destination node incurring the maximum number of signal splits (node I ).
r Find the destination node incurring the minimum number of signal splits (node J ).
r Delete node I from the tree.
r Add node I to the path from the root to node J .
r Repeat until no further improvement.

Sparse Multicasting: Placement of Optical Splitters

To accommodate multicast connections in a transparent optical network, it is not nec-
essary that all optical cross-connects in the network are multicast capable. Sparse light
splitting has been addressed in [Zhang+00] and [Ali+00b] with the objective of mini-
mizing the number of multicast capable optical cross-connects without unduly penaliz-
ing the performance of the multicast connections. [Ali+00b] showed that performance
improves noticeably even when no more than 50% of the optical cross-connects are
multicast capable. [Malli+98] also showed that no more than 70%–75% of the nodes
need to be equipped with the splitting capability to obtain almost the same benefit as
having splitting capability at all the nodes.
In the sparse light-splitting case where only a few nodes have splitting capabilities, a
multicast forest is required to guarantee that all the destinations in a multicast session
can be reached. A multicast forest consists of one or more multicast trees that have a
common root at the source. (The source must transmit copies of the information into all
trees in the forest.) Figure 6.40 shows an example of a multicast session from node 1 to
nodes 5, 6, and 10 accommodated using two multicast trees (shown as heavy and dashed
lines) when nodes 2, 3, 5, and 6 do not have splitting capabilities.
An algorithm that can be used to construct a multicast forest is presented in
[Malli+98]. This is a recursive algorithm that initially tries to include as many of
the destinations as possible in an initial tree. If some destination nodes cannot be in-
cluded in this initial tree the algorithm is recursively called to construct a second tree
with as many of the remaining destinations as possible. This procedure is repeated until
all destination nodes are included in the multicast forest. Clearly, using this approach the
506 Multiwavelength Optical Networks

4
5
10

2 3 8 9

Multicast-Capable OXC

Node with no splitting capability

Figure 6.40 Multicasting in a network with sparse splitting capabilities.

path between source and destinations may not be the one with the least cost (i.e., may
not be the shortest path). This fact can be used to compare the efficiency of multicast
RCA algorithms (e.g. in terms of number of wavelengths required and bandwidth saving
percentage) when sparse splitting is available, to the same multicast RCA algorithms
when full splitting is present at all network nodes.
Two classes of algorithms for constructing multicast forests are presented in
[Zhang+00]. The first class constructs a multicast forest by modifying a multicast
tree in which it is assumed that multicast-incapable nodes can be replaced by splitting
points. The second class of algorithms constructs a multicast forest one tree at a time,
similar to the approach presented in [Malli+98].
Wavelength assignment in the multicast forest case is identical to wavelength as-
signment for multicast trees. Random, first-fit, and most-used wavelength assignment
schemes are some of the approaches that can be utilized. As in the case of multicast trees,
every subtree of the forest has to be assigned the same wavelength on all its branches.
In sparse multicasting, once a decision is made on the number of multicast-capable
OXCs (splitters) to be used, the next problem is their placement in the network. [Ali+00b]
takes an initial look at this problem, but additional work is required on this subject.
Splitter placement in WRNs (SP-WRN) is clearly an NP-complete problem, because it
is a special case of the NP-complete problem of MC-RCA when the number of multicast-
capable cross-connects equals the total number of nodes in the network. In fact, it is
shown in [Ali+00b] that the hardness of the splitter placement problem comes from the
RCA problem, which is just one part of the joint splitter placement and RCA problem.
 Wavelength/Waveband-Routed Networks 507

An ILP developed in [Ali+00b] for the SP-WRN problem requires an incredibly long
time to solve when general-purpose ILP solvers such as CPLEX are used. A heuristic
approach is to solve the multicast routing and wavelength assignment problem first
assuming that all nodes are multicast capable. This solution is then used for the splitter
allocation problem [Ali+00b]. A simple greedy heuristic for placing the splitters after
the MC-RCA problem is solved, is to allocate the splitters to the nodes in the original
network that are used the most as splitting nodes. A more involved heuristic utilizing
simulated annealing is given in [Ali+00b].

6.5 Linear Lightwave Networks: Static Routing Rules

The basic structure of the LLN was described in Section 3.4. Recall that the difference
between a wavelength-routed network and an LLN lies in the functionality of the network
nodes. In the former, the node is a wavelength-selective permutation switch (optical
cross-connect), whereas in the latter, it is a waveband-selective linear divider-combiner
(LDC) (i.e., a generalized switch).44 As illustrated in Section 3.4, the extra functionality
of the LDC offers both opportunities and challenges in allocating resources to optical
connections.
In an LLN, an optical path consists of a fiber path carried on a designated waveband.
An optical connection is created on a designated optical path by allocating a λ-channel
within the path’s waveband to that connection. Thus, connection establishment in an LLN
is a resource allocation problem at three levels: (1) waveband selection, (2) fiber path
allocation (routing), and (3) channel assignment (including λ-channels and transmission
channels).
As we shall see, this three-level separation is needed because of the extra complications
at the optical path level caused by multicast paths and the inseparability constraint.
This section treats LLNs operating under a static routing regime (see Section 6.6 for the
dynamic routing case). Under static routing rules, optical paths, acting as fixed “pipes”
on designated wavebands, are established on selected routes. Once the optical paths are
in place, optical connections can be directed through the pipes by assigning them to
λ-channels within the corresponding wavebands. This amounts to waveband routing of
the connections. This approach applies equally well to dedicated and demand-assigned
optical connections. In other words, we can have static routing rules in the optical path
layer supporting either dedicated (static) or on-demand (dynamic) optical connection
assignments. Furthermore, several logical connections can be multiplexed on a given
optical connection by assigning them to distinct transmission channels.
Resource allocation in LLNs is viewed best from the bottom up, starting in the
optical path layer. As shown in Figure 2.6(c), the LLN can be visualized as several
independently controlled copies of a network with the same physical topology, one copy
for each waveband. The optical spectrum is divided into W wavebands with C λ-channels
per waveband. Establishing an optical path between NASs consists of choosing a route

44
The LLN is assumed to have no wavelength interchange.
508 Multiwavelength Optical Networks

from source to destination stations on a selected waveband and setting the LDCs in the
network nodes to direct signals in that waveband along the chosen route. Thus, creation
of an OP is a routing and waveband assignment (RWA) problem. For point-to-point
OPs this is similar to the RCA problem in wavelength-routed networks, except that we
are now routing a “bundle” of λ-channels instead of one. (We distinguish routing and
waveband assignment from routing and wavelength [channel] assignment in this book
by using the acronym RWA for the former and RCA for the latter. However, the reader
should note that in the literature RWA is commonly used for routing and wavelength
assignment.) In LLNs, as opposed to WRNs, optical paths, optical connections, and
logical connections may be multicast as well as point-to-point. This means that the LLN
has considerably more connectivity options. In particular, a large number of LCs (both
point-to-point and multicast) can be carried on each OP. (The other side of the coin
is that the routing, waveband, and channel assignment problems become considerably
more complex.)
Examples of static routing in LLNs were presented in Section 3.4.2 in the form of
multipoint subnets (MPSs). The optical paths for the MPSs were embedded on the
physical topology in the form of directed trees. These trees were capable of supporting
multipoint logical connectivity using wavelength division techniques (at the optical
connection level) and/or time division techniques (at the transmission channel level).45
Once an OP is established on a given waveband, one or more optical connections
can be activated on the path by assigning λ-channels to them within the waveband and
tuning the optical transceivers in the source and destination stations to the selected
wavelengths. Finally, each LC is established by assigning a transmission channel to it
on the supporting optical connection and configuring the TPs/RPs in the source station
and destination stations to operate on that channel.46
Because of its complexity, it is helpful to partition the three-part resource allocation
problem into two subproblems: RWA and channel assignment. In this section RWA
is considered a static problem and channel assignment is treated both statically and
dynamically. (Section 6.6 considers the case of dynamic routing.)
Let us use the Petersen network of Figure 6.41 as a point of departure to illustrate
various facets of the resource allocation problem. Its physical topology has a number
of properties that are useful in the discussion that follows. It is a regular graph with
N = 10 vertices (network nodes) and L = 15 edges (network links). Each vertex has
degree  = 3. It has diameter D = 2, and it is completely symmetric (i.e., its topology
looks the same as seen from each node). Also, it is a Moore graph. The symmetry of the
network enables us to deduce the important performance parameters without extensive
computation and provides some insight into the effects that limit performance.

45
It should be observed here that even in the case of spectrum partitions with C = 1 (i.e., single-wavelength-
per-waveband LLNs), bundling of LCs can take place at the transmission channel level. Thus, using
TDM/TDMA on a multipoint OP, several sources can share a single λ-channel to produce multipoint
logical connectivity.
46
In WRNs the distinction between an optical connection (carried on a λ-channel) and an LC (carried on a
transmission channel) is unimportant, because there is no opportunity for sharing λ-channels.
 Wavelength/Waveband-Routed Networks 509

1 A

f
e a
2
6 F
6 B 2
k g 3
E
j G 3
J h
i
n
P

d b
H
7
m 7 l

5 4
D c C

5 4

Figure 6.41 Petersen network.

6.5.1 Routing of Optical Paths

In selecting routes for OPs, we must observe the four routing constraints in LLNs.
As first presented in Sections 3.3 and 3.4, these are (1) wavelength continuity, (2)
inseparability, (3) distinct channel assignment, and (4) distinct source combining
(DSC).
Once the LDCs in a network are set to establish an OP, the laws of physics (embodied
in the first two constraints) determine how signals propagate over it. When several OPs
are established on the same waveband, signals may propagate in unintended ways, as
we first encountered in Section 3.4. The last two constraints are imposed to ensure that
when several signals are present on one or more OPs, they do not interfere and thus
destroy useful information.
In LLNs, as in WRNs, it is important to separate station/access constraints from
network constraints. To this end we once again distinguish between nonblocking and
elementary NASs. Figure 6.42 illustrates a possible structure for a nonblocking NAS in
an LLN. The station is attached to network node C of degree 3 through an access link
containing three fiber pairs: c1 , c2 , and c3 (see Figure 6.42[a]). It is equipped to operate
on two wavebands, w1 and w2 , with four λ-channels in each waveband. The station can
be “resolved” into three elementary stations – C1 , C2 , and C3 – one for each access
510 Multiwavelength Optical Networks

Four Transmitters

c1 T w1 ci,
c1 C1
T w2
c2 c2
C C2 Ci
c3
c3 R w1 ci
C3
R w2

Four Receivers
(a) (b) (c)

Figure 6.42 Structure of a nonblocking access station for an LLN.

fiber pair, as shown in Figure 6.42(b). Within an elementary station Ci , there are two
arrays of four OTs, one operating on waveband w1 and the other operating on w2 – (see
Figure 6.42[c]). Each transmitter is fixed-tuned to a λ-channel in its waveband. Signals
from all transmitters in an array for one waveband are first multiplexed onto a fiber, and
then the combined signals for the two wavebands are multiplexed onto the access fiber
ci′ , serving elementary station Ci .47
Similarly, on the receiving side of elementary station Ci , optical power from fiber ci is
waveband demultiplexed, with signals on each waveband sent to a different array of ORs.
Note that the elementary components of a larger station behave, from an optical point
of view, exactly as if they were independent NASs. Therefore, no distinction is made
between an elementary (single access fiber pair) station and an elementary component
of a larger station. Following this pattern, a nonblocking station in an LLN with F fiber
pairs on its internodal links contains F elementary stations, each on its own access
fiber pair, where  is the degree of the node being accessed. Each elementary station
contains an array of W transceiver sets, one set operating on each waveband, and each
set contains an array of C fixed-tuned transceivers, one for each channel in its waveband.
In this way, the station is capable of terminating all optical connections arriving at or
departing from its ONN.
We define the source (destination) of an optical path operating on a waveband w to be
the set of all transmitters (receivers) operating on waveband w in an elementary NAS.
Thus, for example, the nonblocking NAS in Figure 6.42 can act as the source of up to
six different OPs, one for each waveband and each outbound access fiber.
In contrast to sources of OPs, the source of an optical connection is one OT generating
a signal on that connection. The destinations of the connection are all of the ORs reached
by the source signal.
The differences in definitions of optical paths and optical connections are due to
the fact that an optical path routes the aggregate power on one waveband on a fiber,
which could originate from several transmitters within the waveband, whereas an optical

47
Following the terminology first introduced in Section 3.4.2, we use unprimed link labels to designate fibers
carrying signals in the direction of the link reference arrow and primed link labels for fibers carrying signals
in the opposite direction.
 Wavelength/Waveband-Routed Networks 511

connection defines a relation between one transmitter and one or more receivers, all
operating on the same wavelength.
Using these definitions, each station may contain several entities that serve as sources
and destinations of optical paths and connections. Nevertheless, to avoid cumbersome
notation, a source or destination is identified using the label of its station whenever there
is no risk of ambiguity.
Having defined our terms, we now present an example that illustrates the difference
between feasible and infeasible routes for OPs, using the network of Figure 6.41. In our
first look at this network, we assume that it has seven attached elementary stations, with
single fiber pairs on all internodal links. It is also assumed initially that the network
operates on a single waveband, so the RWA problem reduces to OP routing only.
The notation pi = si , Di is used to denote an OP from a source si to a destination
set Di . The fiber path for an OP takes the form of a directed tree composed of all fibers
carrying signals from si to the set Di .
Suppose the following optical paths are to be routed on the network:
p1 = 1, {5, 7}
p2 = 2, 6
(6.52)
p3 = 4, 1
p4 = 4, 5 .
Assuming that shortest path routing is used, Figure 6.43 shows the evolution of the
fiber paths and LDC settings as each OP is established. The role played by the LDCs in
linking the fibers is indicated by intranodal connection matrices for the network nodes
in Figure 6.43.48 Figure 6.43(a) shows that the multicast OP p1 is produced by setting
node A to split the signal arriving on fiber 1′ to fibers f ′ and e′ . Figure 6.43(b) shows
the state of the network after addition of p2 . Now, p2 shares fiber e′ with p1 , and thus
the two OPs coalesce into a directed tree that contains all sources and destinations of the
constituent OPs. As a result of inseparability, fortuitous destinations are added to both
OPs, as explained in Section 3.4.1.
When OP p3 is added, the coalesced optical paths are extended as shown in
Figure 6.43(c). At this point the three OPs become
p1 = 1, {5, 7, 6}
p2 = 2, {6, 1, 5} (6.53)
p3 = 4, {1, 5, 6} .
The notation used in Equation (6.53) represents expanded OPs, of the form pi = si , Di ,
where Di is the destination set of the original OP, expanded to include its fortuitous
destinations (underlined). The coalesced OPs are still in the form of a directed tree.
However, without looking “inside” the LDCs, it is impossible to determine the way
the signals are routed. In particular, the setting of node A indicates that the OPs from
sources 2 and 4 continue on to fibers 1 and e′ without splitting to fiber f ′ , whereas the
48
Each entry in an intranodal connection matrix corresponds to a path within the LDC from an input fiber to
an output fiber, where the rows and columns correspond to inputs and outputs, respectively.
512 Multiwavelength Optical Networks

At A
a e′ f ′ 1′
A
1′
a
f′

F
e′

E
h′

d′
H

7
D
5

5
(a)

Figure 6.43 Optical paths.

OP from source 1 is split between e′ and f ′ . Thus, to specify each optical path uniquely,
the internodal and intranodal connections are required.
Now when we attempt to add OP p4 , originating at station 4 and routed on fiber c
(the shortest path), two things happen. First, because station 4 is an elementary station
operating on a single waveband, it can be the source of only one path, and thus paths
p3 and p4 must coalesce to a single path originating at station 4. Second, examining the
settings of nodes C and D, as shown in Figure 6.43(d), we find that the complete set
of coalesced OPs no longer forms a tree: There are two parallel paths from source 4 to
destination 5, violating the DSC condition.49 Therefore, the new OP cannot be routed
on fiber c. The reason for the DSC violation is that the intended destination for the
49
This type of difficulty was first brought out in the example of Figure 3.23.
 Wavelength/Waveband-Routed Networks 513

1
At A
a e′ f ′
1′
1′
A
a′

f′

F
e′ a′ 2
6

2′
6

At E
B
E 6 d′
e′
h′

d′
H

7
D
5

5
(b)

Figure 6.43 (cont.)

new OP has already been included in p3 as a fortuitous destination. As a result, there

is no need to add another OP to create a path from station 4 to station 5. Because the
labeling of a destination as intended or fortuitous is arbitrary, we simply redesignate
destination 5 as intended, so p3 in Equation (6.53) becomes an intentionally multicast
OP: p34 = 4, {1, 5, 6} , which replaces the two point-to-point OPs p3 = 4, 1 and
p4 = 4, 5 .
In general, to ensure that all optical paths on the same waveband satisfy the DSC
conditions, the graph of all coalesced optical paths within one waveband (including
intranodal paths) must form a set of directed trees. (The trees need not be disjoint.)
514 Multiwavelength Optical Networks

1
At A

1 1 e′ f ′
1′ 1′
A
a′
f′

e′ F a′
6 2

6 2′

B
E
g′
h′

At B
a′ g′
2′
d′ b′ b′

7
D C
5 4′

5 4
(c)

Figure 6.43 (cont.)

That is, there can be no directed cycles or parallel paths. The graph of the coalesced
OPs defines a transformation from the set of original OPs si , Di to their expanded
counterparts si , Di , where Di ⊆ Di . In si , Di , the destination set Di consists of all
destinations “downstream” from si in its optical path tree. Note that there are generally
fewer OPs in the coalesced set than in the original set because all OPs originating from
the same source coalesce to a single OP rooted at that source. For example, in this case,
the two OPs from source 4 coalesced into one.50

50
When two OPs coalesce into a single, expanded OP, the naming of intended and fortuitous destinations
becomes ambiguous. We shall adopt the convention that when two OPs pi and p j , with intended destination
sets Di and D j , respectively, coalesce into a single OP pk , the set of intended destinations for pk is
Dκ = Di ∪ D j .
 Wavelength/Waveband-Routed Networks 515

1
1′
A

f′

e′ F a′
2
6
2′
6

E B

h′

d′ b′

D c C
5 4′

At D
5 4
5 At C
c b′ c
d′ 4′
(d)

Figure 6.43 (cont.)

Because all routing discussions up to this point concerned a single waveband, the
waveband assignment problem was absent. Extending these ideas to a network with
several wavebands, the RWA problem consists of choosing a route and a waveband
for each required optical path so the coalesced OPs assigned to each waveband con-
sist of collections (forests) of directed trees. In making routing and waveband assign-
ments, it is generally desirable to keep the number of fortuitous destinations to a min-
imum. As we shall see, this tends to minimize the number of distinct wavelengths
516 Multiwavelength Optical Networks

required in allocating λ-channels to the optical connections supported by the OPs. Static
routing rules should be designed to achieve this and other performance objectives in-
dependent of the type of optical and logical connectivity embedded onto the optical
paths.

6.5.2 Optical Connections: λ-Channel Assignment

Having established optical paths, the next issue is channel assignment. This problem is
examined in detail for both optical and logical connections in Section 6.5.6. To lay the
groundwork for the channel assignment problem, this section focuses on certain basic
relations between optical path selection and channel (wavelength) assignments for optical
connections. Because the DSC condition must be observed at the optical path level, the
only remaining constraint that needs attention now is distinct channel assignment. We
examine this issue now, limiting attention once again to connections within a single
waveband. (In the multiwaveband case, the channel assignment procedures are executed
independently for each waveband.)
Following the notation for optical paths, we specify optical connections (OCs) in the
form ci = (si , Di ), where Di is the intended destination set. The OC has the expanded
form ci = (si , Di ), where Di is the expanded destination set of its supporting OP. Thus,
all destinations of the OP that are not intended destinations of the carried OC become
the optical connection’s fortuitous destinations.
Suppose the following optical connections are to be established in one waveband on
the Petersen network:

c1 = (1, {5, 7})

c2 = (2, 6)
c3 = (4, 1)
(6.54)
c4 = (4, 5)
c5 = (6, 3)
c6 = (7, 1).

Except for the last two, these can all be carried on the OPs of Equation (6.53). By adding
two more paths – p5 = 6, 3 and p6 = 7, 1 – all connections can be accommodated.
A list of the routes used for these OCs appears in Table 6.3. Each OC uses one
of the five OPs defined earlier, and each table entry indicates the sequence of fibers
carrying a signal from a source to one of its destinations (including the access fibers).
Correspondence between OCs and OPs is determined by their sources. Thus, both c3
and c4 are carried on p34 . There is one entry in each row for each intended destination
and one entry (underlined) for each fortuitous destination. The fortuitous destinations
are “inherited” from the destination set of the OP carrying the OC. (The signal power
on each OC fans out to all the destinations of its supporting OP, and the destinations it
reaches unintentionally become its fortuitous destinations.)
 Wavelength/Waveband-Routed Networks 517

Table 6.3 Routing table.

Destination

Connection Source 1 2 3 4 5 6 7

c1 1 — — — — 1′ e′ d ′ 5 1′ e′ 6 1′ f ′ h ′ 7
c2 2 2′ a ′ 1 — — — 2′ a ′ e′ d ′ 5 2′ a ′ e′ 6 —
c3 4 4′ b′ a ′ 1 — — — 4′ b′ a ′ e′ d ′ 5 4′ b′ a ′ e′ 6 —
c4 4 4′ b′ a ′ 1 — — — 4′ b′ a ′ e′ d ′ 5 4′ b′ a ′ e′ 6 —
c5 6 — — 6′ ea3 — — — —
c6 7 7′ h f 1 — — — — — —

With the inclusion of the fortuitous destinations, the OCs of Equation (6.54) expand
to
c1 = (1, {5, 7, 6})1
c2 = (2, {6, 1, 5})2
c3 = (4, {1, 5, 6})3
(6.55)
c4 = (4, {5, 1, 6})4
c5 = (6, 3)2
c6 = (7, 1)1 .
A useful way to visualize these connections is as a directed optical connection hyper-
graph (OCH), whose tripartite representation is shown in Figure 6.44. In the OCH,
each hyperedge represents an optical connection, and the fortuitous destinations are
indicated by dashed lines. (Sources are listed on the left, destinations are on the right,
and connection labels are in the middle.) Because our sources and destinations are
assumed to be contained in elementary stations in this example, all OCs terminating
at the same destination station (fortuitous or not) are carried on the same access fiber
and may therefore interfere with each other. (A station may contain several optical re-
ceivers, and thus it may terminate more than one OC.) Because more destinations tend
to produce more potential interference, it is important to keep the number of fortuitous
destinations small. To do so requires intelligent routing decisions at the optical path
level.
Note that the list of optical connections (or, equivalently, the OCH) defines completely
the optical connectivity of the network within a given waveband while suppressing
the details of the network’s physical topology and the specific routes taken by the
connections. Thus, in moving up from the optical path layer to the optical connection
layer, the details of the former are left behind.
Once the expanded destination sets for all prescribed OCs are determined, these con-
tain the information necessary to determine λ-channel assignments that satisfy the
DCA condition. A systematic procedure for making these assignments is given in
Section 6.5.6.1. As we shall see, the connections in the example of Equation (6.55)
518 Multiwavelength Optical Networks

C
1
1 1

2
2 2

3
3 3

4
4 4

5
5 5

6
6 6

7 7

Figure 6.44 Optical connection hypergraph.

can be realized using four distinct wavelengths, with a possible λ-channel assignment
indicated by the subscripts on the connections.

6.5.3 Significance of Nonblocking Access Stations in LLNs

We have seen that the feasible optical paths in an LLN, and the feasible optical con-
nections and channel assignments, are constrained by the access links as well as the
internodal network topology, just as they were in WRNs. This introduces an added and
sometimes unnecessary level of complexity into the network model. For example, OP
coalescing, fortuitous destinations, and potential interference of OCs all depend on how
the stations access the nodes. It was for this reason that a distinction had to be made
between general and elementary NASs.
To understand the significance of a nonblocking station, let us return to the question
of OP routing on the Petersen network operating on a single waveband. Recall that we
had attempted to use shortest path routing for all OPs in the list of Equation (6.52) but
were unable to route p4 via the shortest path without violating the DSC constraint. Now,
suppose we replace elementary station 4 by a nonblocking station resolved into the three
elementary stations C1 , C2 , and C3 , as shown in Figure 6.42. This allows us to split the
single source at station 4 into two separate sources, say C1 and C2 . Using these for the
two paths originating at station 4, we now obtain the following OPs:
p1 = 1, {5, 7, 6}
p2 = 2, {6, 1, 5}
(6.56)
p3 = C1 , {1, 5, 6}
p4 = C2 , 5 ,
 Wavelength/Waveband-Routed Networks 519

where p4 is now shortest path routed on fiber c. The change in routing is possible
because the two OPs that originated at source 4 are no longer superimposed on the same
access fiber, so they do not coalesce at the source.51 This is a simple illustration of the
fact that nonblocking stations increase routing options by eliminating the inseparability
constraint on the access fibers.
When analyzing a large network, it is often convenient to assume that all traffic
sources and destinations are contained in nonblocking access stations, with one station
at each node. (This is true for WRNs as well as LLNs.) In this way, a “stripped down”
network model can be used that consists only of the ONNs and internodal links. Because
there is a one-to-one correspondence in this case between nodes and stations, all optical
paths and optical connections can be identified in terms of source and destination nodes
rather than stations. In making this identification, the assumption is that in extending
each internodal path backward to a source and forward to a destination, paths on the
same waveband are always routed on different access fibers. Consequently, using one
nonblocking station per network node, (1) optical paths on the same waveband coalesce
only if they share a common internodal fiber and (2) optical connections are potentially
interfering only if they share a common internodal fiber.
As we shall see in subsequent examples, the use of nonblocking stations can often
lead to significant improvement in network performance.

6.5.4 Local Access to LLNs

As in traditional networks, good designs for optical networks are often based on hierar-
chical principles. We suggested this first in Section 5.1, in which a star network was used
as a backbone for a larger network and clusters of end systems accessed the backbone
nodes via local access subnets (LASs). The same hierarchical principle applies to any
wide area LLN. The backbone might consist of a network with general mesh topology,
and the LASs would be in the form of trees, as in Figure 5.1.
An example of three LASs concentrating traffic onto nodes A and C in the Petersen
network is shown in Figure 6.45. Each LAS is assumed to connect to a “gateway”
network node via a single fiber pair access link. Of particular interest here is the purely
optical version of the LAS; see subnet 2 in Figure 5.1(b), in which the access subnet
consists of two parallel directed trees. Traffic from the stations is routed upstream to the
gateway on one tree, and traffic from the gateway is routed downstream to the stations
on the other tree.
In an LLN, the nodes within the LAS might be either fixed optical splitters and
combiners or controllable LDCs. The net effect of the LAS is to concentrate the traffic
from all attached stations onto the link accessing the gateway. We denote each LAS
using a subscripted node label: A1 for the LAS attached to node A, and C1 and C2 for
those accessing node C in Figure 6.45. Superscripts are used to identify stations within
(1) (2) (3)
an LAS. Thus, LAS C1 contains stations C1 , C1 , and C1 .

51
They still coalesce at destination station 5 if it is an elementary station. However, this does not violate the
DSC constraint. Why?
520 Multiwavelength Optical Networks

(3)
(4)
A1 A1
(1) (2)
A1 A1
A1

ONM
C

C1 (1) C2
C1
(3)
(2) C1
C1

Figure 6.45 Local access subnets on the Petersen network.

Because signals from all stations within the LAS are multiplexed onto a common
fiber accessing its gateway node, some sort of control must be exercised to coordinate
transmissions within the LAS. For example, the gateway node might be equipped with
an optical network manager (ONM) as shown at node C in Figure 6.45, which processes
requests for demand-assigned logical connections from stations within the LASs access-
ing its node. This might include managing admission control, assigning transmission
channels, providing synchronization, and executing various other functions required for
efficient utilization of the optical bandwidth.
To illustrate the operation of the ONM, suppose an optical path C1 , A1 is set up
to carry LCs from stations in LAS C1 to those in LAS A1 . (Because all stations in
the LAS are attached to the same access fiber, the complete set of stations constitutes
the source of the optical path.) If the waveband supporting the path contains several
λ-channels, fixed-frame TDM/T-WDMA might be used for multiplexing the LCs onto
the optical path, where each station is assigned the number of channel–slots needed
for its traffic. Node C then acts like a directed star coupler, broadcasting all traffic
 Wavelength/Waveband-Routed Networks 521

from C1 to all stations in A1 .52 Because stations do not see their own transmissions,
a separate bidirectional channel might be set aside for signaling and synchronization
purposes. In this case the ONM would receive connection requests from the stations on
the upstream signaling channel and would monitor the transmissions from all stations
to verify synchronization. It would broadcast commands to all stations on the down-
stream channel, which would include channel–slot assignments as well as synchroniza-
tion signals. Other optical paths on other wavebands would allow C1 to communicate
with other gateway nodes. Provided that the necessary coordination is exercised by the
ONM, all LASs accessing a gateway are seen by the network as one elementary access
station.

6.5.5 Routing Waveband and Channel Assignment on the Petersen Network

In the previous section, we focused on routing and channel assignment in LLNs in the
case of a single waveband. We now examine static RWA in more elaborate contexts,
continuing with the Petersen network as an illustrative example.
Any routing scheme based on a single waveband runs out of wavelengths rapidly
as the number of connections increases. (Full connectivity among all 10 nodes in the
Petersen network using a spanning tree to route the optical paths requires 90 wavelengths
if one λ-channel is used for each connection.) Thus, to accommodate larger numbers of
connections within a limited optical spectrum, more efficient use must be made of the
fibers and the spectrum. We propose an approach now, based on the use of overlapping
nonspanning trees and multiple wavebands, illustrating it using the Petersen network
once more. Of course, this network is a special case because of its density and regularity.
However, many of the conclusions we draw from this simple example can be extrapolated,
at least qualitatively, to larger networks.
The problem we pose is to determine static routing, waveband, and channel assign-
ments in the Petersen network to support full connectivity among all the nodes in cases
in which several wavebands are available. It is assumed throughout that each internodal
link contains a single fiber pair. We concentrate first on waveband routing. Using static
routing rules, the internodal LDC settings remain fixed. However, usage of the optical
paths may vary dynamically depending on station activity. We make the following three
assumptions:

1. Each LC is point to point and uses the full capacity of a λ-channel.

2. A single access station is attached to each network node.
3. The network operates on W wavebands, with C λ-channels per waveband.

The second assumption allows some simplification of notation; we use the same label
for a station and the node it accesses.

52
If the LDC at the gateway node is set to combine all signals from C1 and C2 , then the transmitting set for
the directed star includes both LASs, and all stations in both subnets constitute the source of the optical
path.
522 Multiwavelength Optical Networks

Networks using two different values of W will be compared. The routing rules to be
explored are all based on multiple trees, chosen to exploit network resources efficiently.
Each tree, Ti , is a star centered at node i and consisting of three links as follows (refer
to Figure 6.41):

T A = {a, f, e}
TB = {b, g, a}
TC = {c, l, b}
TD = {d, m, c}
TE = {e, d, k}
(6.57)
TF = { f, h, i}
TG = {g, n, j}
TH = {h, l, p}
TI = {i, n, m}
TJ = { j, p, k}.

Although these are not spanning trees, their union “covers” the network in the sense
that each node can reach every other node using paths confined to one of the trees to
which it belongs. Furthermore, routing on these trees is necessarily shortest path. Two
cases are considered: (1) W = 3 (multicast optical paths) and (2) W = 5 (point-to-point
optical paths).

Case 1: W = 3, Multicast Optical Paths

In this case we assume that all stations are nonblocking. Because the trees are not fiber
disjoint, we cannot route connections independently on each tree. However, if each tree
is assigned its own waveband, the trees can be made “waveband disjoint” (i.e., if two
trees share a fiber, they are assigned different wavebands so that they lie on different
“copies” of the network). It turns out that the minimum number of wavebands needed
to accomplish this “tree coloring” is W = 3. A possible assignment is T A , TG , and TH
on waveband 1; TB , TD , TF , and TJ on waveband 2; and TC , TE , and TI on waveband 3.
We now configure the exterior nodes of each tree as an MPS, as described in Sec-
tion 3.4.2. As shown in Figure 6.46(a) for T A , a broadcast star joining nodes B, E, and
F is embedded on the tree by setting the LDC at the central node to emulate a star
coupler. The emulation can be in the form of either a reflecting or a nonreflecting star.
In the former case, illustrated in the intranodal connection matrix in Figure 6.46(b), the
LDC is set to combine all signals arriving at node A on fibers a ′ , f , and e and then
“reflect” them back along all outbound fibers. In the nonreflecting star, illustrated in the
intranodal connection matrix in Figure 6.46(c), the set of signals on an outbound fiber
includes all signals except the signal on the corresponding input fiber; that is, there are
no loopback connections. As we shall see, network performance changes significantly
depending on which option is used. Note that all OPs in Figure 6.46(b) are two-hop
paths (excluding the access links), and the LDC settings apply to waveband 1 only. For
reasons to be explained later, station A attached to the central node has been excluded
 Wavelength/Waveband-Routed Networks 523

e a
f

E F B

(a)

B B a′ a B B
At A At A
a e′ f ′ a e′ f ′
e e′
E E A E E a′ a′
e e
f f′ f f
F F F F

(b) (c)

Figure 6.46 Embedded star on tree T A .

from the broadcast star. The resultant OPs on T A are then

p1 = B, {E, F, B}
p2 = E, {B, F, E} (6.58)
p3 = F, {B, E, F}

for the reflecting star and

p1 = B, {E, F}
p2 = E, {B, F} (6.59)
p3 = F, {B, E}

for the nonreflecting star.

A number of options are possible for channel assignment, depending on station
equipment and LDC settings. These are explored in detail in Section 6.5.6. For now, let
us assume that each node is configured as a nonreflecting star. Under our assumptions,
a full point-to-point OC must be dedicated to each LC. Then, to provide full logical
connectivity among the exterior nodes, the following optical connections are established
524 Multiwavelength Optical Networks

Table 6.4 Waveband routing in the Petersen graph: W = 3.

Destination

Source A B C D E F G H I J

A — 3 2 3 2 3 2 2 2 3
B 3 — 1 3 1 1 3 3 1 1
C 2 1 — 1 2 1 2 2 2 1
D 3 3 1 — 1 3 1 1 1 3
E 2 1 2 1 — 1 2 2 2 1
F 3 1 1 3 1 — 3 3 1 1
G 2 3 2 1 2 3 — 2 2 3
H 2 3 2 1 2 3 2 — 2 3
I 2 1 2 1 2 1 2 2 — 1
J 3 1 1 3 1 1 3 3 1 —

on tree T A :
c1 = (B, {E, F})1
c2 = (E, {B, F})1
c3 = (B, {F, E})2
(6.60)
c4 = (F, {B, E})2
c5 = (E, {F, B})3
c6 = (F, {E, B})3 .
Because these are all to be used as point-to-point connections even though they
each have two destinations, only one destination is shown as intended. (Only the three
connections – c1 , c2 , and c4 – would have been required if each optical connection
was shared between two LCs.) The subscripts on the connections indicate λ-channel
(wavelength) assignments obeying DCA, found using the procedures of Section 6.5.6.1.
This requires C = 3 λ-channels to support the six connections. Thus, there is a twofold
spectrum reuse within the tree. In contrast to the two-hop exterior node paths, the paths
from the central node, A, to the other nodes on the tree are all one-hop paths. Although
they could have been included in the embedded star, it is more efficient to route them
outside the tree. This can be done using no additional spectrum. Note that the waveband
assignments used for all trees allocate only two wavebands to each fiber, leaving the third
free. We use this for the single-hop connections. For example, referring to Figure 6.41
and Equation (6.57), fibers a and f both have waveband 3 free because T A operates on
waveband 1, and TB and TF operate on waveband 2. Similarly, fiber e has waveband 2
free. The unassigned wavebands can be used to route the one-hop connections on these
fibers.
Using the RWA rules just described, a possible waveband assignment for this case is
constructed as shown in Table 6.4. (Note that the routing assignments are symmetric.)
 Wavelength/Waveband-Routed Networks 525

Using Table 6.4 and the list of trees in Equation (6.57), we can deduce how optical
connections are distributed on each path and waveband. For example, focusing on node
A and continuing with the assumption of a nonreflecting star at each node, we find from
row A in Table 6.4 that the OPs originating at station A are

pC G = A, {C, G} 2
p H I = A, {H, I } 2
p E = A, E 2
(6.61)
p D J = A, {D, J } 3
p B = A, B 3
p F = A, F 3 ,

where a subscript ∗ j indicates that the OP is carried on waveband j. Although all of

these paths originate from the same NAS, the paths on the same waveband do not coalesce
because they use disjoint internodal fiber paths, and the source station is nonblocking.
For similar reasons, paths assigned to the same waveband and converging on the same
destination do not coalesce.
There are nine point-to-point optical connections originating at station A:

cC = (A, {C, G})

cG = (A, {G, C})
c H = (A, {H, I })
c I = (A, {I, H })
c E = (A, E) (6.62)
c D = (A, {D, J })
c J = (A, {J, D})
c B = (A, B)
c F = (A, F).

The OP carrying each connection is determined by matching the path subscript in

Equation (6.61) and the connection subscript in Equation (6.62). Thus, for example, cC
and cG are both carried by pC G . Note that the two-hop connections are on multicast OPs,
whereas the one-hop connections are on point-to-point OPs. All nine OCs originating
from A are supported by six OPs, divided equally between two wavebands. We shall see
in Section 6.5.6.1 that all 90 connections in the network can be supported using three λ-
channels per waveband. (For example, all two-hop connections on T A can be supported
on three channels on waveband 1, as indicated in Equation [6.60].) Each single-hop
OP has the capacity of a complete waveband available to it (three λ-channels) but
supports only one connection. Thus, the capacity of the one-hop LCs can be tripled by
assigning all available channels to them without requiring any increase in the optical
spectrum.
526 Multiwavelength Optical Networks

E C
3 2

TA 5 TB
2 3
a

4 1
F G

Figure 6.47 Waveband assignments: W = 5.

Case 2: W = 5, Point-to-Point Optical Paths

In the previous example, the use of both multicast and point-to-point optical paths
resulted in a nonuniform capacity available for LCs. We change that situation now
by making all OPs point-to-point, thus eliminating all fortuitous destinations. Each
internodal link again contains a single fiber pair, and we initially use nonblocking access
stations.
It turns out that it is possible to create point-to-point optical paths if three wavebands
are allocated to the six two-hop paths on each tree, with a total waveband complement of
W = 5 sufficient for the network. (The original trees no longer exist as entities carried
on a single waveband; instead they are broken down into point-to-point paths, i.e.,
smaller trees.) An example of a suitable RWA for two overlapping trees, T A and TB , is
shown in Figure 6.47. The wavebands are assigned symmetrically. For example, optical
connections (E, B) and (B, E) use the same waveband, and all waveband assignments
on each fiber are distinct. For example, on fiber a, wavebands 3 and 4 support two-
hop connections on T A ; wavebands 1 and 2 support two-hop connections on TB ; and
waveband 5 supports the single-hop connections ( A, B) and (B, A). A similar pattern
is followed throughout the network, with the complete waveband assignment shown in
Table 6.5.
It is easily verified from Table 6.5 and Figure 6.41 that each fiber carries five distinct
wavebands. Using nonblocking stations (each containing three elementary stations), nine
separate OPs can be set up from each source station to all others, with the destination
station associated uniquely with an elementary source station–waveband combination.
For example, station A contains elementary stations A 1 , A 2 , and A 3 , which can be
used for paths to B, E, and F, respectively, all using waveband 5. Thus, to reach
station E, for example, an optical connection is made from A 2 using a λ-channel on
waveband 5. Because its optical path is point to point, the optical connection has no
fortuitous destinations. Similar arrangements must be made at destination stations. Thus,
for example, station B is the destination for paths on waveband 4 from stations F, G,
and H . To avoid violations of the DCA constraint, these paths should be routed to
different elementary stations: say, B1 , B2 , and B3 , respectively. In this way, station B
can select a signal from any other station using a unique elementary station–waveband
combination. Because all optical connections are either fiber or waveband disjoint, one
 Wavelength/Waveband-Routed Networks 527

Table 6.5 Waveband routing in the Petersen graph: W = 5.

Destination

Source A B C D E F G H I J

A — 5 2 1 5 5 1 3 1 4
B 5 — 1 5 3 4 4 4 5 2
C 2 1 — 2 3 4 3 3 4 5
D 1 5 2 — 4 3 1 1 5 5
E 5 3 3 4 — 2 4 5 2 3
F 5 4 4 3 2 — 2 5 5 1
G 1 4 3 1 4 2 — 3 3 5
H 3 4 3 1 5 5 3 — 4 4
I 1 5 4 5 2 5 3 4 — 4
J 4 2 5 5 3 1 5 4 4 —

λ-channel per waveband is sufficient to support all connections without violating the
DCA condition.
With a single λ-channel per waveband, the LLN is equivalent to a WRN because
all connections are point-to-point (there is no signal dividing or combining), and a
waveband is equivalent to a wavelength. All connections are on shortest paths, and
the average optical path length is H̄ = 5/3. Using this value in the bound, W Netcap , of
Equation (6.11) we find that W = 5 is the minimum possible number of wavelengths
for full connectivity in this network. The fact that this assignment achieves the bound
indicates that all links are fully loaded.
The situation becomes more interesting if the nonblocking stations are replaced by
elementary stations. Now all OPs originating at a common station on a common wave-
band coalesce to a single OP. For example, the 24 point-to-point paths on waveband 5
coalesce to 10 ( mostly) multicast paths:

p A = A, {B, E, F}
p B = B, {A, D, I }
pC = C, J
p D = D, {B, I, J }
p E = E, {A, H }
(6.63)
p F = F, {A, H, I }
pG = G, J
p H = H, {E, F}
p I = I, {B, D, F}
p J = J, {C, D, G} .
528 Multiwavelength Optical Networks

The optical connections on each multicast OP now inherit fortuitous destinations from
the OP. For example, p E supports optical connections (E, {A, H }) and (E, {H, A}). The
fortuitous destinations produce limitations on the feasible λ-channel assignments, with
the result that three wavelengths instead of one are now required to satisfy the DCA
constraints (see Section 6.5.6.1).

6.5.6 Channel Assignment

In an LLN, channels must be assigned at two levels: (1) each optical connection must
be assigned a λ-channel and (2) each LC must be assigned a transmission channel.53
The objective of a good channel assignment procedure is to satisfy traffic requirements
while making efficient use of the available optical bandwidth. We now give a systematic
approach to channel assignment that achieves that objective in the case when optical
paths are in place based on static RWA rules. It is important to recognize at the outset
that channel assignments in an LLN are made independently for each waveband. The
discussion that follows applies to channel assignments within one waveband.

6.5.6.1 λ-Channel Assignment

We begin at the λ-channel level, focusing on wavelength assignments for optical connec-
tions that are assigned without any channelization at the transmission channel level (i.e.,
without dividing a λ-channel into subchannels). This implies that there is no rapid tuning
of transceivers while a connection is active, so the system is purely WDM/WDMA.
To determine the feasible wavelength assignments, we must revisit the DCA con-
dition, which was first discussed in the context of WRNs in Section 3.3.1.2. To
this end, a definition is needed to clarify the notion of interference between optical
connections:
Two optical connections are potentially interfering if they are on the same waveband and
the intersection of their destination sets contains an element that is an intended destination
for at least one of the connections.54

The DCA condition can now be restated as follows:

All potentially interfering optical connections must be assigned λ-channels on distinct wave-
lengths.

A few explanations are in order:

1. Two connections do not interfere when their only common destinations are fortuitous.
(In this case no problems are caused by assigning them the same λ-channel.)
53
Depending on how the wavebands and optical connections are channelized, assignment alternatives may
or may not exist at each level. For example, if each waveband contains only one λ-channel, as in a WRN,
there are no λ-channel alternatives. Similarly, if each LC uses the full capacity of a λ-channel, there is no
channelization of optical connections and therefore no transmission channel alternatives.
54
Note that two connections that are on the same waveband and share a fiber necessarily have identical
destination sets and therefore are potentially interfering according to this definition.
 Wavelength/Waveband-Routed Networks 529

1 2
1 2 3 3

4 4 2 5 6 1

Figure 6.48 Connection interference graph.

2. Two optical connections destined for the same nonblocking NAS on the same wave-
band do not interfere if they do not share any internodal links. (The extra access
fibers on nonblocking stations eliminate interference.)
3. Two optical connections originating from the same elementary station on the same
waveband are always potentially interfering. Because they share a common access
fiber, they are necessarily supported by the same optical path, which makes their
destination sets identical except for the labeling of the intended destinations.

We now present a procedure for assigning λ-channels to a prescribed set of op-

tical connections within a common waveband to minimize the number of required
channels while satisfying the DCA constraints. (In a multiwaveband network, the pro-
cedure is applied independently on each waveband.) It is based on the construction
and vertex coloring of a connection interference graph G CI , similar to G PI defined in
Section 6.3.3.
Each optical connection on a given waveband is represented by a vertex of G CI , and
a pair of vertices is joined by an edge if their connections are potentially interfering.
Then, a feasible λ-channel assignment is equivalent to a vertex coloring of G CI .
To illustrate the procedure, we apply it to the optical connections of Equation (6.55),
shown graphically in the OCH of Figure 6.44. These represent six connections on a
spanning tree in the Petersen network. The graph G CI for these connections is shown
in Figure 6.48. Because we have assumed that the access stations are elementary, the
existence of an edge between two vertices i and j of G CI is determined by checking
whether the connections ci and c j have a common destination. If so, and if the destination
is not fortuitous for both of them, an edge is placed in G CI . For example, there is an
edge joining vertices 1 and 2 in G CI because c1 and c2 both have the destination 6, and
it is an intended destination for c2 .
This graph is colored easily by inspection using four colors, indicated by the numbers
next to the vertices. (This is the channel assignment that appears in Equation [6.55].)
Furthermore, it is a minimum coloring because the largest clique in G CI is of size four
(see Appendix A).
The connection interference graph is a useful tool in comparing different scenar-
ios for operating an LLN because it reveals how different routing configurations (e.g.,
reflecting versus nonreflecting stars), different station structures (e.g., nonblocking ver-
sus elementary), and different optical connection alternatives (point-to-point versus
530 Multiwavelength Optical Networks

1 1 1 2 3 2

2 4 5 3 6 3

Figure 6.49 Connection interference graph for Equation (6.60).

multicast) affect channel assignments. We now illustrate its application using several of
the RWA examples treated in the previous section.

Petersen Network: W = 3
In the Petersen network for the case W = 3, six two-hop point-to-point optical con-
nections are routed on each tree using a nonreflecting star. Equation (6.60) lists the
connections routed on tree T A using waveband 1. They produce the interference graph
of Figure 6.49. Note that G CI can be colored using three colors, leading to the λ-channel
assignments indicated as subscripts in Equation (6.60). Because nonblocking stations
are used in this example, each tree can be treated in isolation. The same analysis on each
tree indicates that C = 3 λ-channels per waveband is sufficient for all connections in
the network. It is left as an exercise for the reader to show that with reflecting stars, six
wavelengths are required.
It is interesting to reexamine the channel assignments in the case in which multicast
optical connections are used for each pair of destinations on the tree. In this case
Equation (6.60) is replaced by
c13 = (B, {E, F})
c25 = (E, {B, F}) (6.64)
c46 = (F, {B, E}),
where each optical connection carries two LCs, which now must share one λ-channel.
For example, c13 carries LCs [B, E] and [B, F], which could share a common λ-channel
using TDM. This connection set clearly requires three different wavelengths. Although
the same number of wavelengths are used in both the point-to-point and multicast cases,
the latter makes less efficient use of the optical spectrum because there is no channel
reuse. As a result, each LC receives only half as much capacity as in the point-to-point
case. However, an advantage of optical multicast is that only half as many OTs are
required for these connections.

Petersen Network: W = 5
All optical paths are point to point in this case, provided that the stations are nonblocking.
Thus there are no potentially interfering optical connections, and channel assignment is
trivial: one λ-channel per waveband is sufficient.
As explained in Section 6.5.5, elementary stations can also be used here, but they
produce multicast OPs. The required number of λ-channels per waveband then increases,
 Wavelength/Waveband-Routed Networks 531

because there are less opportunities for spectrum reuse. In the case of multicast OPs,
there are a number of channel assignment options depending on the performance criteria
that are most important. We consider two possibilities: (1) multicast optical connections
(most economical) and (2) point-to-point optical connections (highest throughput).
Considering the first option, and focusing on waveband 5, we note that there are 24
optical paths, which coalesce to a set of 10, listed in Equation (6.63). We identify each
coalesced OP with one optical connection, where the optical connection is multicast
whenever its supporting OP is multicast. For example,
c F = (F, {A, H, I }) (6.65)
is supported by p F in Equation (6.63). It is left as an exercise for the reader to show that
in this case three wavelengths are sufficient to support all 10 optical connections (see
Problem 22).
Because we are interested in setting up full point-to-point logical connectivity, each
multicast optical connection is required to carry several point-to-point LCs. For example,
c F carries [F, A], [F, H ], and [F, I ]. Because multiplexing of LCs onto an optical
connection is a transmission channel issue, it can be separated from the issue of optical
connection assignments (see Section 6.5.6.2).
This approach does not offer the full capacity of a λ-channel to each logical channel.
If that is required, we must use the option of point-to-point optical connections. For
example, p F now must carry (F, {A, H, I }), (F, {H, A, I }), and (F, {I, A, I }) – each
assigned a distinct wavelength and each originating at a separate OT. To determine the
minimum number of wavelengths required in each waveband in this case, a minimal
coloring of G CI for the resultant connections on that waveband must be determined.
(One more exercise for the reader!) The complexity of the channel assignment (graph
coloring) problem will become apparent to the courageous reader because there are now
24 connections on waveband 5.
Although an optimal λ-channel assignment is more difficult when each multicast
optical connection is replaced by a set of point-to-point connections, it is easy to find
an upper bound on the required number of wavelengths in this case, together with a
suboptimal channel assignment meeting the bound. We note that each threefold multicast
connection that previously required one λ-channel now needs three λ-channels on distinct
wavelengths, because it is broken down into three point-to-point optical connections
sharing the same OP. Because threefold multicast is the worst case in this example,
an upper bound on the number of λ-channels on waveband 5 for point-to-point optical
connections is C = 9 (three times the number of channels used in the multicast OC
case). This reasoning can be applied generally whenever comparing multicast optical
connections and their point-to-point counterparts.55
Summarizing these examples, we see that more throughput is obtained with point-
to-point optical connections because λ-channels do not have to be shared. However,
multicast connections require fewer OTs and usually use less optical bandwidth.
55
The worst case can be overly pessimistic. For example, we found that the point-to-point connections of
Equation (6.60) required no more wavelengths than the twofold multicast connections of Equation (6.64),
both of which were derived from the OPs of Equation (6.59).
532 Multiwavelength Optical Networks

6.5.6.2 Shared Channels: Transmission Channel Assignment

There are many opportunities in LLNs for sharing λ-channels among several LCs using
multiplexing/multiple-access techniques at the transmission channel level. The special
channel-sharing properties of the LLN facilitate the grooming of transmission channels
into λ-channels for efficient utilization of fiber capacity. (See Chapter 7 for further
discussion of grooming.) These possibilities exist because of the multipoint nature of
the optical paths. (The broadcast star, which was explored thoroughly in Chapter 5, is the
simplest and the most effective example of the power of multipoint optical connectivity.)
Channel sharing provides a number of advantages:
r More flexibility in allocating capacity to LCs
r Finer granularity of LCs (see Section 6.5.6.4)
r More efficient use of the optical spectrum
r Transfer of the burden of equipment replication from the optical level to the electronic
level
In most cases the discussion of multiplexing/multiple access in Chapter 5 applies di-
rectly to channel sharing in LLNs. However, extra complications arise in certain special
situations, so it is helpful to treat the following three cases separately: (1) multicast
optical connections, (2) many-to-one optical connections, and (3) many-to-many optical
connections.

Multicast Optical Connections

An N -fold multicast OC creates an embedded 1 × N shared-channel broadcast medium,
which can be used to support several LCs (either point-to-point or multicast). The LCs
can be multiplexed onto a common λ-channel at the source station using a wide variety
of techniques. One possibility is TDM, in which each LC is assigned a transmission
channel consisting of one or more time slots in a fixed-frame schedule. (Other methods
discussed in Chapter 5 can also be used.) Because multiplexing occurs at the electronic
level in the case of multicast optical connections, there is an N -fold savings of OTs and
of the optical spectrum. Furthermore, the multiplexed connections can share the capacity
of the λ-channel in a way that matches their traffic requirements. There is no free lunch,
however! Whenever there is a saving in transmitters and spectrum there is generally an
equivalent reduction in aggregate throughput. In this case, the total throughput of all
LCs supported by one optical connection cannot exceed the capacity of one λ-channel.
On the receiving side, each receiver demultiplexes and selects the information destined
for it. Thus, the optical equipment in the receivers remains the same as for point-to-point
optical connections, but it is not fully utilized.
Applying these ideas to the case of the Petersen network with W = 5, using elementary
stations we note from Equation (6.63) that there are 10 optical paths on waveband 5,
eight of which are multicast because of the coalescing of paths from common sources.
Each of these can support a multicast optical connection, which in turn must carry
several LCs. For example, station F is the source of the threefold multicast connection
c F in Equation (6.65), which must carry three point-to-point LCs. On the receiving side,
each station must use one receiver for each optical connection destined to it, and extract
 Wavelength/Waveband-Routed Networks 533

the desired LCs using TDM in its reception processor. Thus on waveband 5, station A
requires three receivers tuned to λ-channels 2, 1, and 3 to receive LCs [B, A], [E, A],
and [F, A], respectively (see Problem 22). Following this reasoning for the remainder
of the connections, we find that by using optical multicast only 48 OTs are required for
the whole network, but 90 ORs are needed (as in the point-to-point case).

Many-to-One Optical Connections

The reasoning for many-to-one connections is similar to that for multicast or one-to-
many connections. Recall that in discussing λ-channel assignment in Section 6.5.6.1,
whenever several optical connections had the same intended destination they had to
be assigned distinct λ-channels. This was a consequence of the assumption that each
LC required the full capacity of a λ-channel. We now relax that assumption, admitting
the possibility of using multiple access (i.e., channel sharing) when several optical
connections are destined for the same receiving station. An objective is to save on ORs
and optical spectrum.
Continuing with the example of the Petersen network with W = 5, let us return to
the case of nonblocking stations using point-to-point optical connections. Originally,
this required placing nine ORs in each station, one for each LC from another station.
Now, let us use a multiple-access method to reduce the number of receivers. Suppose
each station is equipped with only one receiver on each waveband it uses. In cases when
several optical connections reach a station on the same waveband, any multiple-access
method (say, TDMA) can be used to distinguish the connections.56 For example, Table 6.5
shows that station A receives signals from stations B, E, and F on waveband 5. Using
a fixed-frame TDMA system, the LCs [B, A], [E, A], and [F, A] can be assigned time
slots according to their traffic requirements, so a single receiver at A, equipped with
demultiplexing electronics, can separate them. The transmitting stations must use one
OT for each LC, as in the point-to-point case. As a result, a total of 90 OTs and 48 ORs
are required in the network when multiple access is used. (Not all wavebands are used
at each station.)

Many-to-Many Optical Connections

In the many-to-many case, we must combine multiplexing with multiple access. One
approach that comes to mind is TDM/T-WDMA. But will this work? Let us try it on the
Petersen network with W = 5. To reduce optical equipment, we employ optical multi-
cast, allowing the use of elementary stations. We assume an FT-TR system using a single
transceiver per station on each waveband. Figure 6.50 shows a portion of the OCH on
waveband 5 for this case. Only transmitting nodes E and F and receiving nodes A and
H are shown. Note that each transmitting station multicasts to both receiving stations
and that additional connections exist (shown as dashed lines in the figure). The trans-
missions associated with the additional connections are ignored. The optical connec-
tions c E = (E, {A, H }) and c F = (F, {A, H, I }) use λ-channels 1 and 3, respectively,
56
TDMA requires synchronization of the different transmissions, which may be accomplished by feeding
back control signals from each destination station to the sources it sees.
534 Multiwavelength Optical Networks

Source Destination
A
cE
E

H
cF
F

Figure 6.50 Optical connection hypergraph.

corresponding to the assignments in Problem 22. Referring to Figure 6.41, it can be

seen that the four physical paths involved are all different and hence have different
propagation delays.
Figure 6.51 illustrates how the time frames might appear in a fixed-frame schedule with
L = 3 slots, where LCs [E, H ] and [F, A] are assigned two units of capacity. The frames
are shown at the sources and destinations, and the relative timing has been adjusted so
that the frames on the two channels are aligned properly at receiver A. Note that all
information can be extracted without conflicts at A. However, the transmissions arriving
at H are misaligned by an amount  because the propagation delays are different. As
a result, there are intervals (shown as shaded areas in the figure) during which conflicts
occur. There is, in fact, no way to schedule these transmissions without conflicts using
the given frame time and three slots.57
This is an example of what happens when many-to-many connections take multiple
physical paths. Rooted routing58 is designed to avoid this problem by forcing all paths
to pass through a common root, at which they can all be synchronized. However, in so
doing, the optical path emulates a broadcast star and hence sacrifices possibilities for
spectrum reuse.

[E,A] [E,H] [E,H] [E,A] From E At

t
At Destination
Source [E,A] [E,H] [E,H] [E,A] [F,H ] [F,A ] [F,A ] [F,H ] From F A
t t
E 0

At
Source [F, H] [F,A ] [F,A ] [F,H ] [E,A] [E,H] [E,H] [E,A] [E,H] From E At
t t
F 0 Destination
[F, H] [F,A ] [F,A ] [F,H ] From F H
t
0

Figure 6.51 Fixed-frame scheduling for four LCs.

57
Some ways out of this dilemma are suggested in Problem 23 at the end of this chapter.
58
The rooted routing approach based on an embedded star involves choosing a spanning tree of the network
and picking an appropriate node on the tree as its “root.” The root node is set to operate as a star coupler
and the other nodes are set to confine all paths to the chosen tree.
 Wavelength/Waveband-Routed Networks 535

We now consider another multiplexing/multiple-access approach for many-to-many

connectivity, continuing with the Petersen network example with W = 5 and elementary
stations. Section 5.2.2 showed how TDM/T-SCMA can be used to move multiple access
from the λ-channel level to the transmission channel level. This can be applied directly
to the problem at hand. We again use one optical transceiver per station per waveband.
Each OR captures all λ-channels within its waveband. The station structure is as shown
in Figure 5.7. As before, each OT has a λ-channel assignment that satisfies the DCA
conditions, and most optical connections are multicast, following the optical paths of
Equation (6.63). (All λ-channel assignments in this illustration follow Problem 22.)
Now on the transmission channel level, let each station be assigned a fixed subcarrier
frequency using the same pattern of frequency assignment as on the optical level. All
outbound LCs are then multiplexed on the subcarrier. For example, station A is assigned
the optical/subcarrier frequency pair (ν2 , f 2 ) on waveband 5, with LCs [A, B], [A, E],
and [ A, F] multiplexed on the subcarrier using TDM. Because three distinct wave-
lengths were required for optical channel assignment on waveband 5, a total of three
optical frequencies (corresponding to the wavelength assignments) and three subcarrier
frequencies are required. (The distinct optical wavelengths are needed only to avoid OBI
in this case, so there are no stringent tolerances on their exact values.)
On the receiving side, we assume that each station contains an array of three subcarrier
demodulators following each OR (for a total of nine demodulators on all wavebands)
to detect each LC. For example, station A receives signals on waveband 5 from three
different sources: B, E, and F. These are carried on subcarriers 2, 1, and 3, respectively.
By following the same pattern with subcarriers as with optical frequencies, the DCA
condition is satisfied at both the optical and transmission channel levels. It is important
to note that a reuse of the subcarrier spectrum is achieved here as well as reuse of the
optical spectrum. This leads to a TDM/T-SCMA system with a higher throughput than
would have been possible using a broadcast star.
When demultiplexing the TDM signals emanating from each source, the receiving
station experiences the same propagation delay problems as occurred in the TDM/
T-WDMA case. It is generally impossible to synchronize all sources and destinations
simultaneously. That is why a subcarrier receiver array is proposed here. In this way,
sources can transmit without mutual synchronization. Each receiver processes the de-
modulated subcarriers from all sources it sees, separately and in parallel, so no conflicts
arise. Could we have done the same thing using TDM/T-WDMA? Yes, but this would
have required replication of tuned ORs – a much more expensive proposition.
Still more savings of ORs are possible if a single wideband receiver covering all
wavebands is placed at each station. In this case, however, nine distinct subcarrier
frequencies are required to satisfy DCA at the subcarrier level, eliminating the possibility
of subcarrier frequency reuse. The same total number of subcarrier demodulators (nine)
is required at each station.

6.5.6.3 Performance Comparisons

The Petersen network served as a basic structure to illustrate RWA and channel as-
signment in LLNs. Because several different configurations were presented, we pull
536 Multiwavelength Optical Networks

them together in this section to indicate how different scenarios affect performance. The
following notation is used:
r W = number of wavebands
r Rt = transmitter bit rate (in bps) = capacity of one λ-channel
r C = λ-channels per waveband
r C f = capacity of one fiber (in bps) = W C Rt
r S = total network throughput normalized to the capacity of one λ-channel
r ̺ = S/WC = spectrum reuse factor
r Bλ = bandwidth occupied by one λ-channel, including interchannel guard band (in
GHz)
r Bg = guard band between wavebands (in GHz)
r α = Bg /Bλ = guard bandwidth normalized to λ-channel bandwidth
r Bop = total optical bandwidth occupied on a fiber (in GHz)
r β = Bop /Bλ = total optical bandwidth normalized to λ-channel bandwidth
r η f = C f /Bop = fiber spectral efficiency (in bps/Hz)
r η N = S Rt /Bop = network spectral efficiency (in bps/Hz).
Using this terminology we have59
C f = W C Rt = W Cηm Bλ , (6.66)
where ηm = Rt /Bλ is the spectral efficiency of the λ-channel modulation scheme, taking
into account the interchannel guard band, and
Bop = W C Bλ + (W − 1)Bg . (6.67)
The fiber and network spectral efficiencies can now be expressed as
ηm
ηf = (6.68)
1 + WW−1C
α
̺ηm
η N = S Rt /Bop = ̺η f = . (6.69)
1 + WW−1
C
α
With all channels within a single waveband (W = 1), Equation (6.68) reduces to
η f = ηm , the best possible fiber efficiency. With several channels in each waveband and
W ≫ 1, efficiency drops to η f ≈ ηm /(1 + α/C). Because α may attain values of 10 or
more, it is desirable to group many λ-channels on each waveband to limit the wasted
spectrum in the guard bands. However, we shall see that grouping channels to increase
η f tends to reduce the routing flexibility, which in turn reduces the reuse factor ̺ and
therefore the network spectral efficiency η N . Proper design of an LLN requires balancing
these two factors.
The performance parameters of particular interest in an LLN are S, ̺, η f , and η N ;
the latter three determine how efficiently the resources are being utilized. We now derive
these parameters for the various cases considered in Section 6.5.5. Instead of assuming
that each LC requires the full capacity of one λ-channel, as was done in most of the
previous examples, we leave C arbitrary, on the assumption that the LCs might require
59
It is assumed that the waveband/channel grouping is the same on each fiber.
 Wavelength/Waveband-Routed Networks 537

varying capacities. Thus, more than one λ-channel might be used for an especially high-
capacity LC or several low-traffic LCs between the same source and destination nodes
might share the capacity of one channel.

Case 1: W = 3
We focus on the scenario using nonblocking stations, point-to-point optical connections,
and nonreflecting stars. In this case, the six LCs routed on two-hop optical paths on each
tree share the C λ-channels on its waveband with a twofold reuse of the λ-channels.
Because there are 10 trees with twofold reuse of a capacity Rt C on each tree, the 60 two-
hop connections in the network are allocated an aggregate capacity of 20Rt C. However,
each of the 30 LCs routed on a one-hop path has the full capacity of a waveband available
to it. Therefore, these LCs are allocated an aggregate capacity of 30Rt C. (We assume
that all allocated capacity is used, which in this case leads to an unbalanced traffic
distribution.) This results in the following performance:
r S = 50C
r ̺ = 16.7
r ηf = ηm
1 + 0.67α/C
r η = 16.7ηm
N .
1 + 0.67α/C

Case 2: W = 5
Again, assuming nonblocking stations, there are no multicast optical connections in this
case, so we have
r S = 90C
r ̺ = 18
r η = ηm
f
1 + 0.8α/C
r η = 18ηm
N .
1 + 0.8α/C
Although the expressions for performance derived here provide useful qualitative
information, the results cannot be compared without assuming some specific values for
the system parameters. Let
r Bλ = 20 GHz
r Bg = 100 GHz (α = 5)
r Bop = 1000 GHz (β = 50)
r ηm = 0.1.

With these values of bandwidths, the values of C in the various cases are no longer
arbitrary. From Equation (6.67), we have
+ ,
β − (W − 1)α
C= . (6.70)
W
538 Multiwavelength Optical Networks

Table 6.6 Comparison of performance of various

configurations of the Petersen network.

Case C S ̺ ηf ηN

W =3 13 650 16.7 0.08 1.33

W =5 6 540 18 0.06 1.08
ST 50 50 1 0.1 0.1

Table 6.6 compares the performance of the representative cases discussed here. Recall
that each case corresponds to full logical connectivity on the Petersen network. Also
included in the table is the case ST, in which full connectivity is achieved using the
spanning tree of Figure 6.41 on a single waveband. In each case there are more channels
available than the number of LCs sharing them. For example, when W = 3, 13 channels
are available on each waveband. But only three stations access each tree, sharing the
capacity of a full waveband. Thus to make full use of the capacity of the 13 channels,
there must be at least 13 transceivers distributed among the three stations, with a suitable
multiplexing/multiple-access method used for sharing capacity.
Table 6.6 illustrates that total throughput S and network spectral efficiency η N are
increased by an order of magnitude by increasing the reuse factor ̺. A more modest
improvement in these parameters is achieved by concentrating the channels into a limited
number of wavebands (three versus five). Of course, different assumptions concerning
parameters such as guard bands will lead to different results.
Although all cases considered in these comparisons used the Petersen topology, it is
interesting to examine what happens when a less dense topology is used. For example,
if we had tried to achieve full logical connectivity on a 10-station ring using point-to-
point optical paths (i.e., operating it as a WRN), it would require W = 13 wavebands
(see Equation [6.16]). But with the values of bandwidths given earlier and W = 13, we
find that Equation (6.70) gives C = 0. Thus, the available spectrum cannot support full
connectivity on a ring operated as a WRN.

6.5.6.4 Demand-Assigned Logical Connections

Until now, our channel assignments were treated on a static basis with the basic unit being
an optical channel. However, additional opportunities for multiplexing and multiple
access at the transmission channel level occur when a network has a hierarchical structure
consisting of a backbone together with local access subnets. In this case it is often possible
to use demand-assigned multiplexing/multiple access within the LASs to concentrate
traffic efficiently for routing through the backbone. We illustrate with the network of
Figure 6.45, in which the Petersen backbone is assumed to be operating with demand-
assigned LCs superimposed on an optical layer in which the optical paths provide a static
connectivity pattern. In this way, connection management and control is concentrated in
the logical layer while ignoring the details of the underlying optical paths. Furthermore,
demand assignments at the optical and transmission channel levels allow for LCs with
arbitrarily fine granularity as well as varying capacities. In this section we investigate
 Wavelength/Waveband-Routed Networks 539

how the random nature of demand-assigned traffic, together with its granularity, affects
system performance.
To model random activity in the network, we assume that many elementary stations,
each with a single tunable transceiver, access each node. (The stations might access the
nodes directly or through LASs.) The station population is assumed to be sufficiently
large so that the connection request arrivals can be approximated as a Poisson process.
Assuming an LCC model, the blocking probabilities can then be deduced using the
Erlang-B formula of Equation (5.91). As we shall see, the blocking performance depends
on the way channels are shared in the LLN.
Let us assume that the collection of stations accessing each node generates an offered
traffic G (Erlangs per node), where the destinations of the connection requests are
distributed uniformly to N other nodes. (Any blocking because of busy destination
stations is neglected.) Suppose that one point-to-point optical connection is in place for
each of the N outbound paths to the other nodes; that is, one λ-channel is available
for each destination. In this case, each path is offered G/N Erlangs, so that the Erlang
formula reduces to PB = G/(N + G), which means that the offered traffic must be kept
extremely low for acceptable blocking probability.
As an example, consider the Petersen network with W = 9 and C = 1, in which many
elementary access stations are attached to each node possibly through LASs, as shown in
Figure 6.45. Each node generates G Erlangs, distributed uniformly to nine other nodes,
with one λ-channel available to each destination. Then we find that G ≈ 0.38 for PB =
0.04. This corresponds to a total normalized carried load (throughput) for the network
of S = 3.6, compared with a maximum throughput of S = 90 in the deterministic case.
The poor throughput for demand-assigned traffic is due to the fact that the number of
channels is too small (the granularity is too large).
A “brute force” way of improving PB is to assign many λ-channels (i.e., many parallel
optical connections) to each path. For example, taking C = 30 in the previous example,
we find that G ≈ 216 for PB = 0.04. This gives a carried load of S = 2074, compared
with S = 2700 in the deterministic case. Although this is much better, the random
nature of the connection activity still degrades throughput by more than 20% with a
nonnegligible blocking probability.
Unfortunately, in multiwavelength networks, large numbers of λ-channels may not
be available. However, blocking probability can still be reduced using a limited num-
ber of λ-channels by refining the granularity of the LCs. Suppose that instead of oc-
cupying a full λ-channel, each LC requires a lower effective bit rate: Rt /K , where
K ≫ 1. In this case up to K LCs can be multiplexed on each λ-channel with each
LC allocated one subchannel of a λ-channel. For example, if TDM/TDMA is used, a
subchannel would correspond to a time slot. With several λ-channels allocated to each
path, TDM/T-WDMA can be used so that the basic subchannel corresponds to one
channel–slot.
This procedure is valid for refining the granularity of any LLN. We shall apply it
to two cases of the Petersen network compared in Table 6.6: W = 3 and W = 5. We
determine the traffic-handling capacity of the “fine-grained” network, assuming that
each LC requires one channel–slot. Let f E (C, PB ) be the solution of Equation (5.91) for
540 Multiwavelength Optical Networks

G; that is, f E gives the maximum offered traffic to C channels for a blocking probability
not exceeding PB . Then for the cases being considered, the offered traffic per node
sustainable by the Petersen network with blocking probability not exceeding PB , using
K subchannels per λ-channel and C λ-channels per waveband, is
G = 9 f E (K C/3, PB ), for (W = 3)
(6.71)
G = 9 f E (K C, PB ), for (W = 5).

The constants multiplying f E in each equation reflect the traffic splitting factor at each
node. The first argument of f E indicates the number of subchannels available on each
path. (For the case W = 3, the argument K C/3 reflects the fact that the C λ-channels
in each waveband must be divided among three sets of connections to satisfy the DCA
condition. For W = 5, we assume full sharing.)
Using the values of G from Equation (6.71), the maximum carried traffic sustainable
by the Petersen network at a blocking probability PB is
S(K , C, PB ) = 10(1 − PB )G/K , (6.72)
where division by K is necessary to normalize the traffic to the capacity of one λ-channel.

6.5.7 Multistar Linear Lightwave Networks

Until now, all routing and waveband assignment problems in LLNs were treated on
networks in which the physical topology was given. We now consider the case when a
physical topology is designed for a network in the form of multiple broadcast stars. The
multistar physical topology was discussed in the context of a WRN in Section 6.3.6, in
which each star was based on a wavelength router, supporting point-to-point connections
only.60 In this section we take a broader view of the multistar network, as a physical
support for a hypernet. Now each star acts as an MPS joining a cluster of network
stations. The MPS is a generalization of a point-to-point network link, allowing all
stations in the cluster to communicate with each other through multicast optical paths.
Each MPS operates as a shared medium. Some of the early work on multistar-type
networks appears in [Birk91, Ganz+92].
Our first illustration of this approach appeared in Section 3.4.3.4. There, a seven-node
hypernet of diameter 1 is described, based on seven MPSs (see Figure 3.30). The LCs
in the network are represented in Figure 3.30(a) as a logical connection hypergraph, in
which each hyperedge represents an MPS containing three stations. Figure 3.30(b) gives
an alternative tripartite representation of the LCH, in which each hyperedge appears in
the form of a star. This representation can be viewed as a picture of a fiber topology
designed specifically to support the given LCs. By identifying each hyperedge in the
LCH with a multicast star in the network realization, we obtain a general methodology
for realizing any logical hypernet on a multistar physical topology.
Our interest here centers on hypernets of diameter 1, which correspond to purely
optical networks. (Recall that in a hypernet of diameter 1, every node can be reached
60
Topologies that are generalizations of multistars, called static light trees, are studied in [Barry93b].
 Wavelength/Waveband-Routed Networks 541

from every other through one MPS [i.e., on one logical hop]). To provide for logical
connectivity among all nodes in larger diameter hypernets, they must be realized as
hybrid LRNs, and these are treated in Chapter 7.
Consider a set of N stations that we wish to connect via an LLN. Suppose we are free
to choose both the logical topology and the physical topology, subject to the resource
constraints and performance objectives. The approach we take is to begin at the logical
layer and choose a hypernet logical topology that consists of sets of stations clustered into
MPSs. The natural representation of the hypernet is an LCH H (ν, ε), in which the vertices
ν represent the stations, and the hyperedges ε represent the MPSs (see Appendix A). If
each station is required to be reachable via a direct LC from any other, the LCH must
have a diameter of 1. After a suitable LCH is chosen, we construct a multistar network
patterned after the LCH, in which each star is realized as an MPS corresponding to a
hyperedge in the LCH. Although the topological parameters of the multistar (star sizes
and station degree) are taken from the LCH, the details of implementation and operation
of each star (i.e., the multiplexing/multiple-access methods used) are a separate issue,
independent of the logical layer design. Thus we have two subproblems that can be
treated (almost) independently: logical-layer design and physical-layer design.
The hypernet (or, equivalently, its LCH) can be directed or undirected. In an undirected
hypergraph, the hyperedges are subsets of vertices. If the hypergraph is of diameter 1,
each pair of vertices must belong to at least one hyperedge. In the undirected case,
the MPSs can be realized as undirected stars, with the tripartite representation of the
LCH defining the multistar network. All stations in an MPS communicate with each
other, fully sharing the capacity of the star using any of the channel-sharing techniques
described in Chapter 5. For example, the LCH of Figure 3.30 is an undirected hypergraph
with vertex set ν = {1, 2, . . . , 7} and hyperedges ε = {E 1 , E 2 , . . . , E 7 }, where

E 1 = {2, 5, 7}
E 2 = {1, 2, 6}
E 3 = {2, 3, 4}
E 4 = {1, 4, 5} (6.73)
E 5 = {1, 3, 7}
E 6 = {4, 6, 7}
E 7 = {3, 5, 6}.

In Equation (6.73) each hyperedge defines a 3 × 3 star.

In the directed case, each hyperedge (now called a hyperarc) is of the form E =
(E − , E + ), where E − and E + are, respectively, the in-set and out-set of E. These
correspond, respectively, to the transmitting and receiving stations of an MPS. In realizing
a directed hypernet, an MPS corresponding to a hyperarc E is implemented as a directed
s − × s + star, where s − = |E − | and s + = |E + | are the sizes of the in-set and the out-set
of E, respectively. This allows all stations in E − to multicast to all stations in E + , fully
sharing the capacity of the star. Note that E − and E + need not be disjoint, nor do they
have to be the same size.
542 Multiwavelength Optical Networks

T R
0 0

E0
1 1

2 2
E1

3 3

4 E2 4

5 5

6 6
E3

7 7

Figure 6.52 Directed hypernet GKH (2, 8, 4, 4).

Figure 6.52 shows a multistar realization of an eight-station directed hypernet of

diameter 1, with transmitters and receivers on the left and right, respectively. Four
4 × 4 directed stars provide full connectivity among the stations, and each station has
two transmitters and two receivers. In the language of hypergraphs, the LCH has the
vertex (station) set ν = {0, 1, 2, . . . , 7}. These vertices are clustered into four hyperarcs
(MPSs), each with in-size and out-size 4, and each vertex has in-degree and out-degree 2.
The in-degree and out-degree of a vertex correspond to the number of receivers and
transmitters, respectively, in the station it represents. The hyperarcs are
E 0 = ({0, 2, 4, 6}, {4, 5, 6, 7})
E 1 = ({0, 2, 4, 6}, {0, 1, 2, 3})
(6.74)
E 2 = ({1, 3, 5, 7}, {4, 5, 6, 7})
E 3 = ({1, 3, 5, 7}, {0, 1, 2, 3}).
Once the multistar realization has been determined, the flows on the stars can be
adapted to a prescribed traffic pattern using suitable multiplexing/multiple-access meth-
ods. For example, if the available spectrum consists of C = 4 λ-channels, each star might
be operated using TDM/T-WDMA in an FT-TR mode, in which the capacity of the four
channels is shared among 16 point-to-point LCs according to traffic demand. Any value
of C < 4 can also be used with a concomitant reduction in the total capacity of the MPS.
In general, each hyperarc in the LCH can be viewed as a fully shared, directed medium
with a capacity of C λ-channels. If the network consists of m stars, it achieves an m-fold
reuse of the spectrum.
It should be noted that any hypernet configuration based on multicast stars, as de-
scribed here, can be converted to a multistar WRN simply by replacing the star couplers
with wavelength routers. Thus, the example of Figure 6.52 becomes a fully connected
WRN if the central nodes of the stars are 4 × 4 permutation routers.
 Wavelength/Waveband-Routed Networks 543

It is of interest to find multistar designs that make efficient use of network resources,
accommodating a maximum number of stations with a minimum use of optical spectrum
and transceivers. To this end, we make use of Moore bounds for hypergraphs.
For the undirected case, let Nu (, D, r ) be the number of vertices in a hypergraph of
maximal degree , diameter D, and maximal hyperedge size r . Then the Moore bound
for undirected hypergraphs, given in Appendix A, reduces in the case D = 1 to

Nu (, 1, r ) ≤ 1 + (r − 1). (6.75)

It turns out that this bound can be attained only for certain pairs (, r ). The existence
of large hypergraphs of diameter 1 was studied thoroughly by [Bermond+84]. A few of
their results are presented here.
For the case  = r , undirected hypergraphs achieving the Moore bound exist when
q = r − 1 is a power of a prime. (The seven-vertex hypergraph in Figure 3.30 is an
example.) The bound is not attainable for certain other cases; for example, r = 7 and
r = 15.
It is also shown in [Bermond+84] that undirected hypergraphs exist with

Nu (2, 1, r ) = 2r − ⌈r/2⌉
Nu (3, 1, r ) = 3r − ⌈r/3⌉ (6.76)
Nu (4, 1, r ) = 4r − ⌈r/4⌉.

In the directed case, let Nd (d, D, s) be the number of vertices in a directed hyper-
graph with maximal out-degree d, diameter D, and maximal out-size s. Then, in
the case D = 1, the Moore bound for directed hypergraphs given in Appendix A
reduces to

Nd (d, 1, s) ≤ 1 + ds. (6.77)

Systematic design procedures for directed hypergraphs exist for wide ranges of pa-
rameters. Because the construction methods yield classes of regular hypergraphs of
arbitrary diameters, we discuss them more extensively in Chapter 7. In this section we
illustrate the design methodology using one class, called generalized Kautz hypergraphs,
considering only the case D = 1 [Bermond+97, Imase+83].
Let GKH(d, n, s, m) denote a generalized Kautz hypergraph containing n vertices
and m hyperarcs, where s is the out-size of each hyperarc and d is the out-degree
of each vertex. A hypergraph in this class exists whenever the parameters obey the
relations
dn ≡ 0 (mod m)
sm ≡ 0 (mod n).
With the vertices labeled as integers modulo n and hyperarcs labeled as integers
modulo m, the incidence rules are the following:

Vertex v is in the in-set of hyperarcs

e ≡ dv + α (mod m), 0 ≤ α < d (6.78)

544 Multiwavelength Optical Networks

and the out-set of hyperarc e consists of vertices

u ≡ −se − β (mod n), 1 ≤ β ≤ s. (6.79)

The diameter of the resultant hypergraph is bounded by

D ≤ ⌈logds n⌉. (6.80)

Equation (6.80) shows that if we take ds = n, these hypergraphs approach the Moore
bound for D = 1 when n is large.
The in- and out-size of each hyperarc has the same value s, and the in- and out-degree
of each vertex has the same value d if and only if dn = sm. The hypernet GKH(2, 8, 4, 4)
of Figure 6.52 is an example of this case. Another example is the wavelength-routed
multistar network of Figure 6.19, which was derived from the hypernet GKH(2, 6, 3, 4)
by replacing the directed MPSs based on star couplers by point-to-point connections
based on wavelength routers.
Finally, it should be observed that the two-layered approach to network design taken
here applies to realizations other than multistars. Thus, once the LCH is defined in terms
of a hypergraph, the hyperedges/hyperarcs (MPSs) can be realized on any given physical
topology by embedding each MPS into the given network topology. An example of this
was shown in Table 3.1, in which a seven-node hypernet was embedded in a mesh
topology. More general examples of embedding appear in Chapter 7.

6.6 Linear Lightwave Networks: Dynamic Routing Rules

In Section 6.5 we treated the case of static routing rules, in which a predetermined set of
optical paths (fiber paths on designated wavebands) are put in place, acting as fixed pipes
capable of supporting one or more LCs. The LCs are realized by allocating channels to
them either on a fixed or demand-assigned basis. The static routing approach is suitable
for cases in which a fixed arrangement of optical paths is sufficient for provisioning
capacity for a known, quasistatic traffic pattern. However, in cases when there is a high
degree of uncertainty in the traffic demand, dynamic routing rules are required, and the
OPs as well as the channels are assigned on demand. We consider the dynamic case
in this section, dealing with point-to-point LCs in Section 6.6.1 and multipoint LCs in
Section 6.6.2.61 More complete presentations of the material contained in this section
can be found in [Bala+91a, Bala92, Bala+95].

6.6.1 Point-to-Point Connections

As in the static case, we break down the problem of establishing connections dynamically
in an LLN into two subproblems: (1) routing and waveband assignment and (2) channel
assignment.

61
To avoid unnecessary complications, it is assumed throughout that each LC uses a full λ-channel, so that
LC [s, d] is equivalent to OC (s, d).
 Wavelength/Waveband-Routed Networks 545

Connection requests in a dynamically routed LLN are assumed to arrive sequentially

at an ONM, which attempts to accommodate the requests as they arrive, given the current
state of activity in the network. Active connections are held for random periods of time
and then released. Requests that cannot be accommodated are blocked.
The principal performance measure for the RWA and channel assignment algorithms
discussed later is blocking probability. Throughout the discussion of dynamic routing
rules, it is assumed that rearrangement of currently active connections is not allowed.
That is, the path, the waveband, and the channel allocated to a connection in progress
cannot be changed to accommodate a new connection.
In keeping with our picture of the LLN as multiple independent copies of the same
network, one for each waveband, as shown in Figure 2.6(c), the RWA problem consists
of first choosing a waveband (i.e., a copy of the network) to support a given connection
request and then attempting to find an optical path for the connection on that waveband.
Once a path is found, a free channel on that path is assigned. Commands are then
issued by the ONM to the various nodes along the path, which must set their LDCs to
configure the path, and commands are issued to the terminating NASs to tune to the
assigned channel, thereby completing the optical connection. The most difficult part of
connection management in the LLN is the routing problem; that is, the selection of a
feasible optical path within a chosen waveband, subject to the special LLN constraints:
inseparability and DSC (see Section 3.4.1). In the approach presented here, routing is
separated from waveband and channel selection to simplify the issue of path selection
as much as possible.

6.6.1.1 Selection of a Waveband

Consider an LLN with K wavebands, w1 , w2 , . . . , w K . A simple, yet quite general, way
of selecting a waveband for a given connection request is to maintain a sorted list of
the K wavebands and attempt to assign connections to wavebands in the order in which
they appear on the list. They can be sorted using various criteria. Two rules that have
been explored are Maxband, in which the list is sorted in decreasing order of usage, and
Minband, in which it is sorted in increasing order of usage. Two wavebands with the
same usage are sorted in ascending numerical order. The rationale behind Maxband is
to attempt to reuse a waveband as much as is feasible before trying another one. Each
waveband is loaded as heavily as possible and the “overflow” is offered to the next on the
list. In contrast to Maxband, Minband attempts to load all wavebands equally by starting
with the most lightly loaded waveband and working up.
Using the “list” approach, the steps in a connection allocation algorithm are as follows:

1. Select the waveband at the top of the list.

2. Try to allocate the connection on the chosen waveband using one of the routing
and channel assignment algorithms presented later. (If allocation is successful, the
connection is activated.)
3. If the connection cannot be allocated on the current waveband, then select the next
waveband on the list and go to step 2. (If the connection cannot be allocated on all
wavebands in the list, it is blocked and the algorithm terminates.)
546 Multiwavelength Optical Networks

4. Update the order of the wavebands in the list whenever a connection is activated or
released.
Simulations of the Maxband and Minband approaches to waveband selection have shown
that the former outperforms the latter by a significant margin (approximately 20%) in
typical cases. Thus, in the discussion that follows, the waveband selection algorithm
used is always Maxband.

6.6.1.2 Routing within a Given Waveband

Having chosen a waveband, w, for a requested connection from source station s to
destination station d, we seek an optical path p = s, d on w such that both the DSC
constraint and the DCA constraints are satisfied throughout the waveband when the new
connection is activated.62
Finding paths in an LLN that satisfy all constraints is complicated by the effects of
inseparability. When the new connection is added, inseparability between the intended
connection and those already active on waveband w may cause a coalescing of the
intended path with other active paths. This in turn converts the intended point-to-
point path p to an expanded point-to-multipoint path p ′ = s, D , where the expanded
destination set D includes unintended (fortuitous) destinations as well as the intended
destination d. The structures and destination sets of all other active paths that coalesce
with p are expanded as well (see Section 6.5.1). We recall from Section 3.4.1 that this
type of coalescing may cause violations of the DSC and DCA constraints in unexpected
ways, affecting the original path and its expansion.
One phenomenon that can cause connection blocking frequently in an LLN is color
clash: an inadvertent violation of the DCA constraint caused by the coalescing of a
new path with preexisting paths on the same waveband. An example of a color clash
produced by three interfering connections on a common waveband was presented in
Figure 3.25, repeated here as Figure 6.53. In Figure 6.53(a), the optical connections
(1, 1∗ ) and (2, 2∗ ) are both active using the same channel, λ1 , when the path for a new
connection (3, 3∗ ) is established by resetting the LDCs at nodes D, E, F, and G. As
shown in Figure 6.53(b), the changes in the settings of those switches cause the signals
S 1 and S 2 from the previously active connections to combine on fiber f , violating the
DCA condition. Note that this violation of DCA occurs even before a channel is assigned
to the new connection.
An optical path p for a requested connection is called feasible if the DSC and DCA
constraints are satisfied on its waveband when it is configured, and if a free channel
exists on p that can be assigned to the requested connection.63 (For an intended path p
to be feasible, the DSC and DCA conditions must satisfied on the expanded path p ′ as
well as on p.)
Unfortunately, determining whether a feasible path exists for a given LC on a chosen
waveband with a known pattern of existing connections is an NP-complete problem.
62
In this section all discussion pertains to activity within the same waveband unless stated otherwise.
63
In the current context, a “channel” is a full λ-channel. In more general situations, a channel might be
realized in some other way. For example, a low-rate LC might use one time slot in a λ-channel with a
capacity that is subdivided into multiple time slots using T-WDMA.
 Wavelength/Waveband-Routed Networks 547

1 1*
S1 (1,1* ) C S1
1

A B F G
f
d
S2 H D E
(2,2* ) 1
2 3*

3 2*
(a)

1 1*
S1 C S1

A B F G
f
d S2 + S3
S2 H D E
2 S2 3*

S3
3 2*
(b)

Figure 6.53 Color clash.

It is much simpler to find a path for a connection while ignoring the DSC and DCA
constraints and check for feasibility later. To determine feasibility of a selected path, it is
necessary to generate its expanded version and check whether it violates the constraints.
A recursive algorithm for performing this task can be found in [Bala92].
Considering the complexity of the various possible approaches, our strategy is to
select a path for a given connection on a chosen waveband that has some desirable
properties and then check to see whether that path is feasible. Because dynamic routing
requires real-time computation, the algorithms for path selection must be simple and
efficient.
We present two routing algorithms, called k-SP and Min-Int. Both of them attempt
to find shortest paths on a chosen waveband (based on arbitrary link weights) while
taking into account potential interference due to path coalescing. To define interference
as it is used here, we return to the example of inseparability previously presented in
Figure 3.21 and repeated for convenience as Figure 6.54. In Figure 6.54(a), suppose the
connection (1, 1∗ ) is active (denoted by signal S1 ), and a new connection (2, 2∗ ) is added
on the intended path p = 2, 2∗ = 2–A–B–D–E–2∗ (denoted by signal S 2 ). Due to
inseparability, a portion of S 2 appears at destination station 1∗ , resulting in the expanded
path p ′ = 2, {2∗ , 1∗ } . The presence of the fortuitous destination 1∗ means that the two
active connections must use distinct channels to avoid interference at the receivers. We
define the interference level experienced by connection (2, 2∗ ) on its intended optical
path p as the number of additional independent signals it encounters on that path. In this
example, the interference level is 1.
548 Multiwavelength Optical Networks

1 1*
C S1
a S2
A B F G

H D E
2 S1 3*

S2
3 2*
(a)

1 1*
S1 C S1
a
A B F G

S2 H D E
2 3*
S2
3 2*
(b)

Figure 6.54 Illustrating inseparability.

The two algorithms considered find paths that are short and that tend to reduce
interference. Interference reduction in turn reduces the required number of channels for
the connections, because a path with an interference level of I requires I + 1 distinct
channels for the connections sharing fibers on the path.

k-SP
The path allocation algorithm k-SP chooses a minimum-interference path from among
a set of k shortest paths from source s to destination d. The steps are as follows:
1. Find k shortest paths from s to d.
2. Check each of the k paths for feasibility. If none of them are feasible, the connection
request is blocked on the chosen waveband.64
3. From the subset of paths that are feasible, choose the one with the least interference
for the connection.65
An efficient way to find k shortest paths is through a generalization of Dijkstra’s la-
bel setting algorithm for finding shortest paths [Tarjan83]. Any meaningful link weight
assignment can be used. For example, if power and noise limitations are an important
consideration (a method known in the literature as routing with physical layer con-
straints), the link weight might be chosen proportional to the attenuation on the link.
(If the link weights are all equal, the shortest paths are minimum-hop paths.) If k = 1,
64
If the connection request is blocked on the chosen waveband, other wavebands are tried according to the
waveband selection algorithm being used.
65
Evaluating the level of interference for a given connection is a simple operation.
 Wavelength/Waveband-Routed Networks 549

then the selected path is always the shortest path. It is important to note that the k
shortest paths are enumerated before feasibility is determined. Thus, all k paths might
be infeasible either because there is no free channel on any of them or because they
violate constraints.66
To illustrate the operation of k-SP, we return to the example of Figure 6.54(a), consid-
ering the problem of selecting a path for the connection (2, 2∗ ), assuming that connection
(1, 1∗ ) is currently active on the indicated path. If minimum-hop routing is used with
k = 1, signals S 1 and S 2 share fiber a, and inseparability causes them both to reach
fortuitous destinations, as indicated by the dashed paths in the figure. Hence, the two
connections must use two distinct channels to avoid DCA violations.
However, if 3-SP is used (assuming equal weight links), the following three paths
for (2, 2∗ ) would be found: p1 = 2–A–B–D–E–2∗ , p2 = 2–A–B–C–F–E–2∗ , and
p3 = 2–A–H –B–D–E–2∗ . The first two have an interference level of 1, whereas p3
has an interference level of 0 and would therefore be the selected path, as shown in
Figure 6.54(b). In this case, (2, 2∗ ) can reuse the same channel used by (1, 1∗ ). Note that
the selected path is not the shortest in this case.

Min-Int
Although the k-SP algorithm tends to find a path of reduced interference, it is not designed
to find a path of minimum interference. Furthermore, it is constrained to examining only
k paths. In contrast to the k-SP approach, the Min-Int algorithm finds a path for a new
connection request that always has minimum interference given the current activity in
the network. The algorithm is not constrained to searching a limited number of paths, nor
does it check for feasibility at the path selection phase. Once a minimum-interference
path is selected, its expanded path is checked for feasibility. The connection is blocked
on the chosen waveband if the path is not feasible. Because only one path is selected,
it is especially important for it to have a good chance of satisfying the DSC and DCA
constraints. Choosing a path of minimum interference tends to improve the likelihood
that it satisfies these constraints.
An example of how the algorithm works is shown in Figure 6.55. Suppose connec-
tions (1, 1∗ ) and (2, 2∗ ) are active (using distinct channels) on the chosen waveband,
and a connection (3, 3∗ ) is requested. Assuming that the links are unidirectional in the
orientation indicated by the arrows, there are four possible paths for the new connection:
p1 = 3–C–D–F–3∗ , p2 = 3–C–B–D–F–3∗ , p3 = 3–C–D–E–F–3∗ , and p4 = 3–C–
B–D–E–F–3∗ in order of increasing length (assuming equal link weights). The respec-
tive interference levels for the paths are I1 = 2, I2 = 2, I3 = 0, and I4 = 2, and therefore
the Min-Int algorithm would choose p3 , which uses three internodal hops, compared
with two hops for the shortest path.67 The total number of channels required for the
66
The performance of the k-SP algorithm can sometimes be improved by removing “saturated” fibers from
consideration before executing the algorithm. Each saturated fiber (one that has all channels currently
active on the chosen waveband) is eliminated from the graph of the network before finding shortest paths.
In this way, paths chosen in step 1 of the algorithm are never rejected because of link saturation (see
Section 6.4.1.4).
67
In this example the minimum-interference path is found easily by inspection. A Min-Int algorithm applicable
to the general case is presented in Appendix E.
550 Multiwavelength Optical Networks

1 1*

S1
S1 + S2
S1 S1 + S2 S1 + S2 S1 + S2
A B D F G

S1 + S2
S2

E
S2 C
3* 2*

2 3

Figure 6.55 Illustrating Min-Int.

three connections in this case is only 2, because connection (3, 3∗ ) can reuse one of the
channels used for the other active connections. (The selected path is clearly feasible in
this case.)
Comparing this to k-SP and assuming that there are at least three channels available
on the chosen waveband, we find that the 2-SP algorithm would choose p1 , a shorter
path with more interference. All three channels are required in the 2-SP case because the
interference on the expanded path p1′ has a level of 2. (If only two channels are available,
the connection would be blocked on the chosen waveband using 2-SP, whereas it would
be accepted using Min-Int.) For k ≥ 3, both Min-Int and k-SP give the same result.
Provided that the path p = s, d found by Min-Int satisfies the DSC constraint, it can
be shown ([Bala92]) that it minimizes the additional interference on the entire waveband
caused by the activation of the new connection (see Appendix E).

6.6.1.3 Channel Allocation

Having selected a feasible path for a connection, it remains to allocate a channel to that
connection on the chosen waveband. (Feasibility implies that given the current state of
activity in the network, at least one free channel exists on the selected path.)
A free channel is one that does not violate the DCA constraint on all fibers of the
selected path. It is assumed that a list of all free channels for the selected connection is
available. It would be produced typically as a by-product of a check for DCA constraint
violations. Two simple allocation strategies are of interest: Max Reuse or Min Reuse
channel allocation. In the former, the free channel that is most used on the chosen
waveband is assigned to the connection. In the latter, the least used channel is assigned.
The rationale here is similar to that behind the Maxband and Minband approach to
waveband allocation. In the Max Reuse case, the objective is to reuse a channel as much
as possible before trying another one. In the Min Reuse case, the objective is to distribute
the load on all channels evenly. Although the two allocation problems would appear to
be similar, it turns out that the best strategy for waveband allocation (Maxband) is not
generally the best one for channel allocation. The Min Reuse channel allocation approach
 Wavelength/Waveband-Routed Networks 551

2 1
20
3 19

4 18

5 17

6 16

7 15

8 14

9 13
10 12
11

Figure 6.56 Random network.

typically outperforms the Max Reuse approach because it creates fewer opportunities
for color clashes.

6.6.1.4 Performance of Dynamic Algorithms: Point-to-Point Connections

As indicated in Section 6.4, the performance of a dynamic routing algorithm is very
difficult to evaluate analytically, even in the simpler case of wavelength-routed networks.
We therefore explore the performance of dynamic routing rules for LLNs in this section
via simulation. More details concerning performance of dynamic algorithms can be
found in [Bala+91a, Bala+91b, Bala92].
The network of Figure 6.56 is used in all simulations presented here. It has 20 nodes
of average degree 2.7 and was generated randomly using the procedure presented in
[Hagouel83]. Each link contains a single fiber pair unless indicated otherwise. Each
node is assumed to serve one or more source stations as well as a nonblocking desti-
nation station. The activity of each source station is modeled as a two-state (idle and
active) Markov chain, which generates connection requests at a rate λ when idle, with
a connection holding time of µ−1 (see Figure C.1). The destinations of the connection
requests are chosen at random and equiprobably. The activity of each source is then
determined by the source load parameter ρ = λ/µ, and the total load offered to the
network (in Erlangs) is given by Sλ/(λ + µ), where S is the total number of sources in
the network.
The study focuses first on activity within a single waveband. Routing and channel
allocation algorithms are compared based on their blocking performance. Networks with
single fiber pair links as well as multifiber pair links are studied. In all routing algorithms,
equal link weights are used.
552 Multiwavelength Optical Networks

30

25

20
Total Blocking (%)

1-SP Max
15
1-SP Min

10

0
0.2 0.4 0.6 0.8 1 1.2 1.4
Load per Source

Figure 6.57 Max Reuse versus Min Reuse channel allocation.

In the single-waveband examples, blocking is high (typically as high as 20% for high
offered loads) because of the restriction of all connections to one waveband. It must
be recalled, however, that in a multiwaveband network, blocking of a connection on
one waveband does not mean that the connection request is blocked. It merely means
that additional wavebands must be tried for that connection. In the final example, the
complete blocking performance of a multiwaveband network is examined.

Case 1: Single Waveband; Max Reuse versus Min Reuse Channel Allocation
Figure 6.57 shows the blocking performance of the two channel allocation algorithms
discussed in Section 6.6.1.3 when there are three channels available. Blocking is plotted
as a function of the source load parameter ρ, with a single source per node and a
single fiber pair per link. Because S = 20, the maximum offered load to the network is
12 Erlangs in this example (when ρ = 1.5). The 1-SP routing rule is used with equal
link weights, corresponding to minimum-hop routing.
As can be seen from Figure 6.57, the Min Reuse rule performs better than the Max
Reuse rule for all values of load, with the blocking reaching approximately 25% for
the highest offered load. It is also of interest to know which effects are dominant in
producing blocking. It turns out that approximately half of the blocking is due to color
clashes and the rest is due to the unavailability of a free channel. Blocking due to DSC
violations is negligible in this example. The blocking due to color clashes is substantially
diminished, especially at lighter loads, by using the Min Reuse rule. This is because the
 Wavelength/Waveband-Routed Networks 553

30

25

20
Total Blocking (%)

1-SP Min
15

2-SP Min
10

0
0.2 0.4 0.6 0.8 1.1 1.2 1.4
Load per Source

Figure 6.58 k-SP routing.

lower degree of channel reuse resulting from the Min Reuse rule decreases the chance
that a color clash occurs. Blocking due to channel unavailability is slightly increased by
using the Min Reuse rule but not enough to offset the reduction of color clashes.

Case 2: Single Waveband; k-SP Routing

Figure 6.58 compares the performance of the k-SP algorithm for the cases k = 1 and
k = 2 using the Min Reuse channel allocation rule. Again, there are three channels in
the waveband. As is to be expected, the blocking performance for k = 2 is considerably
better than for k = 1, with both color clashes and channel unavailability reduced as k is
increased.

Case 3: Single Waveband; 2-SP versus Min-Int Routing

Figure 6.59 compares the performance of the 2-SP and Min-Int algorithms. Again the
Min Reuse channel allocation rule is used with three channels. For this case, the range of
the load parameter ρ is increased to 3.1 to observe differences in relative performance at
high loads. As can be seen, Min-Int outperforms 2-SP at the lower loads, but the reverse
is true at the higher loads. This effect can be explained by the fact that the number of
hops on the selected paths is more tightly constrained in the 2-SP algorithm than in the
Min-Int algorithm. As a result, the latter algorithm sometimes selects unusually long
paths to avoid interference. This in turn tends to increase congestion in the network at
554 Multiwavelength Optical Networks

30

25

2-SP
20
Total Blocking (%)

15

10

5
Min-Int

0
0.5 1 1.5 2 2.5 3
Load per Source

Figure 6.59 k-SP versus Min-Int routing.

high loads, increasing blocking probability.68 In this example, the average path length
using the Min-Int algorithm was approximately 3.6 hops, compared with 2.9 hops for
2-SP, indicating a moderate increase in congestion.

Case 4: Single Waveband; Multifiber Links

The purpose of this example is to compare the performance of the routing algorithms
when more than one fiber pair is available on a link. Each link in the network of Fig-
ure 6.56 now has five fiber pairs. To keep the load on the network roughly commensurate
with the total fiber capacity, the number of sources per node is increased to four. The
total number of available channels is still three. Figure 6.60 compares the blocking per-
formance for three algorithms: 2-SP, 10-SP, and Min-Int, with the source load parameter
ρ varied from 0 to 1.6, corresponding to a maximum offered load of approximately
50 Erlangs. Min Reuse channel allocation is used. The 2-SP algorithm performs very
poorly; in fact, it performs significantly worse than in the case of Figure 6.58 with a
comparable load. By increasing k to 10, performance improves to the range of blocking
exhibited earlier for the single fiber pair cases. By contrast, the blocking for the Min-Int
algorithm is far superior (less than 2%) throughout.
68
A similar effect has been observed in dynamic routing algorithms in both wavelength-routed optical
networks and traditional communication networks. At the higher loads, routing rules that tend to produce
longer paths typically lead to higher congestion and higher blocking probabilities and can eventually cause
instabilities in network operation (see, for example, [Schwartz87, page 630 f ]).
 Wavelength/Waveband-Routed Networks 555

60

50

40
Total Blocking (%)

30
2-SP

20

10 10-SP

Min-Int
0
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
Load

Figure 6.60 Blocking in networks with multifiber links.

The comparison in this example reveals a significant difference in performance be-

tween algorithms that consider only a limited number of paths between source and
destination (e.g., k-SP) and those that have no limitation (e.g., Min-Int). In the case
at hand, the number of possible fiber paths between a given source-destination pair is
increased enormously by increasing the number of fiber pairs per link. For example,
one given three-hop path is replaced by 125 three-hop alternative paths on the same
links when there is a choice of five fibers on each link. The 2-SP and 10-SP algorithms
examine only a small fraction of these, whereas the Min-Int algorithm finds a minimum-
interference path, taking all fibers into consideration. In general, for the multifiber case,
Min-Int delivers better performance with less computational complexity than k-SP.
Another way of viewing this comparison is that the first step of the k-SP algorithm
is static in that it selects paths independent of the state of the network. The Min-Int
algorithm is adaptive in that the path is selected based on transnodal weights that are a
function of network activity (see Appendix E).

Case 5: Multiple Wavebands

The available optical spectrum can be partitioned in various ways into wavebands and
λ-channels. In this example, we compare three different partitions, each containing the
same total number of λ-channels. (The different partitions lead to different spectral
efficiencies because of the guard bands between wavebands. This effect is examined at
the end of the example.)
556 Multiwavelength Optical Networks

40

35

30
Total Blocking (%)

25

20

15

10 1 Wb-6 Ch

2 Wb-3 Ch 6 Wb-1 Ch
0
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
Load

Figure 6.61 Blocking in networks with multiple wavebands.

Figure 6.61 compares blocking probabilities in the case of a six-channel network in

which the following spectrum partitioning is used:
r One waveband, six channels per waveband
r Two wavebands, three channels per waveband
r Six wavebands, one channel per waveband

The latter case corresponds to a wavelength-routed network.

The Maxband rule is used for waveband selection, Min-Int for path selection, and
Min Reuse for channel allocation. Two sources are used at each node to double the load
on the network, consistent with the doubling of the total number of channels from three
to six.
Note that blocking improves dramatically as the number of wavebands is increased.
There is a factor of six improvement at high loads and still more at low loads, when two
wavebands are used instead of one. There is negligible blocking for the case of a single
channel per waveband. This illustrates the penalty incurred through LLN constraint
violations when several channels are grouped into wavebands rather than operating the
network as a WRN, with a single channel per waveband.
Other conclusions can be drawn from comparisons of Figure 6.61 with the Min-Int
curve in Figure 6.59. We note that the latter shows a blocking of approximately 16%
with ρ = 1.5 (for an offered load to the network of 12 Erlangs) when three channels are
 Wavelength/Waveband-Routed Networks 557

used on one waveband. Compare this with the curves of Figure 6.61 at the point ρ = 1.5
(giving 24 Erlangs offered load). In the case in which all six channels are confined to a
single waveband, the blocking increases to approximately 32%. This case corresponds to
“scaling up” both the offered load and the network capacity by a factor of 2. Ordinarily,
this upward scaling would reduce blocking because the laws of large numbers work in
our favor in the larger network. However, the simulation shows that the upward scaling
effect is more than cancelled by additional blocking due to LLN constraint violations
when all traffic is confined to one waveband.
Next consider the two-waveband case with three channels per waveband. Figure 6.61
shows approximately 5% blocking, which is a factor of 3 reduction in blocking compared
with the system of Figure 6.59. In the two cases, the offered load to the network per
channel is the same, but blocking is reduced in the two-waveband system by sending the
overflow from the first waveband to the second one. Again, the advantage of partitioning
into more wavebands is evident. The six-waveband case takes this approach to its ultimate
limit.
It it clear from the last example that the more the spectrum is partitioned into wave-
bands, the better the blocking performance. However, spectrum partitioning results in
an increase in optical spectrum usage (as well as requiring a more complex network
node). Recall that to be able to discriminate between wavebands at the network nodes
(LDCs), a guard band of width Bg is needed between wavebands, and Bg normally
must be considerably greater than the λ-channel bandwidth Bλ (see Section 6.5.6.3). To
quantify this effect, let Bop (W, C) be the spectrum required to support W wavebands
containing C channels each, so Bop (1, W C) is the spectrum required if all W C channels
are grouped into a single waveband. Let r (W, C) be the ratio of spectrum usage in the
first case to that in the second case. Then, using Equation (6.67), we have
Bop (W, C) W −1
r (W, C) = =1+ α, (6.81)
Bop (1, W C) WC

where α = Bg /Bλ .
Using Equation (6.81) and assuming α = 5, we find that the two-waveband case
requires a factor of r = 1.83 increase in bandwidth over the single-waveband case,
whereas the six-waveband (WRN) case requires a factor of 5.16 increase in bandwidth.
Thus, in partitioning the spectrum into an increasing number of wavebands, we incur a
gain in performance in terms of blocking but a loss in terms of spectrum efficiency.

6.6.1.5 Summary of Point-to-Point Dynamic Routing in LLNs

The static and dynamic formulations of the connection allocation problem in LLNs
differ in their design and performance objectives. In the static case, one is generally
interested in allocating resources to a set of known connections to minimize the use
of the optical spectrum; that is, the number of wavebands and channels required to
support a given set of connections. In the dynamic case, because the connection set
is not known a priori, one generally works under the assumption of a fixed available
spectrum (wavebands and channels) and attempts to do the resource allocation in a way
that maximizes performance, typically by minimizing blocking probability for a given
558 Multiwavelength Optical Networks

offered load. The unpredictability of the connection request sequence makes it important
for the dynamic algorithms to adapt to random fluctuations in demand.
Among the three parts of the connection allocation problem under dynamic routing
rules (waveband selection, path allocation, and channel assignment), it is only the routing
problem that is difficult, primarily because of the potential interference among optical
paths sharing the same waveband. Simulation results suggest that routing algorithms
(e.g., the Min-Int algorithm presented here) that tend to minimize interference (the
number of fortuitous destinations) on each new connection perform best under most
conditions. The simulations also show that it is best from the point of view of blocking
to attempt to reuse each waveband as much as possible (the Maxband waveband selection
rule) and to attempt to reuse each channel on a given waveband as little as possible (the
Min Reuse channel selection rule).
The performance of LLNs under dynamic routing rules improves markedly as the
spectrum is partitioned into an increased number of wavebands and as the number of
fiber pairs per link is increased. This is because of the fact that the special constraints
of the LLN (inseparability and distinct source combining) are less troublesome in these
cases.

6.6.2 Routing Multicast Connections in LLNs

In the multicast case, under dynamic routing rules, each LC is in the form [s, D], where
the destination set D = {d1 , d2 , . . .} may contain any number of stations, up to the full
broadcast case. An example of a multicast connection from source 1 to destination set
{2∗ , 3∗ } is shown in Figure 6.62. The links on the path p = 1, {2∗ , 3∗ } are shown as
bold lines. The LDC in node A is set to split the optical signal generated at 1∗ , directing
portions of it (possibly amplified) to nodes B and C.
Generalizing the point-to-point case, a multicast connection allocation algorithm can
be broken down into the following steps:
1. Choose a waveband for the connection according to some given selection rule.
2. Attempt to find a path for the connection on the selected waveband. If one is found
that satisfies the LLN constraints, assign a channel to the connection. (If allocation
is successful at this point, the algorithm terminates.)
3. If the connection cannot be allocated on the current waveband, go to step 1. (If all
wavebands have been tried at this point, the connection is blocked.)
For simplicity in this section we focus on the single-waveband case; that is, the
emphasis is on step 2, which is primarily a routing problem. The waveband selection
rule, embodied in steps 1 and 3, is arbitrary and independent of the routing rule. The
Maxband approach, which was found to be best in the point-to-point case, is likely to
perform best under multicast conditions as well.

6.6.2.1 Routing within a Given Waveband: Multicast Trees

Having chosen a waveband w for a requested connection from source station s to
destination set D, we seek an optical path p = s, D on w such that both the DSC
 Wavelength/Waveband-Routed Networks 559

1 2*

A B

3*

Figure 6.62 Example of a multicast connection.

and DCA constraints are satisfied throughout the waveband when the new connection
is activated. At the stage of path selection, satisfaction of the DCA condition implies
that no color clashes occur and that a free channel exists on p that can be used for the
requested connection. A path that satisfies all of these conditions is called a feasible
multicast path.
In the multicast case, inseparability often causes path coalescing, which expands
the path p to a path p ′ = s, D . The expanded destination set D includes fortuitous
destinations in addition to the intended set D. The structures and destination sets of all
other active paths that coalesce with p are also expanded. For the intended path p to
be feasible, the DSC and DCA conditions must be satisfied on the expanded path p ′
as well as on p. The potential violations of the DSC and DCA constraints caused by
path coalescing are more complex in the multicast case than in the point-to-point case,
making routing more difficult.
The set of fibers forming a path p from source s to the intended destination set D must
contain a directed path from s to each station in D. Furthermore, p and its expanded
form p ′ must both satisfy the DSC condition, meaning they cannot contain any directed
cycles or any directed parallel paths. We call a path that satisfies these conditions a
multicast tree. A multicast tree that satisfies the DCA condition constitutes a feasible
multicast path. Thus a routing algorithm for multicast connections in LLNs must (1)
find multicast trees and (2) ensure that they satisfy the DCA constraint. Because the first
step alone is already an NP-complete problem for the special case of a point-to-point
connection, this is clearly a very difficult task.
560 Multiwavelength Optical Networks

Because of the extra complexity of the multicast case, we add a preliminary step to the
routing procedure that greatly simplifies the problem: The network is initially decom-
posed into a set of edge–disjoint trees, and each optical path is confined to one of these
“preselected” trees.69 The remainder of the connection allocation algorithm is performed
dynamically in response to each connection request. Thus the overall algorithm is a
hybrid static/dynamic procedure. Because the fibers carrying the connection are
confined to a tree, the intended path p as well as its expanded version p′ (if the
connection coalesces with other active connections on the tree) satisfies the DSC
constraint automatically.70
A typical LLN operates on multiple wavebands. In the multiwaveband case, a tree
decomposition is performed independently on each waveband, and the decompositions
may be the same or different on the different wavebands. After the trees are preselected,
the dynamic part of the algorithm consists of choosing a waveband for the optical path
carrying the requested connection and determining whether a tree T , which can support
the connection, can be found on that waveband.71 A multicast connection [s, D] can be
supported on a tree T operating on a given waveband if
r T contains s and D
r The optical path p supporting the connection causes no color clashes on T
r A free channel exists for the connection on path p

We consider each of these issues next.

6.6.2.2 Preselecting Edge–Disjoint Trees on a Waveband

The discussion in this section applies to each waveband separately. On a given waveband,
the network can be described as a multigraph. Each node in the network is represented
by a vertex of the multigraph, and each link in the network is represented by F parallel
edges, where F is the number of fiber pairs on the link. We include only the network
nodes and internodal links in the multigraph model, excluding the access stations and
their access links. Each station is assumed to be able to access all trees containing the
node to which it is attached, by an appropriate setting of the LDC at that node.
In general, any network can be decomposed into a set of edge–disjoint trees in
many different ways. The procedure used here for tree decomposition is as follows:
The multigraph is first decomposed into a (hopefully large) number of edge–disjoint
spanning trees. (It is assumed that the graph of the network is connected, so spanning
trees exist; otherwise, not all multicast connections are possible.) Then the remaining
edges are grouped into the smallest possible number of (nonspanning) edge–disjoint

69
Because the trees are preselected, this part of the routing algorithm is “static,” and it can be performed
off-line when the network is initialized. Therefore, the computational complexity of the tree decomposition
step is not a paramount issue.
70
The DSC condition can be violated on a tree only if a connection is routed so that it doubles back on a
link, forming a loop that circulates endlessly through both oppositely directed fibers on the same link – an
obviously pathological routing decision.
71
It should be noted that this procedure can be applied to dynamic connection allocation in the point-to-point
case as well as in the multicast case. However, the extra constraint imposed by tree preselection makes this
static/dynamic approach less efficient than unconstrained purely dynamic algorithms in the point-to-point
case (see case 5 in Section 6.6.1.4).
 Wavelength/Waveband-Routed Networks 561

trees. (Each tree is necessarily an ordinary subgraph, containing no more than one edge
between each pair of vertices.)
In a tree decomposition of the network graph (or multigraph), there are two conflicting
objectives. First, it is important to find as many spanning trees as possible, because each
spanning tree can potentially accommodate any multicast connection. Second, it is
desirable to keep the trees “dense” (i.e., of small diameter) so the multicast paths are
short, promoting channel reuse and reducing chances of color clashes.
The problem of decomposing a graph into the maximum number of edge–disjoint
spanning trees is solved in [Gusfield83] using the method of partitioning of matroids.
Unfortunately, in decomposing a graph into a large number of spanning trees, each tree
tends to be of a large diameter. This is because the degree of each vertex in a tree must be
kept small to allow other trees to include that vertex, thereby maximizing the number of
trees containing each vertex of the graph. In the worst case, each spanning tree becomes
a long chain, with vertices of maximum degree 2.
Another approach to tree decomposition is to seek the densest spanning trees (those
with the smallest diameter). This, however, tends to produce a smaller number of span-
ning trees for the reason stated in the previous paragraph. Thus, some compromise is
required to produce a “good” tree decomposition: one containing many dense spanning
trees. For this purpose, a useful heuristic algorithm has been developed based on a mod-
ified breadth-first search with parameter d (MBFS-d; see [Bala+93]). The algorithm
starts with one node and adds nodes using a breadth-first search until a spanning tree
is formed. The parameter d represents the maximum number of neighboring nodes that
are considered for adding to the tree at each step. As d is increased, the algorithm tends
to find trees of smaller diameter, but this usually produces fewer spanning trees.
Examples of the results of running the MBFS-d algorithm on an eight-node network
with a single fiber pair per link are shown in Figures 6.63 and 6.64. In each case, the first
spanning tree is shown extracted from the original graph G, leaving a graph G ′ , which
is further decomposed into additional trees.
The examples illustrate the two extreme cases: d = 1 (finding a maximum number
of spanning trees) and d = 4 (producing smaller diameter trees). In Figure 6.63, the
MBFS-1 algorithm finds two spanning trees of diameter 8, and one nonspanning tree.
The MBFS-4 case in Figure 6.64 yields only one spanning tree from a total of four in
the decomposition, but the maximum diameter of the trees is 6.

6.6.2.3 Routing Connections on a Tree

Because each multicast connection is confined to a tree, its optical path is (almost)
unique. We provide two alternative routing techniques in the following pages, each of
which has some positive and negative features. Both techniques yield multicast trees as
defined in Section 6.6.2.1.

Shortest Path Routing

If connections are not allowed to double back on links, then the path for any multicast
connection on a given tree is unique and satisfies the DSC constraint automatically. (It
is also the shortest possible path on the tree.) However, shortest path routing can lead
562 Multiwavelength Optical Networks

1
Trees Obtained by Applying
9 MBFS-d (d = 1) on Graph G
2 1
9
2

8
Graph G 3 8 Spanning Tree 3
Diameter 8

7
7 4
4 6
5
6
5
1
9 2

Leftover Graph G′
8
Spanning Tree 3
1
Diameter 8

9 2 7 4
6 5

8
3 1
9 2

7 8
Nonspanning 3
4 Tree

6 7 4
5
6 5

Figure 6.63 Example of a tree decomposition using MBFS-1.

to a problem we have observed earlier in LLNs: color clashes. To illustrate, consider

the six-node, single-waveband, single-fiber-per-link network of Figure 6.65(a), which
is decomposed into two trees, T1 and T2 , in Figure 6.65(b). Suppose two connections,
[6, {1, 2}] and [2, 3], are active on T1 as shown by signals S6 and S2 in Figure 6.65(c) and
are both assigned the same wavelength, λ1 . Now a new request arrives for connection
[5, 4], which must also be placed on T1 . By inseparability, the intended path p = 5, 4
is expanded to p ′ = 5, {4, 1, 2, 3} , carrying a portion of signals S6 and S2 with it to
stations 3 and 4. Because these are both on the same channel, this results in a color clash
on the last portion of the path.
 Wavelength/Waveband-Routed Networks 563

1 Trees Obtained by Applying

MBFS-d (d = 4) on Graph G
9 1
2
9
2

8 Spanning Tree
8 3
Graph G 3 Diameter 4

7
4

7 6 5
4
1
9
6 2
5

8 Nonspanning 3
Leftover Graph G′ Tree

1 7
4
9 6 5
2
1
2
9
8 3
3 Nonspanning
8 Tree
4
7
5
6
7
4 1
2
6 5 9
3
Nonspanning
8 Tree 4
7 5
6

Figure 6.64 Example of a tree decomposition using MBFS-4.

The situation leading to the color clash in this example is an inevitable result of using
the Max Reuse rule for channel allocation on the tree.72 As indicated earlier, channel
reuse in an LLN is a two-edged sword: It can economize on spectrum use, but it can
also introduce unnecessary blocking (assuming that connection rearrangement is not
allowed). In this example, even if there is a free wavelength available, there is no way to
route the new connection on the tree, and it must be blocked.
72
However, even if some other channel allocation rule is used, it is possible for the network state to evolve to
that shown in Figure 6.65(c). Can you construct an example?
564 Multiwavelength Optical Networks

5 4

E D

6 F C 3

A B

1 2

(a) Network

E D E D

T2

F T1 C

A B B

(b) Tree Decomposition

Figure 6.65 Illustrating routing on a tree.

One way of avoiding color clashes is to require all connections using a given tree to
use distinct channels. Unfortunately, this eliminates any possibility of reusing channels
on the tree. An approach that requires distinct channel allocation, but carries with it
some offsetting advantages, is rooted routing, which is considered next.

Rooted Routing on a Tree

Rooted routing, as described in Section 6.5.6.2 can be used for either point-to-point or
multicast LCs. But it is especially effective in the multicast case.
Using rooted routing in our example, the tree T1 is configured as an embedded star
rooted at node A in Figure 6.65(d). Note that the LDCs at all nodes except the root node
are configured to concentrate all transmitted signals at the root and then broadcast the
combined signals back to the receiving side of all stations – an arrangement used earlier
in several other examples. It is important to note that the internodal link settings of the
 Wavelength/Waveband-Routed Networks 565

5 4

E D

p = 5, 4

6 F C 3

S6 S2

A B
S6

1 2

(c) Color Clash on T1

5 6 6 5

E F F E

1 A 1

D C B B C D

4 3 2 2 3 4
(d) Rooted Routing on T1

Figure 6.65 (cont.)

LDCs can be fixed once the tree decomposition is known, because all stations on each
tree must have access to the root node. The only LDC settings that must be changed
when a connection enters or leaves the network are those connecting the affected access
stations to a chosen tree. This is in contrast to shortest path routing, in which all LDC
settings depend on the current state of the network and thus must be changed each time
a new connection becomes active.
Another advantage of the rooted routing approach is that the total capacity of
the waveband can be shared fully among all connections on the tree using any de-
sired multiplexing/multiple-access technique, as described in Chapter 5. (Recall from
Section 5.4.1.4 that shared-channel systems are readily adaptable to multicast LCs.) This
means that the capacity allocated to each multicast connection need not be quantized to
566 Multiwavelength Optical Networks

units of full λ-channels. However, the downside of rooted routing is that the capacity of
the waveband supporting the tree cannot be reused.

Routing Rules and LDC Complexity

Under the routing rules defined here, the tree decomposition of a network is precomputed
when the network is initialized, and it remains fixed thereafter (barring a failure or
network reconfiguration). This means that on a given waveband, interconnections exist
only among the internodal links belonging to a common tree. If this arrangement remains
static, the structure of the LDCs can be simplified to realize only the required internodal
tree connections, implying that fewer LDC ports and switching elements are needed. A
further simplification may be possible in the case of rooted routing, as mentioned earlier.
Provided that means are included in the nodes for maintaining proper power levels, all
internodal LDC settings can be held fixed, so the only controllable elements in the LDCs
are those required for connecting the access stations to the trees.

6.6.2.4 Completing Multicast Connection Allocation

After setting up preselected trees and choosing a waveband for a requested connection,
it remains to choose a tree on the selected waveband and a channel on the selected tree
for the connection.

Choosing a Tree for a Connection

Suppose a waveband is chosen on which to route a new connection. Intuitively, there are
many criteria that might be used for selecting a tree for the connection on that waveband.
Three reasonable possibilities are (1) choose the smallest tree that can accommodate the
connection, (2) choose the most used tree (one with the most active connections), and
(3) choose the minimum-interference tree.
Results of simulations using shortest path routing on trees show that the minimum-
interference rule is superior to the other two from the point of view of blocking. However,
the margin of superiority decreases with increasing degree of multicasting [Bala92,
Bala+93].
In implementing the minimum-interference rule under shortest path routing, all pre-
selected trees are checked to determine which ones contain a feasible multicast tree for
the connection. Among the feasible trees, the one with the least amount of interference
is chosen for the connection. Testing for feasibility requires checking for color clashes
as well as checking for the existence of a free channel. (If rooted routing is used on each
tree, a color clash check is unnecessary.)

Channel Allocation on a Tree

Recall that regardless of the degree of multicast, a single channel suffices for a multicast
optical connection. Under shortest path routing, it is simplest to assume that each
channel is a λ-channel. The rule used for allocating that channel can be important.73

73
It is irrelevant under rooted routing.
 Wavelength/Waveband-Routed Networks 567

Table 6.7 Sizes of trees generated with the

MBFS-d algorithm.

d Sizes of trees

2 20, 20, 20, 20, 18, 12

3 20, 20, 20, 20, 16, 5, 4, 2, 2, 2, 2, 2, 2
4 20, 20, 20, 18, 18, 7, 4, 2, 2, 2, 2
5 20, 20, 20, 20, 16, 6, 5, 2, 2, 2, 2
6 20, 20, 20, 20, 18, 6, 5, 2, 2
7 20, 20, 20, 19, 11, 7, 6, 6, 2, 2, 2
8 20, 20, 19, 19, 9, 9, 7, 7

Recall that in the case of point-to-point connections allocated under dynamic routing
rules, attempting to reuse λ-channels as much as possible on a given waveband may
be a poor tactic because it leads to color clashes. The same holds true in the multicast
case. Therefore, we shall assume henceforth that the Min Reuse rule is used for channel
allocation. In the context of tree routing, Min Reuse implies allocating the channel that
is used least on the selected tree.

6.6.2.5 Results of Simulations

Given the number of factors interacting in dynamic allocation of multicast connections,
simulation is the only reasonable approach to performance evaluation. This section
examines blocking in a simulated multicast scenario, focusing on a single waveband. As
in the point-to-point case, a multicast connection is blocked in a multiwaveband network
only if it is blocked on each waveband.
In the simulation results presented in this section, the network has a randomly gen-
erated topology with 20 nodes, an average degree of 6, and two fiber pairs per link.
The waveband contains three λ-channels. One source is attached to each node, with
connection request generation governed by a two-state Markov chain with variable load
parameter ρ (see Section 6.6.1.4). Each requested connection selects three receivers
randomly. Shortest path routing is used with minimum-interference tree selection. All
receivers are assumed to be nonblocking.
The MBFS-d algorithm is used for tree decomposition, with the results shown for
various values of d in Table 6.7. The table lists the sizes of the trees found using the
algorithm for each value of d ranging from 2 to the maximum nodal degree, 8. Among
all values of d producing the maximum number of spanning trees (d = 2, 3, 5, 6), the
decomposition for d = 6 gives the minimum average path length (in optical hops) when
used as a basis for connection allocation in the simulation. Figure 6.66 shows that this
value of d also produces the lowest blocking probability over almost the complete range
of the load parameter ρ. (Values of d less than 6 are not shown. They produce higher
blocking probabilities for all values of ρ.) Although the blocking probabilities in this
single waveband example are substantial, blocking using multiple wavebands would be
much less, following the same pattern as in the point-to-point case (see Figure 6.61).
568 Multiwavelength Optical Networks

40

35

30
Total Blocking (%)

25

d 8
d 7
20

15

10 d 6

0
0.5 1 1.5 2 2.5 3 3.5 4 4.5
Load per Source

Figure 6.66 Blocking probability for multicast connections.

6.7 Problems

1 Suggest a way of adding links sparingly to the network of Figure 6.8 to reduce the
average distance between nodes.
2 Suggest a way of adding links sparingly to the network of Figure 6.9 to improve its
survivability.
3 Suppose that full logical connectivity is to be set up in the deBruijn network of
Figure 6.4. Assuming a single fiber pair per link, find W Limcut .
4 Solve the RCA problem of Figure 6.12 without the constraint of minimum-hop
routing.
5 Determine an RCA for full connectivity on an eight-node bidirectional WDM ring
using the approach in Section 6.3.4.2.
6 Repeat the previous problem for a nine-node ring.
7 Full logical connectivity is to be established in the six-node bidirectional WDM
ring. According to Equation (6.16), this can be done with five wavelengths if the access
stations are nonblocking.
 Wavelength/Waveband-Routed Networks 569

(a) Find an RCA for this case using five wavelengths. (See, for example, [Ellinas 98b].)
(b) Now suppose that elementary stations are used. Find an RCA for this case using as
few wavelengths as you can. (It is still possible with five wavelengths!)
8 Show that with one WIXC placed anywhere in a bidirectional ring wavelength-routed
network, any set of connections can be established using a number of wavelengths equal
to the maximum number of connections traversing any fiber.
9 Devise a multistar network with eight stations, based on permutation routers and
using four 4 × 4 stars. Compare your answer with Figure 6.52.
10 Program RCA-I and RCA-II.
(a) Try out your programs on some small problems and compare results.
(b) See if you can get as good or better results by inspection.
(c) Compute W Limcut in each case and compare it to the number of wavelengths used.
11 Reformulate RCA-I to include a prescribed set of permissible paths.
12 Consider a three-node unidirectional bus network with nodes A, B, and C. Poisson-
distributed connection requests arrive at node A at rate λ, with exponentially distributed
holding times. There are two wavelengths available. Each connection is destined ran-
domly for nodes B and C with equal probabilities. Calculate the blocking probability
for each of the following cases.
(a) Nodes A, B, and C are WSXCs in a WRN.
(b) Nodes A, B, and C are WIXCs in a WRN.
(c) Nodes A, B, and C are LDC nodes using a single waveband containing both wave-
lengths.
Are there any differences? Explain.
13 In Problem 12, Poisson-distributed connection requests arrive at node A at rate λ.
Each connection is destined randomly for nodes B and C with equal probability. Also,
Poisson-distributed connection requests arrive at node B, destined for C, at a rate of λ/2.
Calculate the blocking probability for each of the following cases using approximations
when necessary.
(a) Nodes A, B, and C are WSXCs in a WRN.
(b) Nodes A, B, and C are WIXCs in a WRN.
(c) Nodes A, B, and C are LDC nodes using a single waveband.
Explain the results.
14 Give an exact Markov chain model for each case of Problem 13.
15 For each case of Problem 13, write a simulator that computes the blocking proba-
bility. Compare the simulations with the calculated results.
16 Derive Equation (6.44) from the Erlang-B formula.
17 Derive Equation (6.46).
18 Using a three-state Markov chain derive Equation (6.43).
570 Multiwavelength Optical Networks

19 Use Equation (6.45) to generate some points matching the simulation points in
Figure 6.28. Assume H = 5 (the worst case traffic distribution), and compute ρ from
the Erlang load per node. How good is the formula in this case?
20 Write a simulator to study dynamic RCA in wavelength-routed networks. Use it to
exercise the various algorithms discussed in this chapter and/or some of your own.
21 For the Petersen network example with W = 3, consider the following variants:
(a) Reflecting stars are used instead of nonreflecting stars.
(b) Nonreflecting stars are used with elementary stations.
(c) Reflecting stars are used with elementary stations.
In each case, find the minimum number of required λ-channels per waveband, and give
a feasible channel assignment for waveband 2. In the last case there is a violation of the
DSC condition. Comment on its significance.
22 Consider the Petersen network example with W = 5, using elementary stations.
(a) In the case of multicast optical connections, let c X denote the optical connection
from station X . Show that a feasible channel assignment for the connections on
waveband 5 is the following: c E , c D , c H , and c J use channel 1; c A , c B , and cG use
channel 2; and cC , c F , and c I use channel 3.
(b) If a point-to-point optical connection is used to carry each LC, find the minimum
number of required λ-channels on waveband 5 and give a feasible channel assignment
for waveband 5.
23 Show by an example that a possible (but not necessarily practical) solution of the
propagation delay problem illustrated in Figure 6.51 is to choose the frame time to be a
common divisor of the relative propagation delays to each receiver. Show by an example
that another approach is to accept less than 100% scheduling efficiency.

24 Consider a five-node four-fiber (two-working fiber) bidirectional ring network with

the full mesh connectivity scenario described in Section 6.3.4. Suppose the ring is now
operated as an LLN using a single waveband instead of as a WRN. The nodes are now
LDCs instead of WADMs.
(a) Is it possible to use shortest path routing for all connections in this case? Explain.
(b) Find a routing and wavelength assignment using as few wavelengths as possible.
(c) Compare with the WRN case.

25 Write a simulator to study dynamic routing and wavelength assignment in LLNs.

Use it to exercise the various algorithms discussed in this chapter and/or some of your
own.

26 Find the time complexity of the k-SP and Min-Int algorithms for allocating point-to-
point connections in an LLN. Include the complexity of the waveband and the wavelength
allocation algorithms.

27 In Figure E.1 add a new transmitter, 4, at node C and a new receiver, 4∗ , at node
G. Now show the new image network. Assign the connection from 3 to 3∗ along the
shortest path as determined by the Min-Int algorithm. Show the new weights on the
 Wavelength/Waveband-Routed Networks 571

intranodal links. Now, allocate a connection from transmitter 4 to receiver 4∗ using the
shortest path as determined by the Min-Int algorithm. Show the upgraded values of the
link weights.

28 Consider the results shown in Figure 6.61. Starting with the plot showing the
blocking for one waveband containing six channels, estimate and plot the blocking
probability for the same network under the assumption that two wavebands containing
six channels each are now available. Assume the Maxband allocation policy and use
reasonable approximations.

29 Find the complexity of the MBFS-d algorithm for tree decomposition.

30 Find the complexity of the overall multicasting algorithm. Include the waveband
selection process, tree selection, the check for color clash, and channel allocation.

31 Consider the plot in Figure 6.66, which shows the blocking for dynamic multicast
connections for d = 6 in an LLN operating on a single waveband with three λ-channels.
Suppose that the same network operates under the same conditions, except that it now
has two wavebands containing three channels each. Estimate the blocking probability as
a function of load using simplifying assumptions when necessary.

32 Construct an example of a connection arrival and departure sequence in the network

of Figure 6.65(a), which leads to the state of the network shown in Figure 6.65(c) when
the Min Reuse channel allocation rule is used.

Bibliography

[Aarts+97] E. Aarts and J. K. Lenstra. Local Search in Combinatorial Optimization. New York:
Wiley 1997.
[Acampora94] A. S. Acampora. The scalable lightwave network. IEEE Communications Mag.,
32(12):36–43, 1994.
[Ali+00a] M. Ali and J. Deogun. Cost-effective implementation of multicasting in wavelength-
routed networks. IEEE/OSA J. Lightwave Technol., 18(12):1628–1638, 2000.
[Ali+00b] M. Ali and J. Deogun. Allocation of splitting nodes in wavelength-routed networks.
Photonic Net. Commun., 2(3):245–263, 2000.
[Arden+91] B. W. Arden and K. W. Tang. Representations and routing for Cayley graphs. IEEE
Trans. Commun., 39(11):1533–1537, 1991.
[Ayanoglu+96] E. Ayanoglu and E. Karasan. Wavelength routing and selection algorithms for
WDM optical networks. In Proceedings of the IEEE Int’l Conf. Commun. (ICC): Workshop on
WDM Network Management and Control. Dallas, TX, June 1996.
[Bala+91a] K. Bala, T. E. Stern, and K. Bala. Algorithms for routing in a linear lightwave network.
In Proceedings of the IEEE Conference on Computer Communications (Infocom), Miami, FL,
April 1991.
[Bala+91b] K. Bala, T. E. Stern, and K. Bala. A minimum interference routing algorithm for a
linear lightwave network. In Proceedings of the IEEE Global Telecommunications Conference
(Globecom), Phoenix, AZ, December 1991.
572 Multiwavelength Optical Networks

[Bala92] K. Bala. Routing in Linear Lightwave Networks. Ph.D. thesis. Columbia University, New
York, 1992.
[Bala+93] K. Bala, K. Petropoulos, and T. E. Stern. Multicasting in linear lightwave networks. In
Proceedings of the IEEE Conference on Computer Communications (Infocom), San Francisco,
CA, March 1993.
[Bala+95] K. Bala and T. E. Stern. Routing in linear lightwave networks. IEEE/ACM Trans.
Network., 3(4):489–500, 1995.
[Bala+96] K. Bala, F. R. K. Chung, and C. A. Brackett. Optical wavelength routing, translation
and packet/cell switched networks. IEEE/OSA J. Lightwave Technol., 14(3):336–343, 1996.
[Bala+97] K. Bala, E. Bouillet, and G. Ellinas. The benefits of minimal wavelength interchange
in WDM rings. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Dallas, TX,
February 1997.
[Banerjee+96] S. Banerjee and C. Chen. Design of wavelength-routed optical networks for circuit
switched traffic. In Proceedings of the IEEE Global Telecommunications Conference (Globe-
com), London, UK, November 1996.
[Baroni+97] S. Baroni and P. Bayvel. Wavelength requirements in arbitrary connected
wavelength-routed optical networks. IEEE/OSA J. Lightwave Technol., 15(2):242–252,
1997.
[Barry93a]
 R. A. Barry. An all-optical non-blocking M × M switchless connector with
O( M log M) wavelengths and without wavelength changers. Electron. Lett., 29(14):1252–
1254, 1993.
[Barry93b] R. A. Barry. Wavelength Routing for All-Optical Networks. Ph.D. thesis. Mas-
sachusetts Institute of Technology, Cambridge, 1993.
[Barry+93] R. A. Barry and P. A. Humblet. Latin routers: Design and implementation. IEEE/OSA
J. Lightwave Technol., 11(5/6):891–899, 1993.
[Barry+94] R. A. Barry and P. A. Humblet. On the number of wavelengths and switches in
all-optical networks. IEEE Trans. Commun., 42(2/3/4):583–591, 1994.
[Barry+95] R. Barry and P. Humblet. Models of blocking probability in all-optical networks
with and without wavelength changers. In Proceedings of the IEEE Conference on Computer
Communications (Infocom), Boston, MA, April 1995.
[Bermond+82] J.-C. Bermond, C. Delorme, and J. J. Quisquater. Tables of large graphs with
given degree and diameter. Information Processing Lett., 15(1):10–13, 1982.
[Bermond+84] J.-C. Bermond, J. Bond, and J. F. Sacle. Large hypergraphs of diameter 1. In
Graph Theory and Combinatorics, pp. 19–28. New York: Academic Press 1984.
[Bermond+89] J.-C. Bermond and C. Peyrat. deBruijn and Kautz networks: A competitor for
the hypercube. In Hypercube and Distributed Computers, pp. 279–293. Amsterdam: Elsevier
Science 1989.
[Bermond+97] J.-C. Bermond, R. Dawes, and F. O. Ergincan. deBruijn and Kautz bus networks.
Networks, 30(3):205–218, 1997.
[Birk91] Y. Birk. Fiber-optic bus-oriented single-hop interconnections among multi-transceiver
stations. IEEE/OSA J. Lightwave Technol., 9(12):1657–1664, 1991.
[Birman95] A. Birman. Routing and wavelength assignment methods in single-hop all-optical
networks with blocking. In Proceedings of the IEEE Conference on Computer Communications
(Infocom), Boston, MA, April 1995.
[Bouillet98] E. Bouillet. Monte Carlo techniques for design of wavelength-routed all-optical
networks. Internal Report. New York: Center for Telecommunications Research, Columbia
University, 1998.
 Wavelength/Waveband-Routed Networks 573

[Brain+88] M. C. Brain and P. Cochrane. Wavelength-routed optical networks using coherent

transmission. In Proceedings of the IEEE Int’l Conf. Commun. (ICC), Philadelphia, PA, June
1988.
[Brown+96] G. Brown and A. Acampora. Systems and methods for employing a recursive mesh
network with extraplanar links. U.S. Patent 5,530,575, June 1996.
[Cerf+74] V. G. Cerf, D. D. Cowan, R. C. Mullin, and R. G. Stanton. A lower bound on the
average shortest path length in regular graphs. Networks, 4:335–342, 1974.
[Chen+96] C. Chen and S. Banerjee. A new model for optimal routing and wavelength assignment
in wavelength division multiplexed optical networks. In Proceedings of the IEEE Conference
on Computer Communications (Infocom), San Francisco, CA, April 1996.
[Chlamtac+89] I. Chlamtac, A. Ganz, and G. Karmi. Purely optical networks for terabit com-
munication. In Proceedings of the IEEE Conference on Computer Communications (Infocom),
Ottawa, Canada, April 1989.
[Chudnovsky+88] D. V. Chudnovsky, G. V. Chudnovsky, and M. M. Denneau. Regular Graphs
with Small Diameter as Models for Interconnection Networks. IBM Technical Report RC-13484,
1988.
[Ellinas+98a] G. Ellinas and K. Bala. Wavelength assignment algorithms for WDM shared
protection rings. In Proceedings of the IEEE Int’l Conf. Commun. (ICC), Atlanta, GA, June
1998.
[Ellinas+98b] G. Ellinas, K. Bala, and G. K. Chang. Scalability of a novel wavelength assignment
algorithm for WDM shared protection rings. In Proceedings of the IEEE/OSA Optical Fiber
Commun. Conf., San Jose, CA, February 1998.
[Fleischner90] H. Fleischner. Eulerian Graphs and Related Topics Part I: Vol. 1. Annals of
Discrete Mathematics, 45. Amsterdam: Elsevier Science 1990.
[Ganz+92] A. Ganz and Y. Gao. Traffic scheduling in multiple WDM star systems. In Proceedings
of the IEEE Int’l Conf. Commun. (ICC), pp. 1468–1472, Chicago, IL, June 1992.
[Garey+79] M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to NP-
Completeness. New York: W. H. Freeman 1979.
[Gerstel+97] O. Gerstel, R. Ramaswami, and G. H. Sasaki. Benefits of limited wavelength con-
version in WDM ring networks. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf.,
Dallas, TX, February 1997.
[Ghafoor85] A. Ghafoor. A New Database Allocation Scheme for a Class of Networks. Ph.D.
thesis. Columbia University, New York, 1985.
[Gopal82] I. S. Gopal. Scheduling Algorithms for Multibeam Communications Satellites. Ph.D.
thesis. Columbia University, New York, 1982.
[Gusfield83] D. Gusfield. Connectivity and edge-disjoint spanning trees. Information Processing
Lett., 16(2):87–89, 1983.
[Hagouel83] J. Hagouel. Issues in Routing for Large and Dynamic Networks. Ph.D. thesis.
Columbia University, New York, 1983.
[Halsall96] F. Halsall. Data Communications, Computer Networks, and Open Systems. Reading,
MA: Addison-Wesley 1996.
[Harary72] F. Harary. Graph Theory. Reading, MA: Addison-Wesley 1972.
[Hill88] G. R. Hill. A wavelength routing approach to optical communications networks. In
Proceedings of the IEEE Conference on Computer Communications (Infocom), pp. 354–362,
New Orleans, LA, March 1988.
[Hoffman+60] A. J. Hoffman and R. R. Singleton. On Moore graphs with diameter 2 and 3. IBM
J. Res. Dev., 5(4):497–504, 1960.
574 Multiwavelength Optical Networks

[Hwang+92] F. K. Hwang and D. Richards. The Steiner tree problem. Networks, 22(1):55–89,
1992.
[IEEE91] IEEE. Standard 802.6, Distributed Queue Dual Bus (DQDB) Sub-network of a
Metropolitan Area Network (MAN). New York: IEEE 1991.
[Imase+83] M. Imase and M. Itoh. A design for directed graphs with minimum diameter. IEEE
Trans. Comput., c-32(8):782–784, 1983.
[Karasan+96] E. Karasan and E. Ayanoglu. Effects of wavelength routing and selection algo-
rithms on wavelength conversion gain in WDM optical networks. In Proceedings of the LEOS
Summer Topical Meeting, Keystone, CO, August 1996.
[Kou+81] L. Kou, G. Markowsky, and L. Berman. A fast algorithm for Steiner trees. In Acta
Informatica, 15(2):141–145, 1981.
[Liu+99] G. Liu and T. E. Stern. An algorithm for optimal routing and wavelength assignment
on a bi-directional ring. Bell Laboratories Technical Memorandum BL0112120-990610-10TM,
1999.
[Malli+98] R. Malli, X. Zhang, and C. Qiao. Benefits of multicasting in all-optical networks. In
Proceedings of the SPIE Conference on All Optical Networking., pp. 209–220, Boston, MA,
November 1998.
[Mokhtar+98] A. Mokhtar and A. Azizoglu. Adaptive wavelength routing in all-optical networks.
IEEE/ACM Trans. Network., 6(2):197–206, 1998.
[Mukherjee+96] B. Mukherjee, S. Ramamurthy, D. Banerjee, and A. Mukherjee. Some principles
for designing a wide-area optical network. IEEE/ACM Trans. Network., 4(5):684–696, 1996.
[Nagatsu+95] N. Nagatsu, Y. Hamazumi, and K. Sato. Optical path accommodation designs
applicable to large-scale networks. IEICE Trans. Commun., E78-B(4):597–607, 1995.
[Pankaj92] R. K. Pankaj. Architectures for Linear Lightwave Networks. Ph.D. thesis. Mas-
sachusetts Institute of Technology, Cambridge, 1992.
[Pankaj+95] R. K. Pankaj and R. G. Gallager. Wavelength requirements of all-optical networks.
IEEE/ACM Trans. Network., 3(3):269–280, 1995.
[Ramaswami+94] R. Ramaswami and K. N. Sivarajan. Routing and wavelength assignment in
all-optical networks. In Proceedings of the IEEE Conference on Computer Communications
(Infocom), pp. 970–979, Toronto, Canada, June 1994.
[Ramaswami+95] R. Ramaswami and K. N. Sivarajan. Routing and wavelength assignment in
all-optical networks. IEEE/ACM Trans. Network., 3(5):489–500, 1995.
[Rayward-Smith+96] V. J. Rayward-Smith, I. H. Osman, C. R. Reeves, and G. D. Smith. Modern
Heuristic Search Methods. New York: Wiley 1996.
[Reeves93] C. R. Reeves. Modern Heuristic Techniques for Combinatorial Problems. New York:
Wiley 1993.
[Rose92] C. Rose. Mean internodal distance in regular and random multihop networks. IEEE
Trans. Commun., 40(8):1310–1318, 1992.
[Ross89] F. E. Ross. An overview of FDDI: The fiber distributed data interface. IEEE J. Select.
Areas Commun., 7(7):1043–1051, 1989.
[Rouskas03] G. Rouskas. Optical layer multicast: Rationale, building blocks and challenges.
IEEE Network, 17(1):60–65, 2003.
[Sahasrabuddhe+99] L. Sahasrabuddhe and B. Mukherjee. Light-trees: Optical multicasting for
improved performance in wavelength-routed networks. IEEE Communications Mag., 37(2):67–
73, 1999.
[Sahasrabuddhe+00] L. Sahasrabuddhe and B. Mukherjee. Multicast routing algorithms and
protocols: A tutorial. IEEE Network, 14(1):90–102, 2000.
 Wavelength/Waveband-Routed Networks 575

[Schwartz87] M. Schwartz. Telecommunication Networks: Protocols, Modeling, and Analysis.

Reading, MA: Addison-Wesley 1987.
[Subramaniam+96a] S. Subramaniam, M. Azizoglu, and A. K. Somani. All-optical networks
with sparse wavelength conversion. IEEE/ACM Trans. Network., 4(4):544–557, 1996.
[Subramaniam+96b] S. Subramaniam, M. Azizoglu, and A. K. Somani. Connectivity and sparse
wavelength conversion in wavelength-routing networks. In Proceedings of the IEEE Conference
on Computer Communications (Infocom), San Francisco, CA, March 1996.
[Subramaniam+97] S. Subramaniam and R. Barry. Wavelength assignment in fixed routing WDM
networks. In Proceedings of the IEEE Int’l Conf. Commun. (ICC), Montreal, Canada, June
1997.
[Sun+01] Y. Sun, J. Gu, and D. H. K. Tsang. Multicast routing in all-optical wavelength-routed
networks. Optical Networks Mag., 2(4):101–109, 2001.
[Takahashi+80] H. Takahashi and A. Matsuyama. An approximate solution for the Steiner prob-
lem in graphs. Math. Japonica, 24(6):573-577, 1980.
[Tarjan83] R. E. Tarjan. Data Structures and Network Algorithms. Philadelphia: Society for
Industrial and Applied Mathematics 1983.
[Wang+00] B. Wang and J. Hou. Multicast routing and its QoS extension: Problems, algorithms,
and protocols. IEEE Network, 14(1):22–35, 2000.
[Wauters+96] N. Wauters and P. Demeester. Design of the optical path layer in multiwavelength
cross-connected networks. IEEE J. Select. Areas Commun., 14(5):881–892, 1996.
[Winter87] P. Winter. Steiner problem in networks: A survey. Networks, 17(2):129–167, 1987.
[Xin+04] Y. Xin and G. Rouskas. Light-tree routing under optical layer power budget constraints.
OSA J. Optical Networking, 3(5):282–302, 2004.
[Yates+96] Y. Yates and S. Everitt. Limited range wavelength translation in all-optical networks.
In Proceedings of the IEEE Conference on Computer Communications (Infocom), San Fran-
cisco, CA, March 1996.
[Zhang+00] X. Zhang, J. Wei, and C. Qiao. Constrained multicast routing in WDM networks
with sparse light splitting. IEEE/OSA J. Lightwave Technol., 18(12):1917–1927, 2000.
7 Logically-Routed Networks

In this chapter we explore the structure, design, and performance of purely optical
networks with electronically switched overlays. These are the logically-routed networks
(LRNs) that were introduced in Section 3.5. Typical examples of LRNs are networks
of SONET digital cross-connects (DCSs), networks of IP/MPLS routers, and ATM
networks carried on a SONET DCS layer. To provide maximum flexibility, the LRN
should be carried on top of a reconfigurable optical network. Although we generally
refer to the underlying infrastructure as “purely optical” (that is, transparent), we shall,
from time to time, relax that requirement to include optical networks having some degree
of opacity on their transmission links.

7.1 Introduction: Why Logically-Routed Networks?

The rationale for using logical switching on top of a purely optical infrastructure has
been discussed at various points throughout the book. The number of stations in a purely
optical network cannot be increased indefinitely without running into a connectivity
bottleneck. The sources of the bottleneck are the resource limitations within the network
(fibers and optical spectrum) and within the access stations (optical transceivers).
Figure 7.1(a) illustrates the bottleneck in a purely optical network. Network access
station (NAS) A has established logical connections (LCs), shown as dashed lines in
the figure, that fan out to stations B, C, and D. If this is a wavelength-routed network
(WRN), each LC is carried on a separate point-to-point optical connection; that is,
three optical transceivers and three distinct wavelengths are required (assuming that
the stations have single fiber pair access links). If it is an LLN, fewer transceivers and
wavelengths could be used for the same three connections. For example, using an optical
multicast connection on an optical path in the form of a tree, a transmitter at A can reach
stations B, C, and D using one wavelength time shared via TDM to create the three
LCs.
If we try to extend the purely optical approach to large networks, however, we soon
run into problems. For example, if a WRN is to support full logical connectivity among
N stations, this would require a fan-out of N − 1 point-to-point optical connections
from each station. Given the finite available optical spectrum, the number of fibers in the
network would have to grow as N 2 to provide the necessary capacity, and each station
would require N − 1 transceivers. For large N , this is an obvious impossibility.
 Logically-Routed Networks 577

B
A
C

(a)

B E

C F
A
D
G

(b)
NAS

LSN with access station

ONN

Figure 7.1 Why logically-routed networks?

Using the optical multicast and channel-sharing capabilities of an LLN helps. Full
logical connectivity could be supported at an NAS using some mixture of fixed and
rapidly tunable transmitters and receivers, combined with multicast optical connections.
However, eventually the routing constraints of the LLN, the overhead associated with
the multiplexing/multiple-access methods required for channel sharing, and the tuning
requirements on the transceivers limit the number of stations that can be accommodated.
Consequently, none of the purely optical approaches is practical for N on the order of
103 or more.
An additional limitation of both the WRN and the LLN approaches is the geograph-
ical “reach” limitation. A transparent optical connection accumulates noise, distor-
tion, and cross-talk over long distances and eventually becomes unusable. Thus, signal
578 Multiwavelength Optical Networks

regeneration (usually electronic) eventually becomes necessary.1 Regeneration offers a

trade-off, sacrificing the advantages of transparency to extend the reach of the optical
path. However, the problem of limited connectivity remains. Although most applications
do not require full logical connectivity among a large number of end systems, aggrega-
tion of traffic from many users at one access station tends to increase the connectivity
requirements at that station. This is the case, for example, if a telephone Central Office
aggregating traffic from many individual callers or an IP router aggregating packet traffic
from many users accesses an optical network through one NAS. Every destination must
be potentially reachable from that aggregation point. The high connectivity requirement
carries with it a problem of low average usage of each optical connection. Many con-
nections fanning out from one station necessarily generate a low average throughput per
connection. This in turn leads to extreme inefficiency if the full capacity of one λ-channel
is allocated to each connection, as is the case in WRNs, for example. Furthermore, the
enormous number of required connections leads to an overwhelming optical resource
requirement (fibers, wavelengths, transceivers, optical switch ports, etc.).
For a concrete illustration, let us revisit the gateway interconnection example of
Section 1.4. A network that serves 10,000 gateways was considered, so the total number
of connections for 10% connectivity is 10,000,000. Now if each gateway transfers an
average aggregate traffic of 2 Gbps, distributed equally to all other gateways, then the
total throughput is 20 Tbps, but the average flow on each connection is only 2 Mbps!
(These numbers do not get much better if one postulates a connectivity requirement
reduced by two orders of magnitude.)
The most efficient way of meeting this combined requirement of high connectivity
and low throughput per connection is through an LRN, as in Figure 7.2. If the gateway
interconnection application is supported by an LRN, each connection between a pair
of gateways is made at the virtual connection level (the virtual topology [VT] layer
in the figure). Recall from Section 3.5 that traffic from each virtual connection moves
through an LRN on a logical path (LP), composed of a succession of LCs interconnected
through electronic logical switching nodes (LSNs). The LSNs perform a routing function,
ensuring that the traffic associated with a given source-destination virtual connection
follows a prescribed LP through the LRN. Because each LC in general supports many
multiplexed virtual connections, the load on a logical connection (and its supporting
optical connection) is normally substantial even though each virtual connection might
carry a mere trickle. At the same time, each LSN processes only the flow on LCs
terminating on it. Thus, the electronic processing load on each LSN is, in general, much
lower than the optical throughput at the ONN to which it is attached. In this way the LRN
distributes the various processing and communications functions between the logical and
physical layers in a manner that is matched to their capabilities.
The effect of this approach on virtual connectivity is illustrated in Figure 7.1(b). Now
the LCs terminate on LSNs accessing the purely optical network through the NASs.
Switching node A reaches nodes B, C, and D through the same optical paths as before.
1
Although regeneration can be performed in the optical domain (Section 4.11), it still renders the optical
connection opaque, eliminating many of the advantages of a transparent optical connection (see Section 1.3).
Furthermore, at this writing optical regeneration is still in the exploratory stage.
 Logically-Routed Networks 579

VT

LT

PT

LTN with access station

Figure 7.2 A schematic of a point-to-point LRN.

But now, the virtual connections carried on those paths may extend still farther on the
next set of LCs that fan out from B, C, and D. In this way, an LP tree is built up,
with the number of potential end-to-end virtual connections growing exponentially with
the depth of the tree. The tree can be extended over any geographical distance without
the accumulation of signal impairments, and the hardware and spectrum limitations
of the purely optical network can be observed by choosing a logical topology that is
consistent with the optical constraints.
Three important features of the LRN are
1. The end systems communicate with each other using virtual connections, each of
which typically carries relatively low traffic when compared to the capacity of a
λ-channel.
2. Typical LRN topologies provide alternate LPs between all source-destination pairs
for load distribution and survivability.
3. The logical topology can be designed almost independently of the supporting physical
topology.
These characteristics suggest that sophisticated routing, traffic engineering, and pro-
tection strategies can be used in the logical layer to adapt to changing load conditions
580 Multiwavelength Optical Networks

and to address fault recovery in the logical layer.2 Furthermore, there are several options
for fault recovery in an LRN: at the virtual connection level by rerouting traffic around
a failed LC or LSN, or at the logical level by rerouting LCs supported by a cut fiber or
failed ONN onto different optical paths (optical layer protection). All of these functions
are facilitated in an LRN because the format of an LC as defined in Section 2.1 makes the
information content of the signals it carries available to the switching nodes for a variety
of purposes, including routing, grooming, and fault detection/recovery. For example,
overhead information in a SONET/SDH frame or header information in IP packets can
be read for monitoring and control purposes in a logical switch.3
In this chapter we study logically-routed networks in detail, focusing first on point-to-
point LRNs (multihop networks) and then on LRNs with multipoint logical connectivity
(hypernets). As in Section 3.5, it is usually convenient to envision the traffic in these
networks as asynchronous, in which case the LSNs would be packet or cell switches.
All of the essential features of LRNs apply to stream (synchronous) traffic as well,
except that addressing and routing are handled differently. In an asynchronous traffic
LRN (e.g., an IP or MPLS network), routing is executed on a packet-by-packet basis.
In a synchronous traffic LRN (e.g., a network of SONET cross-connects), routing is
performed on a synchronous transport signal (STS) basis (with addressing implicit in
the position of slots in a SONET frame). Because packet-/cell-based networks offer
more flexibility and generality in traffic handling, we tend to use this as our prototypical
example.
The logical topology (LT) in an LRN can be viewed as an interface for matching a
prescribed set of traffic requirements to a given physical infrastructure. Thus, based on
some expected flow distribution among its nodes, an LRN must be constructed with a
logical topology and LC capacities that are sufficient to support those requirements. But
to complete the picture, the chosen LT and the logical connection capacities must be
realizable using the resources of the given physical infrastructure: its physical topology
(PT) and its fiber, node, and access station functionality.
Because of the complexity of the design problem, involving both logical and physical
layer constraints, it is helpful to separate it into two subproblems:
1. Logical-layer design (including routing and capacity assignment)
2. Physical-layer design: embedding the LT on the PT
This chapter emphasizes performance and design issues. Thus, for a given logical
topology we are interested in determining performance in terms of throughput and
connectivity, as well as other related features such as average LP length, adaptability
to traffic imbalance, fault tolerance, and so forth. Concerning physical-layer design,
2
For example, bifurcated routing (i.e., assigning several parallel paths for traffic flow between a given source-
destination node pair) makes it possible to distribute traffic loads evenly throughout a network. It also allows
for “graceful” rather than catastrophic performance degradation in the event of a single failure on one of the
paths, providing time for the affected traffic to be accommodated on alternate paths. Also, redundant LPs
can be used for fault recovery (see Section 8.4).
3
Because of the availability of the bit streams in electronic form in opaque optical switches, these switches can
implement certain information-based functions normally associated with the logical layer (e.g., grooming
and fault detection). This is not possible in transparent optical switches.
 Logically-Routed Networks 581

the basic problem is realizing a specified logical topology on a purely optical network
infrastructure. As discussed in Chapter 1, the physical-layer topology is typically fixed
and may support several independent and disparate logical networks of which our LRN
is but one. In this case physical-layer design becomes a topological embedding problem,
encountered previously in Section 3.5. In other cases, the physical topology might be
chosen to match the logical topology of the LRN. A good choice in this case may be
a multistar realization. Multistars were presented previously as physical supports for
wavelength- and waveband-routed networks in Sections 6.3.6 and 6.5.7, respectively.
Multihop LRNs and hypernets are treated separately in the sections that follow.
We referred above to the traffic in the LRN as a “flow,” suggesting that a fluid-flow
model is sufficient for representing the traffic demands at the virtual connection level.
Although this is appropriate in many of the design problems treated in this chapter,
it ignores one very important aspect of LRN operation: the problem of efficiently
multiplexing fine-granularity virtual connections originating with end users onto coarse
granularity logical paths carrying aggregated traffic in the core network; that is, the
grooming problem. The cost of a network design as well as the efficiency of its operation
depends in a significant way on how well this task is carried out. This aspect of LRN
performance and design is introduced in the next section and treated in more depth later
in the chapter.

7.1.1 Multitier Networks: Grooming

As illustrated in Section 2.6, there are many ways an LRN architecture may be orga-
nized. Several LRN overlays may be supported as clients of an optical network, and
several LRNs can be stacked in their own client-server relationships; e.g., IP over ATM
over SONET over the DWDM optical infrastructure. We refer to these architectures as
multitier networks. In each of these cases the equipment in the LRN overlays must be
configured to deal with the granularity mismatches that must be resolved through groom-
ing. (Recall that grooming within the optical layer of a waveband-routed network was
already encountered in Figure 4.89, where a purely optical hybrid waveband/wavelength
switch is used to pack fine granularity λ-channels into coarse granularity wavebands
to achieve economies in optical switching equipment.) We examine some technological
aspects of the LRN problem now.
The basic point-to-point transmission unit in an LRN (in the LT layer of Figure 7.2)
is our generic logical connection. It is supported by an optical connection or lightpath
(routed through the PT layer in the figure). Recall that each LC is a unidirectional
connection whose electrical bit stream is converted to an optical signal carried on a
λ-channel and terminating on electrical ports of its source and destination NASs. In
a point-to-point optical network, there is a one-to-one correspondence between these
entities, so in this section we use LC, optical connection, λ-channel, and lightpath
interchangeably depending on the context to relate the discussion to terms used in the
references. Each LC typically operates at a bit rate determined by the standard for the
LRN of which it is a part, so it normally acts as a unidirectional fixed bit-rate pipe.
Different LCs in the core network might run at different speeds, all within a standard
582 Multiwavelength Optical Networks

STS-48 granularity
switch fabric
Core (STS-48/192)
STS-1 granularity
switch fabric E OC-48 circuits
OC-1 circuit
A B

D C
OC-1 circuit

Figure 7.3 Two-tier architecture.

hierarchy; e.g., STS-48 (2.5 Gbps) or STS-192 (10 Gbps) for SONET. However, the end
systems feed traffic into the network on virtual connections (VCs), which may be running
at a wide range of bit rates. For ease of multiplexing onto the core network, the bit rates
of the VCs accessing the network are usually quantized at a finer granularity: typically
some basic subrate of the LC rates in the core, say, STS-1 (about 50 Mbps) in SONET.
For efficient operation of the network, there is a need for packing these smaller VC rates
into the larger core network bit pipes so that each LC is filled as closely as possible to its
capacity. This requires (electronic) multiplexing/demultiplexing at the entrance/exit of
the network, and possibly within the network core. A single-tiered (purely optical) core
network cannot effectively groom subrate bit streams accessing the network to LC rates
on links in the core, because these operations must be done in the logical (electronic)
layer. Instead, a second tier in the form of an electronic overlay is required. The overlay
consists of LSNs containing the electronic multiplexing and switching equipment to
perform the grooming function. These LSNs together with the LCs interconnecting
them form an LRN configured over the core WRN. The LRN carries the VC traffic
generated by the end systems in the VT layer using the resources of the WRN. The
grooming problem then consists of efficiently assigning logical paths in the LT layer to
the traffic demands in the VT layer.
An example of how grooming is implemented in a typical two-tier network is shown
in Figure 7.3. Each core network node consists of an OXC, and each edge node also
includes a SONET DCS and associated multiplexing equipment. The DCSs execute the
functions of subrate switching, grooming, and add/drop to end systems (not shown). Note
that we are using a detailed symbolic description of the nodes in Figure 7.3 to replace
the simple abstractions in Figure 7.2. (In this case the SONET DCSs at the edge of
the network are represented by boxes containing stars and the OXCs in the core are boxes
containing crosses.) The OXCs and the DCSs in Figure 7.3 play the respective roles of
the ONNs and LSNs in Figure 7.2. The OXC used in this case is not our usual transparent
ONN but rather an opaque switch in the O-E-O form of Figure 4.87; the purpose being to
align our example with current practice. At this writing the much anticipated deployment
 Logically-Routed Networks 583

of purely optical (O-O-O) switching nodes in a transparent core network has not yet
materialized to any significant degree. Instead it has been replaced by the deployment of
opaque O-E-O switches for reasons stated in Section 4.12.2.2. Although some important
benefits of purely optical switching are sacrificed in this case, networks built around these
switches offer a number of advantages in two-tiered architectures, including ease of
grooming, fault recovery, and other control and management operations. These benefits
accrue from the fact that the constituent bit streams in each λ-channel are converted to
electronic form within an O-E-O switch, so the bits used for control and management
become directly accessible, and multiplexing/demultiplexing and switching are possible
at subchannel rates within the switch fabric [Bala+95]. Thus opaque networks currently
remain the most practical and cost-effective way to build a dynamic, scalable, and
manageable core backbone network.
Although Figure 7.3 suggests a clear separation of optical-layer switching on the one
hand and logical-layer multiplexing and grooming on the other, the actual structure of
a typical O-E-O network node, shown more clearly in Figure 4.87, allows for a much
closer integration of switching, multiplexing, and grooming operations.4 This type of
switch is naturally adapted to grooming because the electronic switch fabric generally
operates at a much lower bit rate than the transmission links; a 10- to 40-Gbps (STS-192
to STS-768) transmission channel carried on a single wavelength might be electrically
demultiplexed down to 2.5 Gbps (STS-48) bit streams or less for switching purposes,
and these might be demultiplexed still further for purposes of traffic grooming. As
illustrated in Figure 4.87, subrate tributaries from various end systems (SONET, IP,
ATM) are add/dropped from the switch, and the various multiplexing and grooming
functions to pack/unpack these bit streams into high rate λ-channels can be included
with the switching functions in the electrical core. The inclusion of the subrate grooming
functions is typically implemented by interposing a subrate switch (e.g., a SONET ADM
or cross-connect) between the electrical core switch fabric and the end systems accessing
the core.
Another version of a network node that includes the grooming function is shown in
Figure 7.4, where full wavelengths are add/dropped to an LSN (the grooming switch)
from a transparent OXC. In this way grooming is combined with optical bypass. As
the figure indicates, this arrangement is more in keeping with our layered architecture,
where there is a clear-cut separation of the LSN (in the logical layer) and the OXC
(in the optical layer). Furthermore, optical bypass keeps the transiting traffic out of the
electronic domain, reducing the problem of an electronic bottleneck.
Returning to the illustration of Figure 7.3, note that the edge node functions in the
SONET DCSs are executed in an STS-1 switch fabric while the electrical core of the
opaque OXC operates at STS-48. The figure shows an OC-1 connection between nodes
A and B (dotted line) and another OC-1 connection between nodes A and C (dashed
line). Both are multiplexed (groomed) into higher rate OC-48 connections. The former
follows a one-hop logical path, whereas the latter uses a two-hop LP consisting of two
LCs: [A,D] and [D,C]. Note that [A,D] is carried on two optical hops: (A,E) and (E,D).
4
It also blurs the boundary between the optical layer and the logical layer. The real world is not as simple as
our neatly layered multiwavelength network architecture suggests.
584 Multiwavelength Optical Networks

TO N
D
E SY STEMS

STS-1

R
G OOMING SN
L
SWITCH

e
OC-48 o

SX C
W

OXC

Figure 7.4 The architecture of a grooming node with optical bypass.

At node D, the OC-1 connection is switched from one LC to another, which requires
demultiplexing it from the OC-48 connection that carries it from node A to node D
and then switching and remultiplexing it on another OC-48 connection from node
D to node C. This operation, known as “hairpinning,” is an inevitable and undesirable
consequence of the packing and unpacking of many lower rate connections onto a higher
rate connection in an attempt to “fill” the high rate optical connection transporting them.
One of the objectives of a grooming algorithm is to keep hairpinning to a minimum
because it burdens the electronic switching equipment.
Considerable performance improvement can be obtained in multitiered networks using
these general-purpose grooming and switching nodes. Algorithms for grooming and
resultant network performance are discussed in detail in Sections 7.3.3 and 7.5.3.
Combination of physical-layer switching and logical-layer switching, multiplexing,
and grooming functions into an “all-purpose” core switching node has long been a goal
of network providers. It would accommodate all services at all rates and be scalable
and low cost. Scalability and manageability of software and hardware, however, are
practical issues that have led to the deployment of mostly hierarchical (tiered) network
architectures of the type just described, where each layer is optimized independently.
In particular, fast shared-path fault protection is difficult in a one-tier mesh network
architecture because of the need to reconfigure a large number of cross-connects when
a failure occurs [Ellinas+03] (see Chapter 8).
As opposed to a single-tier opaque network, the protection model in a two-tier archi-
tecture is scalable to large networks with high traffic because protection is always done
for the coarsest granularity traffic flows. Therefore, fast, and capacity-efficient shared-
path protection can be implemented. The multitier opaque architecture does have some
inefficiencies associated with grooming and logical routing, however. These are due to
hairpinning of subrate traffic whenever there is a need to multiplex/demultiplex and/or
 Logically-Routed Networks 585

switch subrate bit streams, as described in the example of Figure 7.3. Still, hairpinning
can be controlled and in fact traded for other performance objectives. For example,
by limiting switching within the network to coarse granularity traffic units whenever
possible, use of the lower rate logical switches is minimized, but this generally leads to
less efficient usage of the optical bandwidth (see Section 7.3.3.2).
A consensus has emerged in the industry, that optical networks will evolve to multitier
architectures where purely optical (O-O-O) switching nodes will coexist with electro-
optical (O-E-O) switching nodes. The former will usually be used for high-speed bypass
services, whereas the latter will provide the control and management as well as the
aggregation and grooming functions that are essential to the efficient operation of the
network. Chapter 11 describes some of the recent deployments of grooming and all-
optical switches and discusses the different configurations of these deployments.

7.2 Point-to-Point Logical Topologies: Multihop Networks

In this section we return to a more abstract view of the network examining the LRN
concept in the context of point-to-point logical topologies. From a top-down point of
view, the transport network is seen as a collection of LSNs (henceforth called nodes)
joined by unidirectional internodal LCs. Thus, the logical connection graph (LCG) for
the LRN is a digraph with vertices that represent the nodes and with arcs that represent
the LCs. LCGs were used in earlier chapters to represent logical connectivity in both
purely optical networks and LRNs. In the former case the vertices of the LCG represent
the access stations, which serve as termination points for all LCs. In the case of the
LRN, the vertices of the LCG represent LSNs, which serve as transit nodes for virtual
connections.5
The virtual connections in an LRN are routed on generally multiple-hop LPs between
end systems. A schematic view of a point-to-point LRN is shown in Figure 7.2. The
virtual topology (VT) represents virtual connections in place among the end systems.
These would typically be a mix of demand-assigned and dedicated connections, so the
VT might change on a fairly short timescale (e.g., minutes). The logical topology is
typically static (or quasistatic), changing when changing traffic distributions or network
faults necessitate reconfiguration. The physical topology is determined by the fiber cable
layout and remains fixed except when faults occur in the cables or the ONNs. Note that
each of the topologies supports the one above it, but they all differ from each other.
Although the LRN must be realized on some physical (optical) infrastructure, the
physical topology of the supporting network and its functionality (e.g., WRN or LLN)
can be ignored in discussing logical-layer performance characteristics. Of course when
it comes to network design, the structures of the logical and physical layers interact
closely with each other.
The earliest proposal for an optically based LRN (using a perfect shuffle logical
topology) appeared in [Acampora87], in which the term multihop was coined to denote
5
At times we shall also include logical terminal nodes (LTNs) without any switching functions. Examples of
LTNs appear in Figures 7.2 and 7.14.
586 Multiwavelength Optical Networks

a point-to-point LRN. Multihop networks based on a more general shuffle exchange

logical topology (called ShuffleNets) are explored thoroughly in [Hluchyj+88, 91]. Ex-
tensions of these ideas to deBruijn graph LTs were presented in [Sivarajan+91]. Irregular
topologies constructed to match traffic distributions were proposed in [Bannister90] and
[Labourdette+91].
Although there is no necessity for LRNs to have regular topologies, almost all of the
known results on these networks focus on the regular case, partly because it is difficult
to draw any general conclusions concerning performance for irregular networks. Some
of the advantages of regularity in both point-to-point and hypergraph LRNs are
r Simplified routing and congestion control mechanisms
r Potentially small network diameters and average internodal distances
r Highly symmetric topologies, simplifying load balancing, and congestion reduction
r Identical structures for each network node, simplifying network implementation

Using arguments similar to those used for WRNs in Section 6.2, we can find certain
important structural bounds on regular directed LCGs. Consider a regular digraph,
G(V, E), of maximum out-degree δ > 1 and diameter D. Then its maximum order is
d δD +1 − 1
NMoore (δ, D) = 1 + δ + δ 2 + · · · + δ D = . (7.1)
δ−1
Equation (7.1) is just the directed version of the Moore bound given in Equation (6.1).6
The average internodal distance in an LRN represented by a digraph G(V, E) of order
N is defined as
1 
d¯ = d(vi , v j ), (7.2)
N (N − 1) i = j

where d(vi , v j ) is the length (in logical hops) of the shortest (directed) path from vertex
vi to v j . For a regular directed network with in- and out-degree δ > 1 and order N , d¯ is
bounded below by
δ − δ D + 1 + N D(δ − 1)2 + D(δ − 1)
d¯ ≥ , (7.3)
(N − 1)(δ − 1)2
where D is the smallest integer such that
δD +1 − 1
≥ N. (7.4)
δ−1
Now let us examine the relation between d¯ and network throughput. Consider an LRN
with N nodes and L unidirectional point-to-point LCs in which each logical connection
has a capacity of one unit of flow. Its average nodal in- and out-degree is δ̄ = L/N .
Suppose ρ units of traffic are injected into each node for a total network throughput
of ρ N units, with uniform traffic distribution, and the average number of logical hops
¯ To maximize the network throughput, the traffic should
incurred by the traffic is H̄ ≥ d.
be routed over logical paths in a way that simultaneously minimizes H̄ and loads all

6
In contrast to undirected graphs, there are no directed graphs that achieve the Moore bound.
 Logically-Routed Networks 587

LCs equally. By flow conservation this would lead to a load on each LC of ρ H̄ /δ̄. This
reasoning gives us the maximum throughput per node for the network:
¯
ρmax = δ̄/d, (7.5)

which can be achieved only if all virtual connection routing is over the shortest LPs.
Note that to realize the bound, the traffic distribution must be “matched” appropriately
to the logical topology and routed optimally. Generally, the bound is not achievable. This
bound is useful in evaluating network performance in the sections that follow.

7.2.1 ShuffleNets
As mentioned, the first regular topologies considered for LRNs were ShuffleNets. The
LCG for a (δ, k)-ShuffleNet is a regular digraph of in- and out-degree δ and of order
N = kδ k . As illustrated in Figure 7.5 for the case δ = 3, k = 2, N = 18, it is conveniently
represented as an arrangement of k columns, each containing δ k nodes. Each column
connects to the next one on its right via δ k+1 arcs, in a generalization of a perfect shuffle
pattern. As shown by the dashed line repetition of the first column in the figure, the LCG
is closed back on itself in a cylindrical fashion to complete the regular structure. The
complete network requires kδ k+1 point-to-point LCs. Although the regular structure of

0 9 0

1 10 1

2 11 2

3 12 3

4 13 4

5 14 5

6 15 6

7 16 7

8 17 8

Figure 7.5 ShuffleNet: δ = 3, k = 2, N = 18.

588 Multiwavelength Optical Networks

the ShuffleNet is useful in many ways, it carries with it a disadvantage: ShuffleNets do

not exist for all values of N . Thus, a ShuffleNet (as well as all other network topologies
possessing a high degree of symmetry) cannot be “grown” in a modular fashion, one
node at a time.
End systems (not shown in Figure 7.5) accessing the switching nodes of the ShuffleNet
communicate with each other by injecting sequences of packets into the nodes, which
then sort and forward the packets over LPs to their destination nodes. For example, using
shortest path routing, traffic from node 0 to node 7 would take the path 0–11–7, whereas
traffic from 0 to 12 would use path 0–9–1–12, 0–10–4–12, or 0–11–7–12. (Shortest
paths are not always unique.)
The diameter of a (δ, k)-ShuffleNet is D = 2k − 1, and thus the order of a ShuffleNet
can be expressed as

D + 1 ( D+1 )
N= δ 2 , (7.6)
2
which shows that it is a moderately dense network, containing a large number of nodes
for a given diameter and nodal degree.
The average internodal distance is given by
kδ k (δ − 1)(3k − 1) − 2k(δ k − 1)
d¯ = . (7.7)
2(δ − 1)(kδ k − 1)
From Equation (7.7) it can be seen that for large ShuffleNets (large δ and/or k), we
have d¯ ≈ 3D/4. Substituting Equation (7.7) into Equation (7.5) leads to the following
bound on throughput:
2δ(δ − 1)(kδ k − 1)
ρmax = . (7.8)
kδ k (δ − 1)(3k − 1) − 2k(δ k − 1)
Figure 7.6 shows a plot of ρmax as a function of N with δ as a parameter. Note that ρmax
decreases relatively slowly with increasing network size due to the roughly logarithmic
dependence of d¯ on N . Note also that ρmax increases with δ due to the increased capacity
available in the network. However, this increase in ρmax is obtained at the cost of increased
usage of physical-layer resources – both transceivers and fiber capacity. As we shall see
(Section 7.4), by using shared rather than dedicated channels in the physical layer, it is

max
10
8
67
4 5
3
2
1

.1 N
1 10 100 1000 10,000
Number of Users

Figure 7.6 Maximum throughput per node for ShuffleNet. (From [Hluchyj+88, Figure 18]. Copy-
right
c 1988 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers,
Inc.)
 Logically-Routed Networks 589

possible to control network throughput and network connectivity (i.e., size) more or less
independently, giving considerably more flexibility in network design.

7.2.2 Families of Dense Logical Topologies

In Section 6.2 we suggested a number of reasons for choosing dense physical topologies
for purely optical networks. A similar case can be made for choosing dense logical
topologies for LRNs. Dense logical topologies lead to fewer average logical hops and
more alternate paths. These in turn produce a number of desirable results:
r The load on each LC is reduced.
r The processing load in each LSN is reduced.
r Opportunities for load balancing and congestion reduction are improved.
r Probability of information loss due to congestion and buffer overflow is reduced.
r Survivability is increased.

Logical topologies are typically specified as digraphs rather than undirected graphs.
Two useful classes of regular digraphs, the deBruijn and Kautz digraphs, can be defined
in terms of the state transition relations in finite state machines (shift registers).
A deBruijn digraph with in- and out-degree d and diameter D, denoted B(d, D), has
N = d D vertices. Its vertex labels can be defined as words of length D from an alphabet
A containing d letters. If (a1 , a2 , . . . , a D ) is the label of vertex v, then there are arcs from
v to all vertices with labels (a2 , . . . , a D , α), α ∈ A, representing a left shift of the vertex
label. Note that a deBruijn digraph has self-loops for all vertices with labels that contain
a D-fold repetition of a single letter. Omitting the self-loops results in an irregular LCG
with the same order and diameter but with fewer arcs. Figure 7.7(a) shows B(2, 3).
A Kautz digraph with in- and out-degree d and diameter D, denoted K (d, D), has
N = d D + d D−1 vertices. It is constructed in a manner similar to the deBruijn digraph.
The vertices are labeled using all words of length D from an alphabet A containing d + 1
letters such that no two consecutive letters are identical (thus it has no self-loops). There
is an arc from any vertex (a1 , a2 , . . . , a D ) to vertices (a2 , . . . , a D , α), where α ∈ A,
α = a D . Figure 7.7(b) shows K (2, 3). Clearly both the deBruijn and Kautz digraphs
are considerably denser than ShuffleNets. The density of the Kautz digraph comes very
close to the Moore bound.
Families of digraphs that can be considered as generalizations of the deBruijn and
Kautz digraphs were proposed by [Reddy+82] and [Imase+83]. The generalized ver-
sions have the useful property that they can be defined for any number of nodes. However,
they are not necessarily regular.
The generalized Kautz digraph, denoted by G K (d, n), has n vertices with integer
labels modulo n where each vertex has out-degree d. There is an arc from vertex v to
vertices u ≡ −dv − α (mod n) for 1 ≤ α ≤ d. The diameter of the resultant digraph is
at most ⌈logd n⌉. If n = d D + d D−k for a positive integer k, then the diameter is D, and
if k = 1, then G K (d, n) is isomorphic to K (d, D).7

7
Two graphs are isomorphic if one can be converted to the other by relabeling its vertices.
590 Multiwavelength Optical Networks

100 110

000 010 111

101

001 B(2, 3) 011

(a) deBruijn diagraph

210

121 101
012

212 010
120 201

021 102
202 020

K(2, 3)
(b) Kautz digraph

Figure 7.7 deBruijn (a) and Kautz (b) digraphs.

Not only are these digraphs useful as LCGs of point-to-point LRNs, but they can also
be used to generate families of logical connection hypergraphs (LCHs) for hypernets
(see Section 7.4).
The digraphs B(2, 3) and K (2, 3), of order 8 and 12, respectively, have the same
diameter as a (2,2)-ShuffleNet, which has eight nodes. However, for larger networks the
density of the ShuffleNet falls far short of the deBruijn and Kautz networks. Thus, for
example, the ShuffleNet, deBruijn, and Kautz digraphs of in-/out-degree 6 and diameter
5 have orders of N = 648, 7776, and 9072, respectively, compared with the Moore bound
of 9331. The deBruijn and Kautz digraphs can be converted to undirected multigraphs
by removing the arc orientations and then to graphs by coalescing parallel edges and
removing self-loops. For example, the undirected graph versions of the digraphs in
Figure 7.7 are shown in Figure 6.4.
It turns out that for large deBruijn and Kautz digraphs, d¯ is very close to D, indicating
that all nodes are almost the same distance apart, as is the case for large random digraphs
(see Section 6.2). For example, B(4, 5), B(5, 5), and B(6, 5) have d¯ = 4.56, 4.70, and
4.77, respectively, compared with the (4,3)-, (5,3)-, and (6,3)-ShuffleNets (of diameter
5) with d¯ = 3.69, 3.76, and 3.81, respectively. This means that deBruijn and Kautz
networks have slightly less maximum throughput than ShuffleNets with the same degree
 Logically-Routed Networks 591

and diameter. Of course, the ShuffleNet being compared has far fewer nodes than the
deBruijn and Kautz networks.
One of the useful features of logical topologies based on arithmetic operations is
that routing rules can be defined conveniently in terms of node labels [Jiang+95,
Sivarajan+91]. Thus, for example, in deBruijn and Kautz digraphs a simple short-
est path “address shift” routing rule can be defined. Let A = (a1 , a2 , . . . , a D ) and B =
(b1 , b2 , . . . , b D ) be, respectively, the source and destination addresses of a packet, and let
i > 0 be the smallest integer for which B = (b1 , b2 , . . . , b D−i ) = (ai+1 , ai+2 , . . . , a D ).
Then we define the shortest concatenation of A and B as
(AB)∗ = (a1 , a2 , . . . , a D , b D−i+1 , b D−i+2 , . . . , b D ). (7.9)
The value of i gives the length of the shortest path from A to B, and the i + 1
successive substrings of length D starting from the left in ( AB)∗ give the sequence of
nodes on that path. For example, in the Kautz digraph of Figure 7.7(b), the shortest path
from node 210 to 021 is found by constructing the shortest concatenation, 21021, for
i = 2, which gives the two-hop LP 210 → 102 → 021.

7.3 Multihop Network Design

Ideally, design of an LRN should proceed simultaneously in both the logical and physical
layers. However, as mentioned in the introduction to this chapter, the complexity of
the combined design problem makes this exceedingly difficult. Instead, we treat the
logical-layer design problem in the next section, followed by physical-layer design in
Section 7.3.2.

7.3.1 Logical-Layer Design

The problem of logical-layer design can be attacked by starting with a prescribed (typi-
cally highly symmetrical) LCG, say a ShuffleNet or deBruijn or Kautz digraph. A more
general approach is also possible in which only certain basic constraints on the LCG are
assumed, such as the in-/out-degree of the vertices (LSNs). In either case, the objective
of logical-layer design is to match the LCG to a given traffic distribution. The problem
can then be treated as two subproblems:
1. Configuring the logical topology: In the case in which the LCG has been prescribed,
completion of the LT configuration requires placement of stations (i.e., mapping of
stations onto the vertices of the prescribed LCG). If the structure of the LCG is left
free, configuring the LT requires determining an LCG that is feasible in the sense
that it can support the prescribed traffic, given the design constraints.
2. Routing: This entails finding a routing assignment for the prescribed traffic on the
chosen LT that is “good” or best according to some performance criterion.
These subproblems are closely coupled. Determining whether an LCG will support
a given traffic distribution subject to capacity constraints on the LCs generally requires
solving the routing problem for the traffic flows, that is, solving a multicommodity flow
592 Multiwavelength Optical Networks

problem. Furthermore, if an optimal solution is to be found, both the LT configuration

problem and the routing problem must be solved simultaneously, resulting in a very
complex combinatorial optimization problem.
Some solution techniques for the logical-layer configuration and routing problems
are presented in [Labourdette+91] and [Bienstock+93]. They assume that the optical
infrastructure is a WRN, so all logical and optical connections are point-to-point and
there is no channel sharing. Given a set of LSNs, a prescribed traffic distribution among
the nodes, and a fixed number δ of LCs to and from each node, a regular LCG of
in-/out-degree δ is sought together with a routing assignment that minimizes the maxi-
mum flow on the LCs. The degree δ indicates the required number of optical transceivers
at each node. Note that the structure of the nodes and their access stations remains fixed
during the optimization, but the LCG is configured to match the offered traffic.
So far, link capacities have not been included in this formulation. Assuming all LCs
have equal capacity, the flows corresponding to a solution of the minimax problem will
accommodate a version of the prescribed traffic distribution, scaled up (or down) until
the flow on the maximally loaded connection equals the link capacity. The smaller the
maximum flow, the more aggregate throughput can be obtained by scaling upward.
The combined LCG design and routing problem can be cast as a mixed integer program
(MIP) using the following variables [Labourdette+90, 91]:

r xi j : A binary variable set to 1 if an LC is placed from node i to j and set to 0 otherwise

r tsd : The traffic injected into node s and destined for node d
r f isd
j : The traffic flow from source s to destination d, carried on connection [i, j]
r z: The maximum of the flows on all connections.

The MIP is given by the following equations:

Minimize z with

j= i xi j = δ, for all i (7.10)

j= i x ji = δ, for all i (7.11)
sd

s,d f i j ≤ M xi j , for all i = j (7.12)

tsd , if i = s
sd sd
 
j= i f i j − j= i f ji = −tsd , if i = d (7.13)

0, otherwise
sd

s,d f i j ≤ z, for all i = j (7.14)
0≤ f isd
j , for all s, d and all i = j.

Equations (7.10) and (7.11) incorporate the degree constraints. Making M in Equa-
tion (7.12) a sufficiently large quantity allows arbitrarily large flows but only on con-
nections that exist. Equation (7.13) provides the flow conservation relations, and Equa-
tion (7.14) sets z equal to the largest link load.
 Logically-Routed Networks 593

13.17
1 8
13.17 10.69
11.24
13.17

2 7
9.76 8.96
0 10 0.9 1 0.8 1 1.1 0.3
13.17
0.9 0 11 0.9 1 1 0.8 1
13.17
1.1 0.9 0 8 0.8 0.9 1 0.9 11.69 13.17
1.1 1 1 0 9 0.9 0.9 1 3 6
0.8 1 0.8 0.9 0 10 1.1 0.9
0.9 0.9 1 1 0.9 0 12 1.1 11.65
13.10
0.8 1.1 1.1 1 0.8 0.9 0 9 9.64 11.96
4 5
11 0.9 0.8 0.7 1 1.1 0.8 0 13.17
(a) (b)

Figure 7.8 A traffic matrix and matched LCG. (From [Labourdette+91, Figure 5]. Copyright

c 1991 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

As might be expected from our earlier forays into the world of combinatorial optimiza-
tion in Chapter 6, this MIP problem is extremely difficult computationally.8 Recognizing
this, [Labourdette+91] decomposes the problem into two subproblems: a preliminary
heuristic for determining a good LCG followed by solution of the (relatively easy) mul-
ticommodity flow problem (a linear programming problem). In the first subproblem the
heuristic attempts to find an LCG that maximizes the single-hop traffic. The routing
problem is then solved on this fixed LCG.
An example of a solution of an eight-node logical layer design problem is shown in
Figure 7.8. Each node has δ = 2.
A prescribed traffic matrix is shown in Figure 7.8(a). Note that the dominant terms,
shown with bold lines, indicate that most of the traffic follows a unidirectional ring.
The LCG optimized for the prescribed traffic matrix, together with traffic flows, found
using the heuristics and routing of [Labourdette+91], are given in Figure 7.8(b). The
LCG is in the form of a chordal ring, and the maximum flow values of 13.17 are
underlined.
The LCG of Figure 7.8(b) can be viewed as an eight-node ShuffleNet [see
Figure 3.33(a)] that has been “rearranged” by changing the source–destination inter-
connections of its stations to match a prescribed traffic distribution. This provides a
useful comparison between a fixed and an optimized LCG. If the traffic matrix of
Figure 7.8(a) had been routed on the ShuffleNet of Figure 3.33(a), the flow on the
maximally loaded link would have increased to 22.1. Thus, the matched LCG with a
maximum flow of 13.17 gives a 41% improvement in achievable throughput.

8
The intractability of the problem is thoroughly documented in [Bienstock+93].
594 Multiwavelength Optical Networks

7.3.2 Physical-Layer Design

In focusing on the logical-layer design problem, the question of realization of the LT on
an optical infrastructure was set aside temporarily. In the most general setting, optical
realization may include design of the physical (fiber) topology as well as embedding
of the LT onto the physical topology. More commonly, the PT is given and it is the
embedding (i.e., optical path routing and wavelength/waveband assignment) that is
of interest. Embedding of point-to-point LRNs has been considered by [Bannister90,
Gerla+92, Kovacevic93]. We consider the problem of embedding on a given fiber
topology here. Multistar realization is discussed in Section 7.5.4.
Three possible classes of purely optical architectures have been examined in previ-
ous chapters: static, WRN, and LLN. In a typical static network based on a tree fiber
topology, as exemplified by Figures 2.7, 2.10, and 2.11, the embedding problem is
straightforward. Assuming that each logical connection in the LT is supported by a dedi-
cated λ-channel using a fixed-tuned transmitter and receiver, there is no routing problem
(source-destination paths are unique). An embedding of the eight-node ShuffleNet of
Figure 3.33(a) on a folded bus is shown in Figure 3.33(c). It requires 16 wavelengths,
one for each of the LCs. If, instead, a star topology based on an 8 × 8 static router of the
form of Figure 2.10 was used, with the wavelength routing rule given in Section 2.3.1.2,
only 5 wavelengths would be required. (With a different static routing rule the number of
wavelengths could be reduced to 2.) Note that because there are no controllable ONNs
in static networks, reconfiguration of the logical topology is simply a matter of retuning
the optical transceivers in the access stations.
In a wavelength-routed network with a given physical topology, embedding of the
logical topology becomes an RCA problem, which was explored in some depth in
Section 6.3. Optical paths and wavelengths are allocated to each LC using static RCA
rules with some objective in mind, such as minimizing the number of wavelengths used
or the optical path lengths.
In cases when LSNs connect to the network through elementary access stations, the
DCA condition requires that all LCs originating on the same node be assigned distinct
wavelengths, and the same holds true for connections terminating on a common node. In
view of these DCA constraints, a lower bound on the number of wavelengths necessary
to embed a given LCG on any physical topology is simply the largest vertex in- or
out-degree of the LCG. This bound is not necessarily realizable because the structure of
the physical topology may impose additional constraints.
As an example of WRN embedding, Figure 7.9 shows the eight-node ShuffleNet
embedded on the Atlantis PT of Figure 3.36(b). Nodes 1 through 8 in the ShuffleNet
LCG of Figure 3.33(a) are identified, respectively, with the NASs A through H in the
optical network. Elementary access stations are assumed, which means that the minimum
wavelength requirement is 2 (see Section 6.3.2). A feasible routing and wavelength
assignment for the connections among the subsets {A, C, E, F} and {B, D, G, H } is
indicated in Table 7.1. To simplify the routing, the four optical paths for the connections
in each subset are confined to directed trees {1, 2, 3, 4, 5} and {29′ , 18, 20′ , 26′ , 22, 7},
respectively, as shown in Figure 7.9.
 Logically-Routed Networks 595

L
l

J K 11 12 A
j S M
10 k 13
27 r
28
15 1 B
9 14 29 ’
R m
q Q 17 b
I 16 V
i
8 23 T p
24 18
h P
H
25 2
7 N
22
n 26
O
G g 21
o
U e 19 c
6
5 20 C
f E
4 3
d

F
D

{A, C, E, F }
{B, D, G, H }
{F, H, C, D }
{E, G, A, B }

Figure 7.9 ShuffleNet on Atlantis.

The fiber orientations (primed or unprimed) on the paths in Table 7.1 are defined with
respect to the link reference arrows in Figure 7.9. Four wavelengths are required to satisfy
DCA for the four connections in subset {A, C, E, F} because they all share fibers 3 and
4. The same wavelengths are reused in subset {B, D, G, H } because the trees for the two
subsets are fiber disjoint. The optical paths among the remaining subsets {E, G, A, B}
and {F, H, C, D} are also confined to fiber-disjoint trees, as indicated in Figure 7.9.9
Thus, each subset of four connections can reuse the same four wavelengths.10
In this example, the 16 connections on the eight-node ShuffleNet are realized in a
WRN using two transceivers per station and four wavelengths, for a spectrum reuse
factor of 4. The maximum network throughput, normalized to the capacity of one
9
There are cases when two directed trees use the same link. However, they are fiber disjoint because they
use oppositely directed fibers on that link.
10
Although a lower bound on the number of wavelengths necessary to realize this LCG on any physical
topology is 2, the additional wavelengths were required in this case because routing was confined to trees. It
is left as an exercise for the reader to determine an embedding of the ShuffleNet on Atlantis with elementary
access stations that requires fewer than four wavelengths.
596 Multiwavelength Optical Networks

Table 7.1 ShuffleNet routing on Atlantis.

Connection Path Wavelength

(A, E) 1, 2, 3, 4 1
(A, F) 1, 2, 3, 4, 5 2
(C, E) 3, 4 3
(C, F) 3, 4, 5 4
(B, G) 29′ , 18, 26′ , 22 1
(B, H ) 29′ , 18, 26′ , 22, 7 2
(D, G) 20′ , 26′ , 22 3
(D, H ) 20′ , 26′ , 22, 7 4

λ-channel, is S = N δ/ H̄ = 8, a value that is achieved if and only if shortest path

routing is used and all LCs are loaded to saturation. (This corresponds to a maximum
normalized throughput per node of ρmax = 1.) Another possibility for embedding using
shared channels is considered in Section 7.4.
The approach we have taken – separating the design problem into logical- and physical-
layer design – may not always lead to a feasible physical-layer embedding. If it turns
out that the designed LT cannot be embedded on a given physical topology without
violating physical constraints, such as available wavelengths, available transceivers, and
path length limitations, all is not lost! It may be possible to modify the logical-layer
design by taking the physical-layer constraints that were violated into account. This
amounts to a “loose coupling” of logical- and physical-layer design problems. Following
this approach, several iterations of the two-step problem can be made with a goal of
“fitting” the LT into the PT, if, indeed, such a fit is possible.
An essential feature of any large network is that it be reconfigurable in response
to changing traffic conditions. As mentioned at the beginning of this chapter, LRNs
are especially flexible in this respect because they can be reconfigured both at the vir-
tual connection level and at the logical level. If the LCG of an LRN must be changed
while it is in operation to adapt to changing traffic, it is important to be able to ef-
fect the reconfiguration without disturbing the active connections. This is known as
hitless reconfiguration. An example of hitless reconfiguration of the LRN and the asso-
ciated problem of physical layer embedding (channel assignment) for an ATM network
embedded on a WRN infrastructure is analyzed in [Bala+96]. In that particular sce-
nario, changes in traffic patterns are known in advance and are assumed to change in
a cyclic order. Cyclic changes in logical topologies occur in a number of applications.
For example, in a network that interconnects financial institutions, traffic patterns and
intensities typically vary periodically on a daily basis. One LCG might be used in the
morning when the data traffic is relatively heavy, another in the early afternoon when
the traffic subsides, and yet another at night when the institutions exchange data for the
day’s transactions and the data traffic is very heavy. [Bala+96] presents some require-
ments on logical switch ports and optical connections so that a known cyclic pattern
of LCGs can be realized in this manner. It also includes some simple heuristics for
 Logically-Routed Networks 597

solving the embedding problem while attempting to minimize the number of wave-
lengths used.

7.3.3 Traffic Grooming in Point-to-Point Logical Topologies

Having examined various aspects of performance and network design for point-to-point
logical topologies, we are now in a position to revisit the grooming problem in these
networks from the point of view of network design and optimization.
The power of the grooming function depends on how it is integrated with the switching
function, and where it is located in the network. The simplest form is source grooming,
where subrate traffic streams (VCs) accessing the network at a common point are
multiplexed onto a higher rate LC and remain on that LC to their destination. In other
words, all VCs traverse the LT layer of the network in a single logical hop. In this case
any switching of the data in the core network is effected at intermediate nodes in the
physical layer along the underlying optical path of the corresponding LC, treating all
subrate streams as a common bundle. (The optical path may include O-E-O switches
operating at subrate speeds, but no grooming will take place at these intermediate
nodes.) Alternatively, full grooming nodes add more flexibility by allowing for multiple-
hop logical paths wherein grooming (adding and dropping of subrate connections) is
executed at intermediate nodes of the LP. If there is no direct LC with enough available
capacity for a traffic demand, or if it is inefficient to devote a single hop LC to it, the
traffic is routed on a multiple hop LP. At each intermediate LSN along the LP the traffic
hairpins into an LS performing both grooming and switching operations. As mentioned,
too much hairpinning is detrimental because it uses up the (expensive) resources of
the electronic switch fabrics in the LSs. However, some hairpinning is necessary to
efficiently fill the capacity of the LCs. In practice trade-offs are involved. Dense logical
topologies offering more opportunity for short logical paths require less hairpinning,
saving on grooming switch cost but resulting in poorer utilization of the capacity of the
LCs.
Figure 7.10 demonstrates the benefit of traffic grooming in a network. Four network
nodes are shown with three subrate traffic demands ([A,B], [A,C], and [A,D]). If no form
of grooming is allowed at node B, then three wavelengths will be required to satisfy
these demands (Figure 7.10[a]). (This would be the case if node B is equipped with an
OXC only.) If node B is now grooming capable (of the form of Figure 7.4), then two
wavelengths are sufficient, with one demand terminating on node B, another taking two
logical hops to node D, and the third optically bypassed to node C (Figure 7.10[b]).
Finally, if full grooming is used at node B, then one wavelength will be sufficient to
establish all three subrate traffic demands with the connections to nodes C and D taking
two logical hops (Figure 7.10[c]). Traffic grooming allows for better resource utilization
and shorter optical paths. The penalty is in terms of loss of transparency, increased
electronics cost, and increased latency.
Of course, multipexing/demultiplexing lower speed traffic onto/from higher speed
connections is a concept that is not new (e.g., aggregating lower speed SONET
598 Multiwavelength Optical Networks

A B D

(a) C

A B D

(b) C

A B D

(c)

Figure 7.10 Benefit of traffic grooming. (From [Cinkler03, Figure 5]. Copyright
c 2003 IEEE.
Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

signals into a higher-speed SONET signal). However, with the introduction of intel-
ligent switching nodes in WDM networks, traffic grooming is now used to automatically
provision end-to-end connections, a concept that is indeed new for telecommunication
networks.

7.3.3.1 Traffic Grooming in WDM Rings

Most of the early work on traffic grooming in WDM networks focused on SONET
over WDM rings [Gerstel+00, Modiano+98, Simmons+99]. In a traditional WDM
SONET ring, end systems at a ring node are interfaced to each wavelength on the ring
through a SONET ADM (Figure 7.11[a]). This is essentially a stripped-down version
of the O-E-O switch of Figure 4.87 with a single input and output. Each λ-channel is
demultiplexed/multiplexed and add/dropped to its own ADM, where it is converted to
an electrical signal that is demultiplexed/multiplexed and add/dropped at some subrate
of the λ-channel rate. For example, in the figure the rate on each wavelength on the
ring is OC-48, and this is demultiplexed in the ADM down to STS-1 and add/dropped
to end systems (not shown) at that rate. Note that some subrate channels may bypass
 Logically-Routed Networks 599

STS-1

STS-1
ADM LSN

ADM

WADM OXC

OC-48

(a) (b)

Figure 7.11 A node in a SONET over WDM ring.

the node, but they must continue on the same wavelength. Thus, there is no wavelength
interchange in these systems, in contrast to the WIXC functionality in the O-E-O optical
switch. (Wavelength interchange can be added by interconnecting subrate ports on
different ADMs via an electrical cross-connect.) Because of the finer granularity of their
operations, the SONET ADMs provide for grooming the subrate channels onto higher
rate λ-channels.
In the configuration of Figure 7.11(a), every wavelength on the ring requires its
own ADM at every node, which drives up cost as the wavelength count in the system
increases. A more economical approach uses optical bypass at the nodes as shown in
Figure 7.11(b), where only the λ-channels containing subrate connections that are to
be add/dropped are diverted (hairpinned) to an ADM, and the remaining wavelengths
bypass the node without further processing. The number of ADMs can be significantly
reduced in this case provided that an efficient grooming algorithm is used to confine
the add/dropped subrate connections to as small a number of wavelengths as possible at
each node. (In the case shown only one wavelength is add/dropped to an ADM.) This is
the objective of the work on grooming in SONET over WDM rings. Note that the con-
figuration in Figure 7.11(b) is of the form of an LSN/NAS combination overlaid on an
OXC.
When optimal grooming is the objective, the problem is usually cast in a static form.
In one form of the static problem, we seek an optimal assignment of a given traffic
distribution to a ring with nodes of the form of Figure 7.11(b), where the granularity
of the offered subrate traffic is some fraction of the bandwidth of a λ-channel. We
wish to assign the traffic (using grooming and routing and channel assignment [RCA]
techniques) to minimize the total network cost. The cost now may be a function of the
number of ADMs and the number of wavelengths. Although it would be desirable to
minimize both of these numbers simultaneously, this is generally impossible because
grooming and wavelength assignment are strongly related. Furthermore, minimizing the
600 Multiwavelength Optical Networks

combined cost is usually intractable for rings of any significant size. It is equivalent to
a joint logical layer (grooming) and physical layer (RCA) optimization problem. In the
case of even small mesh networks, optimization of the logical-layer design alone has
been found to be intractable, as noted in Section 7.3.1. Nevertheless, examples of small
rings have been found [Gerstel+99] in which joint optimization leads to up to 20%
reduction in network cost compared to the case where the two processes (grooming and
RCA) are performed sequentially.
Because of the difficulty of the ring grooming problem, various heuristics have been
proposed for static traffic on both unidirectional rings [Modiano+98] and bidirectional
rings [Simmons+99]. In each case the objective is minimization of ADMs. Both works
demonstrate that the number of ADMs can be significantly reduced (and even approach
the lower bound in the case of unidirectional rings) if effective grooming algorithms are
used. The results of applying the grooming heuristic of [Modiano+98] are illustrated in
Figure 7.12.
The results discussed above were based on uniform, distance-dependent or hub traf-
fic.11 In a more general approach [Zhang+00] proposes generic traffic grooming algo-
rithms for SONET/WDM rings that can be applied to both unidirectional and bidirec-
tional rings with arbitrary numbers of nodes and both uniform and nonuniform traffic.
In this case, grooming (minimizing ADMs) and RCA (minimizing wavelengths) are
treated as separate subproblems.
The RCA problem is solved by packing nonoverlapping subrate traffic demands
into circles. The first part of the algorithm constructs enough circles to include all the
connections in the traffic demand. If a minimum number of circles is used, this effectively
minimizes the required number of wavelengths, because every subrate connection in
a circle will be assigned the same wavelength, and subrate connections in distinct
circles can be groomed into a common wavelength until its capacity is reached. For
uniform demand (a fully connected logical topology) on a unidirectional ring, the optimal
approach is to combine each connection and its return connection on a single circle.
In the bidirectional ring case, the techniques discussed in Section 6.3.4.2 are optimal.
Other algorithms, including those applicable to nonuniform traffic, are discussed in
[Zhang+00].
The second part of the algorithm grooms circles into wavelengths. Circles are assigned
to each wavelength so as to bundle together those that have as many common end nodes12
as possible. This ensures that the total number of ADMs will be kept to a minimum.
The general approach is to distribute circles among wavelengths as evenly as possible,
thereby minimizing the number of wavelengths, and to utilize a good end node matching
algorithm to reduce the number of ADMs on each wavelength. The following greedy
algorithm produces good end node matching. It is performed for each wavelength until
all circles are accommodated.

11
Hub traffic is traffic that is destined to a single node (hub) of the ring.
12
The number of end nodes involved equals the number of ADMs needed for a specific wavelength.
 Logically-Routed Networks 601

Number of ADMs vs. nodes

160

No grooming
140

120
Number of ADMs

100

80 heuristic
algorithm
60

40 Lower bound

20

0
2 4 6 8 10 12 14 16 18
Number of nodes

Figure 7.12 Advantage of grooming static traffic in SONET over WDM rings. (From
[Modiano+01a, Figure 4]. Copyright
c 2001 IEEE. Used by permission of The Institute of
Electrical and Electronics Engineers, Inc.)

r Let m w be the number of circles to be assigned to wavelength λw .

r Find the ungroomed circle that contains the largest number of end nodes and groom
it onto wavelength λw .
r Groom another m w − 1 circles on λw by sequentially finding ungroomed circles
which if groomed on the same wavelength have maximum overlapping of the end
nodes.

Results presented in [Zhang+00] show that this two-step approach yields good per-
formance (in terms of minimizing the number of wavelengths and the number of
ADMs used) for both uniform and nonuniform traffic especially when the grooming
factor is large (i.e., a large number of subrate connections are groomed on a single
wavelength).
Grooming dynamic traffic in ring architectures has also received some attention from
researchers, as the importance of being able to efficiently accommodate changes in
the traffic demand has been recognized. [Berry+00] addresses the task of dynamically
changing traffic in SONET/WDM ring networks by formulating the problem in terms of
a bipartite graph and developing algorithms to minimize the number of wavelengths that
are processed at each node (which reduces the number of ADMs required at each node).
Dynamic traffic in rings has also been studied for architectures using a cross-connect
at one or more nodes on the ring [Gerstel+98, Modiano+98, Modiano+01b]. By switch-
ing traffic from one wavelength to another, the cross-connects add considerable flex-
ibility to the network, reducing the number of ADMs required and facilitating the
dynamic reconfiguration of traffic. [Gerstel+00] compares a number of different ring
602 Multiwavelength Optical Networks

Table 7.2 Relative costs for different ring networks.a

Asymptotic relative costs

W Q H
Wmin Q min Hmin

N
PPWDM 1 4
1
n
Single hub 4 2 2
n
Double hub 2 2 2
N 1 0.5
Hierarchical (α = 4
) 1 N 0.5 4
N
a
(From [Gerstel+98, Table 2]. Copyright
1998
c IEEE. Used
by permission of The Institute of Electrical and Electronics
Engineers, Inc.)

architectures with varying degrees of cross-connect capabilities, including a single or

double hub architecture, an architecture where all nodes have cross-connect capabil-
ities, and a hierarchical ring design. The objective in that work was to minimize the
overall network cost13 for static, dynamic, or incremental traffic patterns, when cross-
connect capabilities are introduced at the ring nodes. The single hub architecture is a
good choice when there are enough wavelengths on the ring and the goal is to min-
imize the transceiver cost. At the other end of the spectrum, an architecture where
all nodes have cross-connecting capability14 uses the wavelengths in the most efficient
manner, but it has maximum transceiver cost. The hierarchical ring design can allevi-
ate some of the transceiver cost, provided that there are spare wavelengths available.
Finally, the double hub architecture is usually a better choice for static traffic, as it
requires half as many wavelengths as the single hub case and has approximately the
same transceiver cost. Table 7.2 compares the asymptotic relative costs for different
ring architectures assuming uniform traffic with parameter g (there are g traffic streams
between the nodes, i.e., the traffic from node i to node j is g/c if c is the number of
traffic streams that can be supported in a lightpath) and N network nodes, where N
is large and a power of 2. The parameter α for the hierarchical ring design signifies
that the nodes in that architecture are arranged such that there are at most α − 1 “access
nodes” (hierarchy 2 nodes) between any consecutive pair of “backbone nodes” (hierarchy
1 nodes).

7.3.3.2 Traffic Grooming in WDM Mesh Networks

Although most of the original work on grooming focused on ring networks, more recent
work has been directed to the more general problem of traffic grooming in WDM mesh
networks. Much of that work poses the problem in much broader terms to include the

13
The network costs that are taken into account in [Gerstel+00] include the number of wavelengths W , the
transceiver cost Q, and the maximum number of hops of a lightpath H.
14
Termed in [Gerstel+00] a point-to-point WDM (PPWDM) ring.
 Logically-Routed Networks 603

complete network design and optimization problem from top (grooming) to bottom
(physical-layer design). In terms of the layered architecture of Figure 7.2, the general
problem can be viewed as the design of an LRN (the logical connections in layer LT
in the figure) to match the traffic demand (the virtual connections in layer VT) to
the physical topology (the optical nodes and links in layer PT). It includes the LCG
design and routing problem discussed in Section 7.3.1, coupled with the physical-layer
RCA problem described in Section 6.3.7, taking into account the granularity of traffic
demands and core network link speeds. A general ILP formulation for this problem is
presented below [Dutta+02]:

Definitions:
r bi j : Integer lightpath count from node i to node j.
r b (l, m): Number of such lightpaths that traverse the physical link from node l to node
ij
m.
r C (k) (l, m) = 1 if a lightpath from node l to node m uses λ over the physical link from
ij k
(k)
l to m, Ci j (l, m) = 0 otherwise.
r plm : Link indicator, indicating the presence of a link from l to m.
r C: Bandwidth of a single λ-channel.
r t (sd) : Traffic demand from node s to node d.
r T = [t (sd) ]: Traffic matrix.
r ti j : Aggregate traffic carried on lightpaths from node i to node j.
(sd)
r ti j : Amount of aggregate traffic carried on lightpaths from node i to node j that is
due to demand t (sd) .
r The domain of i, j, s, d, l, m is 0, 1, 2, . . . , N − 1, where N is the number of nodes
in the physical topology.
r The domain of k is 0, 1, 2, . . . , W − 1, where W is the number of wavelengths sup-
ported on each direction of a bidirectional physical link.

Various cost functions can be chosen as a basis for optimal design:

Minimize total number of lightpaths:

bi j (7.15)
i, j

or
minimize amount of electronic switching:
 (sd) 
ti j − t (sd) (7.16)
s,d,i, j s,d

or
minimize the maximum number of lightpaths at a node:
  
 
maxi max  b ji , bi j  (7.17)
j j
604 Multiwavelength Optical Networks

with
bi j (l, m) ≤ bi j plm , ∀i, j, l, m (7.18)
(k)
ci j (l, m) ≤ plm , ∀i, j, k, l, m (7.19)
 
N −1
 N −1
  bi j , for m = i 
bi j (m, l) − bi j (l, m) = −bi j , for m = j ∀m, i, j (7.20)
 
l=0 l=0 0, for m = i, m = j

bi j (l, m) ≤ W, ∀l, m (7.21)
i, j
W
 −1
(k)
ci j (l, m) = bi j (l, m), ∀i, j, l, m (7.22)
k=0
 (k)
ci j (l, m) ≤ 1, ∀k, l, m (7.23)
i, j
 
N −1
 N −1
  ≤ bi j , for m =i 
(k) (k)
ci j (m, l) − ci j (l, m) ≥ −bi j , for m = j ∀i, j, k, m (7.24)
 
l=0 l=0 = 0, for m = i, m = j
 (sd)
ti j = ti j , ∀i, j (7.25)
s,d

ti j ≤ bi j C, ∀i j (7.26)
 (sd) 
N −1
 N −1
  t , for i = s 
(sd) (sd)
ti j − t ji = −t (sd) , for i = d ∀m, i, j. (7.27)
 
j=0 j=0 0, for i = s, i = d
Among the three possible cost functions mentioned above, the total number of
lightpaths, and the maximum number of lightpaths that are terminating (originating)
at a node might apply to any LCG design problem, either as costs or as constraints.
(Note that a lightpath in this formulation is the same as a logical connection [LC] in the
logical-layer design problem of Section 7.3.1.) The second cost function captures the to-
tal amount of electronic switching in the network and indirectly the transceiver cost. It is
here that the cost of grooming appears. As in any constrained optimization problem, this
cost could alternatively be included as a constraint; e.g., the total number of lightpaths
might be minimized subject to a limitation on the total amount of switching equipment.
Equations (7.18) and (7.19) correspond to the physical topology constraints, Equations
(7.20) and (7.21) relate to lightpath routing constraints (continuity and wavelength lim-
itation, respectively), Equations (7.22), (7.23), and (7.24) are the constraints for the
lightpath RCA subproblem, and Equations (7.25), (7.26), and (7.27) are the constraints
for routing the traffic on lightpaths (grooming).
Because the traffic grooming problem is NP-complete, a number of heuristic ap-
proaches have been proposed to groom lower speed traffic streams onto higher speed
connections in arbitrary WDM mesh networks. [KZhu+02] proposes two simple heuris-
tic algorithms for the traffic grooming problem: maximizing single-hop traffic (MST)
 Logically-Routed Networks 605

and maximizing resource utilization (MRU). In MST the algorithm tries to establish
lightpaths (logical channels) that satisfy as much as possible the aggregate end-to-end
connection requests. It basically tries to pack different connections that have the same
end nodes in the same lightpath, maximizing as much as possible the number of con-
nections that are routed through a single logical hop. If there are not enough resources
in the network to establish lightpaths for all the connections, the algorithm attempts
to route the connections that were not accommodated using the spare capacity of the
currently established LCs. In MRU the algorithm first tries to establish lightpaths be-
tween end nodes with the maximum resource utilization values; i.e., those using the
most LC capacity. The leftover connections are again routed using the spare capacity of
the established lightpaths. Effectively the MRU algorithm uses the same approach as the
MST algorithm with the difference being that the end-node pairs and the connections not
accommodated are now sorted according to their resource utilization values. MRU tries
to utilize wavelengths more efficiently and is shown to perform better than MST in terms
of network throughput when tunable transceivers are used in the network. However, if
fixed transceivers are utilized, MST outperforms MRU.
A general approach to solving the dynamic traffic grooming problem utilizing a
generic graph model is presented in [HZhu+03]. This model consists of an auxiliary
graph and a dynamic traffic grooming algorithm. An example illustrating construction
of the auxiliary graph is shown in Figure 7.13.
Given a network represented by a directed graph G(V, E), where V is the set of
vertices (nodes in the network) and E is the set of directed edges (unidirectional fiber
links in the network), a layered auxiliary directed graph G ′ (V ′ , E ′ ) for that network, with
W wavelengths per link, is constructed in the following manner: The auxiliary graph will
have W + 2 layers, with W wavelength layers (one for each wavelength), the lightpath
layer (layer W + 1), and the access layer (layer W + 2). Figures 7.13(a) through (c)
show an example of the network graph G, its vertex set V , and the auxiliary graph G ′ ,
respectively. The access layer is where the traffic originates and terminates. Each node

Node 0
Node 1 Node 2
0 I 0
Access layer I 0 I 0

1 2 I 0
Lightpath layer I 0 I 0
(a)
I 0
I 0 I 0

0 0
I
I 0 I 0
1 2
(b)
(c)

Figure 7.13 Construction of an auxiliary graph for grooming. (From [HZhu+03, Figure 1]. Copy-
right
c 2003 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers,
Inc.)
606 Multiwavelength Optical Networks

in G is represented by two nodes in each layer of G ′ , corresponding to the input (I) and
output (O) port. Edges are then placed to interconnect these ports as follows:

r A wavelength bypass edge is placed from the input to the output port on every
wavelength layer for every node in the graph.
r A grooming edge is placed from the input to the output port on the access layer for
every node that has a grooming capability.
r A mux edge is placed from the output port on the access layer to the output port on
the lightpath layer for all nodes.
r A demux edge is placed from the input port on the lightpath layer to the input port on
the access layer for each node.
r A transmitter edge is placed from the output port on the access layer to the output port
on the wavelength layer for each wavelength that has transmitters available for every
node in the graph.
r A receiver edge is placed from the input port on a wavelength layer to the input port
on the access layer if there are receivers available for that wavelength for every node
in the graph.
r A converter edge is place from the input port on wavelength layer i to the output port
on wavelength layer j at node k if node k can convert wavelength i to wavelength j.
r A wavelength-link edge is placed from the output port on wavelength layer i at node
k to the input port on wavelength layer i at node j if there is a physical link from node
k to node j and wavelength i is available on that link. The wavelength edges define
the given physical topology.
r A lightpath edge is placed from the output port on the lightpath layer at node i to
the input port on the lightpath layer at node j if there is a lightpath from node i to
node j. The lightpath edges represent the edges (LCs) of a logical connection graph
constructed during the solution of the design and grooming problem.

The auxiliary graph is the framework for a class of algorithms that can be used for the
combined logical-layer design, physical-layer design, and grooming problem under both
static and dynamic traffic demand conditions. All of the edges in the above list except
the lightpath edge represent resources in the physical layer (e.g., links, wavelengths, and
transmitters). Each lightpath edge represents a logical connection set up on a physical
path through the network during the course of the design and grooming algorithm.
During problem solution, new lightpath edges are added one by one, so a complete LRN
is formed when the algorithm terminates, the traffic demands are routed and groomed
onto lightpaths in the LRN, and the lightpaths are assigned wavelengths and routed on
the given physical topology.
Every edge in the auxiliary graph is assigned a capacity c and a weight w. The weight
can place emphasis on different costs, such as number of transceivers or wavelengths.
The capacity parameter c indicates unused capacity of the various network resources
and is modified as the algorithm progresses. For dynamic traffic grooming, each new
connection request is routed on the shortest path in the auxiliary graph, based on the
assigned weights. The way the weights are assigned to the edges in the auxiliary graph
 Logically-Routed Networks 607

will determine how a connection is carried in the network. Four different grooming
policies are analyzed in [HZhu+03]. These are:

r Minimize the number of hops for each new demand, starting with the current LCG.
This is accomplished by attempting to groom a new demand onto an existing lightpath
directly connecting the source and destination nodes of the new demand (single-hop
approach). If this is not possible, try to find a new lightpath joining that source-
destination pair (new lightpath approach). If this is also not possible try routing
the demand through multiple existing lightpaths (multihop approach) or through a
combination of existing lightpaths and at least one new lightpath chosen to complete
the source-destination connection (hybrid multihop approach).
r Minimize the number of traffic hops on the physical topology. This is accomplished by
comparing the number of wavelength-links used in the four approaches discussed in
the previous step and choosing the one with the minimum number of wavelength-links.
r Minimize the number of lightpaths by trying to set up as few new lightpaths as possible.
To achieve this the single-hop, multihop, and hybrid multihop approaches are tried
sequentially before the new lightpath approach is attempted.
r Minimize the number of wavelength-links by trying to minimize the number of extra
wavelength-links that are occupied by each new demand. In this case, if single and
multihop approaches fail, the algorithm compares the hybrid multihop and the new
lightpath approach and picks the one that uses the least number of wavelength-links.

Different grooming policies result from applying different weight-assignment func-

tions in the auxiliary graph. A comparison of the different grooming policies in
[HZhu+03] shows that each is adapted to different network configurations. For ex-
ample, when more transceivers are available in the network the policy that minimizes
the number of hops on the physical topology performs the best, as it utilizes wavelength-
links more efficiently than the other policies. The reader is referred to the literature (e.g.,
[Hadjiantonis+04, Khalil+06]), for an extensive analysis of a number of other grooming
policies and heuristics.

7.4 Multipoint Logical Topologies: Hypernets

The hypernet paradigm was introduced through several examples in Sections 3.4.3 and
3.5.2 – see Figures 3.30 and 3.36(a). A hypernet is a generalization of a multihop LRN,
with point-to-point connections replaced by multipoint connections. The former can be
represented by an LCG and the latter by an LCH (see Appendix A for hypergraph defi-
nitions and terminology). The hypernet, together with its optical network infrastructure,
is similar to the multihop network, as shown in the layered view of Figure 7.14. A virtual
topology is supported by a logical topology, which is now in the form of a hypergraph,
and the LT is in turn supported by the physical topology. The basic transport element in
the hypernet is a multipoint LC, which may join together any number of network nodes.
(Three multipoint LCs are shown in Figure 7.14.)
608 Multiwavelength Optical Networks

VT

LT

PT

Figure 7.14 Layered view of a hypernet.

Multipoint networks and hypergraphs have been studied extensively as models for
bus interconnection networks for parallel processors [Bermond+83, Bermond+86,
Bermond+96, Birk86]. For applications to optical networks see [Birk91, Dowd92,
Jiang+93, Jiang+95, Jiang95, Ofek87]. Hypergraphs have been applied in many other
fields as well. (See, for example, [Shi92] for an application to integrated circuit layout.)
In this section we study the basic structure and performance of networks with hypergraph
logical topologies.
We define an undirected (directed) hypernet to be an LRN with connections that can
be represented by an undirected (directed) LCH H (ν, ε). Figures 7.15(a)–7.15(c) show
LCHs that represent undirected hypernets of orders 7, 12, and 24, respectively. The
vertices of the LCH represent LSNs, and the hyperedges (hyperarcs in the directed case)
represent the LCs, which are carried on multipoint subnets (MPSs).15 The MPSs support
multipoint logical connectivity and channel sharing through multipoint optical paths.
This necessitates the generalized switching capabilities of a linear lightwave network
(LLN). Thus it will be assumed henceforth that the underlying optical infrastructure for
a hypernet is an LLN.
In the undirected case, a hyperedge E = {a, b, . . .} of size r = |E| represents an
undirected MPS containing r LSNs and providing full logical connectivity among all

15
As indicated earlier, we classify the nodes of an LRN as either LTNs, which belong to a single MPS, or
LSNs, which belong to at least two MPSs. Because LTNs are “dead ends” that do not participate in the
routing functions of the network, we shall normally exclude them from the discussion that follows. (In the
illustrations of Figure 7.15, all nodes are LSNs.)
 Logically-Routed Networks 609

(a) (b)
3 2
r 3 r 4
D 1 D 2
N 7 N 12

210

1210 2101
121 101
012
0210 2102

1212 0121 1012 0101

2121 1010
212 010
120 201
2120 0120 2012 2010
1201
0212 0201 0102
021 1202

2020
102
2021 0202 1020

202 020

1021

(c)
2
r 4
D 3
N 24

Figure 7.15 Hypernets.

610 Multiwavelength Optical Networks

nodes in the subnet. Any node in the subnet is reachable from any other in one logical
hop. A hyperedge of size r = 2 therefore represents an ordinary bidirectional point-
to-point link, and a hyperedge of size r provides a logical connectivity equivalent to
r (r − 1) point-to-point connections sharing the capacity of a common medium. A node
of degree  transmits to and receives from  MPSs (hyperedges). In the hypernets of
Figures 7.15(a) through 7.15(c), the hyperedges are of sizes 3, 4, and 4, respectively, and
the nodes are of degrees 3, 2, and 2, respectively.
In the directed case, a hyperarc E = {E − , E + } represents a directed MPS, wherein the
receiving set of the MPS, R = {R1 , R2 , . . .}, corresponds to E + (the out-set of E) and is
reachable in a single logical hop from any node in the transmitting set, T = {T1 , T2 , . . .}
(the in-set, E − , of E). The in-size, s − = |T |, of E indicates the number of nodes
transmitting into E and the out-size, s + = |R|, of E indicates the number of nodes
receiving from E. A hyperarc of in-/out-size 1 represents a single (unidirectional) LC,
and a general hyperarc is equivalent to |T ||R| point-to-point connections from the
transmitting set to the receiving set, sharing the capacity of a common MPS.
A directed MPS with |T | = M and |R| = N is called an M × N MPS. A node of
out-degree d + transmits into d + MPSs (hyperarcs), and a node of in-degree d − receives
from d − MPSs. (In the development that follows, we shall often use the terms MPS and
hyperedge/hyperarc interchangeably.)
To illustrate, consider the undirected hypernet of Figure 7.15(c) with nodes and
hyperedges denoted by labels of lengths 4 and 3, respectively. It can be converted to
a directed hypernet by assigning directions of transmission to each node (see arrows
assigned in Figure 7.20). For example, MPS 101 has transmitting set T = {2101, 0101}
and receiving set R = {1012, 1010}. Each hyperarc is of in-/out-size 2 and each node
of in-/out-degree 1.

The Underlying Graph of a Hypergraph

The connectivity in any hypergraph can also be achieved in an equivalent graph. Thus,
an undirected hypergraph H (ν, ε) can be replaced by a graph G(ν, ε∗ ), in which each
hyperedge E ∈ ε of size r is replaced by r (r − 1)/2 edges in ε∗ , providing full connec-
tivity among the vertices of E. The equivalent graph, G(ν, ε∗ ), is called the underlying
graph of H (ν, ε). Similarly, a directed hypergraph can be replaced by an underlying
digraph, in which each hyperarc of in-size s − and out-size s + is replaced by s − × s +
arcs in the digraph, providing full bipartite connectivity between the in-set and out-set
of that hyperarc.
A multihop LRN with logical connection graph G(ν, ε∗ ) therefore has connectiv-
ity equivalent to a hypernet with LCH H (ν, ε) whenever G(ν, ε∗ ) is the underlying
graph/digraph for the hypernet. The difference between the two is in the way capacity
is allocated to the LCs. In the point-to-point case, each connection has a fixed capacity
dedicated to it, determined by the bandwidth of its supporting optical channel and the
constraints of the optical transceivers. In the hypernet, each MPS has a total capacity
determined by the aggregate optical spectrum allocated to it, as well as the transceiver
constraints. This capacity is shared among all LCs in the MPS. This difference between
 Logically-Routed Networks 611

dedicated channels and shared channels is an important consideration in the development

that follows.

7.4.1 Capacity of a Multipoint Subnet

To study throughput and routing issues in hypernets, it is important to have a measure of
the capacity of an MPS. Because of the ability of an MPS to share its channels among
several LCs, the definition of the capacity of an MPS is more complex than it is for
point-to-point links, depending on the optical spectrum allocated to the MPS as well as
the number of stations and transceivers accessing the MPS.
Each LSN in the hypernet accesses two or more MPSs through an access station,
which may contain several optical transmitters/receivers. For simplicity, we assume here
that each transmitter/receiver is capable of tuning over all optical channels available in
its MPSs and that all transceivers operate at a common bit rate R t . Furthermore, we
assume that the optical spectrum allocated to an MPS consists of one waveband. It may
be a thin waveband consisting of a single λ-channel or a thick waveband consisting of
several λ-channels.
The physical mechanism of channel sharing is irrelevant at the logical level. How-
ever, the total available capacity in the MPS as well as any constraints on how it is
shared do have an effect on the traffic-handling ability of the MPS. To define the key
channel-sharing parameters for an MPS, we adopt the general model of a shared medium
studied in Section 5.3. A directed M × N MPS is assumed to have the channel-sharing
capabilities of a general shared medium with M transmitting and N receiving stations.
In an undirected MPS, M = N and a pair of like-numbered transmitting and receiving
stations represent the transmitting and receiving sides of one node in the MPS.
We define the capacity of an MPS as the maximum possible value of the aggregate
traffic that the MPS can carry. Recall from Section 5.3 (see Equation [5.52]) that the
maximum value of the aggregate traffic, γ̄ , is

γ̄max = η R t C, (7.28)

where C is the number of λ-channels contained in the supporting waveband, and η,

defined in Equation (5.51), is the traffic balance factor for the traffic distribution within
the MPS. The parameter η reflects the “match” between the traffic distribution and the
number of transmitters and receivers in the access stations. The traffic-bearing capability
of an MPS therefore depends not just on the aggregate traffic but also on its distribution.
This dependence is to be expected because of the constraints imposed by the number
and speed of the transceivers.
It was shown in Chapter 5 that, given a traffic distribution with aggregate traffic γ̄ and
balance factor η, a channel-sharing schedule can always be found, irrespective of the spe-
cific multiplexing/multiple-access techniques being used, that accommodates the given
traffic distribution using a bit rate R t arbitrarily close to γ̄ /ηC (see Equations [5.61] and
[5.62]). This suggests using η R t C as the capacity of the MPS. However, it is preferable
612 Multiwavelength Optical Networks

to eliminate the balance factor from the definition of capacity. There are two additional
observations that allow us to do so:
1. η = 1 when C = 1 independent of the traffic distribution.16
2. For any given traffic distribution and any value of C, η can be made equal to unity
by providing a sufficient number of transmitters/receivers in each station.
Another way of stating these observations is that if an MPS is carried on a thin
waveband, it is the channel that limits its capacity, whereas for a thick waveband the
transceiver constraints limit the capacity. In light of this, it is clear that an MPS with
a sufficient number of transmitters/receivers in its access stations can accommodate
any traffic distribution provided that γ̄ < R t C. Therefore, we shall henceforth assume,
without any loss of generality, that η = 1 for each MPS in our LRNs. In this way each
MPS can be considered to be a fully shared medium with capacity R t C.17
If the MPSs supporting an LRN are not fully shared, then the network loses its
hypernet characteristics. For example, suppose each MPS in a hypernet is operated on a
thin waveband using TDM/TDMA on a single λ-channel. Suppose further that each LC
between a pair of nodes in this MPS is allocated a fixed capacity by assigning it one slot
in a fixed-frame system. In this case, when viewed at the logical level, the connectivity
of the LRN is point-to-point rather than multipoint.
We find, therefore, that there are subtle differences in defining logical connectivity
for LRNs supported by multipoint optical connections, depending on the way in which
capacity is allocated. An LLN has the capability to create multipoint, shared-channel
logical connectivity, in which case it provides support for a hypernet. However, if that
capacity is channelized to create fixed-capacity point-to-point connections, the resultant
logical topology is more properly described as a graph rather than a hypergraph. By
channelizing in this way, a hypernet is reduced to a point-to-point network with a logical
connectivity that is described by the underlying graph for the hypernet.
Another complicating issue is the matter of optical resource usage. Typically, a point-
to-point LC in a multihop network would be supported by a dedicated λ-channel, requir-
ing a dedicated optical transmitter/receiver pair. However, in a multihop network realized
over an LLN, a single optical transmitter or receiver might support several multiplexed
connections, each using only a portion of a full λ-channel. If the multiplexed connections
are allocated fixed capacities, this configuration still behaves like a point-to-point LRN,
but if dynamic capacity allocation is used (e.g., by packet switching in the optical layer),
we have a shared-channel system that behaves more like a hypernet.18 The conclusion
is that in an LRN with an underlying optical connectivity that is multipoint, there are
many options for assigning capacity to LCs. Thus, the distinctions between fixed- and
shared-capacity logical connectivity are blurred.
16
With C = 1, all stations accessing an MPS can share a single channel using TDM/TDMA. In that case,
a single transceiver in each station is sufficient to share the channel fully, with η = 1 and R t C = γ̄
irrespective of the traffic distribution.
17
In cases in which the assumptions leading to this definition are violated, so η < 1, the capacity must be
reduced to an effective value of η R t C.
18
In the literature these cases are often referred to as shared-channel multihop networks (see, for example,
[Hluchyj+91, Kovacevic+95]).
 Logically-Routed Networks 613

7.4.2 Families of Dense Hypernets

The performance benefits of dense point-to-point logical topologies that were listed in
Section 7.2.2 are even more evident in hypernets, because hypergraphs can be made far
more dense than graphs. Therefore, it is of interest to identify hypergraphs with large
orders for given diameter D, maximum degree, and hyperedge size. For this purpose it is
useful to have bounds on orders of hypergraphs similar to the Moore bounds for graphs.
We call an undirected hypergraph of diameter D, maximum vertex degree , and
maximum hyperedge size r a (, D, r )-hypergraph and denote its order by N (, D, r ).
A bound on the maximum number of vertices in a (, D, r )-hypergraph is
D−1

N (, D, r ) ≤ 1 + (r − 1) ( − 1)i (r − 1)i . (7.29)
i=0

Equation (7.29) is known as the Moore bound for undirected hypergraphs, and the
hypergraphs that attain it are known as Moore hypergraphs.19 Note that this reduces to
the Moore bound for graphs (Equation [A.3]) when r = 2 and that the potential density
of a hypergraph increases rapidly with hyperedge size r . Figure 7.16(a) illustrates the
fan-out from a vertex in an undirected hypergraph, leading to the Moore bound.
A directed hypergraph of diameter D, maximum vertex out-degree d, and maximum
out-size s is called a (d, D, s)-directed hypergraph, and its order is denoted by n(d, D, s).
A bound on the maximum number of vertices in a (d, D, s)-directed hypergraph is given
by
D

n(d, D, s) ≤ (ds)i . (7.30)
i=0

s=2

r=5
∆=3 d=3

(a) Undirected (b) Directed

Figure 7.16 Illustrating fan-out in hypergraphs.

19
It is known [Bermond+96] that for D > 2 (with the exception of rings of odd order [i.e., r = 2]), Moore
hypergraphs cannot exist.
614 Multiwavelength Optical Networks

Equation (7.30) is the Moore bound for directed hypergraphs, and the hypergraphs
that attain it are known as directed Moore hypergraphs.20 Note that Equation (7.30)
reduces to the Moore bound for digraphs (Equation [A.3]) when d = δ and s = 1.
Figure 7.16(b) illustrates the fan-out from a vertex in a directed hypergraph, leading to
the Moore bound.
Dense hypergraphs have interested both the mathematical and engineering commu-
nities for some time. A survey of results up to 1983 can be found in [Bermond+83].
Hypergraphs can be constructed directly, or indirectly via a related graph. Two tools for
construction of hypergraphs from graphs that are used here are (1) edge grouping and
(2) duality (see Section 7.4.3.1).
Edge grouping is a simple way of deriving a directed hypergraph from a digraph.
The arc set of the digraph is grouped into dicliques (see Appendix A for a definition),
and each diclique is replaced by a hyperarc in the derived hypergraph. For example,
the shuffle hypernet with the LCH shown in Figure 7.17 was derived by grouping the

0 0

1 1

2 2

3 E1 3

4 4

5 5

6 E2 6

7 7

8 8
E3
9 9

10 10
E4
11 11

12 12
E5
13 13

14 14
E6
15 15

16 16

17 17

Figure 7.17 Shuffle hypernet.

20
It is known [Bermond+96] that directed Moore hypergraphs cannot exist for ds > 1 or D > 1.
 Logically-Routed Networks 615

54 arcs of the 18-node (3, 2)-ShuffleNet of Figure 7.5 into 6 dicliques, resulting in a
(1, 3, 3)-directed hypernet. In this transformation, the in-/out-degree of each node has
been reduced from 3 in the ShuffleNet to 1 in the directed hypergraph, with a consequent
3:1 reduction in optical hardware requirements in the nodes.

7.4.3 Kautz Hypernets

In Section 7.2.2 we showed that Kautz and deBruijn graphs are examples of dense
families of regular graphs. It is also possible to define families of Kautz and deBruijn
hypergraphs and their generalizations, which have excellent topological properties, in-
cluding very high density and good survivability. These hypergraphs also retain the
simple addressing and self-routing properties of the corresponding graphs. We focus
here on Kautz and generalized Kautz hypergraphs as the LCHs for Kautz hypernets.
Although the emphasis is on basic performance indices such as throughput and process-
ing load in the switches, it is important to note that hypernets with a high degree of
symmetry, such as the Kautz families, also possess other useful properties such as even
load distribution, ease of routing and multicasting, scalability, and fault tolerance (see
[Jiang95]).
The families of graphs and hypergraphs that are of interest here are
r K (d, D): The Kautz digraph, a regular digraph of in-/out-degree d and diameter D,
defined for arbitrary positive integers d and D.
r G K H (d, n, s, m): The generalized Kautz hypergraph, a directed hypergraph of out-
degree d, order n, out-size s, and number of hyperarcs m, where d, n, s, and m are
positive integers.
r D K H (d, D, s): The directed Kautz hypergraph, a regular directed hypergraph of in-/
out-degree d, diameter D, and in-/out-size s, where d, D, and s are arbitrary positive
integers.
r K H (, D, r ): The Kautz hypergraph, a regular hypergraph of degree , diameter D,
and hyperedge size r , where  and r are arbitrary even positive integers.

Table 7.3 summarizes the key parameters for the corresponding Kautz hypernets.
(In the case of directed hypernets, the node degree refers to the in-/out-degree.) As
will be explained in Section 7.4.3.3, K H (, D ∗ , r ) is derived from G K H (d, n, s, m)
for special values of its arguments. The parameter D and the number of hyperedges

Table 7.3 Kautz hypergraphs.

Node Nodes Diameter Node label No. of No. of

Network degree in MPS D∗ length nodes MPSs

K H (2, D ∗ , r ) 2 r D∗ ≤ D D ( r2 ) D + ( r2 ) D−1 ( r2 ) D−1 + ( r2 ) D−2

D K H (1, D, s) 1 2s D D s D + s D−1 s D−1 + s D−2
K H (, D ∗ , r )  r D∗ ≤ D — ( r4 ) D + ( r4 ) D−1 —
∗
G K H (d, n, s, m) d 2s D ≤ ⌈logds n⌉ — n m = dn/s
D K H (d, D, s) d 2s D — (ds) D + (ds) D−1 —
616 Multiwavelength Optical Networks

N
10
0

D=6 D=5 D=4

0
5

r
2 4 6 8 1 0 1 2

Figure 7.18 Orders of K H (2, D, r ).

in K H (, D ∗ , r ) are determined by the parameters of the corresponding generalized

Kautz hypergraph. The table entry for G K H (d, n, s, m) represents the special case of
symmetric G K H s (see Section 7.4.3.3).
Note that these networks can be “grown” to large orders by increasing the size of
the hyperedges (MPSs) while maintaining a small number of optical transceivers per
node (the nodal degree) and a small logical diameter. An illustration of the effect of
hyperedge size on order is shown in Figure 7.18, in which the order of K H (2, D, r ) is
plotted against r with D as a parameter. For comparison, the diameters D = (2k − 1)
of (2, k)-ShuffleNets must be much larger to produce networks of the same order as
K H (2, D, r ). For example, the Kautz hypergraph K H (2, 4, 8) has an order of 1280 with
a diameter of 4, whereas the (2, 7)-ShuffleNet has an order of 896 with a diameter of 13.
Each of the Kautz families can be defined in various ways. In particular, D K H (1, D, s)
and K H (2, D, r ) are derived conveniently from K (d, D) by duality. The more general
Kautz hypergraphs can be constructed from G K H (d, n, s, m) by selecting special values
of the parameters d, n, s, and m. These relations are studied in the following sections.

7.4.3.1 From Graphs to Hypergraphs: Duality

The dual of a hypergraph H is a hypergraph H ∗ with vertices that correspond to the
hyperedges of H and with hyperedges that correspond to the vertices of H . A vertex e∗j
of H ∗ , corresponding to a hyperedge E j of H , is a member of a hyperedge V j∗ in H ∗ if
and only if the vertex v j in H , corresponding to V j∗ , is a member of E j in H .
Bermond, Bond, and Peyrat [Bermond+86] observed the following relationship be-
tween a hypergraph and its dual:

Proposition: If H is a (, D, r )-hypergraph, then its dual hypergraph H ∗ is an

(r, D ∗ , )-hypergraph where D − 1 ≤ D ∗ ≤ D + 1.

In particular, if G is a graph (or multigraph) of maximum degree r and diameter D,

then its dual is a (2, D ∗ , r )-hypergraph.
 Logically-Routed Networks 617

K H (2, D, r ) by Duality
The duality relations between Kautz hypergraphs and Kautz digraphs work as follows.
By removing arc orientations, the Kautz digraph K (r/2, D − 1) becomes a multigraph of
degree r with a dual that is K H (2, D ∗ , r ), with D ∗ ≤ D. The hypergraph K H (2, D ∗ , r )
is of order (r/2) D + (r/2) D−1 with a degree of 2, whereas the digraph K (r/2, D − 1) is
of order (r/2) D−1 + (r/2) D−2 with a vertex in-/out-degree of r/2. Thus we have gained
a factor of r/2 in number of vertices while saving a factor of r/4 in degree. Although
duality only produces hypernets of degree 2, this special case has considerable practical
importance. It is economical in usage of optical transceivers (only two per LSN), and
despite the limited number of transceivers these hypernets can still be grown to large
sizes by increasing the size, r , of the MPSs.
Figure 7.19 illustrates the duality relation between the hypergraph K H (2, 3, 4) and the
digraph K (2, 3). The latter is first converted to a regular multigraph with degree  = 4
by ignoring the orientation of the arcs. The dual of the multigraph is K H (2, 3, 4).
Note the labeling convention in Figure 7.19. Each hyperedge in K H (2, 3, 4) inherits
its label, a word of length 3, from the dual vertex in K (2, 3). Similarly, each vertex in
K H (2, 3, 4) derives its label, a word of length 4, from its dual arc in K (2, 3), using the
shortest concatenation of the labels of the vertices defining the head and tail ends of
the arc. For example, vertex 1012 in K H (2, 3, 4) obtained its label from the shortest
concatenation of vertices 101 and 012.
The Kautz hypergraphs K H (2, D ∗ , r ) can also be defined directly using alphabets.
The vertex labels are words of length D with no two identical consecutive letters,
constructed on an alphabet A of (r/2 + 1) letters. The hyperedge labels are the words
of length D − 1 with no two identical consecutive letters. The vertex (a1 , . . . , a D )
belongs to two hyperedges: (a1 , . . . , a D−1 ) and (a2 , . . . , a D ). Each hyperedge E =
(b1 , . . . , b D−1 ) contains r vertices of the form (∗, b1 , . . . , b D−1 ) and (b1 , . . . , b D−1 , ∗).

D K H (1, D, s) by Duality
The duality construction described earlier for undirected hypergraphs is extended easily
to the directed case. Starting with a Kautz digraph K (s, D − 1), we obtain a dual directed
Kautz hypergraph D K H (1, D, s) of order s D + s D−1 , gaining a factor of s in number
of nodes while saving a factor of s in nodal degree.
Figure 7.20 illustrates the construction based on K (2, 3) and yielding D K H (1, 4, 2).
The labeling convention for the undirected case is followed here as well. The assignment
of hypergraph vertices to hyperarc in- or out-sets is inherited from the orientations of
their dual arcs in the digraph and is indicated by arrows pointing in or out of each
hyperarc. For example, the vertices 2101, 0101, 1012, and 1010 in the hyperedge E 101
−
are duals of the four arcs incident on vertex 101 in the digraph. The first two are in E 101
because their dual arcs are incident to vertex 101 in the digraph, and the last two are in
+
E 101 because their dual arcs are incident from vertex 101. This construction also suggests
a multistar realization of D K H (1, D, s) based on K (s, D − 1)21 (see Section 7.5.4).

21
The hypergraph D K H (1, D, s) can also be defined directly, using alphabets, following the same general
procedure as for K H (2, D ∗ , r ).
618 Multiwavelength Optical Networks

210

121 101

012

212 010

120 201

021 202 020 102

K(2, 3)
Duality
210

1210 2101

121 101
012
0210 2102
1212 0121 1012 0101

2121 1010
212 010
120 201
2120 2010
0120 2012
1201
0212 0102
021 1202 0201

2020
102
2021 0202 1020

202 020

1021

KH(2, 3, 4)

Figure 7.19 Duality construction.

 Logically-Routed Networks 619

210

121 101

012

212 010

120 201

021 202 020 102

K(2, 3)
Duality
210

1210 2101

121 101
012
2102
0210

1212 0121 1012 0101

2121 1010
212 010
120 201
2120 2010
0120 2012
1201
0212 0102
021 1202 0201

2020
102
0202
2021 1020

202 020

1021

D K H(1, 4, 2)

Figure 7.20 Directed hypergraph construction via duality.

620 Multiwavelength Optical Networks

210

121 101

012

212 010

120 201 Edge Grouping

021 202 020 102

K(2, 3) 201

101

121
012 E01

010
212

120 201

202 020

021 102
D K H(1, 3, 2)

Figure 7.21 Directed hypergraph construction via edge grouping.

7.4.3.2 Edge Grouping

The duality construction of Section 7.4.3.1 is limited to undirected hypergraphs of
degree 2 and directed hypergraphs of out-degree 1. For more general cases, we must
turn to other techniques. One that is not limited in degree is edge grouping. This
technique can be used to construct a directed Kautz hypergraph from an underlying
Kautz or generalized Kautz digraph. Using edge grouping, the vertex set of the derived
hypergraph is the same as that of the underlying digraph.
 Logically-Routed Networks 621

The edge grouping construction was illustrated for ShuffleNets in Section 7.4.2.
Figure 7.21 shows an application of edge grouping to Kautz hypergraphs. The edges
of the digraph K (2, 3) are grouped into dicliques, with each diclique replaced by a
hyperarc in the resultant hypergraph D K H (1, 3, 2). For example, in the diclique with
the shaded vertices, the four arcs – 201 → 010, 201 → 012, 101 → 012, and 101 →
− +
010 – correspond to the hyperarc E 01 , with in-set E 01 = {201, 101} and out-set E 01 =
{010, 012}. Although this example is of degree 1, edge grouping also yields hypergraphs
of higher degrees.

7.4.3.3 Generalized Kautz Hypergraphs

The generalized Kautz hypergraphs G K H (d, n, s, m) were defined in Section 6.5.7.
The objective there was to find dense LCHs of diameter 1. These were implemented as
purely optical multistar networks (LLNs) configured to provide connectivity among all
pairs of nodes in one logical hop. We use the same family of hypergraphs here in the
more general context of LRNs. The definition is repeated here for convenience.
Let n be the number of vertices and d be the vertex out-degree. Choose the number
of hyperarcs m and the hyperarc out-size s such that
dn ≡ 0 (mod m)
(7.31)
sm ≡ 0 (mod n).

The vertices are labeled as integers modulo n and the hyperarcs are labeled as integers
modulo m. The incidence rules are as follows. Vertex v is incident to the hyperarcs

e ≡ dv + α (mod m), 0≤α<d (7.32)

and the out-set of the hyperarc e consists of the vertices

u ≡ − se − β (mod n), 1 ≤ β ≤ s. (7.33)

The diameter of G K H (d, n, s, m) is at most ⌈logds n⌉. If n = (ds) D + (ds) D−k for a
positive odd integer k, then the diameter is D. Thus for k = 1 and D = 2, the orders of
these hypergraphs come within 1 of the Moore bound.
In general the in- and out-size of a hyperarc of G K H (d, n, s, m) will differ, as will the
in- and out-degree of a vertex. However, there is a symmetric case, which is of particular
interest here: The in- and out-size of all hyperarcs equals s, and the in- and out-degree
of all vertices equals d if and only if dn = sm.
To simplify the discussion that follows, we focus on the symmetric case of G K H ,
assuming henceforth that dn = sm. Adding the condition that n = (ds) D + (ds) D−1 ,
G K H reduces to the directed Kautz hypergraph D K H (d, D, s), which has the max-
imum possible order for its diameter D. (Recall that D K H [1, D, s] was derived in
Section 7.4.3.1 using duality.)
To illustrate, the tripartite representation of G K H (2, 42, 3, 28) is shown in
Figure 7.22. Because it obeys the previous symmetry conditions and has n = (ds)2 + ds,
this is also D K H (2, 2, 3). Its order is 42 and the Moore bound for these parameters is
43. Note how it forms a 28-fold multistar structure. Each star provides full connectivity
622 Multiwavelength Optical Networks

Vertices Hyperarcs Vertices

0 0 0
1 1
1
2 2
3 2 3
4 4
3
5 5
6 4 6
7 7
5 8
8
9 6 9
10 10
11 7 11
12 8 12
13 13
14 9 14
15 10 15
16 16
17 11 17
18 12 18
19 19
20 13 20
21 14 21
22 22
23 15 23
24 16 24
25 25
26 17 26
27 18 27
28 28
29 19 29
30 20 30
31 31
32 21 32
33 22 33
34 34
35 23 35
36 36
24
37 37
38 25 38
39 39
26
40 40
41 27 41
Connections from Vertices to Hyperarcs Connections from Hyperarcs to Vertices
Note: All edges are directed from left to right.

Figure 7.22 Tripartite representation of G K H (2, 42, 3, 28). (From [Bermond+97, Figure 5]. Copy-
right
1997.
c Reprinted by permission of John Wiley & Sons, Inc.)

between the three vertices in its in-set and the three vertices in its out-set, corresponding
to nine LCs, or a 3 × 3 diclique in a digraph. This suggests an equivalence between the
generalized Kautz hypergraph and an underlying digraph. In fact, the following relation
exists: The underlying digraph of D K H (d, D, s) is K (ds, D).
 Logically-Routed Networks 623

Thus, each directed Kautz hypergraph is functionally equivalent to a Kautz digraph

of the same order, with degree ds and the same diameter. This relation has important
practical implications, showing more clearly how the hypernet approach economizes
on optical hardware. The hypernet based on D K H (d, D, s) can be realized using d
transceivers in each LSN. Each transmitting node multicasts optically to s receiving
nodes in the same MPS, and each receiving node is accessed by s transmitting nodes.
Now suppose the same logical connectivity is realized as a point-to-point multihop
network using the underlying digraph K (ds, D) as its LCG. This would require ds
transceivers in each node, for a factor of s increase in transceiver cost. (The increased
number of transceivers in the point-to-point case carries with it the potential for higher
throughput, provided that the necessary optical spectrum is available.)
By ignoring the orientation of the hyperarcs in G K H (d, n, s, m) (i.e., by replacing
each hyperarc E with a hyperedge with vertex set E − ∪ E + ), the directed hypergraph
can be converted to an undirected hypergraph. In the important special case when
dn = sm and n = (ds) D + (ds) D−1 , G K H (d, n, s, m) is converted to an undirected
Kautz hypergraph, K H (, D ∗ , r ), with degree  = 2d and edge size r = 2s. The
diameter is D ∗ ≤ D. Thus, the order of the undirected version can be expressed as
N = (r /4) D + (r /4) D−1 . For example, G K H (2, 42, 3, 28) has the aforementioned
properties, with D = 2, and its undirected version is K H (4, 2, 6).
The Kautz hypergraph family provides a wide range of design alternatives for re-
alizing large and dense hypernets with a variety of parameter values and hardware
requirements. It is also worth mentioning in passing that these represent but one ex-
ample of a dense family that can be grown to very large sizes with modest hardware
requirements. Even in the special case of undirected hypergraphs of degree 2, there
are many other possibilities. For example, there exists a regular (2, 10, 16)-hypergraph
of order 5,368,709,120 [Bermond+86]. This compares with the corresponding Kautz
hypergraph K H (2, 10, 16) of order 1,207,959,552.
In our discussion so far, high density has been the property of primary interest
when considering hypergraphs as possible logical topologies for large networks. As is
demonstrated next, hypernets – and especially families of regular hypernets – have other
attractive features, including ease of routing and multicasting.

7.4.3.4 Routing
When node labels are defined in terms of alphabets, routing algorithms in Kautz hyper-
nets can be defined and implemented conveniently using these labels. The algorithms
to be described next are in that category. Thus, we now focus on the classes of Kautz
hypernets that can be constructed using alphabets; that is, the undirected and directed
Kautz hypernets of degree 2 and 1, respectively: K H (2, D ∗ , r ) and D K H (1, D, s). The
labeling is assumed to be defined either directly or through duality construction (see
Section 7.4.3.1).

Address Shift Routing

Shortest path routing for directed Kautz hypergraphs can be implemented by ad-
dress shifting in essentially the same way as for Kautz digraphs in Section 7.2.2.
624 Multiwavelength Optical Networks

Distributed Routing

b1, b2, ... , bn x

Destination Counter
Address

Figure 7.23 Routing header.

If A = (a1 , . . . , a D ) and B = (b1 , . . . , b D ) are, respectively, the source and destina-

tion node labels, then the shortest path from source to destination through the di-
rected hypernet D K H (1, D, s) is determined from the shortest concatenation ( AB)∗ =
(a1 , a2 , . . . , a D , b D−i+1 , b D−i+2 , . . . , b D ), where i > 0 is the length of the shortest path
from A to B in logical hops [see Equation (7.9)]. The sequence of nodes and hyperarcs
along the path is given by
(a2 ,...,a D ) (a3 ,...,a D ,b D − i + 1 )
(a1 , a2 , . . . , a D ) −→ (a2 , . . . , a D , b D − i + 1 ) −→
(a D ,b D − i + 1 ,...,b D − 1 )
(a3 , . . . , a D , b D−i+1 , b D − i + 2 ), . . . , −→ (b1 , . . . , b D ).

Note that the routing decision at any node along the shortest path is based only on
a comparison of the destination address with the current node’s address. This suggests
implementing the algorithm in a distributed fashion with the aid of a routing record, as
shown in Figure 7.23. We assume that the unit being routed is a packet that contains a
routing header to provide self-routing information for transmission through the network.
The routing header is generated at the source station and it consists of the destination
address and a counter. The initial value of the counter, computed by the source node, is
x = i. Each node I = (i 1 , i 2 , . . . , i D ) on the path, including the source node, performs
the following operations:

1. If I is the same as the destination address, accept the packet.

2. If I is not the same as the destination address, check the value of the counter, route
the message to node (i 2 , i 3 , . . . , i D , b D−x+1 ), and decrement the counter by one.

To illustrate, consider D K H (1, 4, 2) shown in Figure 7.20. Suppose a packet is to

be routed from source A = 1012 to destination B = 2010. We have ( AB)∗ = 1012010
with path length i = 3 hops. The routing path then becomes
012 120 201
1012 −→ 0120 −→ 1201 −→ 2010.

Using self-routing, node 1012 places the destination address (b1 , b2 , b3 , b4 ) = 2010 in
the header in Figure 7.23, and the counter is initialized to x = 3. When the packet arrives
at the next node on the path, (i 1 , i 2 , i 3 , i 4 ) = 0120, the counter has been decremented to
x = 2, so the packet is forwarded to (i 2 , i 3 , i 4 , b3 ) = 1201.
This procedure can also be implemented without a counter by performing a shortest
concatenation operation at each intermediate node. Thus, when a packet arrives at
a node I , a shortest concatenation of the current node and destination addresses is
formed, and the packet is forwarded to the next node on the path determined by the
shortest concatenation. Continuing with the previous example, when a packet with
 Logically-Routed Networks 625

destination B = 2010 arrives at node I = 0120, the expression (I B)∗ = 012010 is

computed, indicating that the next node on the address shift path is 1201.

Routing in the Undirected Case

In undirected hypernets K H (2, D ∗ , r ), both (AB)∗ and (B A)∗ can be evaluated, the
former corresponding to left address shifts and the latter to right shifts. The source node
then chooses the path defined by ( AB)∗ if the length of ( AB)∗ is less than or equal to that
of (B A)∗ and chooses (B A)∗ otherwise. Although this may lead to shorter paths than
in the directed case, it does not necessarily produce shortest path routing. We provide
three examples to prove the point. In each case the network is K H (2, 3, 4), derived from
D K H (1, 4, 2) by removing hyperarc orientation.
r Routing from source A = 1012 to destination B = 2010 as in the previous directed
case, we now have (B A)∗ = 201012. Comparing (AB)∗ and (B A)∗ , the latter is chosen
as the shortest path, giving a path length i = 2. The routing path is
101 010
1012 −→ 0101 −→ 2010.
r Routing from source A = 1012 to destination B = 2012, we find that the shortest
path via address shift routing is
012 120 201
1012 −→ 0120 −→ 1201 −→ 2012.

However, referring to the hypergraph in Figure 7.19, it can be seen that both source
and destination nodes belong to the same hyperedge, so the true shortest path is
012
1012 −→ 2012

The problem here is that in K H (2, D ∗ , r ), vertices such as (b1 , b2 , . . . , b D−1 , x1 )

and (b1 , b2 , . . . , b D−1 , x2 ), where x1 = x2 belong to the same hyperedge – as do
(x1 , b2 , . . . , b D ) and (x2 , b2 , . . . , b D ) – so that the shortest path between them is one
hop. However, the common hyperedge is not detected by the address shift algorithm.
r Routing from source A = 1012 to destination B = 0201, we find that the shortest
path via address shift routing is given by (B A)∗ and is three hops. However, the true
shortest path is two hops:
012 201
1012 −→ 2012 −→ 0201,

where the first hop is not an address shift hop.

These examples suggest that it is the hyperedge labels rather than the vertex labels
that should be considered in the routing algorithm. In K H (2, D ∗ , r ) each vertex belongs
to two hyperedges, so four potentially shortest paths can be examined between any pair
of source and destination vertices. This leads to a shortest path-on-edge algorithm (see
[Jiang95]). Because K H (2, D ∗ , r ) is the dual of an undirected Kautz multigraph, the
computation of a shortest path between a pair of hyperedges in the former is equivalent
to shortest path vertex routing in the latter (see the shortest path algorithm for Kautz
graphs given in [Bermond+89]).
626 Multiwavelength Optical Networks

7.4.3.5 Performance Comparisons

In this section we provide some comparisons of the principal performance indices for
families of Kautz hypernets. The quantities that are compared include
r D: The network diameter
r H̄ : The average logical hop count
r γavg : The average hyperedge load
r µavg : The average load processed per node
The average logical hop count H̄ for the traffic in an LRN is the key performance
parameter. It determines three other performance indices:
r Throughput: Because the communication link loading is proportional to the logical
hop count, the maximum sustainable throughput is inversely proportional to the hop
count.
r Time delay: As explained in Section 5.4.2, the total delays in a wide area optical
network operating in a packet-/cell-switched mode are dominated by propagation
delays. (Queueing delays are minimal.) Thus, the average time delay through an LRN
is essentially the sum of the propagation delays on each logical hop. Assuming that
each hop has approximately the same propagation delay, the average time delay is
proportional to the logical hop count.
r Processing load: The hop count acts as a multiplier, determining how much processing
(e.g., IP packet routing) is required in the switches. Because switch processing power
can be the limiting factor in an LRN, this also limits throughput.
The hyperedge load, γ , is the aggregate traffic injected into the hyperedge by all nodes
transmitting into it. The load processed per node, µ, includes only that traffic forwarded
by the node from an upstream node to a downstream node (i.e., it excludes traffic either
entering or exiting the network at that node).
Performance of hypernets is examined under uniform carried traffic, with 1/(N − 1)
units of traffic flowing between each source-destination node pair in a network with N
nodes. Thus, each source injects one unit of traffic into the network regardless of the
network’s size, and the total throughput is N . Under these conditions, flow conservation
dictates the following relations among the performance indices:
r H̄
γavg = (7.34)

for undirected hypernets K H (, D ∗ , r ),
s H̄
γavg = (7.35)
d
for directed hypernets D K H (d, D, s), and
µavg = H̄ − 1 (7.36)
for both the directed and undirected case.
Note that a uniform injected traffic distribution does not necessarily produce uniform
hyperedge load. The exact hyperedge loading depends on the routing rule. In this case, we
 Logically-Routed Networks 627

Table 7.4 Performance of K H (2, D ∗ , r ).

Node label
length r N D∗ H̄ γavg µavg

2 6 12 2 1.18 3.54 0.18

2 8 20 2 1.31 5.24 0.31
3 4 12 2 1.5 3 0.5
3 6 36 3 1.87 5.67 0.87
3 8 80 3 2.11 8.44 1.11
4 4 24 3 2.18 4.36 1.18
4 6 108 4 2.74 8.22 1.74
4 8 320 4 3.04 12.16 2.04
5 4 48 4 2.98 5.96 1.98
5 6 324 5 3.68 11.04 2.68
6 4 96 5 3.86 7.72 2.86

use address shift routing and resort to simulation to determine H̄ and the maximum values
of hyperedge load and processing load: γmax and µmax , respectively. (The performance
results are 10 to 15% better when shortest-path-on-edge routing is used. See [Jiang95].)
Table 7.4 gives the performance parameters for hypernets K H (2, D ∗ , r ) with orders
ranging from 12 to 324. Their diameters D ∗ are either D − 1 or D, where D is the node
label length.
It can be seen from Table 7.4 that H̄ is approximately 0.75D ∗ in most of these cases,
which is the same relation that holds for large ShuffleNets using shortest path routing (see
Section 7.2.1). Both the hyperedge loads and the processing loads are essentially uniform
in the smaller networks and are roughly uniform in the larger ones. (In K H [2, 5, 4], the
network exhibiting the least uniformity in loading, γmax and µmax were found to be 18%
and 37%, respectively, above their average values.)
How do these numbers influence the optical resource requirements? Suppose that
each hyperedge is realized as an MPS supported by a “thick” waveband containing
C = r λ-channels. In this case, the MPS might be operated as a multicast star, using
TDM/T-WDMA in FT-TR mode, with one of the λ-channels dedicated to a transmitter
in each node. Because the nodes are of degree 2 in all of these networks, each node must
be equipped with two optical transceivers. Assuming that one unit of traffic corresponds
to a bit stream running at some fixed basic bit rate R 0 , and that the balance factor
for the traffic in each MPS is η = 1, Equation (7.28) indicates that the N units of
traffic can be carried by the network provided that its transceivers and λ-channels
run at bit rates R t ≥ R 0 (γmax /r ). Note that γmax /r represents the worst-case load on
an MPS normalized by the number of transmitters injecting traffic into it. We shall
refer to γmax /r as the normalized maximum hyperedge load or speedup factor in a
regular undirected hypernet.22 It indicates how much the transmitter bit rate R t must
be increased over the basic rate R 0 to accommodate the hyperedge load, assuming
that one λ-channel is available for each transmitter (i.e., C = r ). (As indicated earlier,
22
The normalized maximum hyperarc load (and speedup factor) for a regular directed hypernet is γmax /s,
where s is the hyperedge in-size.
628 Multiwavelength Optical Networks

under uniform carried traffic the normalized maximum hyperedge load is typically
close to the normalized average hyperedge load γavg /r .) If it is not possible to run the
transceivers/channels at the required speed, there are two other possibilities: either scale
down the traffic requirements proportionately to reduce the speedup factor, or scale up
the optical resources by providing extra transceivers and channels. These are general
rules applicable to all LRNs – either point-to-point or multipoint.

7.4.4 Hypernet versus Multihop

Because hypernets require an LLN infrastructure but point-to-point (multihop) LRNs
can operate over wavelength-routed networks, it is worthwhile to determine what (if
anything) is to be gained by going with the hypernet/LLN approach rather than the
simpler multihop/WRN approach. To this end, we make some performance comparisons
between hypernets and multihop networks. First, the undirected Kautz hypernet and the
ShuffleNet are compared, then comparisons are made between some larger size multihop
networks and directed hypernets.
Figure 7.24 shows a comparison between K H (2, D, 6) and the (2, k)-ShuffleNets of
the same order operating under uniform traffic with one unit of traffic injected into the
network by each source. We assume the same general configurations as in the previous
example. Each node is equipped with a number of transceivers equal to its degree,
so each of these networks requires two transceivers per node. Each transmitter in the

D
10

8
H
7

5
D
4
H
3

0
0 50 100 150 200 250 300 350 400
Order N
(a) Diameter, D, and Average Hops, H
ShuffleNet
KH(2, D, 6)

Figure 7.24 Comparison of hypernets and multihop networks.

 Logically-Routed Networks 629

3.5

2.5

1.5

0.5

0
0 50 100 150 200 250 300 350 400
Order N
(b) Normalized Average Link/Hyperedge Load

Figure 7.24 (cont.)

0
0 50 100 150 200 250 300 350 400
Order N
(c) Average Processing Load

Figure 7.24 (cont.)

630 Multiwavelength Optical Networks

Table 7.5 Performance comparison of DK H (d , D , s) to multihop networks.

Case d D s N H̄ γavg /s µavg

ShuffleNet 2 21 1 22,528 15 7.5 14

Kautz digraph 2 14 1 24,576 10.5 5.25 9.5
DK H 2 5 4 36,864 3.75 1.88 2.75
DK H 2 5 5 110,000 3.75 1.88 2.75
DK H 1 5 8 36,864 3.75 3.75 2.75

ShuffleNet is allocated a full λ-channel, and the six transmitters operating in each MPS
of the hypernet share the capacity of six λ-channels.
In Figure 7.24(a), diameters D and average logical hops H̄ are compared using shortest
path routing. The hypernet is superior to the ShuffleNet by a factor of approximately 2.
Figure 7.24(b) compares the normalized average hyperedge load, γavg /6, for the hypernet,
with average link load in the ShuffleNet, again showing a factor of 2 reduction in load
for the hypernet compared with the ShuffleNet. Figure 7.24(c) shows an improvement
factor of more than 2 in processing load for the hypernet.
What about the relative costs of the two approaches in optical hardware and spectrum
usage? Both systems use the same number of transceivers and the same total number of
λ-channels: one for each transmitter. The actual spectrum usage depends on the degree
of spectrum reuse, which in turn depends on the structure of the physical topology
and the way in which the logical topologies are embedded in the physical layer. In the
absence of any information on the physical layer, let us assume that there is no spectrum
reuse. In that case, the number of distinct λ-channels is the same in both systems.
However, the total spectrum occupied by those channels is less in the hypernet because it
bundles its channels into N /3 thick wavebands, whereas the ShuffleNet requires 2N thin
wavebands. A more complete comparison of resource utilization in multihop networks
and hypernets is undertaken in Section 7.5.2, in which embedding is examined in more
detail.
As a final comparison, Table 7.5 illustrates the performance of larger size hyper-
nets and multihop networks. Two multihop networks: a (2,11)-ShuffleNet (of diameter
21) and a Kautz digraph K (2, 14) are compared with three directed Kautz hypernets
D K H (d, D, s). The orders of the networks range from 22,528 to 110,000. The aver-
age number of logical hops H̄ , normalized average hyperarc load γavg /s, and average
node processing load µavg are listed for each case. Because simulation is out of the
question for these sizes, we assume shortest path routing ( H̄ = d)¯ to determine H̄ . The
known formula (Equation [7.7]) for average internodal distance d¯ is used in the case of
the ShuffleNet, and the approximation H̄ = 0.75D is used for the other networks. The
remaining quantities are derived using Equations (7.35) and (7.36).
In these large networks, the advantage of the hypernet over the multihop approach
is more pronounced. All of the networks require two transceivers per station except
D K H (1, 5, 8), which requires only one. The hypernets serve a significantly larger
number of nodes than the two multihop networks, with a factor of 4 improvement in
logical hop count over the ShuffleNet and an improvement of somewhat less than 3 over
 Logically-Routed Networks 631

the Kautz digraph. The improvement in H̄ carries with it corresponding reductions in

normalized hyperarc load and processing load. As a consequence, the optical transceivers
and LSNs can operate at significantly lower speeds in the hypernets.

7.4.5 Multicast Virtual Connections

Multicast connections are becoming increasingly important in large LRNs. For example,
considerable effort has been invested in designing multicast protocols on the Internet
and making them more efficient. In a multihop LRN, a multicast virtual connection
is implemented by forming a multicast tree in the network, rooted at the multicasting
node and replicating the information transmitted from the root on each link in the tree.
Because all links are point to point, the aggregate load generated in the network by the
multicast connection equals the total number of links in the tree.
In a hypernet, all transmissions into an MPS are always multicast optically to all
receiving stations in that MPS. This suggests that hypernets should be natural supports
for multicast virtual connections.23 Rather than doing a detailed analysis of multicasting
in hypernets, we confine ourselves here to a simple example that illustrates the savings
in usage of network capacity.
Suppose node 1021 in the directed hypernet D K H (1, 4, 2) of Figure 7.20 wishes to
broadcast to all other nodes in the network. A possible approach is shown in Figure 7.25.
A multicast tree is formed with links that are hyperarcs (MPSs) rather than point-to-
point connections. Each node in the tree (except the leaf nodes) logically multicasts to
all nodes in the receiving set of the MPS to which it transmits. In an IP network, for
example, this could be implemented simply by using a multicast local address on each
packet. Because D = 4 and s = 2 in this case, an incomplete tree of degree 2 and depth 4
is formed, containing the other 23 nodes. For example, node 1021 multicasts to the two
nodes 0210 and 0212 in the receiving set of MPS 021, and node 2102 unicasts to node
1020 in MPS 102. Note that in the latter case node 1021 has been “pruned” from the
receiving set because it is already in the tree.

1021

021
0210 0212

210 212
2101 2102 2120 2121

101 102 120 121

1010 1012 1020 1201 1202 1210 1212
X
010 012 020 201 202
X X
0101 0102 0120 0121 0201 0202 2010 2012 2020 2021

Figure 7.25 Multicast tree in D K H (1, 4, 2).

23
We have already seen an example of the economies attainable when multicast LCs are implemented on a
broadcast star, which is just a hypernet composed of a single hyperedge (see Section 5.4.1.4).
632 Multiwavelength Optical Networks

Although Figure 7.25 resembles an ordinary multicast tree, there is an important

difference: Each connection from a transmitting node in the tree to all receiving nodes in
the same MPS uses no more network capacity than if it had been unicasting to a single
node. As a result, the total capacity usage in this example is roughly 50% of what it
would have been in a multihop network.
A more complete discussion of multicasting in hypernets can be found in [Jiang95].
In general, a multicast connection in a directed hypernet can be compared with the same
connection implemented in its underlying digraph. The two networks are equivalent to
each other in terms of connectivity, but the underlying digraph represents a point-to-point
LRN and thus uses network resources less efficiently than the hypernet. The network
capacity used for a multicast virtual connection in a directed hypernet is approximately
a factor of s less than the usage in a digraph. Similar results hold for the undirected
case. Thus, hypernets with large hyperedge sizes are especially efficient in supporting
multicast virtual connections.

7.5 Hypernet Design

As in the case of multihop LRNs, the hypernet design problem can be separated into a
logical-layer design subproblem followed by realization of the chosen logical topology
on an optical infrastructure. The physical topology may be given or left free to the
designer. In the case of hypernets, the division into subproblems is even more important
than in the multihop case because hypernet design is more complex than point-to-point
network design (which is already exceedingly complex).

7.5.1 Logical-Layer Design

When viewed in its full generality, the logical-layer design problem involves finding an
LCH, together with a routing assignment on the LCH, which is capable of supporting
a prescribed traffic distribution. This is a direct generalization of the multihop network
design problem presented in Section 7.3. Following the approach in that section, the
design problem can be cast as a combinatorial optimization problem in which the
objective is to produce an LCH design and a routing assignment that minimizes some
cost function. As in the multihop case, the objective here is to minimize the maximum
flow on the LCs. We formulate this problem now as an MIP, for the case of regular
undirected hypernets with node degree  and maximum hyperedge (MPS) size r . The
variables used are

r xim : A binary variable set to 1 if node i is placed in hyperedge E m and set to 0

otherwise
r yi jm : A binary variable set to 1 if nodes i and j are in hyperedge E m and set to 0
otherwise
r tsd : The traffic injected into node s and destined for node d
 Logically-Routed Networks 633

r f sd : The traffic flow from source s to destination d carried from node i to j on

i jm
hyperedge E m
r z: The maximum of the flows on all hyperedges

The MIP is given by the following equations:

Minimize z with

m xim = , for all i (7.37)

i xim ≤ r, for all m (7.38)
yi jm ≤ 21 (xim + x jm ), for all i, j, m (7.39)
sd

s,d f i jm ≤ M yi jm , for all m (7.40)

tsd , if i = s
f isd sd
  
m j= i jm − j= i f jim = −tsd , if i = d (7.41)

0, otherwise

f isd
 
s,d i, j∈E m jm ≤ z, for all m (7.42)
0≤ f isd
jm , for all s, d, all m and all i = j.

Equation (7.37) is the degree constraint. (It would be replaced by in- and out-degree
constraints in directed hypernets.) Equation (7.38) is the hyperedge size constraint.
Equations (7.39) and (7.40) permit flow only between stations belonging to the same
hyperedge. (M is chosen as a large positive number in Equation [7.40.]) The flow
constraints appear in Equation (7.41), and Equation (7.42) sets z equal to the largest
hyperedge load. Note that no assumptions have been made regarding MPS capacity or
the numbers of optical transceivers in the stations supporting the network nodes. These
quantities could be added as extra constraints.
In view of the difficulty of this problem, the MIP optimization approach will not be
pursued further here. As in the multihop case, it is more prudent to retreat to a sub-
optimal approach, decomposing the logical design problem into the following three
steps: (1) choice of a logical topology, (2) node placement (matching the traffic matrix
to the LT), and (3) routing.
In the first step, a candidate logical topology is chosen from some family having
desirable characteristics (e.g., regularity, symmetry, high density). Following that, the
network nodes are placed on the vertices of the chosen LT, attempting to reduce the
aggregate network load by optimal placement of the nodes.24 Thus, for example, sets of
nodes having high traffic among each other are placed close together in the LT. At the
same time the placements are chosen to spread traffic loads evenly over all hyperedges.
After node placement, a routing algorithm is run with the objective of minimizing
maximum hyperedge loading. These steps can then be iterated to improve the design by
reducing the highest hyperedge loads.

24
More details on the node placement problem can be found in [Jiang+95].
634 Multiwavelength Optical Networks

7.5.2 Physical-Layer Design

We now treat the last subproblem in hypernet design: realization of a given LCH on
an optical infrastructure. We assume that the physical topology (an LLN) is given,
and seek an embedding of the given hypergraph onto the prescribed PT. The (much
easier) problem of realizing the hypernet on a multistar network designed to fit the LT
is considered in Section 7.5.4.
In Section 3.5.2, we showed how a hypernet can be realized on an LLN with a given
physical topology using collections of MPSs, in which each MPS is configured to support
one of the hyperedges. To support full optical connectivity in a hyperedge, its supporting
MPS must be realized as a connected subgraph of the PT, containing all nodes in the
hyperedge. To fix ideas concerning MPS implementation, we continue with the earlier
assumptions of Section 3.4.2 on the structure of the MPSs:
r Each MPS is realized on a tree.
r All transmissions within an MPS use λ-channels within a common waveband.
r The tree is configured as an embedded multicast star.

Although more general configurations of MPSs are possible, there is usually not much
to be gained by violating these assumptions. By specifying the MPS as a multicast star
embedded as a tree, considerable freedom is left in network design. The capacity of the
MPS and the way in which it is shared can be adjusted by using either thin or thick
wavebands, by using various numbers of optical transceivers in each node, and by using
various capacity allocation techniques. Thus, for example, a waveband containing many
λ-channels may be allocated to a given MPS to increase its capacity. Arrayed optical
transceivers may be employed to ensure that the traffic balance factor remains high even
though the traffic distribution seen by the MPS is nonuniform (see Equation [5.51]).
Packet switching in the optical layer may be used to provide for dynamic capacity
allocation in each MPS and so on.
The hypernet embedding problem is a static routing and waveband assignment problem
in an LLN, wherein the optical paths being embedded are multipoint trees. Thus it is a
direct generalization of the embedding problem for multihop networks, which is a static
point-to-point RCA problem in a WRN. In an optimal solution to this problem, one might
wish to minimize the number of wavebands used (maximize waveband reuse), minimize
the sum of the tree sizes, or minimize some combination of the two. In large networks,
this problem is generally intractable, necessitating the use of suboptimal heuristics. A
reasonable approach is to solve the tree-routing and waveband assignment problems
separately. After trees are chosen for each MPS, wavebands are assigned in a fashion
that minimizes the number of wavebands used.

Tree Routing
When selecting trees for a given set of hyperedges, it is desirable to choose fiber
paths with at least two objectives in mind: Making them short and making them fiber
disjoint whenever possible. The first objective promotes economy in the use of network
capacity, reduces congestion, enhances survivability, and improves signal quality. The
 Logically-Routed Networks 635

second objective improves the waveband reuse factor. As in the case of point-to-point
connections, the access links must be considered when determining whether fiber paths
are fiber disjoint.25
In an undirected hypernet, the tree-routing problem is closely related to the problem
of routing multicast optical connections on an LLN with a given physical topology. Any
node in the tree can be considered the source of a multicast optical connection to the
other nodes. Thus, embedding a set of trees is equivalent to a static routing problem for
a set of multicast optical connections. The heuristics of Section 6.6.2.1 can be used here
with some modifications.
Selection of a shortest path (or more generally minimum cost) tree spanning a given
set of vertices of a graph, known as the Steiner tree problem, was first encountered in
Section 6.4.5. Given a set K of vertices of a graph, with edges that have been assigned
costs, a Steiner tree is a tree that spans K while minimizing the sum of its edge costs.
It is well known that the Steiner tree problem is NP-complete. However, some efficient
heuristics are presented in Section 6.4.5. The embedding of a hypergraph onto a graph
can be thought of as constructing a “Steiner forest.”
Embedding based on finding a Steiner forest does not take into account the question
of waveband assignment and waveband reuse. However, it is possible to do combined
routing and waveband assignment for a given LT by updating edge costs in the PT as
the embedding process proceeds. The idea is to embed the hyperedges sequentially and
update the costs of edges in the PT each time a new hyperedge is embedded using those
edges. The cost of an edge is increased by one each time an embedded hyperedge uses that
edge. In this way, successive trees tend to avoid previously used edges, thereby tending
to increase the possibilities for waveband reuse. This is similar to static and dynamic
routing techniques for WRNs that attempt to minimize congestion (see Sections 6.3.8
and 6.4.1).
An example of the application of these tree routing techniques is presented in Fig-
ure 3.36 and Table 7.6. The undirected 22-node hypernet of Figure 3.36(a) is embedded
on Atlantis (Figure 3.36[b]), with the details of the embedding given in the table. Each
hyperedge E i is listed with its tree ti (excluding the access links). We assume elementary
access stations. The chosen trees are Steiner trees and they are edge disjoint whenever
their corresponding hyperedges are disjoint. (Because intersecting hyperedges use the
same access fibers in this case, their trees cannot be fiber disjoint.)

Waveband Assignment
In making waveband assignments, it is generally desirable to minimize the number of
wavebands used. This must be done subject to observing the DCA condition, which
requires that the same waveband can be assigned to two different trees if and only if they
are fiber disjoint. The waveband assignment problem is equivalent to a graph/hypergraph
coloring problem, and thus optimal assignment is very difficult for large networks. Still,
25
The paths for two hyperedges embedded on an LLN with elementary access stations cannot be fiber disjoint
unless the hyperedges are disjoint. (If a node belongs to two hyperedges, its access fibers must be used
in the MPSs supporting both hyperedges.) Similarly, the paths for two hyperarcs cannot be fiber disjoint
unless their in-sets are disjoint and their out-sets are disjoint.
636 Multiwavelength Optical Networks

Table 7.6 Hypernet trees on Atlantis.

Hyperedge
number Tree Waveband

1 1, 2, 3 1
2 5, 21, 25 1
3 12, 13, 17, 18, 25 3
4 29, 28, 27, 15, 16 3
5 7, 8, 9 1
6 28 1
7 19, 26, 21, 22 2
8 4, 5, 6, 7, 23, 15, 27 2
9 18, 17, 25, 24, 14, 9 2
10 12, 11, 10 2

simple heuristics of the type discussed in Appendix A produce fairly good results in
moderate-size networks.
In the case of embeddings in networks composed of elementary access stations, certain
necessary conditions on waveband assignments can be deduced from the logical topol-
ogy alone, even before the embedding is attempted. Following up on our observations
concerning fiber-disjoint trees, we find that
1. For undirected hypernets, two hyperedges containing a common LSN must be as-
signed distinct wavebands.
2. For directed hypernets, two hyperarcs containing a common LSN in their in-sets or
in their out-sets must be assigned distinct wavebands.
This constraint suggests that when using elementary access stations, a minimal edge
coloring of the LCH should be found as a preliminary to a final waveband assignment.
An edge coloring of a hypergraph is an assignment of colors to its hyperedges such
that two intersecting hyperedges are always colored differently. (If nonblocking access
stations are used, the logical topology alone does not constrain waveband assignments,
and thus this preliminary step serves no purpose.)
Applying these ideas to the 22-node hypernet, we note that the hyperedges E 1 , E 2 , E 5 ,
and E 6 are mutually disjoint. Furthermore, their trees, as indicated in Table 7.6, are
disjoint, so they can be assigned the same waveband. Similar observations hold for the
sets E 7 , E 8 , E 9 , E 10 and E 3 , E 4 . As a result, a minimal waveband assignment can be
made with three wavebands, as shown in Table 7.6.
The waveband assignment methodology just described can be formalized in terms of
an auxiliary hypergraph as follows. Given an LT in the form of a hypergraph H l (ν l , εl ),
and a PT defined as a graph G p (V p , E p ), let each hyperedge E il ∈ εl be embedded in
the PT as a tree ti = {ei1 , ei2 , . . .}, where ei j ∈ E p . Let H em (ν em , εem ) be a hypergraph
with vertices that are the edges of the given PT and with hyperedges that are the trees;
that is, ν em = E p and εem = {t1 , t2 , . . .}. Then a minimal waveband assignment for
H l is equivalent to a minimal edge coloring of H em . Note that even if H l (ν l , εl ) is a
graph, H em is generally a hypergraph. Solving this waveband assignment problem by
 Logically-Routed Networks 637

edge coloring the auxiliary hypergraph H em is the counterpart of solving the λ-channel
assignment problem in WRNs by vertex coloring the path interference graph G PI (see
Section 6.3.3).

7.5.3 Traffic Grooming in Multipoint Logical Topologies

In most networks the traffic is a combination of unicast and multicast connections,
therefore traffic grooming includes the grooming of multicast connections as well. An
ILP formulation for grooming multicast traffic was proposed in [Kamal+03], where the
goal of the ILP is to minimize the number of wavelength channels and the number of
SONET add/drop multiplexers (ADMs) used. As in the layered description in Figure 7.2,
[Kamal+03] represents the network as three different layers, the physical, logical, and
virtual connection layers. The authors introduce heuristics to solve the problem by
first obtaining an initial solution utilizing a shortest path tree and first-fit wavelength
assignment (as described in Section 6.4.5) and then improving the solution in an iterative
manner by exploring other routes that are not necessarily the shortest ones. Additional
work on this subject is outlined in [Yang+03]. Here an ILP formulation is used for the
design of optimum light-trees, and then the light-tree-based logical topology is modeled
as a hypergraph over which static multicast connections are routed. Both studies deal
with the grooming of static multicast traffic.
The grooming problem for dynamic multicast traffic in WDM networks is attacked
in [Khalil+04] by designing a light-tree-based hypergraph logical topology. To support
grooming of multicast traffic, the logical switching node should have both multicast and
grooming capabilities. Such a node is referred to as a multicast-capable logical-grooming
switch (MC-LGS) and is shown in Figure 7.26.
In the MC-LGS, a lightpath or light-tree requiring grooming is sent to the grooming
fabric before it is injected back into the network. As mentioned previously, multicast-
capable networks are best modeled by directed hypergraphs. An algorithm is described
in [Khalil+04], based on a directed hypergraph logical topology defined as the set
of all established lightpaths and light-trees, where a hyperarc represents each optical
lightpath or light-tree. When a request for a new multicast connection arrives, the
existing logical topology is checked to see if the new connection can be groomed
onto an existing logical path. This will be possible if there is an established light-tree
in the logical topology that serves the same source and destinations and has enough
residual capacity on its logical links, or if there is a combination of multiple light-trees
that can accommodate the new connection. The former case would be referred to as
a single-hop groomed connection and the latter as a multihop groomed connection in
the logical topology. In the case of multihop grooming, the algorithm searches for a
light-tree that has the same destinations as the requested multicast session. Once such a
tree is found, the algorithm then searches for a single-hop lightpath between the source
of the requested multicast session and the source of that tree. If the algorithm succeeds,
the new multicast connection request is groomed on the combination of the lightpath
and the light-tree. Minimum cost routing algorithms (e.g., Dijkstra’s algorithm) can
be used in the logical topology to determine the route of the new connection if it is
638 Multiwavelength Optical Networks

MC-LGS
DMUX MUX

.
.
.
.
.
. .
Fibers . . Fibers
. Out
In .
. .
. .
.
. . .
.

splitter
Multicast-capable OSW-2
optical switch

OSW-1

Tx Tx Grooming Rx Rx
fabric

... ...
Local add Local drop

Figure 7.26 Multicast-capable logical-grooming switch. (From [Khalil+04, Figure 1]. Copyright


c 2004 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

indeed possible to groom it. If the connection cannot be groomed in the existing logical
topology because a path with enough residual capacity does not exist, multicast routing
and wavelength assignment as described in Chapter 6 will then have to be invoked
to create new lightpaths or light-trees to carry the new connection either alone or in
combination with other existing lightpaths and light-trees. Because [Khalil+04] deals
with dynamic rather than static traffic, the performance criterion of interest is the
blocking probability. The grooming gain of the network (GainG ) is consequently defined
in [Khalil+04] as the gain in network performance (in terms of blocking probability
[BP]) when a grooming scheme is used as compared to blocking without grooming:
(B PN G − B PG ) × 100
GainG = . (7.43)
B PN G
Studies in [Khalil+04] illustrate the effect of grooming on network performance,
which is shown to improve by more than 40% when a grooming scheme is used as
compared to the nongroomed case. As the load increases, it increases the blocking
probability in both the groomed and nongroomed cases, with grooming becoming less
of an issue. The limiting factor is then the network resources (number of wavelengths)
rather than the grooming scheme used to accommodate the multicast requests.
Additional results in [Khalil+04] deal with network performance when a limit is
placed on the number of destinations for multicast connections. The authors studied
blocking performance as a function of a parameter G p , which represents the maximum
 Logically-Routed Networks 639

allowable destinations for a multicast connection, expressed as a percentage of the

total number of network nodes. In their work the number of destinations is uniformly
distributed in the interval [2, d], where d = G p × n/100 and n is the network size. They
found that as G p increases (for fixed load) the blocking probability increases and the
gain obtained using a grooming scheme diminishes considerably. For very high G p a
grooming scheme will yield almost as much blocking as a nongrooming approach or an
approach where a multicast connection is implemented as several independent unicast
connections.

7.5.4 Multistar Realizations

In Section 6.5.7 it was shown that a multistar physical topology is a natural way to
realize a hypernet of diameter 1. Hypernets of diameter 1 are purely optical networks
(LLNs) configured to provide full logical connectivity between all pairs of nodes in
one logical hop. In these multistar realizations, a set of undirected (directed) broadcast
stars are arranged so that each star forms an undirected (directed) MPS, with each MPS
supporting one hyperedge (hyperarc) in the network’s LCH. The design of the multistar
network is facilitated by the fact that its fiber connectivity is an exact replica of the
tripartite representation of its LCH.
These ideas extend directly to LRNs defined by hypergraphs of diameter greater than
1. The only difference is that the connectivity is incomplete on the optical level and must
be completed on the logical level through an overlay of LSNs.
As an example, consider the directed hypergraph G K H (2, 42, 3, 28) with a tripartite
representation that is shown in Figure 7.22. Let each hyperarc be replaced by a cor-
responding 3 × 3 directed star joining the three LSNs in its transmitting set to three
LSNs in its receiving set. For example, hyperarc 0 becomes a star multicasting from
nodes {0, 14, 28} to {39, 40, 41}.26 Because this network has an in-/out-degree of 2, each
transmitter/receiver is connected to two stars. Like-numbered transmitting and receiving
nodes are assumed to belong to the same LSN, and logical paths are set up by routing
traffic through a series of LSNs from source to destination. In this case the LRN has a
diameter of 2, so in the worst case the traffic must transit one intermediate node on a
source-destination path. For example, traffic from source 0 to destination 6 would take
the two-hop path
0 25
0 −→ 40 −→ 6.

In cases in which a hypergraph is constructed from a graph using duality, a multi-

star realization of the corresponding hypernet can be deduced directly from the dual
graph. For example, Figure 7.19 shows the undirected Kautz hypergraph K H (2, 3, 4)
constructed from the Kautz digraph K (2, 3). Each hyperedge (MPS) can be realized by
a star connecting its vertices. The dual digraph gives the exact topology of the multistar
network. Because each vertex of K H (2, 3, 4) is the dual of an arc of K (2, 3), an arc of
K (2, 3) corresponds to an LSN of the Kautz hypernet, and the terminal vertices of that

26
The star connections can be determined from the definition of G K H (d, n, s, m) in Section 7.4.3.3.
640 Multiwavelength Optical Networks

arc represent star couplers to which the LSN is attached. (The LSN can be viewed as a
box inserted in the middle of the arc.)
In the undirected case, each LSN transmits and receives from each star, so it must
have two transceivers, and the links through which it accesses the stars are bidirectional
fiber pairs. The construction can also be used for the directed case in which each arc of
the dual digraph represents an LSN with a single transceiver, with the transmitter and
receiver oriented according to the direction of the arc. For example, node 1012 in the
hypernet of Figure 7.19 is the dual of the arc 101 → 012 in the digraph. Thus, in the
undirected case, LSN 1012 is connected to the two stars 101 and 012. Each of these
is an undirected 4 × 4 star. In the directed case, LSN 1012 transmits into star 012 and
receives from 101, where each is now a directed 2 × 2 star.
The prime advantage of the multistar physical topology is that it provides a very high
degree of spectrum reuse. Each star makes the total usable optical spectrum available to
the transceivers using it, and that spectrum is reused in every star. Thus, for example, in
G K H (2, 42, 3, 28), three transmitting nodes use each star so that the full spectrum is
shared among them. Assuming that the transceivers in the nodes are connected directly
to the stars, they access the network in a nonblocking fashion. That is, a node accessing
several stars can reuse the same wavelengths on different stars.
The number of λ-channels allocated to a star is arbitrary, depending only on the
required throughput. Because a broadcast star acts as a shared-channel medium, any
number of channels can be shared among all nodes accessing the star using any of the
multiplexing/multiple-access techniques described in Chapter 5. A reasonable allocation
would be s λ-channels for a star with in-/out-size s, so nodes equipped with one optical
transceiver can exploit fully the channels and the transceivers.
In a high-throughput application, additional transceivers and channels might be ap-
propriate. In our example, suppose the available spectrum consists of nine λ-channels
per star. Then, to take advantage of the available channels, each node should be equipped
with an array of six transceivers, three for each star it uses. Using these arrays in TDM/T-
WDMA mode, it is possible to utilize the full capacity (nine units) of each star. The
λ-channels in one star are reused in all stars, giving a reuse factor of 28 in this example.
Still more mileage can be obtained from the optical spectrum if the star couplers are
replaced by permutation wavelength routers as in Section 6.3.6. In that case, using s
λ-channels, each s × s star can support s 2 point-to-point LCs among its attached nodes,
each running at the full capacity of one λ-channel. But this requires that each node be
equipped with ds transceivers, where d is its in-/out-degree. The spectrum reuse factor
of such a system is sm, where m is the number of hyperedges (wavelength routing stars).
Using this approach for G K H (2, 42, 3, 28), we obtain the same throughput using three
λ-channels and permutation routers as we had using nine λ-channels and broadcast stars.
The same number of transceivers are used in each case, and the reuse factor is increased
to 84.
The major disadvantage of a multistar is that it uses a very large amount of fiber.
(The number of stars is the same as the reuse factor.) This should come as no surprise
because, as usual, there is no free lunch! Spectrum reuse comes at the inevitable expense
of fiber replication. Another disadvantage of multistars that was cited in Section 6.3.6 is
 Logically-Routed Networks 641

their poor survivability. In a multistar realization of a hypernet of diameter 1, when one

star fails, communication is cut between the transmitting and receiving sets of that star.
However, this disadvantage disappears in hypernets of larger diameter because there are
alternate paths between LSNs in this case.
We have shown through examples that hypernets can accommodate easily on the
order of 105 LSNs with relatively small node degrees, diameters, and hyperedge sizes.
This means that multistar realizations can support hypernets of these sizes using only a
moderate number of λ-channels and using limited optical hardware.
For example, the directed Kautz hypernet D K H (2, 5, 5) has order 110,000. From
Table 7.5 we find that for one unit of uniformly distributed traffic injected at each node,
the average hyperarc load is 9.4 units. Thus, if this network is realized using 10 λ-
channels per star, it can support a throughput of 110,000 (normalized to the capacity of
a λ-channel).27 This requires four optical transceivers per LSN and 11,000 stars.

7.6 Summary

Logically-routed networks consisting of electronic LSNs superimposed on a purely

optical network infrastructure comprise the widest and most versatile classes of multi-
wavelength optical networks. The model used to represent LRNs in this chapter is broad
enough to include virtually any type of hybrid electronic/optical network, including
networks of SONET DCSs, networks of ATM switches or IP/MPLS routers, IP over
SONET, and many more yet to be invented.
The LRNs examined here were divided into two classes: those with point-to-point
logical topologies (multihop networks) and those with multipoint LTs (hypernets). The
optical infrastructure for a multihop network can be either a WRN or an LLN. In the
case of hypernets, the infrastructure must be a network capable of providing multipoint
optical connectivity (i.e., an LLN). It was shown that these networks can be studied at
two more or less independent levels: the logical layer and the physical layer. Desirable
properties of logical topologies were explored, and families of dense, regular LTs were
studied in both the multihop and the hypernet cases.
The families of Kautz graphs and hypergraphs were chosen for special attention as
logical topologies for LRNs. These have desirable properties in terms of density, load
balancing, routing, and survivability. Cost-performance relations were compared within
families of hypernets and for hypernets versus point-to-point LRNs. It was found that
the most important performance parameter in LRNs is the average logical hop count H̄
¯ and that this parameter can be kept
(related closely to the average internodal distance d)
under control cost-effectively by using dense hypernet logical topologies. A hypernet LT
produces the smallest value of d¯ for a given investment in optical resources: transceivers
and optical spectrum.
The design of an LRN can be broken down into logical-layer design and physical-layer
embedding. Although these two subproblems can, in principle, be solved simultaneously,

27
This allows a little slack for uneven traffic distribution.
642 Multiwavelength Optical Networks

this is possible in only very small networks. Logical-layer design consists of finding
a suitable logical topology and routing assignment to accommodate a given traffic
distribution. The performance of a logical-layer design when carrying a prescribed
traffic distribution may be evaluated by determining the communication link loads and
switch processing loads. A well-matched LT will not have any “hot spots,” so the loads
will be distributed fairly evenly throughout the network. Some simulation results were
exhibited to illustrate and compare typical communication and processing loads under
uniform traffic conditions.
In the two-layered approach to the LRN design problem, the traffic distribution was
treated as a fluid flow; i.e., its granularity was ignored. When the disparate granulari-
ties of the end system and core connections were considered, the design problem was
expanded to include grooming. This created a three-layered problem, including the VT
layer. General formulations and heuristics were presented to solve the grooming/design
problem for both rings and general mesh topologies. The key issue that must be ad-
dressed in choosing grooming strategies is the trade-off between the equipment cost and
complexity associated with the grooming operations in the logical layer, and the trans-
mission efficiencies grooming produces in the physical layer by intelligent consolidation
of subrate traffic onto high rate λ-channels.
Physical-layer embedding typically involves routing and wavelength assignment on
a WRN in the multihop case and routing and waveband assignment on an LLN in
the hypernet case. In the case of uniformly distributed traffic, the maximum aggregate
throughput of the embedded LRN is roughly proportional to the total number of optical
transceivers used in the embedding and is inversely proportional to H̄ .
The cost of the LRN in optical resource usage depends, to a large extent, on the optical
spectrum reuse factor, which in turn depends on the richness of the physical topology.
The reuse factor is largest when a multistar PT is used. However, multistars carry with
them certain handicaps, the most significant of which is high fiber usage.

7.7 Problems

1 Show that five wavelengths are sufficient to realize the eight-node ShuffleNet using
a star topology with a permutation router with routing rules defined in Section 2.3.1.2.
2 Show that two wavelengths are sufficient to realize the eight-node ShuffleNet using
a star topology with a permutation router with routing rules different from those defined
in Section 2.3.1.2.
3 Find an embedding of the eight-node ShuffleNet on Atlantis using elementary access
stations and a minimum number of wavelengths.
4 For networks of the same diameter, compare the orders of (2, k)-ShuffleNets to the
orders of the hypernets K H (2, D, r ) plotted in Figure 7.18.
5 This problem concerns the duality relations between graphs and hypergraphs.
(a) Derive K H (2, D ∗ , 6) from K (3, 2) by duality and show its multistar realization.
 Logically-Routed Networks 643

(b) Derive D K H (1, 3, 3) from K (3, 2) by duality and show its multistar realization.
(c) Determine the diameter, D ∗ , of K H (2, D ∗ , 6).
6 Prove that for K H (2, D ∗ , 4) with node label length D, D ∗ < D.
7 Find an embedding of the 18-node shuffle hypernet on Atlantis using Steiner trees
and as few wavebands as possible. Assume elementary access stations, and map the
LSNs onto the stations in Atlantis in numerical/alphabetical order.
8 Find an embedding of D K H (1, 4, 2) on Atlantis using Steiner trees and as few
wavebands as possible. Assume elementary access stations, and map the LSNs onto the
stations in Atlantis in numerical/alphabetical order. (Add stations W and X to nodes i
and j, respectively.)
9 Simulate one or more of the logical topologies used in Table 7.4 under nonuniform
traffic conditions.
(a) Compare your results on average hyperedge load and processing load per node with
those in the table.
(b) Determine the maximum values of these quantities and their percent deviation from
the average.

Bibliography

[Acampora87] A. S. Acampora. A multichannel multihop local lightwave network. In Proceedings

of the IEEE Global Telecommunications Conference (Globecom), pp. 1459–1467, Tokyo, Japan,
November 1987.
[Bala+95] K. Bala, R. R. Cordell, and E. L. Goldstein. The case for opaque multiwavelength
lightwave networks. In Proceedings of the IEEE/LEOS Summer Topical Meeting, Keystone,
CO, August 1995.
[Bala+96] K. Bala, G. Ellinas, M. Post, S. Chien-Chung, J. Wei, and N. Antoniades. Towards
hitless reconfiguration in WDM optical networks for ATM transport. In Proceedings of the
IEEE Global Telecommunications Conference (Globecom), London, UK, November 1996.
[Bannister90] J. A. Bannister. The Wavelength-Division Optical Network: Architectures, Topolo-
gies and Protocols. Ph.D. thesis. UCLA, Los Angeles, 1990.
[Bermond+83] J.-C. Bermond, J. Bond, M. Paoli, and C. Peyrat. Graphs and interconnection
networks: Diameter and vulnerability. In Surveys in Combinatorics, pp. 1–30. Cambridge:
Cambridge University Press 1983.
[Bermond+86] J.-C. Bermond, J. Bond, and C. Peyrat. Interconnection networks with each node
on two buses. In Parallel Algorithms and Architectures, pp. 155–167. Amsterdam: North-
Holland 1986.
[Bermond+89] J.-C. Bermond and C. Peyrat. deBruijn and Kautz networks: A competitor for
the hypercube? In Hypercube and Distributed Computers, pp. 279–293. Amsterdam: Elsevier
Science 1989.
[Bermond+96] J.-C. Bermond and F. O. Ergincan. Bus interconnection networks. Discrete Appl.
Math., 68(1–2):1–15, 1996.
[Bermond+97] J.-C. Bermond, R. Dawes, and F. O. Ergincan. deBruijn and Kautz bus networks.
Networks, 30(3):205–218, 1997.
644 Multiwavelength Optical Networks

[Berry+00] R. Berry and E. Modiano. Reducing electronic multiplexing costs in SONET/WDM

rings with dynamically changing traffic. IEEE J. Select. Areas Commun., 18(10):1961–1971,
2000.
[Bienstock+93] D. Bienstock and O. Gunluk. Computational experience with a difficult mixed
integer multicommodity flow problem. Math. Programm., 68:213–237, 1993.
[Birk86] Y. Birk. Concurrent Communication among Multiple Transceiver Stations via Shared
Media. Ph.D. thesis. Stanford University, Palo Alto, CA 1986.
[Birk91] Y. Birk. Fiber-optic bus-oriented single-hop interconnections among multi-transceiver
stations. IEEE/OSA J. Lightwave Technol., 9(12):1657–1664, 1991.
[Cinkler03] T. Cinkler. Traffic and λ grooming. IEEE Network, 17(2):16–21, 2003.
[Dowd92] P. W. Dowd. Wavelength division multiple access channel hypercube processor inter-
connection. IEEE Trans. Comput., 41(10):1233–1241, 1992.
[Dutta+02] R. Dutta and G. N. Rouskas. On optimal traffic grooming in WDM rings. IEEE J.
Select. Areas Commun., 20(1):110–121, 2002.
[Ellinas+03] G. Ellinas, E. Bouillet, R. Ramamurthy, J.-F. Labourdette, S. Chaudhuri, and K.
Bala. Routing and restoration architectures in mesh optical networks. SPIE Optical Networks
Mag., 4(1):91–106, 2003.
[Gerla+92] M. Gerla, M. Kovacevic, and J. Bannister. Multilevel optical networks. In Proceedings
of the IEEE Int’l Conf. Commun. (ICC), pp. 1168–1172, Chicago, IL, June 1992.
[Gerstel+98] O. Gerstel and R. Ramaswami. Cost effective traffic grooming in WDM rings. In
Proceedings of the IEEE Conference on Computer Communications (Infocom), San Francisco,
CA, April 1998.
[Gerstel+99] O. Gerstel, P. Lin, and G. Sasaki. Combined WDM and SONET network design. In
Proceedings of the IEEE Conference on Computer Communications (Infocom), pp. 734–743,
New York, March 1999.
[Gerstel+00] O. Gerstel, R. Ramaswami, and G. Sasaki. Cost effective traffic grooming in WDM
rings. IEEE/ACM Trans. Network., 8(5):618–630, 2000.
[Hadjiantonis+04] A. Hadjiantonis, A. Khalil, G. Ellinas, and M. A. Ali. A hybrid approach for
provisioning sub-wavelength requests in IP-over-WDM networks. In Proceedings of the IEEE
Canadian Conference on Electrical and Computer Engineering (CCECE’04), Niagara Falls,
May 2004.
[Hluchyj+88] M. G. Hluchyj and M. J. Karol. ShuffleNet: An application of generalized perfect
shuffles to multihop lightwave networks. In Proceedings of the IEEE Conference on Computer
Communications (Infocom), pp. 4B.4.1–12, New Orleans, LA, March 1988.
[Hluchyj+91] M. G. Hluchyj and M. J. Karol. ShuffleNet: An application of generalized perfect
shuffles to multihop lightwave networks. IEEE/OSA J. Lightwave Technol., 9(10):1386–1397,
1991.
[HZhu+03] H. Zhu, H. Zang, K. Zhu, and B. Mukherjee. A novel, generic graph model for traffic
grooming in heterogeneous WDM mesh networks. IEEE/ACM Trans. Network., 11(2):285–299,
2003.
[Imase+83] M. Imase and M. Itoh. A design for directed graphs with minimum diameter. IEEE
Trans. Comput., c-32(8):782–784, 1983.
[Jiang+93] S. Jiang, T. E. Stern, and E. Bouillet. Design of multicast multilayered lightwave
networks. In Proceedings of the IEEE Global Telecommunications Conference (Globecom),
pp. 452–457, Houston, TX, November 1993.
[Jiang95] S. Jiang. Multicast Multihop Lightwave Networks: Design and Implementation. Ph.D.
thesis. Columbia University, New York, 1995.
 Logically-Routed Networks 645

[Jiang+95] S. Jiang and T. E. Stern. Regular multicast multihop lightwave networks. In Proc-
cedings of the IEEE Conference on Computer Communications (Infocom), Boston, MA, April
1995.
[Kamal+03] A. Kamal and R. Ul-Mustafa. Multicast traffic grooming in WDM networks. In Pro-
ceedings of the SPIE/IEEE Optical Networking and Communications Conference (Opticomm),
pp. 25–36, Dallas, TX, October 2003.
[Khalil+04] A. Khalil, A. Hadjiantonis, G. Ellinas, and M. Ali. Multicast traffic grooming in
WDM mesh networks: Multi-hop and hybrid provisioning approaches. In Proceedings of the
IEEE Global Telecommunications Conference (Globecom), Dallas, TX, December 2004.
[Khalil+06] A. Khalil, A. Hadjiantonis, C. Assi, A. Shami, G. Ellinas, and M. A. Ali. Dynamic
provisioning of low-speed unicast/multicast traffic demands in mesh-based WDM optical net-
works. IEEE/OSA J. Lightwave Technol., 24(2):681–693, 2006.
[Kovacevic93] M. Kovacevic. HONET: An Integrated Services Wavelength Division Optical Net-
work. Ph.D. thesis. UCLA, Los Angeles, 1993.
[Kovacevic+95] M. Kovacevic, M. Gerla, and J. Bannister. Performance of shared-channel
multihop lightwave networks. Comput. Commun., 18(1):37–44, 1995.
[KZhu+02] K. Zhu and B. Mukherjee. Traffic grooming in an optical WDM mesh network. IEEE
J. Select. Areas Commun., 20(1):122–133, 2002.
[KZhu+03a] K. Zhu, H. Zhu, and B. Mukherjee. Traffic engineering in multigranularity het-
erogeneous optical WDM mesh networks through dynamic traffic grooming. IEEE Network.,
17(2):8–15, 2003.
[KZhu+03b] K. Zhu and B. Mukherjee. A review of traffic grooming in WDM optical networks:
Architectures and challenges. SPIE Optical Networks Mag., 4(2):55–64, 2003.
[Labourdette+90] J.-F. P. Labourdette and A. S. Acampora. Wavelength agility in multihop light-
wave networks. In Proceedings of the IEEE Conference on Computer Communications (Info-
com), pp. 1022–1029, San Francisco, CA, June 1990.
[Labourdette+91] J.-F. P. Labourdette and A. S. Acampora. Logically rearrangeable multihop
lightwave networks. IEEE Trans. Commun., 39(8):1223–1230, 1991.
[Modiano+98] E. Modiano and A. Chiu. Traffic grooming algorithms for minimizing electronic
multiplexing cost in unidirectional SONET/WDM ring networks. In Proceedings of the 32nd
Annual Conference on Information Sciences and Systems (CISS), Princeton, NJ, February
1998.
[Modiano+01a] E. Modiano and P. Lin. Traffic grooming in WDM networks. IEEE Communi-
cations Mag., 39(7):124–129, 2001.
[Modiano+01b] E. Modiano and R. Berry. Using grooming cross-connects to reduce ADM costs
in SONET/WDM ring networks. In Proceedings of the IEEE/OSA Optical Fiber Comm. Conf.,
Anaheim, CA, March 2001.
[Ofek87] Y. Ofek. The Topology, Algorithms and Analysis of a Synchronous Optical Hypergraph
Architecture. Ph.D. thesis. University of Illinois, Urbana, 1987.
[Reddy+82] S. M. Reddy, J. G. Kuhl, S. H. Hosseini, and H. Lee. On digraphs with minimum
diameter and maximum connectivity. In Proceedings of the 20th Annual Allerton Conference
on Communication Control and Computing., pp. 1018–1026, Urbana, IL, October 1982.
[Shi92] C-J. Shi. A signed hypergraph model of the constrained via minimization problem.
Microelectron. J., 23(7):533–544, 1992.
[Simmons+99] J. Simmons, E. Goldstein, and A. Saleh. Quantifying the benefit of wavelength
add-drop in WDM rings with distance-independent and dependent traffic. IEEE/OSA J. Light-
wave Technol., 17(1):48–57, 1999.
646 Multiwavelength Optical Networks

[Sivarajan+91] K. N. Sivarajan and R. Ramaswami. Multihop lightwave networks based on

deBruijn graphs. In Proceedings of the IEEE Conference on Computer Communications (Info-
com), pp. 1001–1011, Miami, FL, April 1991.
[Yang+03] D.-N. Yang and W. Liao. Design of light-tree based logical topologies for multicast
streams in wavelength routed optical networks. In Proceedings of the IEEE Conference on
Computer Communications (Infocom), San Francisco, CA, March 2003.
[Zhang+00] X. Zhang and C. Qiao. An effective and comprehensive approach for traffic grooming
and wavelength assignment in SONET/WDM rings. IEEE/ACM Trans. Network., 8(5):608–617,
2000.
[Zhemin+03] D. Zhemin and M. Hamdi. Traffic grooming in optical WDM mesh networks using
the blocking island paradigm. SPIE Optical Networks Mag., 4(6):7–15, 2003.
8 Survivability: Protection and
Restoration

Survivability against failures, including failure recovery, is important in any telecommu-

nications network but is highly critical for high-bandwidth optical networks. As more
traffic is concentrated on fewer routes, the number of customers that can be potentially
affected by a failure is increased. An analysis of failures in the Public Switched Tele-
phone Network over a two-year period in the 1990s showed that human error, acts of
nature, and overloads were the major sources of failure [Kuhn97]. The impact of the
failures was measured in terms of how many times a particular failure occurred, dura-
tion of the outage, and number of customers and number of customer minutes affected
during that outage. During that period, the average number of customers affected due to
cable cuts or cable component failures was 216,690, costing 2,643 million in customer
minutes.1 Similarly, the average number of customers affected by each equipment failure
was 1,836,910, costing 3,544.3 million in customer minutes [Kuhn97]. Cable cuts and
hardware/equipment failures account for approximately half of the failures encountered
in the network during that period [Kuhn97].
Fiber cuts are considered one of the most common failures in fiber-optic networks.
Furthermore, the use of WDM over these fibers produces an extremely high volume of
traffic on a cable. Commercially available fiber-optic transmission systems can run at
10 Gbps or more per channel with 80 or more channels (wavelengths) per fiber. This
translates to more than 800 Gbps per fiber. If we consider the fact that on the average
each cable carries 100 fibers, there is a possibility that a cable cut will result in the loss
of several terabits per second. The most prevalent form of communication failure is the
accidental disruption of buried telecommunication cables. Aerial cables are also affected
but not as frequently as those in the ground. Fiber cuts may result from construction
work (“backhoe fade”), lightning, rodent damage, fire, train derailments, vandalism, car
crashes, human error, and so on [Grover89, Grover04]. One of the main reasons why
underground telecommunication cables are so susceptible to failure is the fact that they
are buried in the same public right-of-ways as are other utility transport media (e.g.,
water pipes, gas pipes, television cables [Wrobel90]).
Examples of fiber cuts that affected network operation severely can be found through-
out the short history of fiber-optic networks. During the 1980s when all the major
telecommunications companies were laying their fiber in the ground, cable cuts were an
almost everyday occurrence [Wrobel90]. One of the first serious cable cuts happened

1
A customer minute is defined as the outage in minutes multiplied by the number of customers affected.
648 Multiwavelength Optical Networks

in 1987 when a 12-fiber lightwave cable that was a major backbone carrier facility was
severed. Because of that cut more than 125,000 trunks were put out of service and an
estimated 100,000 connections were lost two seconds after the cut. It took more than
two hours to restore manually some of the capacity on physically diverse routes, at a
cost of millions of dollars [Grover89].
Equipment failures at network nodes, even though not as common, can have dev-
astating consequences. In fact, in general, equipment failures affect many more peo-
ple than cable cuts. If a switching node fails, all the fibers and their connections
through that node fail. Equipment failures can result from natural phenomena (earth-
quake, flood, fire), human error, and hardware degradation [Grover04]. Examples of
well-known equipment failures that have had a devastating effect on the network in-
clude the 1988 fire that destroyed Illinois Bell’s Hinsdale switching office, the 1987
fire that damaged New York Telephone’s Brunswick switching office, and the 1987
AT&T switching computer failure in Dallas. The 1988 Illinois fire was rated at
that time as the worst telecommunications disaster in U.S. history because it hap-
pened on one of the busiest days of the year (Mother’s Day) and took more than
a month to completely repair. A more recent example is the destruction of a key
Verizon call and data-switching center in lower Manhattan on September 11, 2001
[Lee01]. That particular switching center was responsible for providing approximately
200,000 voice lines for homes and small businesses, 100,000 voice lines for several
large businesses, and data services with a capacity equivalent to 3.5 million circuits.
On September 11, the telecommunications switching and computer equipment were
damaged due to falling debris, dust clouds, fire, and flooding, and of the more than
500 copper and fiber-optic cables that originated or passed through the center, approxi-
mately 300 were either damaged or rendered inoperable. It took Verizon technicians two
weeks of restoration work to reconnect two-thirds of the voice lines and about 80% of
the data traffic [Lee01].
All failures result in heavy financial burdens on the telecommunications carriers as
well as revenue loss by their customers. Furthermore, they result in diminished customer
confidence in the companies’ services. Recognizing the importance of protection against
failures, service providers (carriers) are now bound to service-level agreements (SLAs)
with their customers guaranteeing that the customer will be provided with services with a
prescribed network availability [Glossbrenner93] that is typically 99.999% (5 minutes of
downtime per year) [Bell87, Bell90b, Daneshmand+93, Fagerstorm+93, Somani+97].
Most recently, this requirement has become even more stringent, requiring carriers to
meet 99.9999% availability, which translates to 30 seconds of downtime per year.

8.1 Objectives of Protection and Restoration

It is obvious from the enumeration of different failures and their impact on the networks
described earlier that a fast and reliable carrier link and network equipment protection or
restoration method is required. In the current literature, protection and restoration have
different meanings. Protection refers to a preplanned system where a protection path
 Survivability 649

is precomputed for each potential failure (before the failure occurs) and the path uses
preassigned resources for failure recovery (dedicated for specific failure scenarios or
shared among different failure scenarios). In restoration, the recovery route is computed
in real time (after the failure occurs) and spare capacity available in the network is
used to reroute traffic around the failure. (This is typically used in mesh networks
using reconfigurable optical cross-connects [Ramamurthy+99a].) The objective of the
protection or restoration methods is to reroute the affected traffic accurately and rapidly
using the redundant capacity and equipment available in the network.
Even though failures cannot be avoided, quick failure detection, identification, and
recovery make the network more robust and reliable and ultimately increase the level of
customer confidence. For this to be possible, the network’s topology must have inherent
survivability properties, which determine its ability to survive single or multiple link or
equipment failures. For a network to be survivable, its topology must allow rerouting
around a fault condition. This issue falls under the category of network design and
planning. For example, for survivability in the face of single link failures, the graph of
the network must at least be two-edge connected.
Even if a network has a survivable topology, and robust operating conditions have
been established, redundancy is always necessary to ensure that a significant amount
of information is not lost in the case of fiber, equipment, or complete node failure.
Together with redundancy, rapid failure detection, identification, and recovery features
must be present. Service recovery time is important because faster recovery means that
less data are lost during the outage. Recovery speed is important not only because of
the economic impact of the outage time but also because of the vital services that are
currently supported in the network. Whereas previously an outage meant that a telephone
caller had to hang up and try again later, an outage nowadays may affect banks, stock
exchanges, airlines, or public safety.
Much of the early research in fault protection and restoration was concentrated on
point-to-point systems, self-healing rings (SHRs), and centralized/distributed restora-
tion in mesh networks using digital cross-connect systems (DCSs). This chapter first
reviews some of the important protection/restoration techniques currently used for net-
works employing SONET and WDM technologies. It then presents some new methods
developed in the past few years for protection in optical networks with mesh topolo-
gies. The chapter is organized as follows. Section 8.2 describes current fault recovery
techniques used in SONET and DCS networks. These are subdivided into point-to-
point and ring (and ring interconnection) protection techniques, and mesh restoration
techniques. Although the techniques presented in Section 8.2 focus on failure recov-
ery using SONET equipment and DCSs, similar approaches have been studied for
networks employing SDH, ATM, and a number of other technologies [Ayanoglu+96,
Grover04, Kajiyama+92, +94; Kawamura+94, +95; May+95, Nederlof+95, Sato+90,
Veitch+96].
Section 8.3 presents optical-layer protection techniques in networks employing WDM
point-to-point and ring architectures. Optical-layer protection in mesh topologies is
covered in Section 8.4, including shared optical-layer line-based protection, optical
path-based protection, and optical segment protection.
650 Multiwavelength Optical Networks

8.2 Current Fault Protection and Restoration Techniques in the Logical Layer

Because of its well-developed protection mechanisms and widespread deployment, we

use the SONET network architecture as an example of current practice. The SONET
terminology is adopted throughout this section, and the key terms are defined and
explained in Appendix F.
SONET standards specify an end-to-end two-way availability objective of 99.98%
for interoffice applications (0.02% unavailability or 105 min/year maximum down-
time) and 99.99% for loop transport between the Central Office and the customer’s
premises [Bell91]. To conform to these standards, failure recovery times have to be
short. For both point-to-point and SHR systems, automatic protection switching (APS)
is used, enabling the network to perform failure recovery in tens of milliseconds (of the
order of 50 ms to detect the failure and to complete the switching process).

8.2.1 Point-to-Point Systems

Protection in a network can be achieved by a system involving redundant capacity
and using physical-layer APS to switch traffic from failed to protection facilities. APS
is used to improve the reliability and availability performance of SONET transport
systems by switching to standby equipment when failures occur. Systems for APS can
be classified as one-for-one (1:1), one-for-N (1:N; more generally, M:N), or one-plus-one
(1 + 1) [Hall+89]. Any single fiber cut can be recovered using any of these techniques,
as explained later. To protect against equipment failures (e.g., switches or terminals),
spare equipment (protection equipment) is used. The protection equipment provided can
be one-for-one (i.e., one protection system for one working system) or one-for-N (i.e.,
one protection system for N working systems). These are the same techniques as for
link protection, with the difference that spare equipment is used instead of spare fibers.
They can protect against any single node failure in the network. It should be noted that
when dealing with equipment failures (e.g. switches, add/drop multiplexers, and line
interfaces) the particular piece of equipment that fails determines whether a complete
network node goes down or only a single link or channel fails. Normally, complete node
failure would occur as a consequence of a fault in an electronic or optical switch, but
only a single link or channel would go down in the case of a line interface or laser failure
respectively.
SONET linear APS is the standard defined for (1 + 1), (1:1), or (1:N) protection
architectures in point-to-point systems.2 An APS signaling channel is embedded in the
SONET line overhead and carried on the protection facility using the K 1 /K 2 bytes (see
Appendix F).
Although redundancy is used to perform fault recovery, proper design of the fault-
monitoring and protection process ensures that the network is able to recognize and
respond to fault conditions as soon as they occur. In SONET systems, BER thresholds
are established that activate and deactivate the protection switches, time windows are

2
SONET (1 + 1) or (1:N) protection may be unidirectional or bidirectional.
 Survivability 651

B C
A D

Path Protection
Line Protection

Figure 8.1 Path versus line protection.

specified for detecting a fault condition and activating the protection switch, and an APS
protocol is specified to ensure smooth operation of the protection switching process
[ANSI88, Bell90a, 91, 94a].

8.2.1.1 (1 + 1) Facility Protection

In SONET terminology, in the context of point-to-point systems, a facility or link or
line refers to a point-to-point transmission link with two terminals (i.e., without any
intermediate terminals). The route or physical path taken by the facility is the physical
structure on which the link is laid out. It may traverse several Central Offices, being
“spliced” through these offices without terminating any data streams at those points. In
contrast, a path in the SONET world is a tributary (e.g., STS-1, DS3, or VT1.5) within
a SONET frame. Sometimes the word channel is used in place of path. A path may
traverse many nodes in a large network.
Figure 8.1 illustrates the difference between SONET line and path protection. To
clarify the terminology, we use the term logical path to refer to the sequence of nodes
and links in the figure and SONET path to refer to an end-to-end SONET tributary.
Line protection simply reroutes traffic normally carried between the line terminating
equipment at each end of a failed link to an alternate physical structure, detouring
the traffic around the failed link. Path protection (a higher level operation) allocates
an alternate protection SONET path between the terminal nodes (the path terminating
equipment) of a failed path. (Path failure may be caused by the failure of one or more
links or nodes.) Figure 8.1 shows both line and path protection associated with a failed
link, B–C. A SONET path is active on the logical path A–B–C–D between the path
terminating equipment at nodes A and D. In this example, for line protection, all traffic
that normally would have traversed the failed link between nodes B and C is rerouted
over the dashed protection facility, detouring the link failure. This includes the SONET
652 Multiwavelength Optical Networks

Bridge Switch
W
TX RX

TX RX
P

Figure 8.2 (1 + 1) SONET protection.

path in question, as well as any other paths using that link. The reader should note that
even though in this figure all demands affected by the failure follow a single reroute
under line-based recovery, in general, line-based recovery does not limit all the affected
demands to a single reroute. Rather, the affected demands can use many different routes
between the end points of the failure.
In path protection, the SONET path between A and D detects the failure and is rerouted
on an alternate logical path, shown as the solid line. (Other paths using the failed link
would detect the failure independently and be rerouted on their own protection paths.)
Note that path protection requires many more operations than line protection because
many SONET paths generally share each link. This is a basic difference between path
and line protection. In path protection the unit being recovered is a connection occupying
a single channel (which may be anything from a SONET tributary to a full λ-channel)
on a failed link. In line protection the unit being recovered is the aggregate traffic on
the failed link (typically carried on many parallel channels), and it is recovered via a
single reroute or multiple different reroutes around the end nodes of the failed link. As
a result, the signaling and switching operations in path protection are far more complex
than those in line protection.
In a (1 + 1) protection architecture, a protection facility is provided for every working
facility, and the working and protection facilities are on two physically diverse routes
between source and destination. Thus, if a fiber is cut, the signal uses the second facility
to reach the destination. The traffic is bridged permanently at the electrical (i.e., logical)
level to both the working and protection facilities at the transmit end (head end). At the
receive end (tail end), the same data stream arrives from both facilities, and the received
signals are continuously monitored identically and independently for failures [ANSI88,
Bell93, 94a]. The receiving equipment then selects either the working or protection
optical signal based on the switch initiation criteria. In the event of a link failure, an
electrical switch at the receiver end immediately switches to the “other” facility, thus
effectively restoring the service with minimum outage time. In this scheme, because
the signal is bridged at the source, head-end-to-tail-end signaling is not required. The
receiver simply selects the best of the two arriving signals. Figure 8.2 shows an example
of a (1 + 1) electrical layer protection architecture. (In the figure, TX, RX, W, and P
denote transmitter, receiver, working, and protection, respectively.)
The default mode of operation for a (1 + 1) system is the nonrevertive mode. This
means that the receiver does not switch back to the working facility (the path along which
 Survivability 653

Switch Switch
W
TX RX

TX RX
P

Figure 8.3 (1:1) SONET protection.

it was receiving signals before the failure occurred) after the fault condition is cleared.
Operation in the revertive mode, when the system reverts to its original configuration
after the fault is cleared, is optional [Bell93, 94a].
For 100% survivability after a single link failure, each working fiber system needs a
diverse route. Thus, as many diverse routes as working fiber systems are required. The
physical path for the protection facility may be chosen to be the shortest path that is link
disjoint from the working path.

8.2.1.2 (1:1) Facility Protection

A (1:1) APS system provides two diverse routes for the same connection so that if a
fiber cut or a node failure occurs, the signal can reach the destination using an alternate
physical path. The signal at the transmitter end is not bridged permanently but is switched
onto a diversely routed facility depending on the state of the network. On detection of a
failure, the receiver end signals the transmitter end, indicating that a failure has occurred,
and requests the transmitter end to switch to the protection facility. The transmitter end
complies with the request, switches the service to the protection facility, and sends a
message to the receiver end that the service has been switched. The receiver end then
switches to the protection facility and the service is recovered. In the case of (1:1) APS
systems, all switching is revertive (i.e., the head and tail ends switch back to the working
channel after the failure condition is cleared [Bell93, 94a]).
In SONET linear APS, the APS signals are exchanged over the protection facilities
between SONET line terminating elements (LTEs). The (1:1) architecture allows the
protection facility to carry low-priority traffic in the absence of a failure. This low-
priority traffic is preempted in the event of a failure.
As in the (1 + 1) scheme, 100% survivability after a single link failure in a (1:1)
facility protection architecture requires as many diverse routes as working fiber systems.
The route for the protection facility can again be chosen to be the shortest path that is link
disjoint from the working path. Figure 8.3 shows an example of a (1:1) SONET layer
protection architecture. Note that the optical transceivers are included in the links being
protected, so a transmitter or receiver failure is detected as a link failure and is restored.

8.2.1.3 (1:N) Facility Protection

The (1:N) protection method is the same as the (1:1) facility protection method except
that in this case one protection facility protects N working facilities; (N working channels
654 Multiwavelength Optical Networks

Switch Switch
W1
TX RX

W2
TX .. RX
.. . ..
. .
WN
TX RX

P
TX RX

Figure 8.4 (1:N) SONET protection.

can be switched to a single optical protection channel). Permissible values of N are from
1 to 14 [ANSI88]. Again, as in the case of (1:1) APS systems, all switching is revertive.
In the case of two or more simultaneous failures, there is contention for the use of the
protection facility. This contention is resolved using priorities assigned to each working
channel [Bell90a, 93].
Figure 8.4 shows an example of a (1:N) electrical layer protection architecture. As
in the (1:1) case, the protection line can be used to carry low-priority traffic. This
traffic is preempted when one of the working facilities initiates a protection request
[Bell90a].

8.2.2 SONET Self-Healing Rings

The (1+1) and (1:1) protection techniques described in the previous sections use dedi-
cated, diversely routed physical paths to protect traffic against failures. Such techniques,
however, impose penalties because of the extra capacity requirements.
Architectures composed of SONET ADMs interconnected in a ring topology provide
an alternative method of APS that allows facilities to be shared while protecting traffic
within an acceptable recovery time. Because these SONET rings protect automatically
against failures, they are called self-healing rings. For SONET SHRs, the typical pro-
tection process takes of the order of 50 ms. The importance of the SHR architectures
stems from their ability to recover all traffic quickly in the event of a cable cut and part
of the traffic in the event of a node failure.
Ring network architectures provide protection against service interruptions by hav-
ing two diverse facility paths from each SONET ADM. If a cable is cut between any
two ADMs, sufficient redundant capacity exists to reroute the affected circuits to their
destinations. With APS, this rerouting can be done with minimal service interruption.
Ring architectures can also reroute traffic around a failed node. However, traffic origi-
nating or terminating at a failed node cannot be restored unless there is appropriate node
redundancy or diversity.
There are two types of SONET SHRs: the UPSR (unidirectional path-switched ring)
and the BLSR (bidirectional line-switched ring). In a UPSR, in the absence of a fault,
 Survivability 655

the traffic is always transmitted in both directions around the ring (based on the [1 + 1]
concept). In a BLSR, in the absence of a fault, only the working traffic is transmitted
[Bell94a]. Comparison studies between the different architectures show that there are
different areas of application in which each proves economical [Sosnosky+91]. Typi-
cally, unidirectional architectures are used in the collector portions of a network (where
traffic is concentrated before it enters the core), and bidirectional architectures are used
in the core part of a network.

8.2.2.1 Unidirectional Path-Switched Ring Architecture

The UPSR is a (1 + 1) single-ended, unidirectional, SONET path layer dedicated pro-
tection architecture. The nodes are connected in a ring configuration with one fiber pair
connecting adjacent nodes. One fiber on a link is used as the working fiber and the other
one is dedicated for protection. The working and protection fibers operate in opposite
directions, so there is a working ring in one direction (e.g., clockwise) and a protection
ring in the opposite direction (i.e., counterclockwise). For each SONET path, an STS or
virtual tributary path signal is bridged permanently at the head end (i.e., the source or
transmitter) to both outgoing fibers (working and protection) and is transmitted in both
directions around the ring [Bell94a]. The tail end (i.e., destination or receiver) compares
the incoming signals continuously at the SONET path layer and selects the better of the
two based on signal quality. If a fiber is cut (or other transmission defects are detected)
or equipment (a node) fails, the tail end just switches to the data stream coming from
the opposite direction and the failure is recovered automatically. Because each end op-
erates independently, and protection recovery is solely the responsibility of the tail-end
receiver, no notification or coordination with the head end is required. Thus, no APS
signaling channel is required between the ring nodes, and thus the UPSR architecture is
simple to operate and has a fast recovery time. This architecture can work with either
revertive or nonrevertive path switching. Nonrevertive switching minimizes the amount
of information lost during the switching process because it switches only once due to
a failure. However, revertive switching has the advantage that a default position always
exists and is known. Clearly, nonrevertive switching is preferred when SONET systems
with a large bandwidth are used, because even the smallest outage due to the switching
time causes the loss of a large amount of information.
In the case of a bidirectional connection between two nodes on the ring, both directions
of the working traffic travel in the same direction around the ring. The second (protection)
ring carries duplicate copies of the path signals for both connections in the opposite
direction. Thus, the full capacity of the protection ring protects the full capacity of
the working ring. The total capacity available for carrying services around the ring is
therefore limited to that of one fiber and is shared by all nodes on the ring [Shyur+94].
For example, in an OC-48 UPSR, the aggregate traffic injected into the ring by all
nodes cannot exceed 48 DS3s. If a fiber cable is cut, only one direction of the two-way
connection is affected and, for that direction, path switching is performed at the tail end.
The other direction is unaffected by the failure, so no path switching is necessary.
The UPSR architecture can survive a single cable cut. If more than one cable is cut,
the connectivity between the ring nodes could be disrupted and some traffic could be
656 Multiwavelength Optical Networks

lost. The amount of traffic lost depends on the connectivity pattern at the time of the
faults and the location of the multiple faults. Other fault recovery techniques at higher
layers can then be used to provide alternate routes for the connections that could not be
recovered using path switching.

8.2.2.2 Bidirectional Line-Switched Ring Architecture

In a BLSR, a bidirectional connection between two nodes traverses the same intermediate
nodes and links in opposite directions. A channel used in that portion of the ring can
then be reused in another nonoverlapping part of the ring. Thus, the BLSR architectures
have an advantage over the UPSR architectures in terms of the total traffic supported on
the ring.
A BLSR is a bidirectional line-based shared protection architecture. In contrast to
the UPSR, in which the protection capacity is dedicated, the BLSR shares protection
capacity among all the spans on the ring. Failure recovery is performed at the line layer
(a lower layer than the path layer) and involves no path-layer functions. Two types of
BLSRs are defined in the SONET standard: two- and four-fiber BLSRs.

Two-Fiber Bidirectional Line-Switched Rings

In the two-fiber BLSR, adjacent nodes are connected using a single fiber pair. Each fiber
is designated as a working fiber, but half the signal bandwidth on each fiber is dedicated
for protection [Bell95]. For example, a two-fiber OC-48 BLSR has 24 of the STS-1s
assigned to working traffic and the other 24 (50% of the capacity) assigned to protection.
If less than 50% of the capacity is reserved for protection, not all traffic can be recovered
in the event of a failure.
All protection switching is accomplished in the electronics of the SONET ADM
via time slot interchange (TSI). Each working slot on one fiber is preassigned to a
protection time slot on the fiber going in the opposite direction. For example, for an
OC-N ring, time slots 1 through N/2 can be used for working traffic on both fibers
of an internodal link, with time slots N/2 + 1 through N used for protection traffic
[Bell95]. In this case, time slot i of one fiber is protected with time slot N/2 + i in the
other. When a failure occurs, the working traffic from one fiber is looped back onto the
opposite direction, carrying the working traffic in the opposite direction around the ring
away from the failure [Bell95, Harman+94]. As in the case of (1:1) point-to-point APS,
a signaling protocol using an APS channel carried in the SONET line overhead allows
the ring nodes to effect a coordinated line switch when failure conditions occur.
The same procedure works in the case of a node (ADM) failure. In this case, only
traffic that passes through the node can be recovered. Traffic that is added or dropped at
the failed node is lost. A complete node failure has the same effect as the failure of all
the fibers connected to that node. The line switching is performed at both sides of the
failed node to redirect traffic away from that node. Because the traffic that is added or
dropped at a failed node cannot be recovered, it may be misdirected to other erroneous
destinations. To deal with this contingency, squelching has been proposed [Bell95],
wherein a path alarm indication signal (AIS) is inserted into these channels instead of
the real traffic. This procedure can also be used when multiple links or nodes fail, and in
 Survivability 657

general in all cases in which traffic cannot be recovered but may be misdirected because
of failure conditions in the network.
In BLSR architectures, switching is coordinated by the nodes on either side of a failure
in the ring, so that a signaling protocol is required to perform a line switch and to recover
from the failure. Thus, BLSR architectures are more difficult to operate than UPSRs (in
which no signaling is required). The protection switching time with BLSRs depends on
the time it takes for the signaling protocol to be completed.

Four-Fiber Bidirectional Line-Switched Rings

In a four-fiber BLSR, four fibers connect each adjacent pair of nodes, and the nodes are
connected in a ring configuration. Two fibers are now used as working fibers (one in
each direction) and two fibers are designated as protection fibers (one in each direction).
In this case, the fibers are dedicated completely as either working or protection fibers.
Thus, the ring capacity of a four-fiber BLSR is twice the capacity of a two-fiber BLSR.
A bidirectional connection between two nodes travels on separate fibers (opposite di-
rections) through the same intermediate nodes (typically on the shortest path around the
ring).
As in the SONET linear APS architecture, the four-fiber BLSR APS channel is carried
in the K 1 and K 2 bytes on the protection fibers. If both working and protection fibers
are cut on the short path between two nodes, the service is recovered using the long
path. As in the two-fiber BLSR case, traffic that would normally be carried on a working
fiber in the severed link is looped back on protection fibers to direct it away from the
failure. The four-fiber BLSR can carry twice as much traffic as the two-fiber BLSR, but
it requires twice as much fiber and electronics. An additional advantage of the four-fiber
architecture is that it can support both ring and span switching. Ring switching was
described earlier; it occurs when both the working and protection fibers are cut (cable
cut). In this case, the protection switches on both sides of the cut engage, and the traffic
is routed away from the cut with the help of an APS signaling channel. Span switching
occurs when only the working fiber fails. In that case, the traffic is switched from the
working to the protection fiber in the same direction as the working traffic [Bell95]. This
architecture is preferred by long-haul carriers.
If a node fails, only the pass-through traffic can be restored. Squelching is used to
account for the misdirected traffic in the case of node or multiple failures.

8.2.3 SONET Self-Healing Ring Interconnection Techniques

Figure 8.5 shows two possible interconnection configurations for SONET SHRs: single-
and dual-access. For the single-access configuration (Figure 8.5[a]), two rings are joined
by a single-access node, and for the dual-access ring configuration (Figure 8.5[b]),
the ring interconnection point uses two common access nodes [Wu92]. Architectures
consisting of many interconnected rings can be formed by employing several intercon-
nection points on each ring. For the single-access ring interconnection configuration, a
single node failure may disable the communication among the (possibly multiple) rings
that are interconnected through this node.
658 Multiwavelength Optical Networks

(a) (b)

Figure 8.5 Single- and dual-access ring interconnection configurations.

The dual-access ring interconnection configuration avoids this problem by providing

two points of contact between the rings, termed the primary and secondary nodes.
In dual-access SONET SHR architectures, the traffic is dual fed to both rings using
appropriate hardware and software techniques. Because dual access is currently the
most popular method, the discussion that follows focuses on this technique.
One method for interconnecting SONET rings uses a drop-and-continue function,
together with a path selector and/or a service selector function. The drop-and-continue
function drops the traffic at a primary interconnecting node and a copy of that traffic
continues to a secondary node where it is also dropped (i.e., the signal is bridged perma-
nently and dual fed to the interconnecting nodes). The path selector function allows an
interconnecting node to choose the best of two incoming signals independently to for-
ward to the next ring. Drop-and-continue and path/service selector functions allow
inter-ring traffic to survive coincident single fiber and node failures on both rings
[Bell94b, Harman+94].
The following section illustrates the use of these interconnection techniques for
SONET SHR architectures when either the same or different protection schemes (UPSR
or BLSR) are used in each ring. The interconnections ensure that both the intra-ring
traffic (traffic that stays within one ring) and the inter-ring traffic (traffic that crosses
from one ring to the other) is protected in the event of a failure. For a more com-
plete treatment of ring interconnection, see [BeHannesey98, Ester94, Harman+94,
Siller+96].
 Survivability 659

8.2.3.1 Interconnected UPSR–UPSR Architectures

Interconnected UPSRs provide path protection as described in Section 8.2.2.1 for intra-
ring traffic, and protect inter-ring paths using the drop-and-continue feature. The pro-
tection mechanism is independent in each ring. Drop-and-continue together with path
selection ensures that two copies of the signal are maintained between the rings [Bell94a].
In the event of a failure affecting one copy of the signal, traffic between the two rings is
recovered automatically using the other copy, as in the case of single UPSRs.
Figure 8.6 shows how a bidirectional SONET path between nodes in different UPSRs
is protected. The traffic is duplicated at the transmitter as in a single UPSR, and the
two copies are routed on different segments of the ring (in opposite directions). When
it reaches the interconnection nodes, each copy is duplicated once more using the drop-
and-continue function. The two signals reaching each path selector are those that were
routed on different ring segments, so a failure on one link in the ring causes only one of
the two inputs to the path selector to drop out. The path selectors choose the best of their
two input signals and inject them onto the second ring. Thus, two copies of the original
signal enter the second ring on two different fibers traveling in different directions. This
scheme provides protection from single failures on the interconnecting links and nodes.
Recovery from a failure is therefore automatic, because the path selectors respond to a
working signal degradation or failure and choose the signal coming from the protection
path. The path protection technique just described is executed on a path-by-path basis
for each SONET path traversing the two interconnected rings.
Note that the interconnection between the two rings is performed via SONET inter-
faces. This allows the alarms and maintenance signals to be preserved for connections
crossing over from one ring to the other.

8.2.3.2 Interconnected BLSR–BLSR Architectures

As in the case of interconnected UPSRs, the protection mechanisms in interconnected
BLSRs are independent for each ring. A drop-and-continue function and a service
selector (rather than a path selector) are now used to protect the traffic on the rings.
Figure 8.7 illustrates the interconnection of two BLSRs. Again, the drop-and-continue
function is used to duplicate the traffic at the interconnection points, and the two copies of
the signal are injected into the second ring from two different points. A service selector at
the destination node monitors the two duplicate signals. If it detects a failure, it switches
immediately from the current channel to its duplicate. Link and node failures within each
ring are protected by rerouting over protection fibers as described in Section 8.2.2.2.

8.2.3.3 Interconnected BLSR–UPSR Architectures

In this architecture, techniques used in both the UPSR and the BLSR are utilized to
protect the traffic against failures. Drop-and-continue is used in both rings, and service
and path selectors are used (in the BLSR and UPSR rings, respectively) to recover the
traffic if a failure occurs.
Figure 8.8 demonstrates the interconnection between a BLSR and a UPSR. For
a connection from a BLSR to a UPSR, the drop-and-continue function at the
660 Multiwavelength Optical Networks

PS

UPSR

PS
PS

PS

PS

UPSR

PS

PS Path Selector
Drop-and-Continue

Figure 8.6 Two-fiber UPSR-to-UPSR ring interconnection.

 Survivability 661

BLSR

SS

SS

BLSR

SS Service Selector

Figure 8.7 BLSR-to-BLSR ring interconnection.

interconnection points on the BLSR side creates two copies of the traffic that subse-
quently cross over to the second ring. The interconnection nodes in the UPSR transmit
the two copies of the traffic around the ring in opposite directions. A path selector at
the destination node chooses one of the two signals. A slightly different technique is
used when the signal travels from a UPSR to a BLSR. The signal is duplicated, using
662 Multiwavelength Optical Networks

BLSR

SS

PS

PS

UPSR

PS

Figure 8.8 BLSR-to-UPSR ring interconnection.

the drop-and-continue function at both interconnection points on the UPSR side. A path
selector in each interconnection node compares two copies of the signal from the UPSR,
choosing the best to pass on to the BLSR. A copy of the signal enters the BLSR from
each interconnection node, and a service selector at one of the nodes selects one of the
two signals to continue its path in the BLSR.
 Survivability 663

As explained, in the case of interconnected SHRs, the self-healing mechanisms are

independent for each of the interconnected rings. The interconnection is effected in
a way that enables the rings to recover individually from a single failure in each ring.
However, multiple failures may result in the loss of the services that traverse several
rings.
For interconnected SONET SHRs, the problem of the ripple effect has to be taken into
account. This effect occurs due to the propagation of a path alarm indication signal (AIS)
from one ring to another if the self-healing mechanism in the ring in which the failure
occurred does not engage quickly enough. The path AIS then causes the self-healing
mechanism to engage in rings in which a failure has not occurred. This effect can occur
in UPSR–UPSR and BLSR–UPSR interconnected rings. It is not a problem in BLSR–
BLSR networks, in which path AIS is not detected. Because the ripple effect can increase
the service recovery time as well as cause protection switching to engage in rings that do
not have failures, it may compromise the “independence” of the various interconnected
rings. This is especially troublesome when different rings belong to different service
providers [Bell94b, Harman+94].

8.2.4 Architectures with Arbitrary Mesh Topologies

Various methods have been proposed to provide restoration and protection in the logical
layer for networks that have mesh topologies. In this section we provide an overview of
some of them. There are two basic approaches: (1) a generalization of ring protection
techniques using ring covers and (2) dynamic restoration based on reconfiguring DCSs
at the network nodes to respond to failures. The latter approach may be implemented in
either a centralized or a distributed fashion.
A ring cover is defined as a set of closed paths that covers all links collectively in the
network at least once. If these rings are composed of bidirectional fiber links, they can
be made into self-healing structures that can protect the network against a cable cut or
an equipment failure using the APS path or line techniques described in Sections 8.2.2.1
and 8.2.2.2. Failure recovery using the ring cover method can be achieved in a matter
of milliseconds. A negative feature of ring covers is that they may require a significant
amount of extra redundancy to ensure 100% protection against a link or an equipment
failure (see Section 8.2.4.1). The reader must take note, though, that the ring cover
techniques alone do not address the issue of routing the demands on these rings, which
means that a ring cover approach by itself will not produce a minimum total network
cost solution. The joint problem of routing and ring placement for minimizing the total
network cost is more complex and several attempts to address this problem are briefly
discussed in Section 8.2.4.1.
In the DCS restoration approach, when a failure occurs, connections are switched from
the failed links and rerouted over others having spare capacity. Using centralized control,
the reconfiguration necessary for restoration is computed by a central network manager,
which issues commands to the various nodes involved in the restoration process. With
distributed control, the restoration process is handled in an autonomous manner by the
individual DCSs through the exchange of signaling messages. Distributed restoration
664 Multiwavelength Optical Networks

requires complex protocols and high-speed signaling between the nodes as well as
enhanced computational capabilities at each node.
The main disadvantage of DCS restoration techniques is that the reconfiguration time
may not be fast enough to meet user demands for service availability. An advantage is that
they require much less spare capacity than ring-based protection schemes. Typically, the
time required for the restoration of traffic in a mesh architecture using DCS equipment
might be on the order of minutes if a centralized restoration algorithm is used and on
the order of seconds if a distributed restoration algorithm is used. Although centralized
restoration is much slower than distributed approaches, it can restore the network more
efficiently (in terms of overall network capacity requirements) because the network
manager has a global view of the network. Thus, DCS restoration techniques have a
place in networks in which the recovery speed is not as important as the redundant
capacity required for survivability.

8.2.4.1 Mesh Architectures Using Ring Covers for Survivability

A simple choice of a ring cover for a network is a set of fundamental cycles (see
Appendix A). The fundamental cycles “cover” the graph corresponding to the network
in the sense that each edge is included in at least one cycle in the set. Each fundamental
cycle can be treated as an independent SONET SHR. The mesh network then “behaves”
as a collection of SHRs, providing the advantages of simple control logic and speed in
terms of failure recovery.
Any ring cover of the network requires at least 100% redundancy in terms of protection
fibers, where fiber redundancy is defined as the ratio of protection fibers to working
fibers.3 But fundamental cycles, if they are not chosen carefully, may require much
more than 100% redundancy, because a link included in multiple cycles must carry
multiple protection fibers – one for each cycle traversing it.4 A more general approach to
“covering” a network uses cycle decomposition. This is applied to optical layer protection
in Section 8.4.5
Because there are many different sets of fundamental cycles in a network, the design
of “optimized” multiring networks is a complex problem. A number of algorithms have
been developed to find a ring cover for a network topology. The goal of these algo-
rithms is to find the minimum cost ring cover for any given arbitrary mesh network
to minimize the required protection capacity. Various studies have been performed in
planning [May+91] and analyzing multiple SHR networks. Slevinsky [Slevinsky+93]
present an algorithmic approach that is near optimal in minimizing the required pro-
tection capacity. A related problem involves choosing a ring set so that most of the
traffic has its source and destination in the same ring. The problem is addressed but
not solved by [Slevinsky+93]. Gardner, in [Gardner+94], also address the ring cover
3
This assumes the four-fiber ring approach. With a two-fiber ring, the extra capacity would be in the form of
spare channels.
4
In the special case of Eulerian networks, an “ideal” ring cover can be found in the form of bidirectional ring
decomposition of the network. In this case, each link is traversed by exactly one ring (see Section 6.3.5).
5
A cycle decomposition of a graph G is a set S of pairwise edge–disjoint cycles of G such that every
edge e in G belongs to precisely one element of S. If G is Eulerian then G has a cycle decomposition S
[Chartrand+96].
 Survivability 665

problem and describe an algorithm that finds a minimum-cost ring cover for specific
classes of networks, based on the Eulerian approach. The algorithm finds a ring cover
that minimizes the equipment cost required to guarantee 100% recovery from a sin-
gle link failure in arbitrary unidirectional rings. They also propose three heuristics
that solve the same problem for bidirectional rings. A software tool called NetSolver,
introduced by [Gardner+95], finds good ring covers and the routing paths for the point-
to-point traffic in the network. It starts from an arbitrary ring cover and improves on it
through a series of iterations. Other studies have also addressed the problem of analyzing
ring cover designs. In [Wasem91a,b] and [Wasem+92], an algorithm for routing rings
in a network is presented for the case when the topology of the network, the traffic
matrix, and the nodes in each ring are known in advance. This algorithm can create
cost-effective designs and increase network reliability at the planning stage. Finally, in
[Ellinas+00a] an approach similar to [Gardner+94] is developed where a ring cover of
a network is obtained by transforming the graph representing the network to an Eulerian
graph (if the original graph is non-Eulerian) and performing a cycle decomposition
of the transformed graph by adding or removing certain edges from the original net-
work graph. The cycle decomposition of the transformed graph, in conjunction with
the addition of a number of extra edges (required to accommodate the edges removed
or added during the transformation process) results in a ring cover for the original
graph.

8.2.4.2 DCS-Based Centralized Restoration Techniques

Various centralized schemes have been proposed that use on-line computation (assum-
ing global knowledge of the network state) for failure restoration. These techniques
assume a restoration time of a few minutes (compared with a few hours for manual
restoration).
The FASTAR system used in AT&T’s network operates at the DS3 transport network
level and combines centralized and distributed processes to provide failure restoration
[Chao+91]. FASTAR uses a central processor that receives and correlates alarm reports
and calculates, reroutes, and controls the restoration path implementation. The central
processor maintains an extensive database, including the network layout, cross-connect
data, and the status of the network links. Local controllers are also used in each node for
alarm gathering and reporting as well as various test functions. The controllers signal
network conditions directly to the central processor.
For the FASTAR system, the restoration process is as follows. Alarm gathering is
performed at various points in the network (e.g., line-terminating equipment and DCS
equipment), and these alarms are reported to the local controllers, which correlate them
and pass the information to the central processor. The central processor waits until it
receives all possible alarms, prioritizes the failed DS3s, and calculates the appropriate
rerouting for each one. It then activates the restoration paths according to their priorities.
Before transferring services to the restoration path, the central processor makes sure that
the transfer does not disturb any current DS3s that are unaffected by the failure. After
the service transfer, it also checks to verify that the transfer to the restoration path was
successful.
666 Multiwavelength Optical Networks

Distributed intelligence is introduced into the network for fast alarm reporting. How-
ever, the central processor is essential for prioritizing the failed connections and rerouting
them in an efficient manner. This approach is not as fast as the distributed schemes de-
scribed next but is satisfactory for users who do not require extremely fast restoration
capabilities.
In the NETSPAR protocol [Tsai+90], the DCSs reroute the traffic (on the DS3 level)
from failed physical facilities to surviving facilities with spare capacities. After a failure
is detected, the surviving DCSs use a distributed protocol to determine which of the
facilities are working. Based on these results, each DCS loads a new configuration table,
rearranging the network to restore as much of the traffic as possible. These configuration
tables can be calculated online or even off-line. In the case of off-line computation (for
example, for all single link faults) at a central node, the tables will be saved at each
individual DCS and used when a failure occurs. The latter technique will be a mixture
of protection and restoration schemes.
This is a combination of a centralized and a distributed approach. It uses a dis-
tributed topology update protocol to identify a fault and then downloads a precomputed
rerouting table for that fault. Because it uses static, precomputed tables to achieve the
reconfiguration of the network after a failure, the NETSPAR scheme is faster than
the dynamic approaches presented in Section 8.2.4.3. However, the technique becomes
cumbersome when the networks considered have a large number of links and nodes
[Coan+91]. Furthermore, the static nature of the rerouting tables presents problems
when there are network topology changes, because all precomputed tables must be
recalculated.
Centralized restoration requires a complete and accurate database of the state of the
network at all times and has restoration times of the order of minutes. This means that
these techniques cannot be used by themselves to protect networks that carry applications
with large bandwidth requirements.

8.2.4.3 DCS-Based Distributed Restoration Techniques

Faster survivability schemes based on distributed restoration have been introduced to
avoid the problems encountered by the centralized techniques.
A technique, called the SelfHealing Network, was introduced by [Grover87] for dis-
tributed, real-time restoration of high-capacity telecommunication transport networks.
In this system, complex restoration plans are computed and are put into effect in real
time in an entirely autonomous, distributed manner.
The SelfHealing process uses the DCS equipment in a SONET or DS3 transport
network along with additional software features to achieve restoration of complete cable
cuts in 1 to 2 seconds. These numbers correspond to the typical connection drop times.
(If the traffic affected by a failure can be rerouted to the spare capacity in less than the
connection drop times, the connections can be resumed.) This technique was the first
attempt to use DCS equipment to support capacity-efficient restoration in real time and
in a distributed manner without depending on large databases. It uses parallel, distributed
interaction between the DCSs to achieve a very fast restoration mechanism.
 Survivability 667

The restoration process has two main components: (1) a network signature flooding
wave and (2) a reverse signature linking sequence. After a failure occurs (e.g., a cable
is cut), alarms are generated and are detected by the DCS equipment. The two nodes
at the ends of the cable cut are aware of the failure and one becomes the chooser,
whereas the other one becomes the sender (predefined arbitrarily). The sender node
then starts broadcasting appropriately indexed restoration signatures (signature flooding
wave) out of its DCS. These signatures arrive at neighboring DCSs, which in turn rein-
dex the restoration signatures and rebroadcast them to their neighboring nodes. Only
spare DS3s are used for this process. Therefore, one or more signatures arrive at the
chooser using the spare capacity in the network. The chooser then sends out a com-
plementary signature on a link having a preferred arriving signature (reverse signature
linking). This signature goes through a number of tandem nodes and finally arrives back
at the sender. In this manner, a bidirectional connection between the sender and the
chooser is established using spare capacity. This connection is used to restore the failure
condition.
Multiple failures can also be restored using a slightly modified procedure. The basic
SelfHealing protocol remains the same, but additional features are added to ensure
that a node can now perform simultaneous SelfHealing tasks without experiencing any
adverse interactions between concurrent tasks. Studies were performed to show that
two SelfHealing tasks can be executed concurrently in a node to achieve restoration
simultaneously.
The performance of the SelfHealing Network protocol was tested for various networks.
It was shown that the scheme restored successfully any single failure in 2.5 seconds or
less, and it found all the paths of an ideal k-shortest path restoration solution in one
iteration [Grover+90].
The FITNESS (Failure Immunization Technology for Network Service Survivability)
network protocol [Yang+88] uses DCSs in a SONET network to restore failures via
distributed computation in less than 2 seconds. It uses the principle of the sender and
chooser as the SelfHealing protocol and is based on the SelfHealing approach. The
FITNESS protocol uses a selective flooding mechanism to limit the number of messages
exchanged among the nodes. It attempts to find the restoration path with maximum
available bandwidth. Several iterations are tried within a 2-second window, with the
chooser responding only to the best route. In this protocol, the communication between
nodes is done not through signatures but via the SONET embedded operations channel.
As indicated, it requires multiple iterations to find all possible restoration routes. In
contrast, the SelfHealing Network finds all required k shortest restoration paths in a
single iteration (distributed process).
The RREACT distributed restoration protocol [Chow+92] again uses the sender-
chooser method to perform the restoration function. It uses no prior network status
or topology knowledge and supports multiple simultaneous link failure restorations.
RREACT finds all the paths between the disrupted nodes by attaching path information
in the broadcast request messages. This algorithm results in a significant increase in the
number of message exchanges as the network grows. Contrary to the FITNESS protocol,
668 Multiwavelength Optical Networks

it needs only one wave of request messages to perform a link restoration. The protocol
also provides efficient utilization of network spare capacity resources. RREACT outper-
forms FITNESS in terms of both restoration time and spare capacity usage. Simulation
results showed that it outperformed other similar restoration techniques, performing the
restoration in less than 1 second.
Other similar restoration techniques have also been proposed. They are all based on a
simple flooding algorithm to achieve the desired failure restoration. When a node detects
a failure, it broadcasts a restoration message to adjacent nodes to find an alternate route.
Every node except the sender and the chooser responds by rebroadcasting the message.
When the restoration message reaches the chooser, the chooser responds by returning
an acknowledgment to the sender. In this way, alternate restoration paths are found. By
limiting the number of hops for the messages, the process can limit the distance these
messages can travel in the network. [Bicknell+93] compared six algorithms in terms
of restoration time, level of restoration (number of affected working channels restored),
utilization of spare channel resources, range of application (different kinds of failure
cases to which the algorithm can respond), and message volume (number of restoration
messages generated). FITNESS was the slowest of the algorithms because it uses a
“multiwave” approach to find the restoration path. However, it has the lowest message
volume because it chooses the restoration path with the largest bandwidth during each
wave of messages. All algorithms restored most of the working traffic lost, and RREACT
had the best capacity utilization.
In [Chujo+90], a flooding algorithm is used that finds the shortest alternate route. As
in the case of the FITNESS protocol, it requires multiple, successive iterations to find
all the possible alternate routes. Follow-up work in [Komine+90] uses multidestination
flooding (defining multiple choosers) and path route monitoring to address multiple link
and node failures [Bicknell+93].
The NETRATS protocol (Network Restoration Algorithm for Telecommunication
Systems) is described in [Sakauchi+90]. It provides both line and path restoration.
During the line restoration process, a procedure similar to the SelfHealing Network and
the FITNESS protocol is followed. In the path restoration process, a new procedure
similar to that used for line restoration is introduced. Because the path restoration
approach releases each disrupted channel, and the source and destination ends of the
channel reestablish their connections, path restoration requires more time than line
restoration. It is also more complex to implement and requires more network information
and more switching. However, path restoration provides a better solution than line
restoration in terms of the number of connections restored and the number of spare links
used. Also, the path restoration procedure can handle both node and link failures.
Iraschko et al. [Iraschko+96] investigated the problem of capacity optimization for
path restoration networks and attempted to quantify the benefits of path over line restora-
tion in terms of capacity requirements. Results showed that path restoration performed
significantly better than line restoration for several case network studies (up to 19% less
capacity required for recovery).
The distributed mesh restoration techniques presented in this section are better than
the APS point-to-point and SHR techniques presented earlier in terms of capacity
 Survivability 669

requirements, but they are slower than protection techniques. Clearly, there is a trade-
off between restoration speed and spare capacity requirements. Rings require a 100%
duplication of working capacity for full protection. Ring covers generally require quite
a bit more. (The p-cycle approach, presented in Section 8.4.1.3 is the only ring-based
technique that requires less than 100% redundancy.) However, mesh techniques not based
on SHRs can provide full survivability capabilities with redundancy that is typically
proportional to the inverse of the average degree of the nodes in the network. Redundancy
is defined here as the ratio of spare to working capacity. For example, for networks with
an average node degree between 3 and 4.5, 30 to 50% redundancy may be enough for full
survivability capabilities. Case studies presented in [Grover92] demonstrate that a fully
survivable mesh network requires approximately one-third the total link redundancy of
a ring cover design.
Various studies have been made to determine the minimum amount of spare capacity
needed for restoration in mesh-restorable networks. The applications are in networks
where cable cut restoration is possible via rerouting around the fault, using either cen-
tralized or distributed restoration algorithms [Grover+91]. These studies are useful in
designing the network to ensure full survivability with minimum redundancy. When
applied to various test networks, the studies show that 100% survivable mesh networks
are feasible with 50 to 70% redundancy.

8.3 Optical-Layer Protection: Point-to-Point and Ring Architectures

WDM transport systems are being deployed at an increasing rate in backbone net-
works. Currently, optical-layer protection and restoration is being introduced into these
networks. In this section we present several protection techniques for point-to-point
systems and rings. Sections 8.4.1 and 8.4.2, respectively, deal with shared optical-layer
line-based protection and optical path-based protection in general mesh networks. An-
other possible approach to survivability in mesh networks uses distributed restoration
techniques based on reconfigurable optical cross-connects (OXCs) [Ramamurthy+99b].
These techniques are essentially the same as restoration using SONET DCSs, so they
will not be discussed further here.
The point-to-point and SHR protection techniques reuse most of the concepts de-
veloped for SONET protection. Multiwavelength systems add extra complexity to the
protection problem and introduce several additional considerations (mainly at the physi-
cal layer). We shall assume the use of multiple wavelengths per fiber – the case in which
optical protection and restoration techniques are especially useful.

8.3.1 Point-to-Point Systems

In this section we discuss the extension of facility protection to the optical level.

8.3.1.1 (1 + 1) Optical Protection

With (1 + 1) optical facility protection, service is protected from link failures only,
using an optical splitter to bridge the optical signal and dual feed it on diversely routed
670 Multiwavelength Optical Networks

Bridge Switch
W

TX RX

Figure 8.9 (1 + 1) Protection in the optical layer.

working and protection fibers. The protection switching is performed entirely in the
optical domain. In the event of a single link failure, an optical switch at the receiving end
switches to the protection fiber [Bonenfant+94, Wu92]. Because there are no duplicated
electronics, a link failure can be recovered but a transmitter or receiver failure leads to loss
of service. Figure 8.9 shows an example of the (1 + 1) optical protection architecture
for point-to-point systems. This arrangement is not the same as the (1 + 1) SONET
protection architecture. In (1 + 1) SONET protection (Figure 8.2), both incoming signals
are detected and can be compared continuously to select the better of the two. In the
(1 + 1) optical protection architecture, with only one receiver at the tail end, a failure of
the incoming signal is detected at the receiver. Unlike (1 + 1) SONET protection, the
receiver switches to the protection fiber without knowing anything about the state of the
protection fiber. (In the [1 + 1] SONET protection architecture, the duplicate electronics
enable the receiving end to know the state of both signals.)
Other architectures are proposed in [Wu92] that can protect against both equipment
and link failures. They use electrical layer protection to protect against equipment fail-
ures and optical layer protection to protect against link failures. The electrical protection
scheme can use a (1:N) protection technique in conjunction with (1 + 1) optical protec-
tion. (It is unlikely that more than one piece of equipment will fail at the same time.) Of
course, the (1:N) protection procedure is considerably more cost-effective than the (1:1)
method.

8.3.1.2 (1:1) Optical Protection

The (1:1) optical protection scheme is similar to the (1 + 1) method in that diversely
routed facilities protect against link failures. The service is not bridged permanently
to both the working and protection fibers. Instead, a switch occurs once a failure has
been detected [Bonenfant+94, Wu92]. Figure 8.10 shows an example of a (1:1) optical
protection architecture.
An APS signaling channel (on the bidirectional path) is used to coordinate the pro-
tection switch in the event of a failure. In the (1:1) SONET protection architecture,
there is an APS channel between the head and tail ends. This is implemented using the
protection fiber and an APS signaling channel. In a (1:1) optical protection architec-
ture, no communication path exists over the protection fiber because this architecture
does not have duplicate electronics. A communication path is only established when the
transmit and receive ends switch to the protection fiber. When a failure occurs, the re-
ceive end sends a message to the transmit end over the protection fiber without knowing
 Survivability 671

Switch Switch
W

TX APS Signaling Channel RX

Figure 8.10 (1:1) Protection in the optical layer.

whether the transmit end also switched to the protection fiber. Thus, initially it does not
receive any reply. If the transmit end switches to the protection fiber, then the service is
recovered. Otherwise, the traffic is lost. This difficulty can be avoided if a separate
out-of-band optical service channel is provided to support signaling for protection
switching.6
Note that although this is a (1:1) architecture, no spare traffic is supported, because the
system lacks the duplicate transmitters and receivers necessary to support transmission
and reception over the protection facility.

8.3.1.3 (1:N) Optical Protection

The (1:N) optical protection architecture is similar to the (1:1) technique. However, N
working entities now share a single protection fiber. If several working fibers are cut, only
the traffic that is carried on one of them can be recovered. The failure with the highest
priority is recovered first. Figure 8.11 shows an example of a (1:N) optical protection
architecture.
Figure 8.12 shows a combination of (1 + 1) optical and (1:N) equipment protection for
a WDM system. It has significant economic benefits over (1:1) optical and equipment
protection schemes. In this system, a spare wavelength λ P , together with associated
terminal equipment is set aside for equipment protection.

Switch Switch
W1

W2
TX . RX
. . .
. . .
. WN .
TX RX

APS Signaling Channel

Figure 8.11 (1:N) Protection in the optical layer.

6
Currently, service channels are implemented on various wavelengths outside the ITU DWDM grid.
672 Multiwavelength Optical Networks

1 1
1
Bridge 2

N Optical N
(1+1)
P Optical Switch
P Protection P

(1:N) WDM WDM Equipment

Equipment Terminal Terminal
Protection

Figure 8.12 (1 + 1) Optical protection and (1:N) electronic protection for a WDM system.

8.3.2 Self-Healing Optical Rings

WDM ring networks are the subject of considerable interest because of their protection
capabilities. They provide 100% protection for a single cable cut or equipment failure
(when the node that failed is not an originating or terminating node) using a simple but
fast protection-switching scheme [Wu+89]. The WDM SHR architectures proposed in
the literature present various challenges that were not encountered in the SONET SHRs.
For example, if the ring segments separating the WADMs are long, optical amplifiers are
needed between the WADMs on the working fibers. This means that optical amplifiers
are required on the protection fibers as well. But when no faults are present, the protection
fibers are configured in closed loops so that the amplified spontaneous emission noise
from the fiber amplifiers recirculates in this closed protection path unless means are
used to break these closed cycles. Several methods have been proposed to avoid this
condition. They involve a topology based on properties of Eulerian graphs, as well as
technological approaches [Bala+96, Iness+96].
As in the case of SONET SHRs, WDM SHRs can be either path switched or line
switched. The path-switched rings are referred to as unidirectional path-protected ring
(UPPR) architectures (also known as optical channel [OCh] dedicated protection rings
[DPRINGs]), and the line-switched rings are referred to as shared protection ring
(SPRING) architectures (also known as optical multiplex section [OMS] SPRINGs).
Both of these types are discussed in the following sections. For both techniques, optical
layer protection switches typically enable the recovery of failure conditions in a few tens
of milliseconds—a timescale compatible with optomechanical switching technology.
This timescale is reduced considerably when lithium niobate protection switches are
used in WDM SPRINGs [Xin+98].

8.3.2.1 Unidirectional WDM Path-Protected Ring Architecture

In a UPPR the nodes are connected in a ring topology using two fibers that propagate
signals in opposite directions around the ring. One fiber is designated as a working fiber
and the other is designed as a protection fiber. All the working traffic propagates around
the ring in one direction while all the protection traffic propagates around the ring in
 Survivability 673

Optomechanical Switch

WADM P
W
W
Node 1
P

Node 5 Node 2

Node 4 Node 3

Figure 8.13 Four-fiber WDM SPRING architecture.

the opposite direction. This architecture is very similar to the two-fiber SONET UPSR
architecture. The self-healing capability is achieved by using (1 + 1) wavelength path
selection. The signal is bridged on both fibers at the transmit end (using an optical split-
ter), resulting in two copies of the signal propagating in the ring in opposite directions.
The receive end then switches to the protection fiber if a failure occurs. The counter-
propagating signal thus provides network survivability when a link is cut [Elrefaie92,
Wu+90]. All splitting and switching is done entirely in the optical domain. Because the
receiving end does not need to notify the transmitting end that a failure has occurred,
a signaling channel is not required. The decisions are all made at the receiver on a
path-by-path basis.
Because there are no duplicate electronics in this architecture, a transmitter or re-
ceiver failure results in the loss of traffic. A node failure can be protected in the
same way as in the SONET UPSR architecture (i.e., all pass-through traffic can be
recovered). Traffic that originates or terminates at the failed node cannot be recovered.
Depending on the location of the failures, the UPPR may be able to restore multiple
failures.

8.3.2.2 Four-Fiber WDM Shared Protection Ring Architecture

A four-fiber SPRING is a bidirectional four-fiber ring: two fibers are designated as
working fibers and the other two fibers are protection fibers (see Figure 8.13). In this
case, the working traffic travels around the ring in both directions. For example, a
bidirectional connection between two nodes is typically established on the short path
674 Multiwavelength Optical Networks

Link Failure

WADM P
W
W
Node 1 P

Node 5 Node 2 WADM

Node 4 Node 3

Figure 8.14 Four-fiber WDM SPRING surviving a link failure.

connecting the nodes. The working fibers carry all the traffic under normal conditions
and the protection fibers carry no traffic unless a failure occurs. For example, if we
assume bidirectional connections, the traffic from node 1 to node 2 in Figure 8.13 would
be carried in the outer (clockwise) working fiber, and the traffic from node 2 to node 1
would be carried in the inner (counterclockwise) fiber.
To recover from failures the ring uses line protection switching (ring switching)
to perform a loopback function [Elrefaie92, Wu+90], which requires an optical pro-
tection switch (typically optomechanical) at each node (the small white rectangles in
Figure 8.13).7 This protection switching can be revertive or nonrevertive.
When a fiber segment of the ring is cut (a line that interconnects adjacent nodes), the
two nodes at the ends of this segment close their protection switches and interconnect
the working fibers to the protection fibers. The traffic is moved to the protection fibers,
propagating away from the failure. The signals propagate around the ring on the protec-
tion fiber until they reach the other side of the failure, at which point they switch back
to the original working fiber, thus recovering the traffic. For example, in Figure 8.14 the
link between nodes 1 and 2 is cut so that the protection switches close, sealing off the
failed link. Traffic that normally would have entered the link from the WADM in node 1
is switched to the counterclockwise protection fiber and doubles back around the ring
until it enters the closed protection switch in node 2, which directs it to the WADM in
that node. The protection path taken by the traffic replaces the cut (clockwise) working
7
Note that the ring nodes themselves may possess protection-switching capability. In that case, external
optical protection switches are not required.
 Survivability 675

Node Failure

WADM P
W
W
Node 1 P

Node 5 Node 2 WADM

Node 4 Node 3

Figure 8.15 Four-fiber WDM SPRING surviving a node failure.

fiber in the failed link. (A similar operation protects the counterclockwise working fiber.)
An APS channel is required to coordinate the protection switching at both ends of the
failure.
When a node fails, the protection switches in that node switch to interconnect the
working and protection fibers on the incident links. The signal again propagates around
the ring until it reaches the other side of the failed node, where it switches back to
the working fiber, thus recovering from the failure. Figure 8.15 illustrates how the ring
survives a failure of node 1.
As in the SONET SHRs, only traffic that passes through a failed node can be recovered.
Traffic that originates or terminates at a failed node cannot be recovered and is lost.
Also, as in SONET some of the traffic may be misdirected and has to be squelched. (No
specific methods have been standardized to deal with misdirected traffic in WDM SHR
architectures.) Because no duplicate electronics are used, the SPRING architecture does
not protect against a transmitter or receiver failure [Bonenfant+94, Manchester+96].

8.3.2.3 Two-Fiber WDM Shared Protection Ring Architecture

In the two-fiber WDM SPRING adjacent nodes are connected using two fibers. Each fiber
is designated as working fiber but half the signal bandwidth on each fiber is dedicated
for protection [Wu+90, Wu92]. As for the four-fiber SPRING, a cable cut or a node
failure causes the protection switches to interconnect the working and protection fibers
to redirect the traffic away from the failure.
676 Multiwavelength Optical Networks

Working 1: 1 2 3 4
Working 2: 5 6 7 8 Optomechanical Switch

WADM
Working 1

Working 2

Node 1

Node 5 Node 2

Node 4 Node 3

Figure 8.16 Two-fiber WDM SPRING architecture.

In the SONET two-fiber BLSR, 50% of the fiber capacity is reserved for protection,
and TSI is used to switch the traffic from fiber 1 on a link to the oppositely directed
fiber 2. Similarly, in the two-fiber WDM SPRING, only half of the wavelengths are used
in each fiber. However, there is one important difference. In a SONET BLSR, a single
bidirectional connection uses the same time slot in both directions. During protection
switching, a TSI is performed at the protecting node, moving the working slots to the
protection slots before transmission onto the fiber. In the two-fiber WDM SPRING,
an analogous operation on the wavelengths would require the use of wavelength inter-
change, which is currently difficult to achieve while remaining in the optical domain.
However, a simple wavelength assignment rule avoids this problem. It suffices to assign
wavelengths from two disjoint sets to the two directions of a given bidirectional con-
nection (Figure 8.16). For a network with N wavelengths, the working traffic in fiber 1
can use wavelengths 1 through N/2, and the remaining wavelengths can be reserved for
protection traffic. In fiber 2, the assignments would be reversed. If a link failure occurs,
the protection switch at the head end of fiber 1 on the concerned link closes, thereby
looping working traffic onto fiber 2 of the preceding link. This traffic then moves in
the opposite direction around the ring away from the failure [Bala96]. Because of the
method of wavelength assignment, the rerouted traffic is on a vacant wavelength on fiber
2 in each link. Once again, node failures can be recovered in a manner similar to SONET
two-fiber BLSRs.
 Survivability 677

The SPRING architectures require coordination by the nodes on either side of the
failure in the ring. A signaling protocol is required to perform a line switch and to
restore the network [Bonenfant+94].8 SPRING architectures are much more difficult to
operate than UPPR architectures, which require no signaling protocols for protection
switching. However, they offer a substantial increase in the amount of carried traffic
when compared with UPPRs.

8.4 Optical-Layer Protection: Mesh Architectures

Current survivability schemes in optical networks can be classified with respect to three
major characteristics: (a) backup route preplanned prior to occurrence of a fault (protec-
tion) versus real-time (dynamic) backup route computation after the occurrence of a fault
(restoration), (b) line based9 versus path based, and (c) centralized backup route com-
putation with centralized failure restoration protocol implementation versus distributed
backup route computation with distributed failure restoration protocol implementation.
Figure 8.17 shows a taxonomy of the schemes utilized for surviving a failure, focusing
on the preplanned protection schemes. For various reasons, including the characteris-
tics of the physical-layer technology for optical networks, dynamic restoration is less
well-developed in the optical layer than it is in the logical layers; e.g., in SONET or
IP networks. Thus, we limit the discussion of optical-layer survivability to preplanned

Path
protection OChS DPRING

OMS SPRING
Ring 2 fiber
Link
protection OMS SPRING
4 fiber
Preplanned
protection 1+1

Dedicated

Path 1:1
Survivability
protection

Mesh Shared (1:N)

Dedicated
Ring
cover
Ring Node Shared
loopback cover
Link Cycle Double
protection Cover
Generalized
Restoration loopback P-cycles

Figure 8.17 A taxonomy of survivability schemes. (After [Maier+02, Figure 3]. Copyright

c 2002 Kluwer Academic Publishers. Used with kind permission of Springer Science and Busi-
ness Media.)
8
A detailed signaling protocol for the SPRING has not yet been standardized.
9
Also called local-span or link based.
678 Multiwavelength Optical Networks

protection methods. Each of the techniques shown in Figure 8.17 can fall under the
distributed or centralized classification mentioned previously. The figure includes the
various ring-based protection schemes that were discussed earlier in this chapter as
well as those applicable to arbitrary mesh network topologies to be discussed in this
section.

Line-Based versus Path-Based Survivability Schemes

In a line-based approach, whenever a failure is detected, predetermined backup paths
(single or multiple reroutes) around only the failed link are utilized to bypass the failure
(Figure 8.1 shows an example where a single backup path is used for rerouting all
the connections that were affected by the failure.) Although this approach restores the
failure locally, and is thus very fast, it generally requires more redundant capacity
for recovery purposes than path-based techniques, and it cannot be easily adapted for
node failure recovery [Iraschko+96, Iraschko+98, Ramamurthy+99a]. Examples of
line-based survivability approaches include the self-healing rings and the ring covers
that were discussed previously in this chapter and the cycle double covers and p-cycle
techniques presented in Sections 8.4.1.2 and 8.4.1.3, respectively.
In path-based approaches, the backup path is distinct (link/node disjoint)10 from the
working path. Assuming that only one path fails at a time, the backup capacity reserved
on a given link can be dedicated for a specific connection or can be shared among several
working paths if those working paths do not use the same links or nodes [Bouillet+07].
This ensures that for single failures, backup capacity will always be available to recover
all working paths affected by the failure. Path-based approaches provide more efficient
capacity utilization than line-based approaches and can easily protect against node
failures; however, they are generally slower than line-based schemes because of the
number of cross-connect reconfigurations required and the complexity of the recovery
protocols.

Precomputed versus Real-Time Backup Route Computation

In the precomputed case, the backup path for each failure scenario is calculated before a
failure occurs, and the protection capacity for each failure scenario is reserved in advance,
whereas in the real-time case, the backup path is computed dynamically after the failure
occurs, and available redundant capacity is used to restore the failed connection. Clearly
the precomputed approach ensures a much faster failure recovery process than the real-
time approach. However, precomputed backup paths use the network resources less
efficiently than the real-time approach because they cannot make use of information on
the current network state.

Centralized versus Distributed Backup Route Computation

When a centralized protection/restoration scheme is utilized, a central network controller
is responsible for keeping track of the current state of the network and calculating the
10
A link-disjoint protection path is sufficient to protect the network against link failures only. A protection
path that is node disjoint (except for its ingress and egress nodes) protects against both link and node
failures.
 Survivability 679

backup paths either before or after a failure occurrence. For example, in the real-time
restoration case, when a failure occurs, the controller correlates the fault alarms, isolates
the failure, and computes the most efficient restoration path while avoiding the failed link
or node. After the restoration path is computed, the controller notifies all relevant nodes
to execute the cross-connects needed to accommodate the new optical path. On the other
hand, distributed tehniques do not require a central controller with complete knowledge
of the entire network. Rather, local controllers at each node can store information about
their physical links, and their own and neighboring nodes [Johnson+94]. Signaling is
then used to update the nodes with relevant topology and available capacities, to compute
and establish the new routes, and to initiate the cross-connects to accommodate these
new routes.
The centralized survivability approaches may be slower than the distributed ones; how-
ever, they are more efficient in terms of capacity utilization, as the backup paths are cal-
culated based on complete information on the current state of the network [Iraschko+98,
Ramamurthy+01].
We now present methodologies for protection in the optical layer of networks with
arbitrary mesh topologies starting with shared optical layer line-based protection and
following with optical layer path-based protection, which can be either dedicated or
shared. We continue with a newer approach that lies somewhere between line and path
protection, termed segment or island protection and restoration. The section concludes
with a brief look at survivability for multicast optical connections.

8.4.1 Shared Optical Layer Line-Based Protection

A number of methodologies have been developed for APS in the optical layer of networks
with arbitrary topologies. These are shared optical protection (line-based protection)
techniques that can potentially recover from failures within 50 ms by detouring traffic
around a failed link or node. This speed is comparable with the specifications for recovery
at the logical (path) level in SONET rings and similar networks. Although we present
these protection techniques in the context of purely optical networks (WRNs using
optical cross-connects as their switching nodes), they are also applicable to networks
composed of point-to-point optical transmission links terminated by electronic switching
equipment (e.g., digital cross-connects).

8.4.1.1 Why Shared Optical Line-Based Protection?

In traditional mesh networks, recovery from link or node failure is typically imple-
mented in a centralized manner using path protection techniques. The protection is
accomplished on a “fine-grained” basis; that is, the individual traffic units being re-
covered are typically in the tens of megabits-per-second range (e.g., STS-1 or DS3). In
(1 + 1) mesh protection, for example, each signal is carried simultaneously on a working
and a protection path, and a path selector at the receiver chooses one or the other based
on the received signal quality. In a SONET network, there might be many individual
connections using a failed link, each requiring its own protection path. This leads to
considerable wasted bandwidth in the network. In the (1:1) or (1:N) case, the required
680 Multiwavelength Optical Networks

spare capacity is less. The protection paths are not active unless there is a fault, and they
may be used for lower priority traffic, which is preempted when necessary. However,
failure recovery in this case is a slower and more complex operation. Each of these
techniques requires considerable coordination, processing, and delay, because many
connections may be affected and have to be rerouted in the event of a single link or node
failure.
In the currently evolving multiwavelength optical networks, the amount of traffic
carried over each fiber is typically in the range of hundreds of Gbps. To be efficient,
fault recovery techniques in these networks should deal with large units of capacity.
Recovering connections on an STS-1 basis is wasteful in terms of bandwidth, time,
complexity, and cost. To improve performance in the multiwavelength case, the path-
protection techniques described earlier can be implemented using a λ-channel as the
basic unit to be rerouted. Nevertheless, the basic problems associated with coordination,
centralized processing, complexity, and delay are still there but on a somewhat reduced
level.
Several line-based protection schemes for mesh networks have evolved from ring-
based protection approaches. In these architectures, sets of rings (cycles) with special
characteristics are precomputed and when a failure occurs, well-established protection
switching schemes are used to recover the traffic. The three notable ring-based protection
architectures are ring-covers (Section 8.2.4.1), cycle double covers, and p-cycles, of
which the latter two are discussed below. The section concludes with a generalized
loopback approach to protection which is not ring/cycle-based.

8.4.1.2 Cycle Double Covers

The shared optical protection technique discussed next is based on [Ellinas98]. Because
it operates on a line- rather than path-protection basis, the fundamental unit being pro-
tected is a transmission link or switching node rather than an end-to-end connection.
The protection methodology based on cycle double covers is, in a broad sense, a gener-
alization of that used in SPRINGs. Each link carries a protection fiber for each working
fiber, and thus the system requires 100% redundancy.11 This is in contrast to the ring
cover approaches described earlier, where more than one virtual ring may cover the same
link in the network (which implies that more than 100% redundancy is required). Fault
recovery decisions are made locally, in a distributed fashion, and independent of the
state of activity in the network. The implementation uses simple and reliable protection
switches in each network node so that protection is accomplished without significant
processing, transmission, or propagation delays.
Automatic protection switching for recovery from link failures in mesh networks is
considerably simpler than recovery from node failures. We concentrate here on link pro-
tection [Ellinas+96]. The discussion is broken down according to network topologies –
planar and nonplanar. The planar case is considerably simpler.12

11
In this section, all network links are assumed to be bidirectional four-fiber links.
12
Various other special cases, such as Eulerian networks, offer additional possibilities for simplification. The
reader is referred to [Ellinas98] for more details.
 Survivability 681

Link Failure Protection

Consider a wavelength-routed optical network with a general mesh topology supporting a
number of connections with optical paths that are determined by the settings of the optical
cross-connects. Typically, each fiber in these networks carries several WDM connections.
We assume that a network link consists of a pair of working fibers accompanied by a pair
of protection fibers. This is similar to a link in the four-fiber WDM SPRING discussed
earlier. Four 2 × 2 optical protection switches terminate the fibers in each link as in a
SPRING. Under fault-free operation, these switches connect the working fibers to the
OXCs in the network nodes. If a link fails, the protection switches at the ends of the
failed link move to their protection state. This automatically reroutes the affected traffic
around the fault over a path consisting of protection fibers. These fibers are organized
into protection cycles, with a configuration that depends only on the topology of the
network and not on the current network state (i.e., its active connections). Thus, the
protection cycle topology is computed off-line and set up when the network is originally
activated.
Because protection switching is performed at individual switching nodes without
instructions from a central network controller, the APS process is distributed and au-
tonomous. However, a switching protocol implemented locally at each node is required
to coordinate the protection switching process.
Figure 8.18 illustrates recovery from a link failure using protection cycles. A link
terminated by four 2 × 2 protection switches is shown connecting optical switching
nodes 1 and 2. The link consists of two working fibers, W+ and W− , and two protection
fibers, P+ and P− . Before link failure, the switches are set in their normal position
so that traffic between the OXCs traverses the link on the working fibers as shown in
Figure 8.18(a). The protection fibers P+ and P− are interconnected with other protection
fibers to form part of the protection cycles C1 and C2 , respectively. (These cycles are
completed on a sequence of other network links not shown in the figure.) When the link
fails, the protection switches move to the protection position, shown in Figure 8.18(b).
Now all traffic that normally would be routed from node 1 to 2 over the failed link is
detoured around the link using protection cycle C1 , and, similarly, traffic in the opposite
direction is rerouted over C2 .
The key to the successful operation of this protection technique is the creation of
a family of protection cycles that covers the network in a particular way. Using this
approach, protection against all single link failures (and some sets of simultaneous
failures) is possible in all networks with topologies (graphs) that are at least two-edge
connected. A ring is a simple example of a two-edge connected graph (see Appendix A).
Protection against all link failures is not possible in networks with graphs that are one-
edge connected because they possess links called bridges. A fault in a bridge results in
a disconnected graph, and any connections between the two components of the graph
are lost. In [Ellinas+96], a systematic method is proposed for protecting any two-edge-
connected network against link failures. The methods proposed also protect networks
containing bridges against any link failure except the failure of a bridge. (In the following,
whenever we refer to link failure recovery, it is to be understood that bridge failures are
excluded.) The method is based on the following:
682 Multiwavelength Optical Networks

Protection Switch
C1 W
P 1

1 2

P W C2
OXC
(a)

C1

C2

(b)

Figure 8.18 Rerouting around a failed link.

Proposition 1: Recovery from a single link failure in any optical network with arbitrary
topology and bidirectional fiber links can be achieved using APS if a family of directed
cycles using the protection fibers can be found so that

1. All protection fibers are used exactly once.

2. In any directed cycle, both protection fibers in a pair are not used unless they are
in a bridge.

Figures 8.19 and 8.20 show examples of graphs with directed cycles covering each
graph in the manner prescribed in Proposition 1. By interconnecting the protection
switches appropriately in the corresponding network, these cycles can be implemented
as protection cycles on the corresponding protection fibers. Methods are described next
for finding an appropriate set of protection cycles in the planar and nonplanar cases. We
then explain how these cycles can be used to protect against link failures using APS.
 Survivability 683

Figure 8.19 Directed cycles in a planar graph.

The Planar Case

Figure 8.19 shows a plane graph that is covered by five directed cycles. The figure
suggests a systematic procedure for constructing the protection cycles in planar graphs
to fulfill the requirements of Proposition 1. The graph is first embedded on the plane
to form a plane graph that contains f − 1 inner faces and one outer face, where f is
Euler’s number (see Appendix A). The protection cycles are then the f face boundaries,
with each oriented properly. This result is embodied in the following:

Proposition 2: Every planar graph can be decomposed into a family of directed cycles
in which each edge is used exactly twice (once in each direction) and, in each directed
cycle that does not include a bridge, an edge is used once at most. In each directed cycle

Figure 8.20 Directed cycles in a nonplanar graph: K 5 .

684 Multiwavelength Optical Networks

24
9 11
1
23 5
2 10 12
22 8 4
3
21 7 13
6
20 18
17
19 16 15
14

Face[0] = [6 13] [13 12] [12 6] Face[10] = [8 22] [22 17] [17 8]
Face[1] = [6 12] [12 4] [4 6] Face[11] = [7 17] [17 16] [16 7]
Face[2] = [6 7] [7 15] [15 6] Face[12] = [17 22] [22 21] [21 20] [20 19]
Face[3] = [6 4] [4 7] [7 6] [19 18] [18 17]
Face[4] = [14 15] [15 16] [16 17] [17 18] Face[13] = [24 8] [8 9] [9 24]
[18 19] [19 20] [20 21] [21 22] [22 23] Face[14] = [3 8] [8 7] [7 3]
[23 24] [24 9] [9 1] [1 5] [5 10] [10 11] Face[15] = [12 11] [11 10] [10 12]
[11 12] [12 13] [13 14] Face[16] = [9 8] [8 3] [3 9]
Face[5] = [6 15] [15 14] [14 6] Face[17] = [3 7] [7 4] [4 3]
Face[6] = [6 14] [14 13] [13 6] Face[18] = [2 4] [4 5] [5 1] [1 2]
Face[7] = [8 24] [24 23] [23 22] [22 8] Face[19] = [5 4] [4 12] [12 10] [10 5]
Face[8] = [15 7] [7 16] [16 15] Face[20] = [3 4] [4 2] [2 3]
Face[9] = [7 8] [8 17] [17 7] Face[21] = [9 3] [3 2] [2 1] [1 9]
Number of faces = 22

Figure 8.21 Face traversal for a planar national network.

including a bridge, the bridge is used twice (once in each direction) in the same directed
cycle.

Algorithms are available [Even79, Fisher65, Gibbons85] that test a graph for planarity in
linear time. A planarity testing, face traversal algorithm, which is a variation of the path
addition algorithm by [Gibbons85], is described in [Ellinas98]. This algorithm embeds
the planar graph on the plane and traces the faces in the appropriate directions to create
a family of cycles with the characteristics defined in Proposition 1. Figure 8.21 shows an
embedding of a planar graph representing a “national network” for the contiguous United
States, together with its protection cycles. All cycles except the outer one (face [4]) are
traced counterclockwise, on the faces found using the algorithm.
Although the face boundaries of a planar network are a good choice for protection
cycles, they are not the only possibility satisfying the conditions of Proposition 1.
Other choices are generally possible, leading to families that contain smaller numbers
of cycles with longer average lengths.13 However, many short protection cycles are
normally preferred over fewer longer cycles for reasons that will become apparent as we
proceed.

13
It can be shown that the face boundaries are the largest possible set of protection cycles.
 Survivability 685

The Nonplanar Case

Extension of Proposition 2 to the nonplanar case requires certain results from graph
theory dealing with cycle double covers (CDCs). A CDC of a graph G is a cycle
decomposition of G such that each edge appears in exactly two cycles (see Appendix A).
There is a basic conjecture in graph theory on the existence of CDCs: Every bridgeless
graph has a CDC.
The CDC problem has been studied thoroughly by graph theorists [Alspach+85].
However, the conjecture remains an open question, never having been proved or dis-
proved. It is therefore extremely unlikely that a network will ever be discovered that does
not possess a CDC!
A CDC is said to be orientable when it is possible to choose a circular orientation for
each cycle of the CDC in such a way that each edge is taken in opposite directions in
its two incident cycles [Alspach+85]. Figure 8.20 shows an orientable CDC consisting
of four cycles, shown with solid (thin or bold), and dashed (thin or bold) lines. An
orientable CDC is exactly what we seek in constructing our protection cycles: A cycle
decomposition such that each edge appears in exactly two protection cycles and each
edge is used in opposite directions in the two cycles. A number of results in graph theory
are strong evidence for the existence of orientable CDCs in any bridgeless graph. They
obviously exist in all bridgeless planar networks because the face boundaries constitute
an orientable CDC.
A heuristic presented in [Ellinas+97] is designed to find an orientable CDC of any
given graph (planar or nonplanar). The heuristic includes modifications of the CDC to
accommodate bridges, as indicated in Figure 8.19, and thus finds a family of protection
cycles satisfying Proposition 1. It uses a backtrack (branch-and-bound) approach to find
the family of protection cycles.
Figure 8.22 shows an orientable CDC for a nonplanar version of the ARPANet. It was
found using the heuristic in [Ellinas+97]. When applied to planar graphs, this heuristic
generally does not choose the face boundaries as protection cycles, and typically finds
a covering with less than the maximum possible number of cycles (Euler’s number).
However, the number of cycles is usually close to the maximum, indicating that the
cycles obtained are relatively short.
If multiple link failures occur with all failed links on distinct protection cycles, all
failures can be recovered. Additional details on the bounds on the number of restorable
link failures can be found in [Ellinas+00b].
The basic concept of protection cycles can be extended to protect against switching
node failure as well. However, it turns out that node failure protection is a more complex
issue than link failure protection. In particular, the required protection switch hardware
differs depending on the arrangement of the families of protection cycles, which in turn
depends on the network topology. Planar topologies require 2 × 3 protection switches for
link and node protection rather than the 2 × 2 devices used for link protection only. In
nonplanar networks, 2 × 5 switches are required. See [Ellinas98] and [Ellinas+99]
for a complete description of node failure protection in planar and nonplanar
networks.
686 Multiwavelength Optical Networks

20

17
19
4 13
2 11
5
16
15
6 9
14
1
10 18
7 8
12
3

Cycle[0] [5 6] [6 1] [1 2] [2 5] Cycle[6] [13 15] [15 16] [16 19] [19 17] [17 13]
Cycle[1] [16 15] [15 14] [14 18] [18 16] Cycle[7] [3 1] [1 6] [6 7] [7 3]
Cycle[2] [10 8] [8 12] [12 18] [18 14] [14 20] Cycle[8] [2 1] [1 3] [3 8] [8 10] [10 19] [19 16]
[20 19] [19 10] [16 18] [18 12] [12 14] [14 15] [15 13] [13 11]
Cycle[3] [5 2] [2 4] [4 5] [11 4] [4 2]
Cycle[4] [8 9] [9 11] [11 13] [13 17] [17 20] [20 14] Cycle[9] [17 19] [19 20] [20 17]
[14 12] [12 8] Cycle[10] [7 10] [10 9] [9 8] [8 3] [3 7]
Cycle[5] [9 10] [10 7] [7 6] [6 5] [5 4] [4 11] [11 9]
Number of cycles 11

Figure 8.22 Orientable CDC of the ARPANet.

Implementation of APS for Link Failures Using Protection Cycles

Having shown how to determine a family of protection cycles with the characteristics
defined in Proposition 1, we now present an example illustrating how these cycles can
be used to protect a network with an arbitrary topology against a single link failure.
Figure 8.23 illustrates the intranodal connections between protection switches in
a seven-node planar network. Each protection switch is a 2 × 2 device that passes
connections through its attached link when in its default position and loops them back
through its node when a failure is detected on the attached link. Note that with the
switches in their default positions, all protection fibers are interconnected into protection
cycles (shown as dashed lines) coinciding with the face boundaries.
Figure 8.24 illustrates the protection scenario, showing the protection switch settings
in the network of Figure 8.23 after a failure in the indicated link. A pair of bidirectional
connections from S to D is depicted (with heavy solid lines) in Figure 8.23. After the
failure, the path followed by the connections follows working fibers up to the protection
switches, which have sealed off the failed link. The connections then switch to protection
fibers, following the dashed, bold lines until they reach the other side of the failed
link. There they switch back to the working fibers and continue on the same path that
they used before the link failure. Note that the dashed segments of the routes (on the
protection fibers) are carried on the two protection cycles that would normally traverse
the failed link. The APS process recovers all connections using the failed link in both
directions.
As illustrated in Figure 8.24, the path followed by each rerouted connection occupies
the full length of a protection cycle less the one hop on the failed link. As a consequence,
the overall physical path length for a signal using a protection cycle is increased by the
 Survivability 687

S D

D S

Figure 8.23 Seven-node planar network with default protection switch settings.

length of the cycle less twice the length of the failed link. (This is generally a net
length increase.) To maintain acceptable signal quality after a failure, this path length
increment should be kept small. Hence it is desirable to choose protection cycle families
that consist of many short cycles rather than a few long ones.14

D
S

S
D

Figure 8.24 Seven-node planar network after a link failure.

14
For example, if the network happens to be Eulerian, a possible choice for protection cycles would be a
double-traced Eulerian trail (see Section 6.3.5), but this consists of two very long cycles – a poor choice.
688 Multiwavelength Optical Networks

(a) p-cycle (b) Failure of link (c) Failure of chord

on the p-cycle of the p-cycle

Figure 8.25 Failure recovery using the p-cycle approach.

8.4.1.3 p-Cycles
The p-cycles approach extends the ring cover idea by using the fact that a ring can also
be used to protect all of its chords15 or “straddling edges” [Grover+98, Grover04]. In
this approach, preconfigured cycles (called p-cycles) are created using the network’s
spare capacity. Two types of p-cycles are defined: link p-cycles and node-encircling
p-cycles. As their names signify, link p-cycles protect against a link failure and node-
encircling p-cycles protect against a node failure. Specifically, link p-cycles protect all
the channels on the failed link and node-encircling p-cycles protect all the connections
that pass through a node that has failed. The focus in this section will be on link p-cycles.
These cycles can protect against a link failure on the cycle (in which case the p-cycle
contributes one protection path) or against a link failure off the cycle (on a chord of the
cycle). In the latter case, the p-cycle contributes two protection paths (both paths around
the p-cycle that the failed chord belongs to). For both cases bidirectional protection is
provided. Figure 8.25 shows an example of a p-cycle and how it recovers from link
failures on and off that cycle.
The p-cycle designs differ from ring or CDC schemes in that the p-cycle approach is
based on protection of individual channels on a link, in contrast to rings and CDCs where
it is the aggregate traffic on a link that is recovered. Thus, rings and CDCs are line-based
approaches, whereas p-cycles have certain aspects of path protection, in the sense that
individual connections can be designated for protection, and more efficient use can be
made of the network capacity used for protection (see Section 8.4.2). A downside of
the p-cycle approach is that protection switching is more complex than it is in rings and
CDCs, where simple 2 × 2 loopback switches are all that is required at the ends of each
link.
The p-cycles work in either opaque or transparent networks; i.e., in networks with
full, partial, or no wavelength conversion. As in the case of CDCs, special care needs
to be taken in transparent networks to ensure that an optical connection is not too long
after it is switched onto the protection path (to minimize signal impairments as much as
possible). A case study in finding the optimal set of p-cycles presented in [Schupke+02]

15
A chord is a link not on the ring that connects a pair of nodes on the ring.
 Survivability 689

showed that when full wavelength conversion is assumed, the protection capacity was
approximately half of the working capacity (in terms of wavelengths/channels). That
ratio was increased to 71% when no wavelength conversion was available.
An advantage of the p-cycle approach over the ring cover and the cycle double cover
approaches is that it does not have to cover all links in the network with rings (by omitting
the chords that have their end nodes on the p-cycles but are not included in the p-cycles),
thus reducing the spare capacity required (less than 100% redundancy). It has been
shown that with p-cycle protection, the spare capacity required for full link protection is
comparable to that of conventional path-based mesh protection and that the achievable
redundancy (the ratio of spare to working capacity) has a lower bound of 1/(d̄ − 1),
where d̄ is the average node degree in the network [Grover+98, Stamatelakis+00].
A case study on a Pan-European network performed in [Schupke+02] validates these
claims and shows that when the maximum allowable p-cycle length is approximately
4500 km, redundancy less than 60% is achieved. As the maximum allowable p-cycle
length increases, it is also shown that the redundancy approaches the lower bound for
networks with full link protection.
The following ILP finds the set of p-cycles that require the minimum spare capacity
cost for 100% recovery (all channels on the link that has failed are recovered) [Stamate-
lakis97]:
r S = set of network links
r P = set of elementary cycles of the graph
r δk, j is 1 if link k is included in cycle j and is 0 otherwise.
r xi, j is the number of useful paths16 cycle j provides for recovery of link i.
r w j is the number of working channels on link j.
r s j is the number of spare channels on link j.
r n j is the number of unit-capacity copies of cycle j in the design.17
r c j is the cost of link j.


Minimize ck sk (8.1)
k∈S

With

sk = j∈P δk, j n j ∀k ∈ S (8.2)

wi ≤ j∈P xi, j n j ∀i ∈ S (8.3)
nj ≥ 0 ∀j ∈ P (8.4)
sk ≥ 0 ∀k ∈ S. (8.5)

16
Useful paths are defined as the units of capacity that a p-cycle can contribute to the recovery of a specific
link failure.
17
n j specifies the number of channels assigned to the j-th p-cycle, and n j = 0 indicates that the p-cycle is
not used.
690 Multiwavelength Optical Networks

Constraint (8.2) ensures that the spare capacity on link j can support all the p-cycles
that cross that link, while constraint (8.3) ensures that the number of useful paths that
are provided for each path support 100% recovery.
Additional ILPs have been developed for optimization of p-cycle designs, for exam-
ple, finding a set of p-cycles that restores the maximum number of channels on a link
that has failed when the amount and placement of the redundant capacity is given, or
optimizing working path routing and p-cycle placement at the fiber and at the wave-
length level with and without wavelength continuity. For a detailed discussion of p-cycle
ILPs, and p-cycles in general, the reader is referred to [Grover04]. Additional recent
work on flow p-cycles (p-cycles that provide additional protection capability for any flow
segment along a path) and failure-independent path protecting p-cycles (p-cycles that in-
clude end-to-end failure independent protection switching) can be found in [Grover+05,
Shen+03].
It is interesting to note that if the graph of the network is Hamiltonian, only a single
p-cycle is required to protect all the links in the network [Stamatelakis+00]. However,
this will not yield the most efficient design in terms of capacity efficiency, and it was
shown in [Sack+04] that the optimum solution to this problem involves a mix of p-
cycles, some of them Hamiltonian and some not Hamiltonian. (A Hamiltonian cycle of
a graph is a cycle that traverses all vertices of the graph exactly once.)

8.4.1.4 Generalized Loopbacks

A generalized loopback approach was introduced in [Finn+98] and [Medard+99b] and
elaborated in [Medard+02]. It is a line-based protection approach that is not based
on cycles. The main idea behind generalized loopback is to create pairs of digraphs
(primary and secondary) and use one digraph in the pair to protect against any failures
in the other. Each arc in a primary/secondary digraph pair may correspond to a fiber, a
bundle of fibers, or a wavelength on a fiber, with the arc orientation corresponding to
the direction of transmission on the fiber/wavelength. Each primary and corresponding
secondary digraph is a conjugate of the other, meaning that the two digraphs are identical
except for a reversal of the directions of their arcs (see Section A.1.6). This approach
will operate with or without wavelength continuity, that is, a wavelength carried by a
fiber in one direction can be backed up by the same or by a different wavelength carried
by a fiber in another direction.
Assuming a transparent network carrying unidirectional connections, if a link failure
occurs, connections on one wavelength on the primary (secondary) digraph are looped
back on the same wavelength around the failure using the secondary (primary) graph. To
implement this loopback the optical signal on the upstream end of the failed link on the
primary (secondary) digraph is switched to the secondary (primary) digraph creating a
backup signal that propagates through the network by flooding that digraph. When the
backup path reaches the other side of the failure the traffic is switched back onto the
primary (secondary) digraph completing the recovery process.
The principle underlying generalized loopback is that each digraph (primary and
secondary) must have at least one directed path between all pairs of nodes after any
link failure occurs. After pairs of digraphs are identified having the required properties
 Survivability 691

1 2 3 4

8 7 6 5

Figure 8.26 Generalized loopback example. (From [Medard+02, Figure 1b]. Copyright

c 2002
IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

(an off-line procedure), the generalized loopback approach simply provides loopback
protection switching between primary and secondary digraphs similar to that used in self-
healing optical rings and CDCs. An algorithm presented in [Medard+02] demonstrates
how the digraphs are identified, based on finding a directed cycle that passes through
all the nodes in the network. Figure 8.26 provides an example of link failure recovery
in a sample network with predetermined digraphs, the primary shown with solid arcs
and the secondary with dashed arcs. Unidirectional traffic carried from right to left
on the primary digraph on failed link (2,3) is recovered on path 3 − 6 − 7 − 2 on the
secondary digraph. (Redundant copies of the recovered signals that are generated in
the flooding process are eliminated when they arrive at the network nodes during the
recovery procedure.)
When the connections being protected are bidirectional a four-fiber approach can
be used for primary and secondary digraphs (as in four-fiber rings in Section 8.3.2.2),
where each link in the network consists of two pairs of oppositely directed unidirectional
fibers, with one pair included in a primary digraph and the other in the secondary. An
analogous arrangement for protection of bidirectional connections can be set up using a
two-fiber, two-wavelength system (as in two-fiber WDM rings in Section 8.3.2.3). In this
case each unidirectional fiber carries one wavelength assigned to the primary digraph
and a second wavelength assigned to the secondary.
As in rings and cycle double covers this is a scheme that requires 100% redundancy
at the fiber (or wavelength) level on every link. However, generalized loopback handles
node failures more easily than CDCs do. The fact that generalized loopback can provide
protection on a wavelength rather than fiber basis means that different levels of protection
can be offered for different connections (wavelengths).
Details of how the primary and secondary digraphs are identified for the case of
link and node failures and the protection protocol for such an approach can be found
in [Medard+99b] and [Medard+02]. Additional work in [Lumetta+00] shows that
significant reduction of protection capacity requirements can be achieved if the capacity
in protection fibers is assigned to carry unprotected traffic. However, this capacity gain
does not come without a price. Even though the generalized loopback method can still
692 Multiwavelength Optical Networks

Bridge Working Path Switch

X X
X
X X

X
X Node Z
Node A

Protection Path

Figure 8.27 (1 + 1) dedicated protection architecture.

recover from any single link failure, robustness of the network to more than one failure
in this case is unfavorably affected.

8.4.2 Optical Path-Based Protection

In optical path-based protection, an end-to-end alternate path is set up between the source
and destination of each affected optical connection in the case of a failure. The original
(working) and backup (protection) paths must be link or node disjoint to guarantee
successful single link or link and node failure recovery, respectively. The protection
paths should be predetermined so that the recovery process does not take more than a
few tens of milliseconds if low-speed (optomechanical) protection switches are used.
If higher speed optical switches are used, even faster recovery times can be achieved.
Typically using path-based protection, when a working connection is provisioned, its
protection path is determined at the same time.

8.4.2.1 (1 + 1) Dedicated Backup Path Protection

In (1 + 1) dedicated backup path protection the traffic for each connection is bridged at
the source node and is carried by two diverse paths: working and protection, similar to the
case of (1 + 1) point-to-point optical layer protection discussed in Section 8.3.1.1. Diver-
sity of the working and backup paths is attainable at various levels: link (fiber) disjoint,
link and node disjoint [Bhandari99, Suurballe74], or shared risk group (SRG)18 disjoint
[Ellinas+03, Labourdette+02, Ramamurthy+01]. For example, if the connections in
the network are protected only against link failures then link diversity is required.
In (1 + 1) dedicated backup path protection, when a link or node failure occurs, the
failure is detected at the destination (downstream) nodes (both end-nodes are destina-
tion nodes in case of bidirectional traffic), and these nodes switch over to the protection
paths. In this fashion, each destination node performs an operation identical to that ex-
ecuted in point-to-point (1+1) automatic protection switching (APS). This arrangement
is illustrated in Figure 8.27 for a unidirectional connection from node A to node Z.

18
A shared risk group consists of several entities with a shared common resource that can be affected by a
single failure (e.g., optical channels in a single fiber, or fibers in a single cable) (see Section 8.4.2.4).
 Survivability 693

W1

W2

protection

W3

Figure 8.28 (1:3) shared protection in a mesh network. (From [Maier+02, Figure 6b]. Copyright


c 2002 Kluwer Academic Publishers. Used with kind permission of Springer Science and
Business Media.)

The advantage of this technique is that it is extremely fast (on the order of a few
milliseconds), it is simple to implement, it does not require additional signaling, and
it can easily protect against node failures as well as link failures. However, it makes
inefficient use of capacity, because more than 100% redundancy is generally required
for protection, as protection paths are at least as long as working paths and usually longer.

8.4.2.2 Shared Backup Path Protection

To alleviate the problem of inefficient capacity usage in (1+1) dedicated path-protected
mesh networks, shared backup path-based protection can be utilized. This approach
entails using a protection architecture similar to the (1:N) protection method used in
point-to-point optical (or SONET) systems. If a number of working paths are link (or
node or SRG) disjoint then it is possible for their protection paths to share capacity
on one or more links [Bouillet+02a, Bouillet+02b, Datta+01, Doshi+99, Ellinas+03,
Labourdette+02]. The requirement that the paths be disjoint ensures that a single failure
event will not interrupt more than one working path sharing the same protection capacity.
(If a set of nondisjoint working paths shared the same protection resources, there could
be contention for the reserved capacity, in which case only one of the working paths
could be recovered.) An example of (1:3) shared path-based protection is shown in
Figure 8.28.
In (1:N) backup path-based shared protection, even though the protection path is
precomputed prior to the failure, the traffic demand is not bridged onto the protection
path because in this case it is shared among a number of working paths. Therefore,
the protection capacity is reserved but not active prior to the failure and the cross-
connections along the protection paths are not established until a failure occurs. This
means that the OXCs have to be reconfigured after the failure for each working path
694 Multiwavelength Optical Networks

that needs to be recovered [Ellinas+03]. Therefore, additional signaling is required (in

a centralized or distributed manner) to coordinate the reconfiguration of these cross-
connects, which in turn increases the time required to recover from the failure. The
complexity of this operation in terms of signaling and switching can be considerable
because each affected working path must be recovered independently using its own
protection path. Moreover, there may be hundreds of working paths (e.g., wavelength
channels) affected by a single link failure, with still more brought down by a node
failure.
ILP formulations have been developed for the problem of routing and wavelength
assignment for static traffic demand using a number of protection schemes including
dedicated and shared backup path-based protection. Routing of working and protection
paths as well as wavelength assignment are considered together in these optimization
problems. A typical objective is to jointly determine the working and protection path
routing while minimizing the number of wavelengths used on all links in the network
for both paths. As it turns out, these formulations involve a very large number of
variables and equations and become intractable for large networks. They can be made
tractable by limiting the admissible solutions (for example, restricting the number of
allowable working and/or protection paths by allowing only shortest path routing).
Optimization formulations for the (1 + 1) and (1:N) shared backup path-based protection
problems [Ramamurthy+99a] are given below. These problems are similar to the RCA
optimization problems discussed in Section 6.3.3. However, in RCA optimization a fiber
path (route) and wavelength are chosen for each connection (demand), whereas in path-
based protection optimization a pair of working and protection routes and wavelengths
are chosen for each demand.
Let
r N = number of nodes in the network
r E = set of links e1 , e2 . . . , ek , . . . , in the network
r E = |E| = number of links in the network
r W = maximum number of wavelengths on a link
r R i = set of admissible routes, r i , r i , . . . , r i , . . . , for node-pair i
1 2 n
r M i = |R i | = number of admissible routes for node-pair i
r M = Max M i
i
r E i = set of links, e, in route r i
n n
r R i ⊆ R i = set of admissible routes for node-pair i that do not utilize link e
j j
r d = demand for node-pair i (measured in wavelength channel requests)
i
r s = number of spare wavelengths reserved for protection on link e
j j
r w j = number of working wavelengths utilized on link e j
r γ i,n = 1 if r i is utilized as a working route on wavelength w, and is equal to 0 otherwise
w n
r oi,n = 1 if r i is utilized as a protection route on wavelength w, and is equal to 0
w n
otherwise
r δ i,n = 1 if a protection route r i utilizes wavelength w when link k fails and is equal to
w,k n
0 otherwise
r m j = 1 if wavelength w is utilized by some protection route r that traverses link e j
w
and is equal to 0 otherwise
 Survivability 695

The following two ILPs correspond to the cases of dedicated and shared path-based
protection, respectively. In each case the network is protected against link failure only;
i.e., the working and protection paths must be link disjoint. Furthermore, the problem is
cast in the context of transparent networks; i.e., wavelength continuity is maintained.

ILP: Dedicated Path-Based Protection

The goal of this ILP is to minimize the total number of wavelengths, (w j + s j ), used in
the working and protection paths while protecting for all possible single link failures.
The ILP solution determines the routing and wavelength assignment for the working and
protection paths.
E

Minimize (w j + s j ) (8.6)
j=1

With

(w j + s j ) ≤ W, 1≤ j ≤E (8.7)

W
 
di = γwi,n , 1 ≤ i ≤ N (N − 1) (8.8)
n:rni ∈R i w=1
N (N −1) W
  
wj = γwi,n , 1≤ j ≤E (8.9)
i=1 n:e j ∈Eni w=1

N (N −1) W
  
i,n
sj = ow , 1≤ j ≤E (8.10)
i=1 n:e j ∈Eni w=1
N (N −1)
 
γwi,n + ow
i,n
≤ 1, 1 ≤ w ≤ W, 1≤ j ≤E (8.11)
i=1 n:e j ∈Eni
W
  W
 
γwi,n = i,n
ow , 1 ≤ j ≤ E, 1 ≤ i ≤ N (N − 1). (8.12)
n:e j ∈Eni w=1 n:rni ∈R ij w=1

The first constraint limits the number of wavelength channels on each link to a
maximum W , while the second ensures that the demand for each node-pair i is met. The
third and fourth constraints evaluate the number of working and protection connections,
respectively, that are traversing link j. The fifth constraint is the wavelength-continuity
constraint and it ensures that at most one working or protection connection can use
wavelength w on link j, while the sixth constraint ensures that the link-disjoint condition
is met on each working/protection path pair.

ILP: Shared Path-Based Protection

This ILP also minimizes the total number of wavelengths used in the working and
protection paths (w j + s j ) while protecting against all possible single link failures. In
this case, however, the protection paths are shared by a number of disjoint working
696 Multiwavelength Optical Networks

paths. This ILP solution also determines the routing and wavelength assignment for the
working and protection paths.
E

Minimize (w j + s j ) (8.13)
j=1

With
(w j + s j ) ≤ W, 1≤ j ≤E (8.14)

W
 
di = γwi,n , 1 ≤ i ≤ N (N − 1) (8.15)
n:rni ∈R i w=1
N (N −1) W
  
wj = γwi,n , 1≤ j ≤E (8.16)
i=1 n:e j ∈Eni w=1
W

sk = m kw , 1≤k≤E (8.17)
w=1
N (N −1) E
  
m kw ≤ δwi,n , 1 ≤ k ≤ E, 1≤w≤W (8.18)
i=1 j=1 n:rni ∈R i ,ek ∈Eni
j

N (N −1) E
  
N (N − 1)E Mm kw ≥ δwi,n , 1 ≤ k ≤ E, 1 ≤ w ≤ W (8.19)
i=1 j=1 n:rni ∈R ij ,ek ∈Eni
 
N (N −1)
 
 γwi,n  + m wj ≤ 1, 1 ≤ j ≤ E, 1≤w≤W (8.20)
i=1 n:e j ∈Eni
W
  W
 
i,n
δw, j = γwi,n , 1 ≤ j ≤ E, 1 ≤ i ≤ N (N − 1). (8.21)
n:rni ∈R ij w=1 n:e j ∈Eni w=1

The first constraint limits the number of wavelength channels on each link to less
than W . The next six constraints refer to the network before the failure occurs, whereas
the eighth constraint refers to the network after a failure has occurred. Specifically,
the second constraint ensures that the demand for each node-pair i is met, the third
constraint evaluates the number of working connections traversing link j, and the fourth
constraint evaluates the spare capacity required on link k. The fifth and sixth constraints
indicate whether wavelength w is utilized for some protection path on link k, while the
seventh constraint is the wavelength-continuity constraint, specifying that at most one
connection (either working or protection) can use wavelength w on link j. Finally, the
eighth constraint defines the total number of rerouted connections between node-pair i
when a failure occurs (link j fails).
As an example, a test case of a 15-node network in [Ramamurthy+99a] shows that
for 35 connections 84 wavelength-links are required when no protection is required and
that number becomes 228, 126, and 254 wavelength-links when dedicated-path, shared-
path, and shared-line approaches are used, respectively. As expected, shared path-based
 Survivability 697

Figure 8.29 Spanning trees used in optical path protection.

protection achieves the best capacity utilization compared to dedicated path and shared
line-based protection. Additional results obtained in [Ellinas+03] for dynamic traffic
using heuristic approaches also demonstrates the fact that the shared backup path-based
approach is the most capacity efficient. However, the recovery time required in this case
is much slower than dedicated protection (on the order of a few hundred milliseconds).
This is mainly due to the additional time it takes to send messages along the protection
path and the time required to determine and execute the cross-connections at each node
along the protection path [Ellinas+03]. The optimizations presented above are for a
relatively small problem so that all possible working/protection path pairs could be
considered. For large networks with large numbers of connections, the set of feasible
alternate routes R i must be restricted to produce a tractable ILP. Additional ILPs and
their performance results for line-based protection in mesh optical networks can be
found in [Ramamurthy+99a], and ILPs for line-based and path-based protection in
mesh networks with wavelength conversion can be found in [Zhang03].

8.4.2.3 Redundant Trees for Automatic Protection Switching

A novel approach to path protection was presented in [Finn+97] and [Medard+99a].
(The assumption in that work is that when a link fails, both directions of that link fail.)
The protection technique is based on creating redundant trees that provide two-edge
or two-node connectivity to all other nodes. Assuming bidirectional connections, two
directed spanning trees (a working and a protection tree) are found for each node,
such that a failure does not affect the traffic upstream of that failure on the working
tree, whereas the traffic downstream from that failure now uses the protection tree.
Figure 8.29 shows an example of such a tree arrangement, protecting the leftmost node
in the network.19

8.4.2.4 Shared Risk Groups

Failures of multiple optical channels are usually due to fiber or cable cuts. Consider the
six-node optical network of Figure 8.30. Each cylinder in the figure represents a conduit.
19
This work has a close relationship to generalized loopback.
698 Multiwavelength Optical Networks

U V

SRG 2

W X

Bridge
(SRG 1)

A B
Y Z

Figure 8.30 Shared risk groups. (After [Ellinas+03, Figure 3a]. Copyright

c 2003 SPIE. Used by
permission of the International Society for Optical Engineering.)

Optical channels carried on the two links connecting two distinct pairs of nodes traverse
the same conduit. If the conduit and the fibers it contains are accidentally severed all
the optical channels inside the conduit fail. The concept of shared risk group (SRG)
expresses the risk relationship that associates all the optical channels with a single
failure [Ramamurthy+01, Strand+01]. Protection and restoration schemes are typically
developed based solely on the graph of the network, assuming that individual link or node
failures are statistically independent events. This does not take into account the SRGs,
which destroy this independence among network link failures. For example, suppose that
(1 + 1) protection is used for the connection between source-destination nodes A and B
in Figure 8.30, accessing the optical network at cross-connects Y and Z, respectively.
The working path uses the single link Y-Z and the protection path uses Y-W-X-Z. If the
bridge goes down, both working and protection paths fail simultaneously, because they
both traverse SRG1. On the other hand, failure of SRG1 can be successfully recovered
if the protection path takes the route Y-W-U-V-X-Z.
As mentioned above, an SRG may consist of all the optical channels in a single
fiber (e.g., SRG2 in Figure 8.30), the optical channels through all the fibers wrapped
in the same cable, or all the optical channels traversing the same conduit (e.g., SRG1
in Figure 8.30). Because a fiber may run through several conduits, an optical channel
may belong to several SRGs. In the previous discussion of dedicated and shared backup
path-based protection, the implicit assumption was that all the optical channels on a link,
and no others, belong to one SRG. However, this is the elementary case. In general, we
can classify all SRGs as three possible types (or a combination of these) (Figure 8.31):
(a) elementary type (e.g., all the optical channels on a link, and no others, belong to
one SRG), (b) several entities passing through the same node share a larger facility
(e.g., several fibers pass through the same conduit on their way to the same node), and
 Survivability 699

SRG type a

(a) (a’ )

SRG type b

(b) (b’ )

SRG type c

Not an accurate
representation
(c) (c’ )

Figure 8.31 SRG classification. (After [Ellinas+03, Figure 3b]. Copyright

c 2003 SPIE. Used by
permission of the International Society for Optical Engineering.)

(c) several entities with different origins and destinations share a larger facility (e.g.,
several fibers pass through a single submarine conduit).
Graphically we can easily represent SRG types (a) and (b) in the form shown in
(a′ ) and (b′ ), respectively, in Figure 8.31. However, a graphical representation of type
(c) SRGs is more complex. If it is presented in the form shown in (c′ ) in Figure 8.31,
additional paths are introduced in the network that are not physically possible in the
actual network. Thus, special care needs to be taken in the routing algorithms to avoid
generating false results in the routing computations.
If an arbitrary set of links can belong to the same SRG, then the problem of finding
SRG-diverse working and protection routes between a pair of nodes in the network has
been shown to be NP-complete [Ellinas+03]. Essentially, the difficulty of (1 + 1) SRG-
diverse routing arises because the SRGs can be defined in arbitrary and unrealistic ways,
forcing an algorithm to enumerate a potentially exponential number of paths in the worst
case. Furthermore, given an optical network with elementary SRGs, and a set of already
established shared path-based protected optical paths, the dynamic routing problem of
finding a feasible working and shared-protection path for a new optical path request
has also been shown to be NP-complete [Ellinas+03].20 In view of the difficulty of the
dynamic routing problem, heuristics have been developed to find working and protection
paths that are SRG-disjoint in networks with dynamic traffic. One possible heuristic that
uses Suurballe’s algorithm [Suurballe74] is presented in [Zhang03]. However, using this
approach there is no guarantee that SRG-disjoint paths will be found, either because

20
The case of elementary SRGs is the one considered in the (static routing) ILP for the shared path protection
above. (The elementary SRG is taken into account by requiring that working paths sharing the same
restoration capacity be link/node disjoint.)
700 Multiwavelength Optical Networks

none may exist or because it may not be possible to map some SRGs of type (c) into a
suitable link-layer graph required by the algorithm.
ILP formulations for link failure protection in a mesh network with duct-layer con-
straints (a duct is just a representation of a general type of SRG) are presented in
[Zhang+03]. The objective of the ILP is to minimize the total number of wavelength-
links in transparent (wavelength-continuous) networks for the combined routing and
wavelength assignment problem under dedicated and shared path-based protection. The
ILPs developed are found to be too difficult to solve for networks of realistic size because
of the large number of variables involved and the large size of the search space, so the
problem is subdivided into two separate ILPs, one for the routing subproblem and one
for the wavelength assignment subproblem.
Finally, to protect the network against link and node failures, the working and backup
paths need to be SRG and node disjoint. SRG and node-diverse paths are more demanding
in terms of backup resources compared to the SRG-diverse approach. Performance
analysis in [Ellinas+03] for networks with 17 and 100 nodes demonstrated that the
amount of additional resources required to recover from node failures as well as link
failures is relatively small considering the benefits of protecting the network against
node failures as well.

8.4.3 Segment Protection

Segment protection was first introduced in [Ou+01] and [Ho+02]. The scheme presented
in [Ou+01] first partitions the network into several smaller areas and in each area
two fiber disjoint paths are computed for each connection between area border nodes.
However, the technique does not formulate a partitioning scheme nor does it provide for
recovery from node failures. In the segment protection scheme presented in [Ho+02],
called short leap shared protection (SLSP), a working path is subdivided into several
equal-length and overlapped segments, each defined as a protection domain (Figure 8.32).
SLSP then performs the protection process within each protection domain rather than
for the whole path. Because domains overlap, the scheme can recover from node failures
as well. Simulation results reported in [Ho+02] demonstrate that SLSP achieves 100%
restorability with better capacity efficiency than shared path-based protection schemes.
However, the proposed protection process suffers from an increase in the signaling
complexity.
More recent work on segment protection was presented in [Hailemariam+04], where
an island-based protection technique was proposed. It includes a formal way of par-
titioning the network into smaller subnetworks, together with a method of routing the
working and protection paths, and optimizing the protection capacity. It also describes an
implementation of the proposed technique using an “Island Restoration Protocol” (IRP).
In this architecture, the network is first partitioned into subnetworks called “islands,”
protection paths are computed for a given set of working paths, and redundant capacity
is allocated off-line. When a link or node failure occurs, a real-time localized protection
scheme is invoked in the island(s) affected by the failure.
 Survivability 701

Protection domain 1 Protection domain 2 Protection domain 3

Working path

Figure 8.32 SLSP protection scheme. (After [Ho+02, Figure 3]. Copyright
c 2002 IEEE. Used
by permission of The Institute of Electrical and Electronics Engineers, Inc.)

Prior to provisioning the network, an island identification procedure is executed, which

involves decomposing the network into overlapping islands. An island is a subnetwork
consisting of a center node, called the “island node,” its incident links, the nodes adjacent
to the island node and all the links and nodes needed to reroute traffic when the island
node or any of the incident links fail. Several types of islands can be constructed based
on the designer’s needs. Typically, the size of an island can be minimal or it can be
expanded. A minimal island reacts more rapidly to failures, whereas a larger island
is more flexible and has better capacity utilization for recovery. Figure 8.33 gives an
example of different size islands centered on node 22 of the network shown.
When a node fails the recovery procedures and protection path are confined to its
island. For example, the failure of node 22 in Figure 8.33 would be restored via a protec-
tion path confined to island 22. When a link fails, simultaneous protection procedures
are invoked in the two islands centered at its end nodes; e.g., islands centered at nodes
21 and 22 for link (21,22). Thus, recovery procedures are invoked in a given island only
when its center node or one of the links incident on the center node fails. The backup
path used for recovery of a failure within a given island never passes through the island
node. This way, the island protection technique can protect against a link or node failure
using exactly the same procedure and the same back path; that is, the technique is failure
independent.
Identification of the islands allows for the computation of the protection paths and
the reservation of the necessary redundant capacity (this is done prior to the failure).
Various ILP formulations have been proposed for the purpose of optimizing capacity
[Hailemariam+04].
The key features of the island-based protection technique are the fact that it is fail-
ure independent (does not distinguish between link and node failures), it has recovery
times comparable to line-based protection schemes, and capacity utilization better than
or comparable to competing shared protection capacity schemes [Hailemariam+04].
Fast recovery speeds are achieved because it performs localized (island-based) fault
protection resulting in short protection paths and limited signaling overhead. The
fact that the protection paths are confined to islands also means that multiple fail-
ures can be recovered provided that they affect disjoint islands. Precomputation of the
702 Multiwavelength Optical Networks

17 18
8 14
16 20 6 16 19
15 12 15 7
12 20
21 21
22 24 9 12 15 21 24
22 11
23 11 23 22 23
11
10
26 26 27
27
26 27 25
25 25
28

(a) Minimal island I22 (b) Shortest Path island I22 (c) 2-stage island I22

Figure 8.33 Examples of different types of islands centered on node 22. (From [Hailemariam+04,
Figure 1]. Copyright
c 2004 IEEE. Used by permission of The Institute of Electrical and Elec-
tronics Engineers, Inc.)

protection paths also minimizes real-time functions. Furthermore, the short protection
paths together with shared protection techniques result in efficient capacity utilization.

8.4.4 Survivability Techniques for Multicast Connections

As applications for multicast connections gain more prominence, attention has turned to
survivability techniques for multicast connections. [Singhal+03] proposes a segment-
protection and a path-pair-protection approach to protect multicast sessions against
single fiber cuts. The segment-protection scheme computes protection segments disjoint
from the working segments in the multicast tree, and the path-pair-protection approach
finds a link-disjoint path from the source node to every destination node. One would
assume that to protect a multicast session it would be sufficient to find arc-disjoint
working and protection trees using a minimum cost path heuristic.21 However, this
is not the case; arc-disjoint trees often do not efficiently protect a multicast session.
Furthermore, there are even cases where a protection path cannot be found using this
approach even though enough resources are available in the network. The examples in
Figure 8.34 illustrate that the path-pair-protection and segment-protection approaches
21
In a minimum-cost path heuristic, nearby destination nodes are picked one by one using shortest path
routing and the path from source to each destination is added to the partially built tree until all destinations
for a multicast session are accounted for. When a working tree is created in this fashion all the arcs of the
working tree are removed from the network graph and a protection tree is found using the same method. In
this way arc-disjoint working and protection trees are created.
 Survivability 703

S
S S

U V
U

W
d2 V W

d1
d1 d2 d1 d2
X

(a) Optimal path approach (b) Shared segment-disjoint (c) Optimal path-pair approach
succeeds where arc-disjoint approach succeeds where succeeds where shared
tree approach fails arc-disjoint tree approach fails segment-disjoint approach fails

Figure 8.34 Examples of segment and path-pair protection of multicast sessions. (From
[Singhal+03, Figure 2]. Copyright

c 2003 IEEE. Used by permission of The Institute of Electrical
and Electronics Engineers, Inc.)

are superior to the arc-disjoint approach in protecting multicast sessions. In each

case, the source node S is connected to destination nodes d1 and d2 . The solid lines
indicate the available network links (unidirectional or bidirectional as indicated by the
arrows), the working tree is shown with dotted lines and the protection tree with dashed
lines. Finally, the path-pair approach succeeds even in the case where the segment-
protection scheme fails to protect the multicast session. (See [Singhal+03] for the
construction procedures.)
More recent work in [Rahman+05] uses a collapsed-ring dedicated (1 + 1) protec-
tion scheme to protect multicast sessions. The minimum-cost collapsed-ring (MC-CR)
algorithm finds a ring path among the source and all destination nodes and establishes
two optical paths, starting in opposite directions from the source and terminating at
the final node of the destination set after traversing the ring path. An example of this
scheme in comparison to a minimum-cost heuristic (MCH) is shown in Figure 8.35.
Simulation results comparing MC-CR with the arc-disjoint algorithm described above
and with MCH show that MC-CR uses less protection capacity than the others for large
multicast session sizes. Additional work in [Khalil+05] further optimizes the MC-CR
algorithm, obtaining even better performance results in terms of required protection
capacity.

8.5 Summary

Survivability techniques have evolved over the years, from those developed for SONET
networks to techniques for networks that utilize DCSs and more recently to those for
mesh optical networks. The nature of the fault recovery techniques is largely determined
by the topology of the system being protected. Point-to-point systems utilize automatic
protection switching (APS), whereas ring networks utilize self-healing ring architectures
704 Multiwavelength Optical Networks

d1
1

1 1
S 1 d2 Fiber-Link Connectivity
2
2
2 2
d3
1
d1
1

1 1
S d2 1+1 Multicast Tree using MCH
1
2
2
2 2
d3
1
d1
1 1

1
S 1 d2 1+1 Multicast Tree using MC-CR
2
2
2 2
d3
1

Working

Protection

Figure 8.35 Example illustrating the arc-disjoint and MC-CR algorithms. (From [Rahman+05,
Figure 1]. Copyright
c 2005 IEEE. Used by permission of The Institute of Electrical and Elec-
tronics Engineers, Inc.)

to protect their connections against single link/node failure scenarios. Survivability

techniques for mesh networks can be divided in broad terms into two categories: (1)
protection techniques, which are those that precompute the protection path prior to the
failure and use protection capacity reserved in advance for each failure scenario, and (2)
restoration techniques, which are those that compute the protection path dynamically
in real time after the failure has occurred. Fault restoration uses the currently available
redundant capacity to restore a failed connection, but in contrast to fault protection it does
not reserve capacity in advance. These techniques are further subdivided into line-based
and path-based protection/restoration schemes, which can use redundant capacity either
dedicated for each failure scenario or shared among different working connections
depending on some diversity constraints. Furthermore, working and protection paths
can be calculated using a centralized or distributed approach, and the recovery process
itself can also be centralized or distributed. As in the case of networks with DCSs,
centralized restoration approaches for optical networks always yield better results in
terms of capacity efficiency, whereas distributed restoration schemes are more scalable
and have faster recovery times.
 Survivability 705

One approach to line-based protection for mesh networks involves decomposing the
network into rings and using self-healing ring techniques to protect against failures. The
simplest way to do this is by finding a ring cover for a network, an approach that enables
extremely fast failure recovery. However, this approach generally requires more than
100% redundancy. A better ring-based scheme, called cycle double covers, guarantees
single failure recovery with exactly 100% redundancy for networks with general mesh
topologies while still maintaining fast recovery times. p-Cycles is the third ring-based
scheme discussed in this chapter. It is very attractive as it requires redundant capacity
similar to that required for path-based protection in mesh networks while exhibiting
APS-like recovery times.
A fourth line-based protection approach that is not ring-based but uses similar APS
techniques is the generalized loopback method. Generalized loopback creates a primary
and secondary digraph for each network and uses the secondary digraph to protect
against any failures in the primary digraph utilizing an APS process. As in cycle double
covers, this approach requires exactly 100% redundancy but is more flexible for recovery
of node as well as link failures.
Path-based protection schemes analyzed here include a redundant tree approach,
where a working and a protection spanning tree are found for every node, and the
protection tree is used when a failure is detected on the working tree. Dedicated (1 + 1)
and shared path-based protection schemes are also examined and it is shown that shared
schemes always perform better in terms of capacity efficiency while dedicated schemes
outperform shared schemes in terms of recovery speed. In dedicated protection, each
protection path is dedicated to a single working path and these paths are link, node,
or SRG disjoint. In shared protection, multiple disjoint working paths can share a
protection path. In that case the redundant capacity is only soft-reserved and protection
paths are established after the failure event, for each working path affected by the
failure.
Survivability techniques that do not fall into either the line- or path-based categories
were also presented in this chapter. They are called segment recovery techniques and
they have properties of both line- and path-based methods. They demonstrate how
a hybrid approach can exploit the best features of both the line- and path-based ap-
proaches while avoiding their limitations. Some recent work on protection of multi-
cast sessions is also described in this chapter, although this work is still in its early
stages.
Dynamic (real-time) restoration in mesh optical networks has not been discussed here
because it is similar to the restoration approaches for traditional DCS networks that were
covered in Sections 8.2.4.2 and 8.2.4.3. This topic has received considerably less attention
in the literature than preplanned protection approaches. In [Ramamurthy+99b] some
distributed control protocols are described for both the line- and path-based restoration
approaches. The authors use two metrics to compare the two schemes: restoration time
and restoration efficiency (defined as the proportion of the failed connections restored),
and they demonstrate that the restoration efficiency is better in path-based restoration
while the restoration time is better in line-based restoration.
706 Multiwavelength Optical Networks

8.6 Problems

1 A network is represented by a seven-vertex graph with the following edges: (1, 3),
(1, 4), (1, 5), (1, 6), (2, 4), (2, 5), (2, 7), (3, 5), (3, 6), (3, 7), (4, 6), (4, 7), (5, 7), and
(6, 7). Is it planar or nonplanar? If it is planar, embed it in the plane and obtain the
protection cycles for the network as face boundaries. If it is nonplanar, find an orientable
CDC.
2 A number of proposed failure protection schemes are based on the use of protection
cycles using redundant protection fibers. What are some of the physical problems that
can arise in these protection cycles if optical amplifiers are present in the protection
fibers? Propose methods of eliminating these problems.
3 The fault protection/restoration schemes presented in this chapter require a system
that detects the failures and responds appropriately.
(a) Present a block diagram of a detection and control system that might be employed
at the end of each fiber link in a purely optical network.
(b) What are some of the detection problems that might be encountered in such a network
and what are some possible solutions? (Consider the possibility of fiber amplifiers
in the links.)
(c) How can the system be configured so that both sides of a link react properly to a
failure in case of a partial cable cut (e.g., only one working fiber, one protection
fiber, or one working/protection fiber pair is cut).
4 The protection/restoration schemes presented in the latter part of this chapter are
used for the survivability of the WDM transport layer. Assuming that the WDM
layer lies below a SONET layer, which in turn lies below an IP layer (i.e., this is
an IP/SONET/WDM architecture), and that all three technologies have their own
protection/restoration schemes, certain interoperability problems may arise. Describe
some of these problems and outline possible solutions.
5 A network is represented by a seven-vertex graph with edges (1, 2), (1, 3), (1, 4),
(1, 5), (1, 6), (2, 3), (3, 4), (4, 5), (5, 6), (6, 7), and (7, 2). All links are bidirectional.
(a) How many simultaneous link failures can be protected given the scheme of
Section 8.4.1.2? (Consider the two cases of bidirectional and unidirectional
connections.)
(b) Give an example of multiple link failures in a case in which several unidi-
rectional connections are active, and indicate the link failures and connections
restored.
(c) Repeat the previous part for bidirectional connections.
6 Consider a network represented by an 11-vertex graph with edges (1, 2), (1, 3),
(1, 5), (1, 7), (2, 3), (2, 4), (2, 8), (3, 4), (3, 5), (3, 6), (4, 7), (4, 8), (4, 11), (5, 6),
(5, 7), (6, 8), (6, 9), (7, 9), (7, 10), (8, 9), (8, 11), (9, 10), (9, 11), and (10, 11). All
links are bidirectional.
 Survivability 707

(a) Assume that the ring cover method of Section 8.2.4.1 is used to obtain a family of
protection cycles. Find a set of fundamental cycles and calculate the number of extra
protection fibers required so the network can survive any single link failure.
(b) Another method of finding a ring cover is by first obtaining a planar subgraph for the
nonplanar graph. Describe this method and obtain the required ring cover as well as
the number of extra protection fibers required for the network.
7 A 10-node network is fully connected, with each of its links bidirectional. Find two
spanning trees with the characteristics defined in Section 8.4.2.3 so that they can always
protect node 1 of the network.

Bibliography

[Alspach+85] B. R. Alspach and C. D. Godsil. Cycles in Graphs. Annals of Discrete Mathematics,

Vol. 115. Amsterdam: Elsevier Science 1985.
[ANSI88] American National Standards Institute. Optical Interface Rates and Formats Specifi-
cation, ANSI Standard T1.105, 1988.
[Ayanoglu+96] E. Ayanoglu and R. D. Gitlin. Broadband network restoration. IEEE Communi-
cations Mag., 34(7):110–119, 1996.
[Bala96] K. Bala. Multiwavelength optical network architecture. In Proceedings of the Seventh
Workshop on Very High Speed Networks, Baltimore, MD, July 1996.
[Bala+96] K. Bala and C. A. Brackett. Cycles in wavelength routed optical networks. IEEE/OSA
J. Lightwave Technol., 14(7):1585–1594, 1996.
[BeHannesey98] M. BeHannesey. Ring interworking and BLSR. In Proceedings of the National
Fiber Optics Engineers Conference (NFOEC), Orlando, FL, September 1998.
[Bell87] Bell Communications Research. LATA Switching Systems Generic Requirements. Tech-
nical Report No. TR-TSY-0C0064, Issue 2, July 1987.
[Bell90a] Bell Communications Research. Automatic Protection Switching for SONET. Technical
Report No. SR-NWT-001756, October 1990.
[Bell90b] Bell Communications Research. Generic Reliability Assurance Requirements for Fiber
Optic Systems. Technical Report No. TR-NWT-000418, 1990.
[Bell91] Bell Communications Research. Synchronous Optical Network (SONET) Transport
Systems: Common Generic Criteria. Technical Report No. TR-NWT-000253, December
1991.
[Bell93] Bell Communications Research. Transport Systems: Generic Requirements. Technical
Report No. TR-NWT-000499, December 1993.
[Bell94a] Bell Communications Research. SONET Dual-Fed Unidirectional Path-Switched Ring
(UPSR) Equipment Generic Criteria. Technical Report No. GR-1400-CORE, March 1994.
[Bell94b] Bell Communications Research. SONET Ring Interworking Issues and Solutions. Tech-
nical Report No. SR-3039, June 1994.
[Bell95] Bell Communications Research. SONET Logical Ring Services—Network Element and
Operations System Impacts. Technical Report No. SR-3487, April 1995.
[Bhandari99] R. Bhandari. Survivable Networks: Algorithms for Diverse Routing. Norwell, MA:
Kluwer Academic 1999.
708 Multiwavelength Optical Networks

[Bicknell+93] J. Bicknell, C. E. Chow, and S. Syed. Performance analysis of fast distributed

network restoration algorithms. In Proceedings of the IEEE Global Telecommunications Con-
ference (Globecom), pp. 1596–1600, Houston, TX, November 1993.
[Bonenfant+94] P. A. Bonenfant and C. M. C. Davenport. SONET optical layer protection switch-
ing. Bell Communications Research, Technical Report No. TM-24609, December 1994.
[Bouillet+02a] E. Bouillet, J.-F. Labourdette, R. Ramamurthy, and S. Chaudhuri. Enhanced
algorithm cost model to control tradeoffs in provisioning shared mesh restored lightpaths.
In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Anaheim, CA, March
2002.
[Bouillet+02b] E. Bouillet, J.-F. Labourdette, G. Ellinas, R. Ramamurthy, and S. Chaudhuri.
Stochastic approaches to route shared mesh restored lightpaths in optical mesh networks. In
Proceedings of the IEEE Conference on Computer Communications (Infocom), New York, NY,
June 2002.
[Bouillet+07] E. Bouillet, G. Ellinas, J.-F. Labourdette, and R. Ramamurthy. Path Routing in
Mesh Optical Networks. New York: Wiley 2007.
[Chao+91] C.-W. Chao, P. M. Dollard, J. E. Weythman, L.T. Nguyen, and H. Eslambolchi.
FASTAR—A robust system for fast DS3 restoration. In Proceedings of the IEEE Global
Telecommunications Conference (Globecom), pp. 1396–1400, Phoenix, AZ, December 1991.
[Chartrand+96] G. Chartrand and L. Lesniak-Foster. Graphs and Digraphs. New York: Chapman
& Hall 1996.
[Chow+92] C.-H. E. Chow, S. McCaughey, and S. Syed. RREACT: A distributed protocol for
rapid restoration of active communication trunks. University of Colorado, Computer Science,
Technical Report No. EAS-CS-92-18, 1992.
[Chujo+90] T. Chujo, H. Komine, K. Miyazaki, T. Ogura, and T. Soejima. The design and
simulation of an intelligent transport network with distributed control. In Proceedings of the
Network Operations Management Symposium, pp. 927–937, San Diego, CA, February 1990.
[Coan+91] B. A. Coan, W. E. Leland, M. P. Vecchi, and L. T. Wu. Using distributed topology
update and preplanned configurations to achieve trunk network survivability. IEEE Trans.
Reliabil., 40(4):404–416, 1991.
[Daneshmand+93] M. Daneshmand and C. Savolaine. Measuring outages in telecommunications
switched networks. IEEE Communications Mag., 31(6):34–39, 1993.
[Datta+01] S. Datta, S. Sengupta, and S. Biswas. Efficient channel reservation for backup paths
in optical mesh networks. In Proceedings of the IEEE Global Telecommunications Conference
(Globecom), San Antonio, TX, November 2001.
[Doshi+99] B. Doshi, S. Dravida, P. Harshavardhana, O. Hauser, and Y. Wang. Optical network
design and restoration. Bell-Labs Tech. J., 4(1):58–84, 1999.
[Ellinas+96] G. Ellinas and T. E. Stern. Automatic protection switching for link failures in optical
networks with bi-directional links. In Proceedings of the IEEE Global Telecommunications
Conference (Globecom), London, UK, November 1996.
[Ellinas+97] G. Ellinas, A. Hailemariam, and T. E. Stern. Creation of a family of cycles cor-
responding to the orientable cycle double cover. New York: Center for Telecommunications
Research, Columbia University, Technical Report No. CU/CTR/TR 474-97-08, 1997.
[Ellinas98] G. Ellinas. Fault Restoration in Optical Networks: General Methodology and Imple-
mentation. Ph.D. thesis. Columbia University, New York, 1998.
[Ellinas+99] G. Ellinas, A. Hailemariam, and T. E. Stern. Protection of a priority connection
from an optical switch failure in mesh networks with planar topologies. In Proceedings of the
IEEE/OSA Optical Fiber Commun. Conf., San Diego, CA, February 1999.
 Survivability 709

[Ellinas+00a] G. Ellinas, S. Rong, A. Hailemariam, and T. Stern. Protection cycle covers in

optical networks with arbitrary mesh topologies. In Proceedings of the IEEE/OSA Optical
Fiber Commun. Conf., Baltimore, MD, March 2000.
[Ellinas+00b] G. Ellinas, T. Stern, and A. Hailemariam. Protection cycles in WDM mesh net-
works. IEEE J. Select. Areas Commun., 18(10):1924–1937, 2000.
[Ellinas+03] G. Ellinas, E. Bouillet, R. Ramamurthy, J.-F. Labourdette, S. Chaudhuri, and
K. Bala. Routing and restoration architectures in mesh optical networks. SPIE Optical Net-
works Mag., 4(1):91–106, 2003.
[Elrefaie92] A. F. Elrefaie. Self-healing WDM ring networks with an all-optical protection path.
In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper ThL3, pp. 255–256, San
Jose, CA, February 1992.
[Ester94] G. W. Ester. Can you talk to your SONET neighbor? Telephony Mag., November
1994.
[Even79] S. Even. Graph Algorithms. Rockville, MD: Computer Science Press 1979.
[Fagerstorm+93] R. Fagerstorm and J. Healy. The reliability of LEC telephone networks. IEEE
Communications Mag., 31(6):44–49, 1993.
[Finn+97] S. G. Finn, M. M. Medard, and R. A. Barry. A novel approach to automatic protection
switching using trees. In Proceedings of the IEEE Int’l Conf. Commun. (ICC), pp. 272–276,
Montreal, Canada, June 1997.
[Finn+98] S. G. Finn, M. M. Medard, and R. A. Barry. A new algorithm for bi-directional
self healing for arbitrary redundant networks. In Proceedings of the IEEE/OSA Optical Fiber
Commun. Conf., San Jose, CA, February 1998.
[Fisher65] G. Fisher. Computer Identification and Extraction of Planar Graphs. Ph.D. thesis.
Department of Electrical Engineering, Columbia University, New York, 1965.
[Gardner+94] L. M. Gardner, M. Heydari, J. Shah, I. H. Sudborough, I. G. Tollis, and C. Xia.
Techniques for finding ring covers in survivable networks. In Proceedings of the IEEE Global
Telecommunications Conference (Globecom), pp. 1862–1866, San Francisco, CA, November
1994.
[Gardner+95] L. M. Gardner, I. H. Sudborough, and I. G. Tollis. NetSolver: A software tool
for the design of survivable networks. In Proceedings of the IEEE Global Telecommunications
Conference (Globecom), pp. 926–930, Singapore, November 1995.
[Gibbons85] A. Gibbons. Algorithmic Graph Theory. Cambridge: Cambridge University Press
1985.
[Glossbrenner93] K. C. Glossbrenner. Availability and reliability of switched services. IEEE
Communications Mag., 31(6):28–33, 1993.
[Grover87] W. D. Grover. The SelfHealing network. In Proceedings of the IEEE Global Telecom-
munications Conference (Globecom), pp. 28.2.1–28.2.6, Tokyo, Japan, November 1987.
[Grover89] W. D. Grover. SelfHealing Networks: A Distributed Algorithm for k-Shortest Link-
Disjoint Paths in a Multi-Graph with Applications in Real-Time Network Restoration. Ph.D.
thesis. The University of Alberta, Edmonton, 1989.
[Grover+90] W. D. Grover, B. D. Venables, J. H. Sandham, and A. F. Milne. Performance studies
of a SelfHealing network protocol in Telecom Canada long haul networks. In Proceedings of
the IEEE Global Telecommunications Conference (Globecom), pp. 453–458, San Diego, CA,
December 1990.
[Grover+91] W. D. Grover, T. D. Bilodeau, and B. D. Venables. Near optimal spare capacity
planning in a mesh restorable network. In Proceedings of the IEEE Global Telecommunications
Conference (Globecom), pp. 2007–2012, Phoenix, AZ, December 1991.
710 Multiwavelength Optical Networks

[Grover92] W. D. Grover. Case studies of survivable ring, mesh and mesh-arc hybrid networks.
In Proceedings of the IEEE Global Telecommunications Conference (Globecom), pp. 633–638,
Orlando, FL, December 1992.
[Grover+98] W. D. Grover and D. Stamatelakis. Cycle-oriented distributed preconfiguration:
Ring-like speed with mesh-like capacity for self-planning network restoration. In Proceedings
of the IEEE Int’l Conf. Commun. (ICC), Atlanta, GA, June 1998.
[Grover04] W. Grover. Mesh-based Survivable Networks: Options and Strategies for Optical,
MPLS, SONET and ATM Networking. Upper Saddle River, NJ: Prentice Hall 2004.
[Grover+05] W. Grover and A. Kodian. Failure-independent path protection with p-cycles: Ef-
ficient, fast and simple protection for transparent optical networks. In Proceedings of the 7th
International Conference on Transparent Optical Networks, pp. 363–369, Barcelona, Spain,
July 2005.
[Hailemariam+04] A. Hailemariam, G. Ellinas, and T. E. Stern. Localized restoration in mesh
optical networks. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Los Angeles,
CA, February 2004.
[Hall+89] R. D. Hall and S. Whitt. Protection of SONET based networks. In Proceedings of the
IEEE Global Telecommunications Conference (Globecom), pp. 821–825, Dallas, TX, November
1989.
[Harman+94] W. Harman, H. Kobrinski, and R. Kutz. Interworking among Transport Network
Architectures. Bell Communications Research, Technical Report No. TM-24124, June 1994.
[Ho+02] P. Ho and H. T. Mouftah. A framework for service-guaranteed shared protection in
WDM mesh networks. IEEE Communications Mag., 40(2): 97–103, 2002.
[Iness+96] J. Iness, B. Ramamurthy, B. Mukherjee, and K. Bala. Elimination of all-optical cycles
in wavelength-routed optical networks. IEEE/OSA J. Lightwave Technol., 14(6):1207–1217,
1996.
[Iraschko+96] R. R. Iraschko, M. H. MacGregor, and W. D. Grover. Optimal capacity placement
for path restoration in mesh survivable networks. In Proceedings of the IEEE Int’l Conf.
Commun. (ICC), pp. 1568–1574, Dallas, TX, June 1996.
[Iraschko+98] R. R. Iraschko, M. H. MacGregor, and W. D. Grover. Optimal capacity placement
for path restoration in STM or ATM mesh survivable networks. In IEEE/ACM Trans. Network.,
6(3):325–226 1998.
[Johnson+94] D. Johnson, G. N. Brown, S. L. Beggs, C. P. Botham, I. Hawker, R. S. K. Chng,
M. C. Sinclair, and M. J. O’Mahony. Distributed restoration strategies in telecommunications
networks. In Proceedings of the IEEE Int’l Conf. Commun. (ICC), New Orleans, LA, May
1994.
[Kajiyama+92] Y. Kajiyama, N. Tokura, and K. Kikuchi. ATM self-healing ring. In Proceedings
of the IEEE Global Telecommunications Conference (Globecom), pp. 639–643, Orlando, FL,
December 1992.
[Kajiyama+94] Y. Kajiyama, N. Tokura, and K. Kikuchi. An ATM VP-based self-healing ring.
IEEE J. Select. Areas Commun., 12(1):171–178, 1994.
[Kawamura+94] R. Kawamura, K. Sato, and I. Tokizawa. Self-healing ATM networks based on
virtual path concept. IEEE J. Select. Areas Commun., 12(1):120–127, 1994.
[Kawamura+95] R. Kawamura and I. Tokizawa. Self-healing virtual path architecture in ATM
networks. IEEE Communications Mag., 33(9):72–79, 1995.
[Khalil+05] A. Khalil, T. Rahman, A. Hadjiantonis, G. Ellinas, and M. Ali. Pre-planned multicast
protection approaches in WDM mesh networks. In Proceedings of the European Conference on
Optical Communications, Glasgow, Scotland, October 2005.
 Survivability 711

[Komine+90] H. Komine, T. Chujo, T. Ogura, K. Miyazaki, and T. Soejima. A distributed restora-

tion algorithm for multiple-link and node failures of transport networks. In Proceedings of
the IEEE Global Telecommunications Conference (Globecom), pp. 459–463, San Diego, CA,
December 1990.
[Kuhn97] D. R. Kuhn. Sources of failure in the public switched telephone network. IEEE Comput.
Mag., 30(4):31–36, 1997.
[Labourdette+02] J. Labourdette, E. Bouillet, R. Ramamurthy, G. Ellinas, S. Chaudhuri, and
K. Bala. Routing strategies for capacity-efficient and fast-restorable mesh optical networks.
Photon. Network Commun., 4(3–4):219–235, 2002.
[Lee01] C. Lee. When disaster strikes: Putting corporate values to the test. Chief Executive
Magazine’s 4th Annual “CEO2CEO” Conference, New York, NY, October 31, 2001.
[Lumetta+00] S. Lumetta, M. Medard, and Y.-C. Tseng. Capacity versus robustness: A tradeoff
for link restoration in mesh networks. IEEE/OSA J. Lightwave Technol., 18(12):1765–1775,
2000.
[Maier+02] G. Maier, A. Pattavina, S. De Patre, and M. Martinelli. Optical network survivabil-
ity: Protection techniques in the WDM layer. Photon. Network Commun., 4(3–4):251–269,
2002.
[Manchester+96] J. Manchester and P. Bonenfant. Fiber optic network survivability: SONET/
optical protection layer interworking. In Proceedings of the National Fiber Optics Engineers
Conference (NFOEC), Denver, CO, September 1996.
[May+91] G. May and D. Jammu. A distributed architecture for survivable SONET transport
networks. In Proceedings of the IEEE Global Telecommunications Conference (Globecom),
pp. 2013–2017, Phoenix, AZ, December 1991.
[May+95] K. P. May, P. Semal, Y. Du, and C. Herrmann. A fast restoration system for ATM-ring-
based LANs. IEEE Communications Mag., 33(9):90–99, 1995.
[Medard+99a] M. Medard, S. G. Finn, R. G. Gallager, and R. A. Barry. Redundant trees for auto-
matic protection switching in arbitrary node-redundant or edge-redundant graphs. IEEE/ACM
Trans. Network., 7(5):641–652, 1999.
[Medard+99b] M. Medard, S. G. Finn, and R. A. Barry. WDM loop-back recovery in mesh
networks. In Proceedings of the IEEE Conference on Computer Communications (Infocom),
vol. 2, pp. 752–759, New York, NY, March 1999.
[Medard+02] M. Medard, R. A. Barry, S. G. Finn, W. He, and S. Lumetta. Generalized loop-back
recovery in optical mesh networks. IEEE/ACM Trans. Network., 10(1):153–164, 2002.
[Nederlof+95] L. Nederlof, K. Struyve, C. O’Shea, H. Misser, Y. Du, and B. Tamayo. End-to-end
survivable broadband networks. IEEE Communications Mag., 33(9):63–71, 1995.
[Ou+01] C. Ou, H. Zang, and B. Mukherjee. Sub-path protection for scalability and fast recovery
in WDM mesh networks. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf.,
Anaheim, CA, February 2001.
[Rahman+05] T. Rahman and G. Ellinas. Protection of multicast sessions in WDM mesh optical
networks. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Anaheim, CA, March
2005.
[Ramamurthy+99a] S. Ramamurthy and B. Mukherjee. Survivable WDM mesh networks:
Part I – Protection. In Proceedings of the IEEE Conference on Computer Communications
(Infocom), vol. 2, pp. 744–751, New York, NY, March 1999.
[Ramamurthy+99b] S. Ramamurthy and B. Mukherjee. Survivable WDM mesh networks:
Part II – Restoration. In Proceedings of the IEEE Int’l Conf. Commun. (ICC), vol. 3,
pp. 2023–2030, Vancouver, Canada, June 1999.
712 Multiwavelength Optical Networks

[Ramamurthy+01] R. Ramamurthy, Z. Bogdanowicz, S., Samieian, D. Saha, B. Rajagopalan,

S. Sengupta, S. Chaudhuri, and K. Bala. Capacity performance of dynamic provisioning in
optical networks. In IEEE/OSA J. Lightwave Technol., 19(1):40–48, 2001.
[Sack+04] A. Sack and W. Grover. Hamiltonian p-cycles for fiber-level protection in homoge-
neous and semi-homogeneous optical networks. IEEE Network, 18(2):49–56, 2004.
[Sakauchi+90] H. Sakauchi, Y. Nishimura, and S. Hasegawa. A self-healing network with an
economical spare-channel assignment. In Proceedings of the IEEE Global Telecommunications
Conference (Globecom), pp. 438–443, San Diego, CA, December 1990.
[Sato+90] K. Sato, H. Hadama, and I. Tokizawa. Network reliability enhancement with virtual
path strategy. In Proceedings of the IEEE Global Telecommunications Conference (Globecom),
pp. 464–469, San Diego, CA, December 1990.
[Schupke+02] D. A. Schupke, C. Gruber, and A. Autenrieth. Optimal configuration of p-cycles
in WDM networks. In Proceedings of the IEEE Int’l Conf. Commun. (ICC), New York, NY,
April/May 2002.
[Shen+03] G. Shen and W. D. Grover. Extending the p-cycle concept to path segment protection
for span and node failure recovery. IEEE J. Select. Areas Commun., 21(8):1306–1319, 2003.
[Shyur+94] C.-C. Shyur, Y.-M. Wu, and C.-H. Chen. The capacity comparison and cost analyses
for SONET self-healing ring networks. IEICE Trans. Commun., E77-B(2):218–225, 1994.
[Siller+96] C. A. Siller, Jr. and M. Shafi. SONET and SDH – A Sourcebook of Synchronous
Networking. New York: IEEE Press 1996.
[Singhal+03] N. Singhal and B. Mukherjee. Dynamic provisioning of survivable multicast ses-
sions in optical WDM mesh networks. In Proceedings of the IEEE/OSA Optical Fiber Commun.
Conf., pp. 207–208, Atlanta, GA, March 2003.
[Slevinsky+93] J. B. Slevinsky, W. D. Grover, and M. H. MacGregor. An algorithm for survivable
network design employing multiple self-healing rings. In Proceedings of the IEEE Global
Telecommunications Conference (Globecom), pp. 1568–1573, Houston, TX, December 1993.
[Somani+97] A. K. Somani and N. H. Vaidya. Understanding fault tolerance and reliability. IEEE
Comput. Mag., 30(4):45–50, 1997.
[Sosnosky+91] J. Sosnosky, T.-H. Wu, and D. L. Alt. A study of economics, operations and
applications of SONET self-healing ring architectures. In Proceedings of the IEEE Global
Telecommunications Conference (Globecom), pp. 2018–2024, Phoenix, AZ, December 1991.
[Stamatelakis97] D. Stamatelakis. Theory and Algorithms for Pre-configuration of Spare Capacity
in Mesh Restorable Networks. M.Sc. thesis. The University of Alberta, Edmonton, 1997.
[Stamatelakis+00] D. Stamatelakis and W. D. Grover. Theoretical underpinnings for the effi-
ciency of restorable networks using preconfigured cycles (p-cycles). IEEE Trans. Commun.,
48(8):1262–1265, 2000.
[Strand+01] J. Strand, A. Chiu, and R. Tkach. Issues for routing in the optical layer. IEEE
Communications Mag., 39(2):81–87, 2001.
[Suurballe74] J. W. Suurballe. Disjoint paths in a network. Networks, 4:125–145, 1974.
[Tsai+90] E. I. Tsai, B. A. Coan, M. Kerner, and M. P. Vecchi. A comparison of strategies for
survivable network design: Reconfigurable and conventional approaches. In Proceedings of
the IEEE Global Telecommunications Conference (Globecom), pp. 49–55, San Diego, CA,
December 1990.
[Veitch+96] P. Veitch, I. Hawker, and G. Smith. Administration of restorable virtual path mesh
networks. IEEE Communications Mag., 34(12):96–101, 1996.
[Wasem91a] O. J. Wasem. An algorithm for designing rings for survivable fiber networks. IEEE
Trans. Reliabil., 40(4):428–432, 1991.
 Survivability 713

[Wasem91b] O. J. Wasem. Optimal topologies for survivable fiber optic networks using SONET
self-healing rings. In Proceedings of the IEEE Global Telecommunications Conference (Globe-
com), pp. 2032–2038, Phoenix, AZ, December 1991.
[Wasem+92] O. J. Wasem, R. H. Cardwell, and T.-H. Wu. Software for designing survivable
SONET networks using self-healing rings. In Proceedings of the IEEE Int’l Conf. Commun.
(ICC), pp. 425–431, Chicago, IL, June 1992.
[Wrobel90] L. A. Wrobel. Disaster Recovery Planning for Telecommunications. Norwood, MA:
Artech House 1990.
[Wu+89] T.-H. Wu, D. J. Kolar, and R. H. Cardwell. High-speed self-healing ring architectures for
future interoffice networks. In Proceedings of the IEEE Global Telecommunications Conference
(Globecom), pp. 801–807, Dallas, TX, November 1989.
[Wu+90] T.-H. Wu and R. C. Lau. A class of self-healing ring architectures for SONET network
applications. In Proceedings of the IEEE Global Telecommunications Conference (Globecom),
pp. 444–451, San Diego, CA, December 1990.
[Wu92] T.-H. Wu. Fiber Network Service Survivability. Norwood, MA: Artech House 1992.
[Xin+98] W. Xin, G.-K. Chang, H. Dai, J. Young, T. Robe, S. J. Yoo, B. Meagher, M. Rauch,
R. Wagner, J. Jackel, J. Baran, G. Ellinas, and M. Z. Iqbal. Performance and operation of
WDM layer automatic protection switching in a 1177 km reconfigurable multiwavelength ring
network. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper PD25, San
Jose, CA, February 1998.
[Yang+88] C. H. Yang and S. Hasegawa. FITNESS: Failure immunization technology for net-
work service survivability. In Proceedings of the IEEE Global Telecommunications Conference
(Globecom), pp. 1549–1554, Hollywood, FL, November 1988.
[Zhang03] H. Zhang. WDM Mesh Networks: Management and Survivability. Norwell, MA:
Kluwer Academic 2003.
[Zhang+03] H. Zhang, C. Ou, and B. Mukherjee. Path-protection routing and wavelength assign-
ment (RWA) in WDM mesh networks under duct-layer constraints. IEEE/ACM Trans. Network.,
11(2):248–258, 2003.
9 Optical Control Plane

In Chapter 2 we proposed a layered view of the connections in an optical network, focus-

ing primarily on issues associated with optical layer transport but including a discussion
of transport in logical network (e.g., IP network) overlays as well. Then in Section 3.1
we encountered a different way of “slicing” the functionality of an optical network,
distinguishing three planes: transport, control, and management. In general terms, the
transport plane is responsible for the physical transfer of data across an optical network,
the control plane provides the intelligence required for the provisioning and mainte-
nance (e.g., failure recovery operations) of a connection, and the management plane
provides management services such as performance monitoring, fault and configuration
management, accounting and security management. This chapter provides a summary
of the current state of optical network control, which is a broad and rapidly evolving
subject. The reader is referred to texts completely devoted to the subject of control (e.g.,
[Bernstein+04] for a more comprehensive treatment).
The line between management and control is not clearly defined. But roughly speak-
ing, management functions deal with long-term issues and operate on slow timescales,
whereas control functions are associated with rapid changes in network configurations
and operate on short timescales. For example, the repair of a network fault such as a cut
cable would be a management function. It might require days or weeks. On the other
hand, “point-and-click” provisioning, where a network user controls the provisioning
and configuration of a connection, is a control function. The same is true for automatic
fault recovery, where connections are detoured around a fault using spare capacity re-
served for this purpose (Chapter 8). Each is performed in real time, with fault recovery
sometimes operating on millisecond timescales.
The transport plane has been covered in depth in earlier chapters and is the dominant
focus of this book. Transport in both the optical and logical layers has been treated largely
from an abstract point of view, concentrating on methodologies for network design and
performance evaluation. The transport network components were usually presented
in the form of generic optical switching nodes, links, and access stations. Once we
move from network design to network operation, the control and management functions
become paramount. Control and management functions are closely tied to the charac-
teristics of real network elements and architectures so we move away from the abstract
approach toward real-world issues. Thus this chapter is a reflection of the characteristics
of current and evolving equipment (network elements) and standards. The standards are
designed to enable networks involving diverse equipment and multiple administrative
 Optical Control Plane 715

Optical network

IP/MPLS router

Optical node Edge

Router Edge
Core

Figure 9.1 A mesh optical network. (From [Xin+01, Figure 1]. Copyright
c 2001 IEEE. Used by
permission of The Institute of Electrical and Electronics Engineers, Inc.)

domains to operate smoothly, and, conversely, the equipment and control/management

procedures are configured to implement evolving standards. A basic driving force
behind the current developments in optical network control and management is the
need to unify and coordinate these operations in networks containing multivendor
equipment.
A multilayered network typically contains an underlying optical transport network
(the physical layer of Figure 2.1), acting as a server for one or more logically switched
client networks (the logical layer in the figure). A possible configuration might be
a client network consisting of high-speed IP routers interconnected through a WDM
optical core network (Figure 9.1).1 Optical connections in the core network (lightpaths)
support logical connections among the switches and routers in the client network. The
client-server relation can be more general than this. A client of a transport network
may be an end system rather than a logical network overlay and a client network (e.g.,
a SONET network) may also act as a server for a higher layer client (e.g., an IP
network). Furthermore a network may be a client riding over more than one transport
network (e.g., an IP network may be served by two independently administered optical
networks).
A node in the optical core requires not only an optical switching component but
equipment to control it as well. Figure 9.2 shows an example of a possible optical
node implementation. In this example the optical switch is controlled by an IP-based
controller. The node controllers and the communications channels linking them together
and to other control agents constitute the control plane for the optical network. Even
though in this example the controller is shown as a separate router sitting above the
OXC, it could just as well be integrated in the same unit as the optical cross-connect.
Similarly, the control channel between two neighboring OXCs is depicted here as a
separate (out-of-band) channel but an in-band solution could also be implemented.

1
Consistent with the application-oriented approach in this chapter our terminology and graphical network
element symbols shift from the abstract to the concrete; for example, an ONN becomes an optical switch
(e.g., OADM or OXC) and an LS becomes an electronic switch (e.g., a SONET DCS or IP router.)
716 Multiwavelength Optical Networks

Control plane

Control

Control IP/MPLS router

or neighboring
node
OXC

Data

Optical network node

Figure 9.2 Example of an optical node architecture. (From [Xin+01, Figure 2]. Copyright

c 2001
IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

9.1 Introduction to the Optical Control Plane

In general, a fully functioning control plane in an optical transport network will in-
troduce significant advantages, including rapid connection provisioning, rapid recovery
in case of a failure condition, and the introduction of traffic engineering for dynami-
cally allocating network resources to the provisioned connections to optimize network
performance. (Traffic engineering objectives include meeting quality-of-service [QoS]
requirements, distributing traffic flows to utilize network resources optimally, and ensur-
ing that service level agreements are satisfied.) Furthermore, the control plane facilitates
the introduction of new services in the network and allows for the control of connections
in a multidomain and multivendor environment. To achieve these benefits, the following
capabilities are required: (a) a routing protocol to provide for resource and topology
discovery and route computation; (b) a signaling protocol to provide for the setup, main-
tenance, and tear-down of connections; (c) a signaling network that provides for the
communication between the entities in the client and transport networks and among the
entities in the transport network; and (d) a uniquely defined addressing scheme for all
the network entities.2 These capabilities will be described in detail in the sections that
follow.
In previous chapters we dealt with routing and fault recovery in optical networks from
an abstract viewpoint without explaining how these functionalities are implemented.

2
Control planes must have the capability to set up a large number of connections in each domain, with
the added capability to scale to a much larger set of connections. Thus the transport network should have
a naming and addressing scheme that is scalable. A possible scalable addressing scheme could use two
types of addressing: (1) public addresses at the end points (addresses that are globally and uniquely defined
within a network, support a large number of end points, and provide some aggregation information) and
(2) transport network internal addresses that are also scalable and visible to the network operators but not to
the network users [McGuire+01].
 Optical Control Plane 717

Router
Optical Network

Router

Figure 9.3 Provisioning a connection between two routers through an optical network.

This chapter will give an overview of the control and implementation aspects of these
functions. More specifically, it will present an overview of the control plane for optical
networks, realized in the form of signaling, routing, and discovery protocols. Figure 9.3
shows an example of a connection provisioned between two routers through an optical
transport network. To establish this connection, a route has to be found in the opti-
cal network that has enough resources to accommodate the connection, and the nodes
through which the connection is established have to be configured appropriately. There
is an emerging consensus that the best way to achieve this is to adapt control plane
protocols developed for the logical (electronic) network layers, such as Multiprotocol
Label Switching (MPLS), to the optical network environment. Developed by the Internet
Engineering Task Force (IETF), MPLS essentially extends concepts and protocols from
IP networks to broaden the functionality of these and other packet-switched networks,
leading to improved performance and more flexibility. A first step toward optical net-
works by the IETF in 1999 was an extension of MPLS, called MPλS. This was further
extended to generalized-MPLS (GMPLS) supporting devices that perform switching in
time, wavelength, and space [Mannie+04]. The extensions to wider varieties of networks
follow directly from the similarities in control of these networks. For example, when ap-
plying GMPLS to optical networks, an analogy can be made between a Label-Switched
Path (LSP) in MPLS and an optical connection (lightpath) in an optical network. Thus,
the use of GMPLS, the grandchild of the Internet Protocol, results in an IP-centric
approach to optical network control.
The industry forums have proposed several architectural options for interaction be-
tween the routers (the logical layer switches) and the optical layer network elements that
are implemented to achieve end-to-end connectivity. These are the overlay, peer, and
718 Multiwavelength Optical Networks

hybrid models [Bernstein+00, Rajagopalan+00, Rajagopalan+04]. In the overlay

model, the optical layer is completely separate from the logical (e.g., IP) layer for
purposes of control. In one proposed version of an overlay system, the User-Network
Interface (UNI) model, clients request high-bandwidth connections from the optical net-
work, via a UNI, which forwards the request to the optical network. In another version,
the soft permanent circuit (SPC) model, a management system separate from both the
client and transport networks, communicates with both. The management system relays
the connection request from an end system in the client network to a transport network
node that completes the provisioning of the connection in the optical network. The key
feature of both of these overlay models is that client equipment has no knowledge of
the optical network topology or resources, so it cannot participate in the provisioning
process. The overlay model is most appropriate for infrastructures that consist of mul-
tiple administrative domains. The topology and control isolation between the transport
and the client layers, which is maintained in the overlay model, is particularly useful in
cases where the transport network might consist of several subnetworks operating under
different administrative policies, and there may be several client layers. In this situation,
the overlay approach provides the optical transport layer with well-defined interfaces to
its client layers. This permits each subnetwork to scale independently.
Though the overlay model is the simplest of the models enumerated above, it has
its disadvantages. Because it completely separates the state and control information at
the boundaries of the transport and client layers, it tends to use the network resources
inefficiently. Furthermore, if a failure occurs in one administrative domain, this may
cause a number of unrelated failures in other domains.
A more sophisticated model is the peer model that integrates the client and transport
layers. It collapses the two layers into a single integrated layer managed and traffic
engineered in a unified manner.
The peer model has a complete view of the network resources at both the client and
transport layer and unlike the overlay model it supports an integrated routing approach
[Kodialam+01]. For purposes of control, all network elements in both the client and
transport layers are peers and are fully aware of the network topology and resources.
The unified control plane obviously improves control coordination and fault handling
among different network elements and it can use the network resources more efficiently
for connection provisioning and failure recovery. It does, however, present a scalability
problem due to the amount of state and control information to be handled by any network
element within an administrative domain. A practical problem in the implementation of
the peer model is the fact that it is highly unlikely that the administrative authority of
the transport layer would provide information about its network resources to the client.
For example, a service provider would be reluctant to provide a client full access to the
topology and resources of its transport network, which means that the implementation of
such a model is not likely to occur any time soon. Furthermore, some of the equipment in
the transport network (the optical core) may use proprietary signaling for intranetwork
communication rather than a standardized signaling approach such as GMPLS.
A third approach is a hybrid model that combines some of the characteristics of
both the peer and the overlay models. In this case, some of the client network elements
 Optical Control Plane 719

(e.g., IP routers) are peers to the transport network elements and share the same control
plane, whereas others have their own control plane and interface with the transport
network using the UNI [Banerjee+01b].
Although several vendors use proprietary signaling and routing protocols to implement
their connection provisioning and fault recovery functions, it is clear that a standardized
approach to signaling and routing is paramount for the implementation of automatic
“point-and-click” provisioning. This applies as well to the development of automated
recovery mechanisms in networks that span a number of different administrative areas
managed by different service providers and equipped with network elements by different
vendors. Apart from the provisioning and fault recovery advantages across multivendor
networks, a common control plane also reduces the operational cost of the network, as
it simplifies the network operation and management. As mentioned above, a standard-
ized control plane, based on MPLS, has already been developed and is widely used
for IP networks. Implementation of a control plane for optical networks using GMPLS
requires modifications of the MPLS routing and signaling protocols to account for the
presence of transparent cross-connects in the network. Modifications to the routing pro-
tocols include, for example, the advertisement of optical resources such as bandwidth on
wavelengths, fiber identifiers, etc. Modifications to the signaling protocols include the
explicit specification of LSPs across the optical transport network. A Link Management
Protocol (LMP) not previously defined for MPLS is also added under the GMPLS proto-
col umbrella that addresses issues of neighbor discovery, link management, and control
channel management specifically in optical networks. A more detailed description of
the signaling, routing, and link management protocols for GMPLS is presented in the
sections that follow.

9.1.1 Control-Plane Architecture

A large optical network may be divided in a number of administrative or control do-
mains to accommodate multiple service providers with their own administrative and
control procedures. Figure 9.4 shows a network divided into two control domains with
control-plane agents (the cubes in the figure) and control interfaces (UNI and NNI)
that are described below. The connectivity shown defines the control plane adjacency.
These control agents do not need direct (physical) connectivity to provide the desired
control functionalities. What is required is that some sort of communication infrastruc-
ture is available within which all control agents are reachable from each other. The
communications infrastructure among the control agents is commonly called the Data
Communications Network (DCN) for the control plane.

9.1.2 Control-Plane Interfaces

As noted earlier, a network may be partitioned into a number of smaller networks (sub-
networks) each having its own control domain. This may be the case, for example, when
there are a number of different service providers each using their own network tech-
nologies from different vendors. When a connection needs to be provisioned end to end,
720 Multiwavelength Optical Networks

UNI E-NNI UNI

Client Client

Control domain Control domain

Figure 9.4 Control plane architecture. (From [Saha+03, Figure 2]. Copyright

c 2003 IEEE. Used
by permission of The Institute of Electrical and Electronics Engineers, Inc.)

certain information needs to be passed between different control domains to establish

the desired connection. This information is not the complete state of each subnetwork
but rather some summarized data concerning the subnetworks (what is termed an in-
formation abstraction). Figure 9.5 shows an example of two jointly controlled optical
networks, each divided into two subnetworks under separate control domains. The var-
ious control interfaces are also shown. Control interfaces denote the places where the
exchange of information between control domains occurs. A control agent may be inte-
grated with the network elements or may reside in a different device separate from the
network element (as illustrated in Figure 9.23). Note that an administrative framework
providing an addressing method for the controlled elements and a management frame-
work to manage the control protocols are also required for the control plane to function
properly.
The control interfaces that are illustrated in Figure 9.5 are

r User-Network Interface (UNI): UNI is the control interface between a node in the
transport (optical) network and a node in the client (e.g., IP) network.
r Network-Network Interface (NNI): NNI is the control interface between two nodes
(or two subnetworks). Depending on where the interface lies it is denoted as:
– Interior Network-Network Interface (I-NNI): I-NNI is the control interface between
two nodes or two subnetworks in the same control domain.
– Exterior Network-Network Interface (E-NNI): E-NNI is the control interface be-
tween two nodes or two subnetworks in different control domains.

3
Neighbor discovery, which will be described in the following section, is the only control plane function that
is implemented on the transport links rather than in the network nodes or separate control agents.
 Optical Control Plane 721

Optical Network 1

I-NNI

E-NNI
UNI

Control Domain 1 Control Domain 2

Optical Network 2

UNI
Client
E-NNI Network

Control Domain 1 Control Domain 2

Figure 9.5 Control plane interfaces. (After [Bernstein+04, Figure 5-8]. Copyright

c 2004. Addison
Wesley, Inc. Reprinted by permission of the authors.)

9.1.3 Control-Plane Functions

There are three main control-plane functions: neighbor discovery, signaling, and routing.
These functions will be introduced here and described in more detail later in this chapter.
Neighbor discovery is the process by which a network element “discovers” its neigh-
bors. Discovery determines the network element’s connectivity to all of its neighboring
network elements. The connectivity information obtained includes such information as
the neighbors’ IDs, the link termination IDs, etc.
Signaling entails the way control agents communicate among themselves (using stan-
dard communication protocols) to provision and maintain end-to-end connections in the
network.
Routing includes two separate control functions: topology and resource discovery
and path computation. The first function involves the discovery of all resources in
the network and the discovery of the network topology, which is accomplished by the
exchange of information on available resources and topology between the control agents.
Based on this information, a control agent can implement path computation, which is
the determination of a feasible route for a desired connection.
A fourth control plane function of importance is resource management, which involves
representation and keeping track of available local resources to be used in routing
computation.
722 Multiwavelength Optical Networks

The control functions described above as well as the establishment of connections can
be implemented using any of the following control paradigms:

r Centralized control: A central controller communicating with every network element

is used for discovery (network resources, topology), computation of the routes, and
establishment of the connections.
r Distributed control: All control plane functions (resource discovery, path computation,
and path establishment) are distributed among the individual network elements.
r Hybrid control: Some control-plane functions are centralized (a central controller is
present) and some are distributed among the individual network elements. For example,
the discovery of the resources and the path computation can be implemented in a
centralized manner, whereas path establishment can be implemented in a distributed
manner.

9.2 Overview of Multiprotocol Label Switching

To provide the necessary background for optical network control via GMPLS, we first
need a brief description of MPLS. It is an architecture that integrates routing and
switching and provides for the efficient designation, routing, forwarding, and switching
of diverse traffic flows through a packet-switched network. The traffic transported by an
MPLS network may originate in diverse forms, such as IP packets, ATM cells, Ethernet
frames, etc. Originally developed as an extension of IP, MPLS essentially provides a
framework for connection-oriented communications over the connectionless IP network
addressing the issues of speed, scalability, QoS management, and traffic engineering.
It is used to provide the control plane necessary to ensure automated provisioning and
maintenance of connections and management of network resources to meet stringent
service demands of the network users. It also provides the functionality for managing
traffic flows of different granularities remaining independent of, but adaptable to, the
various layer-2 and layer-3 protocols for the networks in which it operates. MPLS
resolves the problems of multiple stacked client-server layers such as IP over ATM (e.g.,
scalability issues, complexity of control and management) and allows for new services
such as multicasting and virtual private networks (VPNs). As its name implies, MPLS
can operate in conjunction with a variety of network protocols in addition to IP. However,
for simplicity our overview will focus on IP.

9.2.1 Packet Transport through an MPLS Network

In an IP network, each packet has its own header containing information on the source
and destination IP addresses. If there is a large amount of traffic destined to the same
address, it would be an unnecessary waste of resources if the IP routers at every hop
between the source and the destination processed every IP header individually. (IP
header processing constitutes a significant computational burden.) MPLS recognizes
such packets and adds a label identifying them for purposes of forwarding and service
 Optical Control Plane 723

differentiation. The efficiency and flexibility of the MPLS protocol is attained through
“swapping” (replacing) labels at every MPLS router traversed by the packet. The MPLS
routers at every intermediate hop now need to look only at this label and not at the
information in the IP header to forward the packets rapidly and efficiently toward their
destination. One can think of general packet routers as “post offices” that mark, classify,
and monitor mail, processing it according to class of service, priority, and destination.
MPLS performs these functions in a particularly efficient manner by virtue of the way in
which labels are assigned and processed. To begin the MPLS transport process, traffic
must be encapsulated in special labels as it enters the network, a function performed at
the first MPLS node the packet encounters.
The MPLS switching nodes (high speed logical switches) are called label-switching
routers (LSRs). Transmission of a packet in an MPLS-based network occurs on a Label-
Switched Path, which is a path through the network defined by a sequence of labels
appended to the packet and swapped at each LSR from the source to the destination.
Label swapping is implemented using a forwarding table, called a Label Information
Base (LIB).
The assignment of packet flows to LSPs is based on the concept of a Forwarding
Equivalence Class (FEC), which is a flow of packets that share the same requirements
for purposes of routing and QoS and are treated similarly en route to their destination.
All packets in the same FEC are transmitted on a common LSP.4 Thus the FEC is
a connection-oriented transport mechanism. Classification into FECs may be based
on a variety of criteria, including destination address (the most obvious and simplest
criterion), delay or loss tolerance, priorities, etc. FECs can also be used for traffic
engineering, wherein traffic flows with the same ingress and egress nodes are split into
different FECs assigned to different LSPs to distribute traffic evenly within the network.
MPLS distinguishes between two types of routers, those at the interface between
an MPLS network and the end systems accessing this network (the ingress and egress
nodes) called label edge routers (LERs), and the nonedge routers, called interior LSRs.
The LERs play a key role in the screening and classification of traffic for assignment of
LSPs as the traffic enters the transport network. On the ingress side, the LER examines
the incoming packet to determine whether it should be labeled and, if so, what label (and
hence what LSP) to assign to it.5
A special database in the ingress LER maps the layer-3 packet header information
into an MPLS label, which is attached as a “shim header” inside the layer-2 header and
acts as an outer envelope of the IP packet (see Figure 9.6[b]). Figure 9.6(a) shows the
format of the 32-bit MPLS header. The first 20 bits are used to specify the label, the next
3 bits specify the CoS (class of service), the next bit (stack bit) indicates the presence of
a label stack, and the 8-bit time-to-live (TTL) field is used to discard packets that loop
around in the network. Thus MPLS acts as a shim layer intervening between layer 2 (the

4
Routing according to FECs is analogous to routing in an ATM network via virtual paths (see Section 2.6.2).
5
It may be that some traffic accessing the network at a given point may be identified for MPLS processing
while other traffic might be processed in its original form. For example, IP packets designated for expedited
transport might be singled out for MPLS labeling while the remaining IP packets might be transported using
conventional IP processing.
724 Multiwavelength Optical Networks

Label CoS S TTL

(20 bit) (3 bit) (1 bit) (8 bit)

(a) MPLS Header

L2 Header MPLS L3 Header L3 Data

Header

(b) MPLS Packet

Figure 9.6 (a) MPLS header format and (b) MPLS packet format.

Network Interface layer of IP) and layer 3 (the Internet layer). Once labels have been
attached, the nonedge routers serve as interior LSRs to route traffic across the network.
The LSRs perform a label swapping and forwarding function using the LIB. Based on the
incoming label and router interface (port), a new outgoing label replaces the incoming
label, and the packet is forwarded to the designated outgoing router interface. In this
manner, the MPLS labels are used to route their packets through the network along a
desired LSP without having to “unwrap” them to expose and process the IP header. At
the egress LER, the MPLS label is removed, exposing the packet’s original IP (layer-3)
header.
An example of MPLS packet forwarding is shown in Figure 9.7, which follows packets
from each of two traffic flows entering the network at LSR A from Host X. The ingress
router, LSR A, consults its database and assigns the labels 21 and 17 to packets from the
upper and lower streams, respectively. At LSR B an LIB is consulted, and based on the
input port and input label, packets from the upper stream are forwarded to the upper port
with their labels replaced by 47, and those from the lower stream are forwarded to the
lower port with their labels replaced by 11. Packets arriving at the egress routers LSR
C and D are stripped of their MPLS labels revealing their IP headers and transmitted to
hosts Y and Z, respectively.
It is important to note that the labels have local significance only, which means that
the label is relevant only on a single link between adjacent LSRs. The fact that labels
are defined locally means that they can be “reused” throughout the network, unlike
source and destination addresses, which have global significance. This greatly reduces
the routing table size and table lookup complexity, promoting network scalability.
An LER performs many packet analysis functions, including mapping the existing
layer-2 protocol to MPLS, mapping MPLS to layer 3, and classifying traffic. In addition,
the LER decides what traffic becomes an MPLS packet. One decision-making method is
called “triggered mode.” Using this method, when a predetermined number of packets are
addressed to a single location within a specified time window, the router will determine
 Optical Control Plane 725

Host Z

LSR D
Egress
LSR A
LSR B 47
Ingress 21

17 11
Host X

LSR C
Egress

Host Y

Figure 9.7 Two LSPs in an MPLS packet-switched network. (From [Harrison+01, Figure 1].
Copyright

c 2001 Data Connection Inc. Used by permission of Data Connection Inc.)

that there is a “traffic stream” and will designate the stream for MPLS processing,
assigning it to an FEC and an LSP.
Depending on how they are established, there are two types of LSPs: (1) control
driven, where LSPs are established prior to data transmission, and (2) data driven, where
LSPs are established on the detection of the presence of flow of data traffic as in the
triggered mode mentioned above. The establishment of LSPs is performed by exchanging
information among the LSRs using an appropriate label distribution protocol. Label-
switched paths are assigned to FECs based on various criteria, including the traffic-
handling capability of the path and traffic characteristics of the FEC such as peak traffic
load, interpacket delay variation, and allowable dropped packet percentage. In this way
the MPLS network is structured to support traffic engineering and QoS requirements.
An LSP can be determined using traditional hop-by-hop routing techniques based on
destination address only or by an explicit (source) routing procedure (also known as
constraint-based routing). In explicit routing the LSP is defined by the ingress node,
and it may be based on a variety of criteria, detouring traffic around congestion points
or adhering to administrative policies based on bandwidth requirements on each link,
packet delay, etc.
Because the LSRs are responsible for forwarding based on packet labels, MPLS
requires a protocol for assigning labels to the LSPs for each traffic flow and distributing
726 Multiwavelength Optical Networks

these labels to the LSRs, thereby “binding” labels for each FEC to an assigned path.
An example of such a protocol is the Label Distribution Protocol (LDP), which assigns
labels to LSPs based on destination address. A refinement of LDP, Constraint-based
Routing Label Distribution Protocol (CR-LDP) includes QoS and CoS considerations.
The path/label assignment protocols necessarily rely on basic routing, traffic engineering,
and resource reservation techniques, which are for the most part extensions of the
protocols used in IP networks. For example, routing, which involves neighbor, link state,
and topology discovery followed by route computation, might use the Border Gateway
Protocol (BGP), Intermediate System to Intermediate System (IS-IS) protocol or Open
Shortest Path First (OSPF) protocol (see Section 9.3.2). All of these were originally
developed for IP networks. The OSPF and BGP protocols have been modified so that
label distribution information for MPLS can be piggybacked on them without requiring
a separate label distribution protocol such as LDP.
Traffic engineering and resource reservation typically are based on the Resource
Reservation Protocol with Traffic Engineering extensions (RSVP-TE) (Section 9.3.3.3),
which is an extension of the basic RSVP protocol used in IP networks. It reserves
resources (e.g., bandwidth), enforces QoS requirements through proper packet schedul-
ing in routers, and supports the assignment of explicitly routed LSPs to attain traffic
engineering objectives.
MPLS can be viewed as a means of forming tunnels across all or part of the back-
bone of an IP or similar network to create shortcuts underneath the IP backbone.6 The
tunnel carries data encapsulated with MPLS headers between the ingress and egress
LSRs, short-circuiting the layer-3 operations in the intervening routers. For example,
in Figure 9.7 above, the LSPs handling packets from ingress LSR A deliver packets to
their egress LSRs C and D without processing the encapsulated IP headers, using the
MPLS labels only. For more efficient operation, the LSP tunneling operations can be
hierarchical: one LSP can tunnel under another. Using a process known as label stacking
several LSPs can be “bundled” over a portion of their paths. An example is shown in
Figure 9.8, where two LSPs follow the same path between LSR C and LSR E so they
can be combined along a higher level LSP tunnel between those two points. Packets
entering the higher level LSP tunnel are given a higher level label to see them through
the tunnel, retaining their first-level labels to distinguish them when they emerge. The
labels are arranged in a stack with the label for the original (lower level) LSPs (from
ingress LSRs A and B to hosts X and Y, respectively) at the bottom and the label for the
combined (higher level) LSP at the top. The topmost label is transmitted first and is the
only label used for routing the packet. As packets enter and leave a tunnel, labels are
pushed and popped from the stack. For example, in Figure 9.8, packets arrive at LSR C
from LSR A and B with single labels 6 and 13, respectively. At LSR C, the label 17 is
pushed onto the stacks of packets from both flows, indicating that they are to be tunneled
through to LSR E. At LSR E, the stack is popped with the original lower level labels

6
The term tunneling originated in the IP standards and literature to describe various techniques of encapsu-
lating packets generated by other architectures and protocols, for purposes of transmitting them through an
IP network.
 Optical Control Plane 727

LSR A Host X
Backbone network

6
LSR D LSR E
LSR C 17
17 6 21 6

17
1713 2113
21
13

LSR B Host Y

Figure 9.8 Label stacking. (From [Harrison+01, Figure 2]. Copyright

c 2001 Data Connection
Inc. Used by permission of Data Connection Inc.)

exposed for processing. At this point, a label swap occurs and the packets are routed to
their respective destinations.
Label stacks allow a finer granularity of traffic classification between ingress and
egress nodes than is visible to the LSRs in the core of the network, which route data
solely on the basis of the topmost label in the stack. This helps to reduce both the size of
the forwarding tables that need to be maintained on the core LSRs and the complexity
of managing data forwarding across the backbone.

9.2.2 MPLS Protocol Stack

The protocol stack for MPLS operation is depicted in Figure 9.9, summarizing the
functions described above. The boxes in the upper layer: Routing, LDP, and CR-LDP
compose the preliminary operations necessary to determine paths for the LSPs and
distribute the label information required to set up LIBs in each LSR. The routing
function might be executed by any one of many existing routing protocols; e.g., OSPF
or BGP, or explicit routing might be used, where the source of an FEC specifies the path
for the traffic flow. (CR-LDP includes provisions for constraint-based explicit routing
as well as label distribution.) These preliminary operations bind labels to FECs, thereby
providing the forwarding information that populates the LIB at each LSR.
Signaling is needed by the LDP for exchanging discovery and control information
among the LSRs. The User Datagram Protocol (UDP) is used in the initial discovery
phase (neighbor, link state, and topology discovery), which includes each LSR signaling
its presence in the network through Hello messages. Once discovery is completed the
more reliable Transmission Control Protocol (TCP) is used for label distribution and
other control functions. Both UDP and TCP are basic components of the TCP/IP Internet
protocol suite. The three layers in the dashed box in Figure 9.9 (which can be imple-
mented in hardware for high speed operation) constitute the real-time packet-forwarding
operations in layer 2 (shown as the network interface layer) and layer 3 (the IP layer), as
well as in the MPLS shim between these two layers.
728 Multiwavelength Optical Networks

Routing
LDP CR-LDP
(OSPF, BGP, ATM’s PNNI)

TCP UDP

IP Fwd

LIB MPLS Fwd

PHY

Figure 9.9 MPLS protocol stack. (From [MPLS07, Figure 8]. Copyright

c 2007 International
Engineering Consortium. Used by permission of the IEC)

9.2.3 MPLS Applications

MPLS-based networks have several effective applications. The most important of these
are outlined below.

9.2.3.1 Virtual Private Networks

In a VPN a virtual topology is set up to provide a specified pattern of connectivity
among the end systems. For example, it might be arranged so that each customer of
an organization can reach other sites in the organization but not any other customer’s
sites. This is done by providing a “private” network for each customer. MPLS networks
offer support for VPN services in a relatively simple and efficient manner, because
in MPLS forwarding decisions are based on labels associated with designated FECs
and not on a destination address. In MPLS networks, a VPN is deployed by simply
provisioning a set of LSPs connecting the different sites in the VPN. The MPLS VPN
model is highly scalable utilizing a simplified routing approach, and has the flexibility
for providing a range of QoS to users by sorting incoming traffic based on specified
policies.

9.2.3.2 Quality of Service/Class of Service

Differentiated services are supported by classifying traffic into service classes. This is
done by provisioning multiple LSPs between pairs of LERs with each traffic flow engi-
neered for different performance guarantees, or by provisioning a single LSP between
pairs of LERs and using priority queueing in the LSRs.
 Optical Control Plane 729

9.2.3.3 Multicasting
MPLS networks provide multicasting simply by mapping a layer-3 multicast tree to an
LSP tree to accommodate IP multicast traffic. Because of its efficiency, MPLS offers
better multicasting performance than IP and other layer-3 protocols.

9.2.3.4 Traffic Engineering

MPLS allows the operator to attain traffic engineering objectives by distributing traffic
flows over various routes in a network to minimize congestion and balance the load on
the network’s links and routers/switches. This is easily achieved using explicit routing,
which allows for arbitrary designation of paths for traffic flows to achieve desired
performance goals. Constraint-based routing similarly selects explicit routes for an LSP
to meet specific performance requirements.

9.2.3.5 Integration of IP and Layer-2 Protocols

MPLS offers ease of integration and better routing performance in IP over ATM and
other networks operating at layer 2. For example, where ATM is used as the backbone
network supporting multiple types of services, MPLS provides for simpler integration
of control information, improved IP routing performance and improved scalability of IP
routing, as compared to the conventional client-server IP over ATM overlay model. The
improvements in routing are possible, because the topology of the ATM network is now
visible to the IP routing procedures.

9.3 Overview of Generalized Multiprotocol Label Switching

Extensions of MPLS to create what is called Generalized Multiprotocol Label Switch-

ing (GMPLS) enable the migration of the MPLS control plane to wavelength-routed
optical networks [Mannie+04] as well as many other types of networks as yet un-
defined. Although MPLS supports labels for specific types of packet-switching proto-
cols (e.g., labels for packets or cells in the cases of IP and ATM, respectively), GMPLS
is essentially open-ended, supporting labels adapted to arbitrary network technologies at
the physical layer. For example, for a TDM network (e.g., SONET/SDH) the labels will
be associated with time slots, for a wavelength/waveband switching network (i.e., an
optical network with wavelength/waveband switching functionality) the association will
be with wavelengths/wavebands, and for a fiber (space) switching network the labels
will be tied to port (or fiber) numbers.
The “G” in GMPLS is the key to its ability to automate network operation in envi-
ronments involving interconnected dissimilar networks. This generality makes it pos-
sible to manage network elements, provision connections, implement traffic engineer-
ing, and provide protection and restoration rapidly and automatically across network
and domain boundaries. For example, it enables the control of connections passing
through interconnected metro, long-haul and local area networks, all involving dissimi-
lar physical-layer facilities (e.g., IP routers, SONET links, and wavelength-routed optical
paths).
730 Multiwavelength Optical Networks

The first step in the generalization of the MPLS protocol, dubbed MPλS
[Awduche+01a], involved the concept of wavelength switching. This was extended
later to other types of physical-layer technologies, where a label was defined as any-
thing that is sufficient to identify a traffic flow. This moved the basic features of label
switching down to the physical layers of very general classes of networks; hence, the
name GMPLS. In this general context, the content of the label defines the bandwidth
that is available for data flow on a path as well as the routing of that path. For example,
if the label is associated with the wavelength of a selected λ-channel, then the available
bandwidth for the flow is the bandwidth of that channel.
Although MPLS has both data and control-plane specifications, GMPLS only deals
with the control plane. It employs extended versions of signaling and routing schemes
developed for MPLS traffic engineering and also uses the label distribution procedures
developed for MPLS. GMPLS supports real-time provisioning of optical channels,
including dynamic routing and protection/restoration capabilities in mesh networks
(a relatively new area of GMPLS). It also supports link management, a new function
that was not defined under MPLS. The main GMPLS components are
r Link management: This includes neighbor discovery and management of signaling
mechanisms for provisioning and fault recovery. These are basic functions underly-
ing all other aspects of network operation. LMP is the entity responsible for these
procedures.
r Routing: Topology and resource discovery are a part of the routing function as is
distributed path computation for both the working and protection paths (where appli-
cable). Extending the routing techniques used in MPLS, the Open Shortest Path First
protocol with Traffic Engineering extensions (OSPF-TE) or the Intermediate System
to Intermediate System protocol with Traffic Engineering extensions (ISIS-TE) with
appropriate enhancements can be used to implement the GMPLS routing function.
Both of these use shortest path routing algorithms based on assigned link costs, and a
topological database is built up in each node by exchanging Link State Advertisements
(LSAs) containing updated values of link costs and other link state information. In
this chapter we focus only on the OSPF-TE protocol with enhancements for routing
in optical networks, as the representative protocol for GMPLS routing.7
r Signaling: The signaling function is used for connection establishment (provisioning)
and fault recovery. A signaling protocol is used for this purpose, whose messages
are carried on a signaling network that may coincide with or be separate from the
data transport network. The Resource Reservation Protocol with Traffic Engineering
extensions (RSVP-TE) and CR-LDP with appropriate enhancements can be used to
implement signaling in GMPLS.
Each of the three functions just enumerated has a client-server relation with the
preceding one. Provisioning and fault recovery are executed by the signaling protocol

7
The OSPF-TE and ISIS-TE protocols are similar except for the fact that the former was developed for IP
networks and uses IP to exchange information between nodes, whereas the latter was originally developed
within the OSI architecture for networks operating under the Connectionless Network Protocol (CLNP).
 Optical Control Plane 731

and make use of the underlying path computation and topology/resource discovery
functions that are executed in the routing function. Topology and resource discovery
in turn use the neighbor discovery information provided by the link management
function to build up a global picture of the network. We now consider each of these
components in more detail.

9.3.1 Link Management in GMPLS

The main link management functions in GMPLS involve the determination of connec-
tivity between nodes via a neighbor discovery procedure and the implementation of that
procedure together with related functions through a suitable protocol. We describe these
briefly in this section.

9.3.1.1 Neighbor Discovery

A network element automatically learns its connectivity or “adjacency” to other network
elements (its neighbors) through a neighbor discovery procedure. For two network
elements to be adjacent, in the sense used here, there must be an operating communication
link between them. (A pair of nodes joined by a physical link whose transmitter has
failed are no longer adjacent.) Automated neighbor discovery is essential for acquiring
the topological information needed for setting up signaling and data paths through a
network. If automated neighbor discovery is not available, the information concerning
interconnection of the network elements must be manually provided to the network
control/management system. This is a very time-consuming and error-prone process
that becomes especially cumbersome when the network connectivity is changing. On the
other hand, when an automated process is present, changes in the network connectivity
can be detected almost instantaneously and all information will be quickly updated.
This in turn ensures that a new connection will be correctly provisioned using up-to-
date information about the topology and connectivity of the network. Although neighbor
discovery would seem to be a simple problem, it turns out to be quite complex, especially
in optical networks. The techniques used for discovery depend on the “level” of the
adjacency of the neighbors: the logical level (called “layer adjacency”), the physical
level (physical media adjacency), or the control-plane level (called control adjacency).
For example, two OXCs terminating a SONET link at the line level would be layer
adjacent, two opaque OXCs terminating a λ-channel on a fiber link would be physically
adjacent, and two OXCs linked together via a control channel (not necessarily carried
on a physical link joining them) would have a control adjacency.
Different approaches to neighbor discovery would be used in each of these cases. We
are primarily interested in physical media and logical-layer adjacency here. For example,
for discovering physical media adjacency between nodes, messages containing node and
link identification information can be exchanged over out-of-band control channels
on fibers joining the nodes. On the other hand, logical-layer adjacency is most easily
detected using an in-band approach. Taking the example of a SONET network, the
trace bytes in the SONET frame (see Appendix F) can be used to carry the neighbor
732 Multiwavelength Optical Networks

discovery information. These examples both use in-fiber control channels. However,
another approach is to carry the neighbor discovery information on a separate out-of-
fiber DCN. In what follows we assume the in-fiber case for simplicity.
The information collected by a neighbor discovery protocol includes the neighbors’
node and port identifiers for each interconnection, the types of links (e.g., fiber, wave-
length, or SONET), the length of each link, its physical parameters (e.g., bandwidth and
other optical characteristics), etc.
The neighbor discovery information must be collated by a provisioning system to
get a full picture of the network topology. If a centralized system is used, each net-
work element can periodically send its own node and link interface information to a
central manager, which correlates the information to build an accurate representation
of the network topology. In a distributed provisioning system, connectivity information
is spread throughout the network from node to node so that each node acquires the
information to construct its own view of the network topology. (Means are incorpo-
rated in the various neighbor discovery protocols and network management systems
to detect inconsistencies in the topological data accumulated at different points in the
network.)
The simplest mechanism for implementing a neighbor discovery protocol is the pe-
riodic transmission of Hello messages from each node, containing a unique link ID
associated with the transmitting node. When a node receives a Hello message from a
neighboring node, containing node/link ID information, it returns a new Hello mes-
sage on that link with both the local and remote IDs.8 (As we have already seen, Hello
messages are the typical mechanism for periodically checking whether communication
channels in various protocol layers are operational.)

9.3.1.2 Link Management Protocol

LMP is the means used in GMPLS for implementing the neighbor discovery functions
described above as well as several other essential tasks at the link level, including control
channel management, link bundling, and link fault isolation. LMP goes beyond logical-
layer neighbor discovery protocols such as those used in SONET networks, providing
extra features such as physical media adjacency discovery in transparent optical networks
[Lang+05].
LMP employs Configuration and Hello messages to identify the neighbors of a network
element as well as the interfaces (ports) at opposite ends of each incident link. Control
channel management deals with the establishment and monitoring of the health of the
LMP session. As indicated above, the control channel can be set up in-band (e.g.,
using the SONET Line/Section overhead bytes) or out-of-band (e.g., on a separate
wavelength dedicated to control on a data link or through a separate IP network). In a
transparent optical network, only out-of-band control is possible. As indicated earlier, a
control channel may or may not be carried on the same fiber as the data link. A control

8
This works only in the case of bidirectional links. Other procedures have been developed for unidirectional
links.
 Optical Control Plane 733

channel that is physically diverse from the data links has the advantage that its health is
independent of the health of the physical data links. Monitoring of the control channel
is carried out by the exchange of Hello messages at specified time intervals to ensure
that the LMP session is operational.
Link bundling is a technique used to combine several parallel links having the same
properties for purposes of routing into a single logical group, called a “bundled” or
“traffic engineering” (TE) link, joining a pair of adjacent nodes. This is important for
simplifying information exchange in routing protocols when large numbers of parallel
links are involved. (see Section 9.3.2.4).9
Bundling minimizes the amount of information that is exchanged in the network by
the routing protocol by advertising only the bundled TE link as a representation of all
of its component links. As only the TE link is advertised by the routing protocol, a
method for identifying the component links is still required for routing purposes. This
is accomplished in LMP via a link verification and a link summarization mechanism.
Link verification utilizes ID mappings at both ends of a component link to determine
which one will be used by a given LSP. It is also used to verify the physical connectivity
of the component links and to ensure that no error occurred during the (manual) cabling
process. Link summarization groups the IDs at each end of individual components to
create a corresponding TE link ID. It also correlates the link properties between adjacent
network elements.
Fault isolation is particularly important and more difficult in the case of transparent
networks. In these networks, a failure propagates downstream across multiple nodes,
so that it is not an easy task to isolate its location. LMP determines the link where the
failure has occurred by “backtracking” upstream until it finds a data link that has not
failed. It then deduces that the next downstream link on that path has failed and reports
the error.
LMP in networks with optical switches is more complex than in O-E-O switches. One
approach to neighbor discovery in opaque O-O-O switches of the form of Figure 4.88
is to transmit and receive “test” signals on add and drop ports, respectively, where
these signals are transmitted between nodes on out-of-band control channels [Bern-
stein+04]. Unfortunately, these discovery procedures require a considerable amount of
manual intervention and are not applicable to transparent O-O-O switches of the form of
Figure 4.86. At this writing, adaptation of neighbor discovery protocols to transparent
switches has lagged, primarily due to the limited deployment of these switches in current
operational networks.
After LMP or a similar protocol is executed and all node-port associations are ob-
tained, the network topology is created automatically either by a centralized management
system or in a distributed fashion. The network is then ready for connection provision-
ing. The first step in provisioning is the determination of an appropriate path for the
requested connection. This is the function of the routing protocols discussed in the next
section.

9
For example, in an opaque DWDM network there may be many fibers, each containing many λ-channels
resulting in hundreds of parallel links joining a pair of neighbors.
734 Multiwavelength Optical Networks

9.3.2 Routing in GMPLS

The GMPLS-based routing protocols currently used for the optical network control
plane are based on those that have been used successfully for decades in IP networks, so
the procedures and terminology we describe here have the flavor of a packet-switched
network rather than, say, a wavelength-switched optical network. For example, the net-
work element doing the routing/forwarding of packets in an IP network is called a router
whereas the element that routes/switches a wavelength channel in an optical network
would be an OXC. Despite the differences between packet-switched networks and optical
networks, GMPLS provides the framework to extend the applicability of “traditional”
IP routing protocols to various types of optical networks. In this section we first discuss
the features of routing protocols and give a brief description of algorithms for path
computation. We conclude with an overview of the OSPF protocol and its enhancements
for GMPLS routing. OSPF was chosen here as the representative routing protocol for
GMPLS as it is the most widely used in current GMPLS implementations. Additional
details on the IS-IS protocol, another popular routing protocol not discussed here, can
be found in [Callon90, Oran+90].

9.3.2.1 General Features of Routing Protocols

Routing protocols are required to determine a “good” or “optimal” path through a
network for each requested connection. As we have seen in various places in this book,
many different optimality criteria may be used in choosing a path. The considerations
involved in selecting a path range from signal quality (e.g., dispersion, noise, and
attenuation) to security and protection from faults. The most important issues will be
treated below.
In large networks, routing can entail traversal of several administrative domains,
known as Autonomous Systems (AS).10 Protocols that deal with routes within an AS
(intra-AS routing) are called Interior Gateway Protocols (IGP). An example of an interior
gateway protocol is OSPF [Moy98a, 98b]. Similarly, routing protocols that deal with
inter-AS routing are called Exterior Gateway Protocols (EGP). The basic EGP used in
the Internet is the Border Gateway Protocol (BGP).
The routing algorithms that were originally used for path computation in IP net-
works and are currently used in GMPLS are all variants of methods for finding shortest
paths through graphs. The path “lengths” or “costs” or “weights” are sums of link
costs/weights.11 Two basic algorithms are frequently used for shortest path computa-
tion: the Bellman-Ford algorithm [Bellman58] and Dijkstra’s algorithm [Dijkstra59].
From a mathematical point of view, they are extremely simple compared to the joint
routing and channel assignment problems discussed in Chapter 6 for optical networks.
However, from a network control point of view the path computation algorithm is just
one small part of the overall routing problem. Another major issue is where the algo-
rithm is implemented in a network and what information it requires. Finally, shortest path
10
Just as a large network may consist of several ASs, a single AS may be broken down in a hierarchical
fashion into several areas to allow the routing protocols to scale easily to large networks.
11
The key engineering feature of the shortest path problem is assignment of appropriate link costs to achieve
a desired performance objective.
 Optical Control Plane 735

algorithms are adapted to opaque networks only. Routing in transparent optical networks
is more complex and requires different techniques (see below).
Before going into the details of shortest path routing, it is important to note that when
discussing routing in terms of shortest path algorithms we are implicitly dealing with a
dynamic routing problem. That is, we are assuming that the network is already in oper-
ation with a number of active connections, and connections are being provisioned (and
deleted) sequentially. A fundamentally different version of the routing problem occurs
when a network is in the planning or initialization stage with a list of all connections
to be provisioned available before the routes are chosen. This is the static version of
routing, discussed in detail in Section 6.3 where various optimization formulations are
presented for RCA in both transparent and opaque optical networks.
Shortest path routing protocols rely on algorithms that utilize either partial or global
network topology and link information to compute a route, and they may be computed at
a central location (for example, a centralized network management system [NMS]) or in
a distributed fashion at each network node. A disadvantage of the centralized approach
is that the NMS is a single point of failure: A fault in the NMS brings the whole network
down. In contrast, in a distributed approach to routing, one or more links or nodes can
go down and the system will still operate.
In GMPLS, link state and topology information originates from neighbor discovery
procedures using LMP. Routing algorithms that need global knowledge of the network
are called link-state algorithms because the state of all links connecting the network
must be known a priori for route computation. The routing procedure used in OSPF,
based on Dijkstra’s algorithm, is an example of a link-state routing protocol. It requires a
database at each node consisting of global network information, from which it computes
a routing tree consisting of shortest paths from itself as the source to all other nodes
as the destinations. (It is distributed in the sense that each node only computes shortest
paths from itself to the other nodes, although it has enough information to compute all
source-destination paths.)
The Bellman-Ford algorithm is an example of a distributed procedure where each
node requires only partial information, composed of the states of its incident links
and the distances of its neighbors to all destination nodes. From this information it
computes a distance vector consisting of the distances on shortest paths from itself
as the source, to all destination nodes. (Procedures of this type are called distance
vector protocols.) Each node maintains a routing table consisting of the latest computed
distance vector and the next node on its shortest path to each destination. After each
distance vector computation, each node transmits the value of its new distance vector
to its neighbors. Computation is iterative, where each node updates its own distance
vector whenever it observes a change in either its own link states or the distances of its
neighbors to their destination nodes. If all link states remain fixed while the iterations are
taking place, this procedure eventually converges to a fixed set of shortest path routing
tables at each node. The original routing algorithm used in ARPANet was a distance
vector protocol based on the Bellman-Ford algorithm, and RIP (Routing Information
Protocol) is a widely used distance vector protocol based on a variant of the Bellman-Ford
algorithm.
736 Multiwavelength Optical Networks

Link-state protocols are usually favored over distance vector protocols because link-
state protocols converge after N steps, where N is the number of network nodes. On the
other hand, elementary distance vector algorithms have erratic convergence properties,
with rapid convergence after “good news” (e.g., a new link has been brought up or
its cost reduced) and slow convergence after “bad news” (a link cost is increased or
it goes down). Furthermore, in a packet-switched network, packets may “loop” around
the network while the algorithm is converging, because the nodes have inconsistent
views of the network state, resulting in erroneous routing decisions. Because of these
differences ARPANet, in its early days, dropped the distance vector approach in favor
of the link-state approach.
There are some advantages of distance vector algorithms, however. They require
smaller databases, less computation at the nodes, and less exchange of information
among the network nodes. In this sense, they have better scalability qualities than link-
state algorithms. For this reason, there has been considerable work on improving the
distance vector approach to conserve its advantages and remove its defects. An example
of such an improvement is the path vector protocol used in BGP [Rajagopalan+91].
In considering the cost of discovering and disseminating topology and resource infor-
mation required in routing it should be noted that this information is also used for other
network operations and planning purposes. Up-to-date information on the state of the
network (in terms of topology and network resources) is also critical for implementing
disaster recovery plans.

9.3.2.2 Operational Considerations in Routing

The most important considerations in the routing problem are those that are hidden from
view when it is abstracted to a shortest path problem in a graph.

Path Protection and Physically Diverse Paths

Fault protection is high on the list of criteria in choosing a path through a network.
The most commonly used protection technique in large mesh networks is path-based
protection (see Section 8.4.2). Using this technique, when a path is provisioned (the
working path), another path between the same end points (the protection path) is also
chosen.12 To protect against any single link failure (the most common type of failure)
the working and protection paths must be link disjoint. Algorithms exist for choosing
pairs of link-disjoint paths subject to various minimum cost criteria. However, a link in
a working path may be carried in the same cable as one in the protection path. In this
case, a cable break (again a common form of failure) will bring down both the working
and protection paths. Sets of entities affected by a common failure are called shared
risk groups (SRGs) (see Section 8.4.2.4). Shared risk link group (SRLG) is the specific
term used for a set of links that share a risk of failure. Clearly, the choice of a protection
path must take into account the existence of SRLGs; that is, the working and protection
paths must be SRLG disjoint or, in other words, physically diverse. Once the SRLGs

12
The protection path may be dedicated, as in (1 + 1) protection, or shared, as in (1:N) protection.
 Optical Control Plane 737

have been identified, the routing algorithm can be adapted (not always easily) to take
them into account. However, the difficult practical problem is identification of the
SRLGs and insertion of this information into the routing protocols. Automatic neighbor
discovery procedures do not find SRLGs, so this information must be determined in the
field and inserted manually into the link state database, a slow and error-prone procedure.

Link Properties
Because of the possible inhomogeneity of the networks that may be controlled, detailed
knowledge of the properties of the links, and their switching interfaces at the terminating
nodes is needed for choosing a feasible path through a GMPLS network. This information
must be disseminated by the routing protocols as link-state information. For example,
a path transiting an OXC must use compatible inbound/outbound links and switching
interfaces; e.g., if an optical connection carried on a λ-channel is being routed through an
opaque OXC, a free channel of sufficient bandwidth must be available on the inbound and
outbound links and the transceivers must be capable of handling the connection’s bit rate
and modulation format. Furthermore, the accumulated effects of optical transmission
impairments on the end-to-end path must be small enough to maintain a satisfactory bit
error rate. Many other properties can be included in the link state; e.g., whether the link
is a component of a bundle or whether it is part of a protection path and designated for
preemption when the paired working path fails.
The requirements on link-state information for routing in transparent networks are
far more severe. The simplest of these is wavelength continuity. A common wavelength
must be available on each link on the chosen path. More difficult is the determination of
the parameters of transmission impairments on each link, the switch cross-talk levels,
and their accumulated effect on signal quality at the optical receiver (see Section 4.9).
Because of the complexity of this problem and the fact that deployment of transparent
networks is still not widespread, this aspect of the routing problem is in the exploration
stage at this writing.
Delegating the routing operations to the individual network elements rather than to
an NMS-type centralized manager enhances network scalability and resilience. How-
ever, distributed routing protocols do not disseminate complete global information to all
nodes, which can lead to suboptimal route computation. For example, distributed routing
approaches involving selection of working and protection paths in path-protected net-
works usually lead to increased network capacity requirements and higher network cost
[Chaudhuri+01]. On the positive side, distributed path computation increases network
robustness and scales better than a centralized approach. There are also some techniques
that can be used to estimate the state of the network based on partial information, thereby
mitigating the deleterious effects of incomplete information [Bouillet+02].

9.3.2.3 Open Shortest Path First

Open Shortest Path First is a distributed link state routing protocol. The word open
indicates that the protocol is open to the public and nonproprietary. OSPF has the
following features that make it attractive as an Interior Gateway Protocol (IGP):
738 Multiwavelength Optical Networks

r Logical separation of the network into hierarchical routing areas for control and
efficiency
r Fast response to topology changes with low overhead
r Authentication of routing updates

In a distributed link state protocol, each node floods the network with its local in-
formation (its incident link characteristics) using Link State Advertisements (LSAs).
Each node receives LSAs from all the other nodes in the network, from which it creates
its own view of the complete network topology and the network resources. Updated
information is sent out by individual nodes either periodically or when a change occurs
in the network (in terms of available resources or in terms of network connectivity).
Each node performs path computations based on its view of the network.
OSPF determines the shortest path to a destination using Dijkstra’s algorithm. Each
router computes a shortest path tree from itself as the source to each destination node.
The routing metric can be chosen arbitrarily to meet traffic engineering criteria and
constraints, and the resultant shortest paths are used for constructing a forwarding
database. The availability of a shortest path tree in each node means that the tree can be
used for explicit routing when that node is an ingress node.
OSPF supports hierarchical routing where a large network is divided into smaller
areas that are all interconnected through a backbone area. Flooding of detailed link-state
information is confined to single areas. This reduces the scalability problems associated
with the flooding process and reduces the size of the database that each node must keep.
OSPF neighbors achieve adjacency when they have exchanged and agreed on key pa-
rameters, they have provided a copy of their databases to their neighbors, and they have
determined that these databases are synchronized; i.e., they are consistent and complete.
(Note that this is a higher level of adjacency than that defined in link management pro-
tocols.) Once adjacency is achieved, the OSPF neighbors can proceed with their routing
computation.
Each OSPF router initiates and propagates LSAs. They are initiated in response to an
activation or deactivation of a link or a change in its link weight, as well as periodically
(typically at 30-minute intervals) to inform the other nodes about its link states. Each
node forwards received LSAs out of all its ports except the incoming port(s).
Because OSPF is used in large networks with many control domains, it is important
that link-state information be authenticated before it is accepted. Therefore, within an
area the LSAs are authenticated using cryptographic techniques so that a “rogue” router
may not send false routing advertisements. Only trusted routers are allowed to introduce
changes.

9.3.2.4 OSPF-TE with Enhancements for GMPLS Support

As noted previously, the protocols used by GMPLS to control optical networks were
originally designed for packet-switched networks. However, there are fundamental dif-
ferences in the control of routing in packet-switched IP networks and routing in circuit-
switched optical networks. IP routing involves functions in both the control and data
plane (it is involved in data plane forwarding decisions). However, in optical networks
 Optical Control Plane 739

control and data planes are completely separate, and connection establishment and re-
source allocation occur prior to the actual data transfer, a basic characteristic of circuit-
switched operation. Furthermore, IP routing is implemented hop by hop while routing
in optical networks is explicit; that is, source directed. An advantage of the explicit
routing approach is that if path protection is used, both working and protection paths
can be established when a connection is provisioned. However, this implies that physical
diversity information must be available to the optical routing protocols.
In addition to these differences, there are some special challenges in trying to use
IP techniques and generalizations of MPLS to control optical networks. Among the
issues addressed in GMPLS are the large number of ports on an optical switch, the large
number of parallel links between adjacent nodes (with each link carrying a large number
of wavelengths), the specific requirements of routing in an optical network (e.g., link type
and bandwidth requirements), the presence of SRLGs, the information requirements for
source routing in optical networks, and the control overhead associated with updates
of link information, which can potentially flood the network with unnecessary control
traffic. Specific problems have been addressed in GMPLS to adapt it to routing in
an optical network using an enhanced version of OSPF-TE. We summarize the most
important problems and solutions below. For details of OSPF-TE see [Katz+03].

Link Bundling
This was proposed as the solution to the large number of OSPF adjacencies between
neighboring nodes in optical networks, which would greatly increase the size of the
LSAs exchanged in the optical network if this information was disseminated in detail.
When link bundling is used between neighboring nodes, the resultant bundle (TE link) is
assigned the aggregated attributes of the individual links, and a single OSPF adjacency
can be maintained for each bundle. This way, the size of the link state database is
significantly reduced, which in turn allows for the link-state protocol to scale gracefully.
Even though some information is lost during the bundling process, the advantages
gained in terms of scalability and reduced amount of control information exchanged are
far more significant than the value of the information lost. A TE link is considered to be
“up” if at least one of the links in the bundle is operational. In the case of link bundling,
OSPF control messages are sent over one control channel and these control channels are
logically different from the bundled links. When explicit path computation takes place
at the ingress node only a link bundle is specified. Which individual link is used to route
the connection is decided locally during the signaling process. Figure 9.10 illustrates
link bundling. Details of the decision process involved in bundling links can be found
in [Kompella+05a].

Link-State Information Dissemination

Opaque LSAs, which provide a generalized mechanism to extend OSPF, are utilized to
disseminate resource information specific to optical networks (number of wavelength
channels, bandwidths, etc.) as well as SRLG information needed for the diverse routing
of working and protection paths. Although OSPF-TE allows a node to advertise TE
links (single or bundled links with specific characteristics), additional characteristics of
740 Multiwavelength Optical Networks

OXC OXC

Component Link TE Link

Figure 9.10 Link bundling illustration.

a single component link needed for optical routing have to be advertised via opaque
LSAs for routing in a single area [Kompella+05b]. Some characteristics particularly
important in optical networks, advertised via opaque LSAs are
r Link protection type: Describes the protection capabilities of the link, such as unpro-
tected, (1 + 1) protected, preemptible, etc.
r Link encoding type: Identifies the framing protocol used on the link such as Ethernet,
SONET, etc.
r Link bandwidth parameter: Identifies the link bandwidth parameters such as maxi-
mum bandwidth, unreserved bandwidth, and the minimum and maximum reservable
bandwidths (minimum and maximum bandwidth that can be allocated to a connection).
r Link cost metric: Used for shortest path routing during path selection.
r Interface switching capability descriptor: Includes the switching capability parameter
that identifies the switching capabilities associated with a link (i.e., TDM-capable,
wavelength-switch capable, fiber-switch capable) and the switching capability specific
information parameter that could include such information as switching granularity,
whether concatenation is supported, etc.
r Link SRLG list: Contains a list of SRLGs for verifying path diversity.

Note that in hierarchically organized networks the additional link attributes adver-
tised by the opaque LSAs described above are confined to a single area. This poses
problems when routing connections across multiple areas, because path computation in
optical networks requires advertisement of sufficiently detailed resource information on
an end-to-end path. This can be accomplished by developing an appropriate resource
summarization scheme. At this writing, work in standards bodies is proceeding on rout-
ing across multiple areas. The various proposals on TE across multiple areas focus on a
number of mechanisms such as restricting the flooding of TE LSAs locally in a single
area, broadcasting TE LSAs across area boundaries, or broadcasting summary TE LSAs
across the entire autonomous system [Cheng03, Kompella+03, Srisuresh+04].
Because any change in the link parameters (such as the number of free wavelength
channels on a link) will trigger LSA updates, and because there are a large number of link
 Optical Control Plane 741

Fiber LSP

TDM LSP
Packet LSP Layer 2 LSP

Figure 9.11 LSP hierarchy in GMPLS. (From [Iovanna+03, Figure 2]. Copyright

c 2003 IEEE.
Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

parameters of interest in optical networks, these updates will tend to inundate the network
with superfluous traffic. A solution to this problem is to impose LSA thresholds. These
thresholds will guarantee that LSA updates are not generated for every minor change in
the network resources but triggered only when the difference between the current and
previous link state exceeds a prescribed threshold.

Nested LSPs
An LSP hierarchy is an additional enhancement in OSPF to account for the fact that
the number of wavelength channels in an optical network is limited (compared to a
much larger MPLS label space) and the fact that optical bandwidth is allocated using
a small discrete set of values (as compared to MPLS LSPs that can be allocated any
bandwidth within a continuous range) [Banerjee+01b, Kompella+05c]. LSP hierarchies
are essentially nested LSPs created by grouping LSPs that enter and exit the optical
transport network at the same ingress and egress nodes to form one aggregated optical
LSP. This is accomplished by treating LSPs as links in the link-state database, which in
turn produces more efficient utilization of the wavelength channels. The LSP hierarchy
can be used for grooming at various levels in a fashion analogous to the grooming
of sub-λ connections through a wavelength “pipe” as described in Chapter 7. The use
of LSP hierarchies also addresses the problem of allocating discrete values of optical
bandwidth, because when the optical LSP is treated as a link, MPLS LSPs with different
bandwidth values can use this optical LSP as long as their aggregate bandwidth does not
exceed the bandwidth allocated for the optical LSP. An illustration of an LSP hierarchy
is shown in Figure 9.11, based on the multiplexing capability of the LSP types.

Transparent Networks
Routing in transparent networks with no wavelength conversion involves additional
constraints, including wavelength continuity and the signal degradation caused by accu-
mulation of physical impairments. Wavelength continuity implies that wavelength avail-
ability must be advertised by the routing protocol [Chiu+03, Saleh+00]. The problem
of signal quality implies that some impairment information (e.g., cross-talk, amplified
742 Multiwavelength Optical Networks

spontaneous emission [ASE] noise, polarization mode dispersion [PMD], etc.) has to
be advertised by the routing protocols, and bounds on each of the impairments have to
be determined in order to include impairment effects in the path calculation decisions
[Chiu+05]. The physical parameters that come into play for routing purposes will vary
depending on a number of factors such as equipment design, fiber characteristics, service
characteristics (e.g., data rates), network size, and engineering and deployment strategies
[Chiu+03].

9.3.3 Signaling in GMPLS

A signaling protocol is required to provision, maintain, restore, and delete connections
in a network. For example, to establish an optical connection from an ingress to an egress
node in an optical network, the cross-connects in each node in the path have to be set
appropriately. GMPLS signaling is a means for conveying the necessary provisioning
information (selected port, channel, etc.) from one node to the next along the chosen path.
To implement a signaling protocol, a DCN is required that is used to transport signaling
messages between the network nodes. In most networks the DCN is packet based (for
example, IP), providing network layer connectivity between the control agents in an
optical network [Saha+03]. The design of the DCN infrastructure must address issues
such as security, reliability, and performance requirements (e.g., limits on packet loss
and delay). An addressing scheme is also required so that nodes in the DCN can be
uniquely identified. Distinct signaling protocols and DCNs may be used for connection
provisioning and connection recovery. Signaling networks may or may not be associated
with the optical data plane.
Signaling consists of two types of communication – the communication between the
end users and the switches at the end points (egress and ingress switches) that normally
takes place at the UNI, and the communication between switches in the network that are
responsible for provisioning and tearing down the connection through the network. This
second type of communication normally takes place at the Network-Network Interface
(NNI).
The first part of this section summarizes the actions at the UNI and NNI inter-
faces, and the general role of signaling in connection provisioning. The second part
describes the RSVP and RSVP-TE protocols and their enhancements for GMPLS sig-
naling [Berger+03a]. We choose RSVP as the representative signaling protocol because
current GMPLS implementations are typically based on RSVP-TE and its enhance-
ments. Additional details on the CR-LDP protocol, another popular signaling protocol
for GMPLS, can be found in [Ashwood-Smith+03].

9.3.3.1 Optical User-Network Interface (UNI) and Network-Network

Interface (NNI)
When the overlay model is used for network control, the transport network provides a
UNI, which is a standard (GMPLS compatible) signaling interface between the client
and transport networks and is used by the client network elements to request connections
across the transport network [OIF01] (see Figure 9.12). The transport network in this
 Optical Control Plane 743

Service Provider Domain

Client Domain

Control Plane
UNI
(signaling)

UNI
(transport)

Optical Transport Network

Figure 9.12 User-Network Interface.

case is optical and clients might be logical-layer electrical devices such as IP routers,
ATM switches, or SONET ADMs. The UNI also provides a neighbor discovery function
that allows for the equipment on both sides of the interface to identify one another and
a service discovery function that allows the client network to discover the services that
are available from the transport network.
When a provisioning request is made across the UNI the following parameters are
requested by the transport network:
r Bandwidth of the connection
r Class of service
r Request for a route that is physically diverse from a set of previous connections. This
is termed the diversity request.13
r Characteristics of the data plane

13
This request can be used, for example, so that a working and a backup path can follow diverse routes.
744 Multiwavelength Optical Networks

Request A->B label Request B->C label Request C->D label

1 4 1 4 1 4 1 4

A B C D
2 5 2 5 2 5 2 5

3 6 3 6 3 6 3 6

A->B label = 1 B->C label = 3 C->D label = 2

Route: A2-A4-B1-B6-C3-C5-D2-D6

Figure 9.13 Provisioning in GMPLS.

The UNI enables the clients to dynamically establish connections across the (optical)
transport network within a few seconds compared to the previous manual provisioning
times of hours and days required to establish a connection across multivendor networks.
It also enables dynamic traffic engineering by allowing the clients to vary the number of
their established connections based on the traffic load.
Large optical networks are often partitioned into smaller control domains, and connec-
tions often span more than one domain. As mentioned in Section 9.1.2, Network-Network
Interfaces are required between nodes within the same domain (I-NNIs) or in different
domains (E-NNIs). Signaling across E-NNI interfaces is required when a connection
crosses domain boundaries. [ITU-T03] describes this in detail when GMPLS RSVP-TE
is used.
Before examining the signaling protocols in detail, it is useful to outline how pro-
visioning and label distribution for optical connections is implemented under GMPLS.
An illustration is shown in Figure 9.13, where the nodes are OXCs, and the links join-
ing the ports could be logical channels (e.g., SONET/SDH), wavelengths, or fibers. A
command for provisioning a connection may be issued by user equipment, in which
case it is called a switched connection, or the command may be issued by a network
management system, in which case it is called a soft permanent connection. In either
case, the provisioning command triggers a signaling exchange as shown by the arrows
in the figure. Here, the connection is provisioned from port 2 on source node A to port 6
on destination node D, with an explicit route assigned at the source along the path
indicated in the figure. This requires configuring the OXCs with the settings A2-A4,
B1-B6, C3-C5, and D2-D6. In GMPLS the labels are distributed to the OXCs and
used for label swapping in a manner analogous to that used in MPLS (see Figure 9.7).
The OXCs now operate as LSRs, with the label-swapping operation now corresponding
to cross-connect settings in the OXCs, and the resultant path analogous to an LSP in
MPLS. Thus, for example, the cross-connection from input port 2 to output port 4 on
OXC A establishes a connection from the source to input port 1 on OXC B. This is
represented as a label-swapping operation, where an incoming MPLS label = 2 (derived
 Optical Control Plane 745

from the input port ID on A) is swapped to an outgoing label = 1 (derived from the
input port ID on B). Note that provisioning requires knowledge of the port IDs at both
ends of each link, information that is acquired by neighbor discovery protocols (see
Section 9.3.1.1). Because it is a circuit-switched connection that is being provisioned
here, the label swapping is required only during connection establishment, in contrast
to MPLS in packet-switched networks, where label swapping is required for processing
each packet.14
The signaling for label distribution works as follows. Labels for the path specified
in the provisioning command are requested via messages sent downstream from source
node A through intermediate nodes B and C to destination node D. Node D responds
with a message upstream to node C specifying label 2 (the input port ID on D) for
the path from C to D, and similar messages are sent upstream to the remaining nodes,
propagating the label assignments back to the source node, after which communication
can proceed. This is the same technique that is used for setting up routing tables in
MPLS.

9.3.3.2 Overview of the Resource Reservation Protocol

The original packet transport procedure at the network layer in IP networks was connec-
tionless, with packet forwarding executed on a best-effort basis. When interest developed
in services beyond data, such as voice, images, and video, connection-oriented transport
with QoS constraints became important. The Resource Reservation Protocol (RSVP)
was developed to address these needs; it is a signaling protocol defined by an IETF stan-
dard [Braden+97] to support point-to-point and multicast traffic in IP networks using
connection-oriented transport with QoS requirements. RSVP is not a routing protocol
and does not support path computation functions. Instead, RSVP relies on paths com-
puted by an underlying routing protocol such as OSPF. The role of RSVP in its original
incarnation was to introduce mechanisms into IP for enforcing QoS requirements for
well-defined traffic flows, called sessions. This is accomplished in RSVP by reserving the
necessary link and node resources for the session, and this in turn is achieved by commu-
nicating the QoS requirements to packet filters and schedulers in the IP routers.15 When
extended to optical networks and other types operating under GMPLS, the properties of
RSVP are modified to adapt to the types of connections (typically circuit-switched rather
than packet-switched) encountered in these networks and to create a clear separation
between the control and transport planes (see Section 9.3.3.4).
There are four basic entities in RSVP: (1) sessions (the traffic flows), (2) messages
(the carriers of signaling information), (3) objects (carried in the message payload), and
(4) states (stored in the IP routers). RSVP runs directly over IP and reservations are
made in one direction only from the source to the destination of the session. Path and
Resv messages are used to reserve the resources along the session path. A source sends a

14
Although the extension of label swapping in packet switches to setting connections in an optical switch is
a bit of a stretch, it is worthwhile in the sense that it allows the methodology of MPLS to be carried over
to more general networks.
15
The filters or classifiers identify the packets designated for a particular QoS, and the schedulers implement
the queueing actions needed to enforce the QoS constraints.
746 Multiwavelength Optical Networks

Path message flow

Resv message flow

Source router Destination

Figure 9.14 Path and Resv message flows in RSVP for resource reservation.

Path message to the destination to indicate the traffic characteristics of the session. The
routers that process this signaling message along the data path create the Path state16 for
the session in their databases and forward the message to the next hop. They also make
a note of the reverse path back to the source. Based on the Path message received and
the traffic characteristics specified in that message the destination sends back a Resv
message along the reverse path. Each router on the reverse path that processes the Resv
message creates a local Resv state.17 It then forwards the message to the previous node in
the path from the sender and it also allocates resources on the downstream link (toward
the destination). Thus resources are reserved for the traffic flow on a unidirectional path
from the source to the destination. Figure 9.14 shows an example of the Path and Resv
message flow to establish a connection. After the reservation is established, packets can
flow following the appropriate path and receiving the appropriate QoS treatment.
RSVP follows a soft state technique in the sense that Path and Resv messages are
not acknowledged and instead are regenerated periodically to refresh the Path and Resv
states of the nodes. A node will eventually delete its Path and Resv states if it does not
receive refresh messages at appropriate time intervals. (A session can also be deleted
when a specific delete message is received as is explained in the description of the
RSVP messages that follows.) This approach allows for the automatic accommodation
of routing changes without any manual intervention.
The RSVP signaling protocol utilizes seven types of messages: (1) Path, (2) Resv,
(3) PathErr, (4) ResvErr, (5) PathTear, (6) ResvTear, and (7) ResvConf. The Path and
Resv messages are the ones used to establish reservations for a connection as explained
above. PathErr and ResvErr are error notification messages. The PathErr message is
sent to the source from a router along the path using the previous hop information. The
Path state of the routers in the rest of the path is not affected by this error message. The
ResvErr message is send to the receiver (destination) from a router along the path. In this
case, routers in the rest of the path might potentially change their Resv state based on this
message. The ResvConf message is sent from the source or an intermediate router to the
destination router confirming the establishment of the connection reservation. Finally,
the PathTear and ResvTear messages are used to tear down a reservation. The PathTear
message is generated from the source node or any intermediate node and is sent to the
16
The Path state contains information about the session, including the sender for this specific session, the
traffic characteristics of the session and the identity of the previous hop.
17
The Resv state includes information about the received flow reservation request, a description of the packets
that will receive the QoS treatment for the specific session and information about the previous hop.
 Optical Control Plane 747

IP Header

RSVP Common Header

RSVP Object 1

RSVP Object 2

.
.
.

RSVP Object n

Figure 9.15 RSVP message format.

receiver node to delete the Path state of a session. The ResvTear message is generated
at the destination node or any intermediate node to delete the reservation for a session.
Note that although a PathTear message removes both the Path and the Resv state at a
node, the ResvTear message removes only the Resv state for the session.
Figure 9.15 shows the format of an RSVP message. In the RSVP common header, a
message type field identifies the message type. The main RSVP objects are

1. Session: Identifies a traffic flow.

2. Sender Template: Identifies the sender of the flow.
3. Sender Tspec: Identifies the traffic characteristics of the flow.
4. Flowspec: Describes the reservation request.
5. Filterspec: Describes the packets for a specific session that will undergo specific QoS
handling at the nodes in the path of that session.

Information carried in these objects is saved in the IP routers’ local databases in the
form of Path and Resv states to ensure the appropriate processing for packets in the
traffic flows transiting the router.

9.3.3.3 RSVP with Traffic Engineering Extensions

Traffic Engineering extensions to RSVP (RSVP-TE) have been created to adapt RSVP to
systems involving label switching by accommodating the formation, maintenance, and
elimination of LSPs [Awduche+01b]. RSVP-TE supports the request and assignment
of labels for the establishment of LSPs, the specification of an explicit route for an LSP,
the specification of bandwidth and other characteristics for an LSP, and the association
of related LSPs (via the use of LSP identifiers). It also introduces a new protocol for the
maintenance of RSVP peer adjacencies.
748 Multiwavelength Optical Networks

New objects defined in RSVP-TE to accommodate the new features described above
are

r Label request object: Used to request a label for the LSP tunnel being set up. It is a
request from an upstream to a downstream neighbor, carried in the Path message.
r Label object: The label that has been assigned in response to a label request, carried
in the Resv message.
r Explicit route object: Carried in the Path message during LSP establishment. Explicit
route objects may specify all nodes on a source-destination path (a Strict Explicit
Route) or they may be incomplete, leaving parts of the route unspecified (a Loose
Explicit Route).
r Record route object: Records all nodes in the path from source to destination. Carried
in the Path message.
r Tunnel identification: Identifies the LSP tunnel and its egress node (carried in
Session object).
r Session attribute object: Describes parameters other than QoS for a session. It is
carried in the Path message.

LSP Tunnel Establishment

To establish an LSP, a Path message is created at the ingress LSR containing the explicit
route, label request, tunnel identification, session attributes, and record route objects;
this message is addressed to the egress LSR. Intermediate LSRs on receiving the Path
message create local Path states, determine the local resources that are available for the
LSP, determine the next hop LSR by examining the explicit route object, determine the
previous hop information, and forward the Path message to the next hop LSR. The Path
message is forwarded from the intermediate nodes to the destination LSR. On reaching
the destination LSR, a Resv message is sent from the egress to the ingress LSR. Each
intermediate LSR that receives the Resv message creates a local Resv state as before
and reserves local resources for the LSP. Each LSR also inserts in the Resv message
that it sends to its upstream node, a label it has assigned to that node for the LSP
being created. Thus, label assignments propagate from the egress to the ingress node,
with each node choosing a label for its upstream neighbor. In this way the intermediate
LSRs that receive the Resv message use the label assigned to them by their downstream
LSR and the label information they have assigned to their upstream LSR to complete
their MPLS forwarding tables. When the Resv message reaches the ingress node the
LSP is established and packets can flow from source to destination. LSP tunnels are
removed either when no refresh messages are sent (soft-state times out) or when explicit
PathTear messages are sent by the ingress node. As in the case of RSVP, a ResvTear
message can also be sent by any intermediate LSR or by the egress LSR, but this results
only in the deletion of the Resv states and not the Path states from the intermediate
LSRs.
Finally, in RSVP-TE a “Hello” protocol is defined between adjacent LSRs (a Hello
message is send every 3 seconds) with the explicit purpose of quickly detecting node
failures [Awduche+01b]. If a Hello message is not received in a prescribed time period,
 Optical Control Plane 749

then an LSR considers that its neighbor has failed. It can then take appropriate action
such as reporting the failure to the higher management layers.

9.3.3.4 RSVP-TE with Enhancements for GMPLS Support

The principal extensions to the MPLS signaling protocols to accommodate signaling
in an optical network deal with the handling of multiple underlying technologies, path
setup for bidirectional connections, and the independence between the control and data
planes.
Optical networks can potentially accommodate a variety of multiplexing and switching
technologies and techniques; for example, SONET/SDH coexisting with wavelength
switching or wavelength and fiber switching and so on. To support diverse types of
switching and multiplexing, new label formats and requests are needed. Thus, RSVP-
TE must be able to support the generalized label concept for establishing LSPs at any
architectural layer in a network [Berger+03b]. To this end, new objects are defined in
GMPLS RSVP-TE as follows:

r Generalized label request object: Replaces the label request object in RSVP-TE and
carries the following information: LSP encoding type, switching type, generalized
protocol ID (the type of payload), source and destination end points, and connection
bandwidth. It is carried in the Path message.
r Generalized label object: Replaces the label object in RSVP-TE. It is carried in the
Resv message.
r Suggested label object: Allows the upstream node to “suggest” to its downstream
node which generalized label to return in the response phase to avoid delays during
the response phase. It is carried in the Path message.
r Upstream label object: The label selected by the upstream node for the reverse direc-
tion of the connection (destination to source). It is carried in the Path message and
is used for bidirectional connections. As mentioned above, while MPLS RSVP-TE is
used for unidirectional LSPs, lightpaths are typically bidirectional, necessitating this
extension.
r Label set object: Used by an upstream node to control the selection of labels by the
downstream nodes. It is carried in the Path message. For example, if the labels are
wavelengths only a set of distinct wavelengths will be available for the connection.
r Acceptable label set object: Generated by a node when that node cannot accept a
specific label. It is carried in PathErr, ResvErr, and Notification messages (defined
below).
r Protection information object: Indicates the protection type on each link in the path.
It is carried in the Path message.
r Administrative status object: Used to signal administrative actions such as testing of
a connection.
r Interface identification object: Identifies the data link on which labels are being
assigned. This is required because in optical networks control and data planes are
separate (logically and possibly physically), and signaling over an out-of-band con-
trol link will cover multiple data links. The separation of control and data transport
750 Multiwavelength Optical Networks

allows for continued functioning of data connections even if the signaling system
fails.
r Notification request object: Indicates the address to which a failure notification should
be sent. This signaling extension for notification messages is needed to accommodate
error handling and provide quick notification of failures to the nodes responsible for
protecting/restoring the failed LSPs. Section 9.3.3.5 presents an example of signaling
for failure protection in optical networks.

Lightpath Establishment and Tear-Down Using GMPLS RSVP-TE

In GMPLS, Path and Resv messages are still used to provision a connection. However,
in contrast to IP and MPLS, the DCN for signaling does not use the same links as those
on which the connection is established. The Path message now carries the LSP tunnel
session object, explicit route object, interface identification object, generalized label
object, sender Tspec, sender template, and upstream label objects. Each intermediate
node receives the Path message, creates a Path state, modifies the explicit route, upstream
label, and interface identification objects, and forwards the message downstream. A Resv
message is then generated when the Path message reaches the destination node. At each
intermediate node processing the Resv message, the cross-connects are set to establish
a bidirectional connection.
PathTear messages alone cannot in this case be used for the deletion of a connection
in optical networks, because when the first cross-connect is reconfigured to tear down
the path the downstream nodes will think that a failure has occurred. This will in turn
trigger efforts to protect the connection using the protection or restoration mechanisms
that have been put in place. To circumvent this problem, a Path message containing an
administrative status object is sent by the ingress node along the connection path, to indi-
cate to all nodes in the path that the connection is in the process of being deleted. A Resv
message is generated in response, also with an administrative status object. On reception
of the Resv message at the ingress node a PathTear message is then generated and as
intermediate nodes in the path receive this message they delete their Path and Resv states.

9.3.3.5 Signaling for Protection and Restoration in Optical Networks

If path protection is provided for a connection, GMPLS signaling is used to provision a
protection path prior to failure events. When a failure occurs and is detected at the source
(via a Notify message), the source node invokes its signaling protocol (RSVP-TE) to
activate the protection path that was precomputed and reserved for the connection that
failed. The Notify message is used in RSVP-TE to notify a session end point (source or
destination) when a failure occurs. Intermediate nodes do not process the message (as
is the case for Path and Resv messages) but rather just forward it to the appropriate end
points of the affected connection. On reception of the Notify message, the session end
point identifies the failed connection using the LSP ID and sends an Ack message to the
node that generated the Notify message.
Figure 9.16 shows an example of a failure event (link failure) and the signaling
messages generated to implement path protection. When the failure occurs at link C-D,
a Notify message is sent to node A (the source node). Node A then generates a Path
 Optical Control Plane 751

Path message sent along the backup path A-F-G-E and

Resv message sent along the backup path from E-G-F-A

G Protection
F path
OXC
OXC

Notify message sent to A

C D

A OXC OXC OXC X OXC OXC E

B failure

Working
path OXC OXC

H I

Figure 9.16 Protection signaling using GMPLS RSVP-TE.

message that is sent along the predetermined protection path and a Resv message is sent
along the protection path in the reverse direction to reserve the network resources.

9.4 Conclusions

A standardized control plane for optical networks is essential for providing automated
connection provisioning and fault recovery in multidomain and multivendor networks.
It also reduces operational cost as it simplifies the operation and management of the
network. At this writing, however, the standardized control plane is still a work in
progress within various industrial organizations and standardization bodies. In view of
the rapid evolution of the optical network this is perhaps inevitable.
The overlay, peer, and hybrid models have been proposed as possible approaches to
the implementation of a control plane, depending on the amount of information that
can be exchanged between a transport network and its client networks. In the overlay
model, the client and transport networks have completely different control planes with
minimal interaction between them via a UNI (similar to the currently deployed IP/ATM
networks), with the transport network hiding its contents from the client network. In
contrast, the peer model uses a single integrated control plane that controls both the
client and transport networks. In that case, the edge devices “see” the contents of the
transport network. A hybrid model has aspects of both the overlay and peer models.
Even though the overlay model is simpler to implement it is obviously not as efficient
752 Multiwavelength Optical Networks

as the peer model in terms of resource allocation, because neither of the two control
planes have complete information. It does scale better than the peer model however,
and it does not face one of the main problems of the peer model: the reluctance of the
service providers of the transport networks to reveal the details of the network to the
clients.
Generalized MPLS has been proposed by several standardization organizations (in-
cluding the IETF and the ITU) as the basis for the control plane for optical networks
(with either opaque or transparent OXCs). Because MPLS is a well-defined and proven
approach to control and enhance data transport for IP networks, a generalization of
MPLS is the natural evolutionary path to control in a wider variety of settings, includ-
ing multiwavelength optical networks. Although MPLS supports only packet (or cell)
types of labels, GMPLS supports any label for entities such as wavelengths, wavebands,
fibers, and time slots. When applied in optical networks, GMPLS derives its routing and
signaling functions from enhancements and traffic engineering extensions of the routing
and signaling protocols used by MPLS. It also utilizes a new link management protocol
that provides for automated neighbor discovery, as well as link and control-channel
management in optical networks. The enhancements to the MPLS routing and signaling
protocols that are implemented for GMPLS are necessary to account for the special
features of the optical network, including the characteristics of optical cross-connects,
the presence of links composed of large numbers of parallel fibers each carrying a large
number of wavelengths, and the need to specify LSPs explicitly over the optical core
network.
Control-plane reliability, a critical subject in the operation of optical networks, is
beyond the scope of this work and is not discussed here. For additional information on
reliable control-plane design, the reader is referred to [Li+02].
There are a number of control-plane protocol standards being developed in various in-
ternational bodies; the main organizations working in this area are the ITU-T, IETF, and
OIF. For example, to standardize the control-plane functionalities of provisioning, main-
taining, and recovering connections in a transport network a number of control-plane
components have been defined by the ITU. Collectively they constitute the framework
for an Automatically Switched Optical Network [ASON01], which is the ITU’s approach
to a unified control architecture for optical transport networks. [Saha+03] describes the
activities of each body and attempts to identify areas of overlap and relations between
the various standards.
Arguably, the most successful protocols and standards are those that are developed
ad hoc to address immediate networking needs; those developed for the Internet are in
this class. Conversely, the least successful are those proposed from the top down as very
general frameworks treating a very broad range of configurations. Many of these have
come and gone; the list is long and forgettable. The GMPLS approach to optical network
control has some aspects of both of these. Its underlying components, derived from
the IP world, provide a firm, specific, and workable foundation despite the differences
between IP and optical networks. However, the generality of the GMPLS approach,
the paramount importance of physical-layer issues (transmission impairments, switch
characteristics, etc.) in optical networks, and the fact that optical networks remain a
 Optical Control Plane 753

rapidly moving target can be roadblocks in the successful application of the control
techniques we described in this chapter.
One of the determining factors in the successful development of standards is eco-
nomics. Deployment of optical network control standards was considerably impeded by
the economic slowdown in the entire telecommunications industry at the turn of the
millennium; more specifically it was impeded by the delayed deployment of intelligent
optical networks. For widespread deployment of the optical control plane to occur, a
mature set of standards is required. Currently, telecommunications service providers
are reluctant to replace their legacy equipment and management infrastructure, which
means that any optical control plane installed in their networks must be able to interoper-
ate with legacy infrastructures (legacy network elements and legacy operations support
systems). Finally, because the proposed optical network control plane is an outgrowth
of the IP network control plane, transport network operators must gain confidence in the
operation of the IP protocols and their extensions before they decide to deploy them in
their networks. Nevertheless, we believe that as the standards mature and the need for a
dynamic optical control plane makes more business sense [Pasqualini+05], the network
operators will have no choice but to deploy an optical control plane in their networks.

Bibliography

[Ashwood-Smith+03] P. Ashwood-Smith (Editor), L. Berger (Editor). Generalized MPLS

Signaling—CR-LDP Extensions. Internet Engineering Task Force (IETF) RFC 3472, January
2003.
[ASON01] Architecture for the Automatically Switched Optical Network (ASON). ITU-T Rec-
ommendation G.8080/Y1304, V1.0, October 2001.
[Awduche+01a] D. Awduche and Y. Rekhter. Multi-protocol lambda switching: Combining
MPLS traffic engineering control with optical crossconnects. IEEE Communications Mag.,
39(3): 111–116, 2001.
[Awduche+01b] D. Awduche, L. Berger, D. Gan, T. Li, V. Srinivasan, and G. Swallow. RSVP-
TE: Extensions to RSVP for LSP Tunnels. Internet Engineering Task Force (IETF) RFC 3209,
December 2001.
[Banerjee+01a] A. Banerjee, L. Drake, L. Lang, B. Turner, D. Awduche, L. Berger, K. Kom-
pella, and Y. Rekhter. Generalized multiprotocol label switching: An overview of signaling
enhancements and recovery techniques. IEEE Communications Mag., 39(7):144–151, 2001.
[Banerjee+01b] A. Banerjee, L. Drake, L. Lang, B. Turner, K. Kompella, and Y. Rekhter. Gener-
alized multiprotocol label switching: An overview of routing and management enhancements.
IEEE Communications Mag., 39(1):144–150, 2001.
[Bellman58] R. Bellman. On a routing problem. Q. Appl. Math., 16(1):87–90, 1958.
[Berger+03a] L. Berger (Editor). Generalized MPLS Signaling—RSVP-TE Extensions. Internet
Engineering Task Force (IETF) RFC 3473, January 2003.
[Berger+03b] L. Berger (Editor). Generalized MPLS—Signaling Functional Description. Inter-
net Engineering Task Force (IETF) RFC 3471, January 2003.
[Bernstein+00] G. Bernstein, J. Yates, and D. Saha. IP-centric control and management of optical
transport networks. IEEE Communications Mag., 38(10):161–167, 2000.
754 Multiwavelength Optical Networks

[Bernstein+04] G. Bernstein, B. Rajagopalan, and D. Saha. Optical Network Control: Architec-

ture, Protocols and Standards. Boston, MA: Addison-Wesley 2004.
[Bouillet+02] E. Bouillet, J. Labourdette, G. Ellinas, and R. Ramamurthy. Stochastic approaches
to compute mesh restored lightpaths in DWDM optical network architectures. In Proceed-
ings of the IEEE Conference on Computer Communications (Infocom), New York, NY, June
2002.
[Braden+97] R. Braden (Editor). Resource Reservation Protocol (RSVP). Internet Engineering
Task Force (IETF) RFC 2205, September 1997.
[Callon90] R. Callon. Use of OSI IS-IS for routing in TCP/IP and Dual Environments. Internet
Engineering Task Force (IETF) RFC 1195, December 1990.
[Chaudhuri+01] S. Chaudhuri, R. Ramamurthy, and S. Sengupta. Comparison of centralized and
distributed provisioning of lightpaths in optical networks. In Proceedings of the IEEE/OSA
Optical Fiber Commun. Conf., Anaheim, CA, March 2001.
[Cheng03] D. Cheng. OSPF Extensions to Support Multi-Area Traffic Engineering. Internet
Engineering Task Force (IETF) Internet Draft draft-cheng-ccamp-ospf-multiarea-te-extensions-
01.txt, February 2003.
[Chiu+03] A. Chiu and J. Strand. Control plane considerations for all-optical and multi-domain
optical networks and their status in OIF and IETF. SPIE Optical Networks Mag., 4(1):26–34,
2003.
[Chiu+05] A. Chiu and J. Strand (Editors). Impairments and other constraints on optical layer
routing. Internet Engineering Task Force (IETF) RFC 4054, May 2005.
[Dijkstra59] E. W. Dijkstra. A note on two problems in connection with graphs. Numer. Math.,
1:269–271, 1959.
[Harrison+01] E. Harrison and B. Miller. Protection and restoration in MPLS networks. Data
Connection White Paper, http://www.dataconnection.com, October 2001.
[Iovanna+03] P. Iovanna, R. Sabella, and M. Settembre. A traffic engineering system for multi-
layer networks based on the GMPLS paradigm. IEEE Network, 17(2):28–37, 2003.
[ITU-T03] Distributed Call and connection Management Mechanism using GMPLS RSVP-TE.
ITU, Recommendation G.7713.2, 2003.
[Katz+03] D. Katz, K. Kompella, and D. Yeung. Traffic Engineering (TE) Extensions to OSPF
version 2. Internet Engineering Task Force (IETF) RFC 3630, September 2003.
[Kodialam+01] M. Kodialam and T. V. Lakshman. Integrated dynamic IP and wavelength routing
in IP over WDM networks. In Proceedings of the IEEE Conference on Computer Communica-
tions (Infocom), pp. 358–366, Anchorage, AK, April 2001.
[Kompella+03] K. Kompella, Y. Rehkter, J. Vasseur, and T. Chung. Multi-area MPLS Traf-
fic Engineering. Internet Engineering Task Force (IETF) Internet Draft draft-kompella-mpls-
multiarea-te-04.txt, June 2003.
[Kompella+05a] K. Kompella, Y. Rehkter, and L. Berger. Link Bundling in MPLS Traffic Engi-
neering. Internet Engineering Task Force (IETF) RFC 4201, October 2005.
[Kompella+05b] K. Kompella and Y. Rehkter (Editors). OSPF Extensions in Support of Gener-
alized MPLS. Internet Engineering Task Force (IETF) RFC 4203, October 2005.
[Kompella+05c] K. Kompella and Y. Rehkter (Editors). LSP Hierarchy with Generalized MPLS
TE. Internet Engineering Task Force (IETF) RFC 4206, October 2005.
[Lang+05] J. P. Lang (Editor). Link Management Protocol (LMP). Internet Engineering Task
Force (IETF) RFC 4204, October 2005.
[Li+02] G. Li, J. Yates, C. R. Kalmanek, and D. Wang. Control plane design for reliable optical
networks. IEEE Communications Mag., 40(2):90–96, 2002.
 Optical Control Plane 755

[Mannie+04] E. Mannie (Editor). Generalized Multi-Protocol Label Switching (GMPLS) Archi-

tecture. Internet Engineering Task Force (IETF) RFC 3945, October 2004.
[McGuire+01] A. McGuire, S. Mirza, and D. Freeland. Application of control plane technology
to dynamic configuration management. IEEE Communications Mag., 39(9): 94–99, 2001.
[Moy98a] J. Moy. OSPF: Anatomy of an Internet Routing Protocol. Reading, MA: Addison-
Wesley Longman, 1998.
[Moy98b] J. Moy. OSPF Version 2. Internet Engineering Task Force (IETF) RFC 2328, April
1998.
[MPLS07] Multiprotocol Label Switching (MPLS). International Engineering Consortium (IEC)
White Paper, Web ProForum Tutorials, http://www.iec.org. 2007.
[OIF01] User Network Interface (UNI) v1.0 Signaling Specification. Optical Internetworking
Forum (OIF), OIF Contribution, December 2001.
[Oran+90] D. Oran (Editor). OSI IS-IS Intra-Domain Routing Protocol. Internet Engineering
Task Force (IETF) RFC 1142, February 1990.
[Pasqualini+05] S. Pasqualini, A. Kirstadter, A. Iselt, R. Chahine, S. Verbrugge, D. Colle,
M. Pickavetand, and P. Demeester. Influence of GMPLS on network provider’s operational
expenditures: A quantitative study. IEEE Communications Mag., 43(7):28–38, 2005.
[Rajagopalan+91] B. Rajagopalan and M. Faiman. A responsive distributed algorithm for
shortest-path routing within autonomous systems. J. Internetwork. Res., 2(1):51–69, 1991.
[Rajagopalan+00] B. Rajagopalan, D. Pendarakis, D. Saha, R. Ramamurthy, and K. Bala. IP over
optical networks: Architecture aspects. IEEE Communications Mag., 38(9):94–102, 2000.
[Rajagopalan+04] B. Rajagopalan, J. Luciani, and D. Awduche. IP over Optical Networks: A
Framework. Internet Engineering Task Force (IETF) RFC 3717, March 2004.
[Saha+03] D. Saha, B. Rajagopalan, and G. Bernstein. The optical network control plane: State
of the standards and deployment. IEEE Communications Mag., 41(8):S29–S34, 2003.
[Saleh+00] A. Saleh, L Benmohamed, and J. Simmons. Proposed Extensions to the UNI for
Interfacing to a Configurable All-Optical Network. Optical Internetworking Forum (OIF) Con-
tribution oif2000.278, November 2000.
[Srisuresh+04] P. Srisuresh and P. Joseph. OSPF-xTE: An Experimental Extension to OSPF for
Traffic Engineering. Internet Engineering Task Force (IETF), Internet Draft draft-srisuresh-
ospf-te-06.txt, March 2004.
[Xin+01] C. Xin, Y. Ye, T. S. Wang, S. Dixit, C. Qiao, and M. Yoo. On an IP-centric optical
control plane. IEEE Communications Mag., 39(9):88–93, 2001.
10 Optical Packet-Switched Networks

Optical networks as described in previous chapters of this book have progressed steadily
since the mid-1980s from point-to-point transmission systems to broadcast stars to ring
networks to fully reconfigurable multiwavelength mesh networks utilizing a wide range
of optical layer equipment: reconfigurable add/drop multiplexers, optical cross-connects,
optical amplifiers, and optical access subnets. The next frontier in optical networking is
the optical packet-switched network.
The present generation of multiwavelength optical networks are circuit-switched in
their core, meaning that they are connection oriented. In these networks, regardless of
the specific scheme used to set up an optical connection, a significant delay (typically
of the order of milliseconds or more) is always incurred during the setup period, dur-
ing which the intermediate switches between source and destination are configured to
support data transport.1 This means that circuit switching is efficient only when the
average duration of the connections is much longer than the setup time; i.e., seconds or
more.
In current applications, exemplified by Internet browsing, a typical source will trans-
mit data in short bursts (on microsecond timescales), possibly changing destinations
with each burst. This is completely incompatible with the circuit-switched approach,
where a source-destination pair holds a dedicated connection for an extended period
of time. Of course, the currently accepted solution to this problem is to maintain the
circuit-switched optical infrastructure and provide a packet-switched logical layer (e.g.,
a network of electronic IP routers) over the optical layer to deal with the bursty traf-
fic. Packet switching provides a transport technology that closely matches the bursty
nature of today’s multimedia traffic and uses the bandwidth efficiently thus ensuring
economical use of the network resources. In this way, a variety of users with different
requirements can be managed in terms of service quality and bandwidth provisioning
through a flexible, scalable, and high-throughput network architecture.
This state of affairs is beginning to show signs of strain and obsolescence, with possi-
ble changes on the horizon, for two reasons: (1) the ever-increasing load on the electronic
layer equipment (e.g., the IP routers) is taxing the capabilities of electronics, leading
network designers to seek optical solutions, and (2) the notion that electronics is smart
and optics is dumb, as we suggested in Section 1.3, is not quite as true now as it used
1
One reason for the long setup delay is the relatively slow timescale on which the current generation of optical
switching devices operate. Another reason is the cumbersome nature of typical connection provisioning
mechanisms.
 Optical Packet-Switched Networks 757

to be. As a result, some proposals for optical, as opposed to electronic packet-switched

networks, which seemed unrealistic in the recent past are now becoming feasible, and
new ideas are proliferating. In optical packet-switched networks, information is trans-
ported in the optical domain in the form of packets, and the data payload generally
remains in optical form from source to destination, avoiding conversion between opti-
cal and electrical forms and avoiding the bottlenecks associated with processing in the
electronic domain. Novel “smart” optical devices capable of high speed signal process-
ing, and fast optical-switching elements are the enabling technology for optical packet
switching (OPS).
It should be noted that optical packet-switching networks are by definition opaque
(that is, nonlinear), because the routing of the optical signal (packet forwarding) is a
function of the information it carries (the addressing information in the packet header).2
Routing based on header information is a nonlinear operation, as opposed to space
switching and wavelength routing in transparent networks, which are linear operations.3
Thus the advantages of transparency that accrue in circuit-switched networks do not
apply here. However, opacity carries its own advantages. As we shall see, the more
nonlinear (opaque) optical signal processing there is in the nodes, the better is the OPS
network performance. (OPS as discussed here bears little resemblance to, and should
not be confused with packet switching in the optical layer, as described in Sections 3.2.2
and 5.6.)
While optical packet switching is the term that is generally used to denote packet-
based switching in optical networks, there are actually three versions of packet-based
switching within this general framework. The first version is basic OPS, in which the data
is switched packet-by-packet providing arbitrarily fine switching granularity. Each packet
header is read, typically electronically, and based on header information the packet (or at
least the payload section of the packet), which remains in the optical domain, is routed
through each switch to its destination. The second version is optical burst switching
(OBS), in which the switching is done on a burst-by-burst basis, where a burst is a string
of packets with the same source and destination. This is similar to a connection-oriented
approach, and it minimizes the need for header parsing and buffering at intermediate
network nodes. Optical burst switching is an attractive alternative to OPS because it
is more efficient than circuit switching while avoiding the buffering issues of OPS.
The third version is optical label switching (OLS), which is essentially a variant of
basic OPS using label switching as the packet-forwarding procedure. In OLS a label
of local significance at each node is used to forward the packets in a manner similar
to MPLS/GMPLS packet forwarding described in Chapter 9. Packets, which may also
contain higher layer (e.g., IP or MPLS) packet headers, are encapsulated with an optical
label as they enter the network. Routing decisions are made at each intermediate node

2
Recall that an opaque network is one which processes the optical signals in a nonlinear fashion. See
Section 1.3.
3
We have seen opaque space switching nodes as well, for example, the opaque O-E-O and O-O-O switches
discussed in Section 2.3.2.4. But the opacity of these switches is due to the (nonlinear) signal regeneration
that takes place at the line interfaces. In this case the nonlinearities are just smart enough to improve the
signal quality, but not smart enough to provide a packet-switching function.
758 Multiwavelength Optical Networks

based on the label, which is erased and rewritten at each node.4 As in MPLS, the use of
labels with local significance means that this approach is scalable, because a given label
can be reused with different significance in different parts of the network. Although
each of these alternatives involves some (typically electronic) processing of control
information in headers, the number of bits involved is orders of magnitude smaller than
the bits in the data payload, which leads to a processing load that is vastly reduced as
compared to completely electronic packet switching in the logical layer.
The aim of this chapter is to offer an assessment of the potential of OPS for core
optical networks, present the underlying principles of optical packet switching, and
document the maturity of the various alternatives.5 The three approaches, basic OPS,
OBS, and OLS will be discussed, presenting typical network elements and protocols
in addition to the issues that arise in implementing these techniques. The focus is on
architectural issues, including switch architectures, signaling methods, contention res-
olution, and the resulting performance parameters, such as end-to-end latency, packet-
blocking probability, and network throughput. The chapter concludes with some ob-
servations on the viability and practical applicability of OPS in current and future
networks.

10.1 Optical Packet-Switched Network Architectures

A convenient way to grasp the essentials of an OPS architecture is to compare it with

the most common current form of packet-switched network, realized as a collection of
network nodes in the form of electronic packet switches (e.g., IP routers) joined by point-
to-point optical links. In this case packets arrive from the links to input ports of a node,
undergo an optical-to-electrical conversion, and are temporarily buffered in electronic
form while their headers are read. The buffers are realized using random access memory
(RAM). The header information in a packet is compared with information in a routing
table, a decision is made determining the output port to be used, and the packet is queued
for transmission on that port. (Queueing is simply a matter of scheduling the order in
which packets are read out of buffers in the RAM.) In some cases (e.g., MPLS) headers
may be written as well as read.
Hidden away in this packet-switching process are a multitude of activities that enhance
network performance. The most important of these are
r Support of a high degree of connectivity between end systems through packet-by-
packet routing

4
While the first implementations of OLS used electronic label processing, recent OLS experiments have used
all-optical processing [Blumenthal+03].
5
The optical packet switch described in this chapter, which is a node of a larger OPS network, should not be
confused with the term photonic packet switch, which has been used in the past to describe a stand-alone
switch with an optical fabric. Typically, the photonic packet switch is an E-O-E structure, accepting and
delivering packets in electronic form and converting them to optical form, where the optical fabric is used
to speed up the switching process. Proposed as early as the 1980s, these switches were usually designed as
the single node in star-type networks [Arthurs+88].
 Optical Packet-Switched Networks 759

r Resolution of contention for outbound transmission links through buffering and pos-
sibly choice of fiber, wavelength, and channel
r Enforcement of quality-of-service (QoS) requirements through appropriate queueing
and routing decisions
The equipment necessary to achieve these goals is essentially the logic used for
reading/writing the header and making the routing decision, the RAM required for
buffering, and the hardware used for transferring packets from the buffers to the output
ports. When translating these generic packet-switching functions to OPS (or to OBS or
OLS) the objective is to maintain the speed and simplicity of optical circuit switching
while achieving the enhanced performance that packet switching can offer. Typically,
header processing and routing decisions are made electronically by tapping a small
amount of optical power from a packet “on the fly” and converting it to an electronic
signal. Based on a table lookup, the packet, which remains in optical form, is switched
(optically) to the desired outbound port. This entails delaying the packet arriving at
the node while header processing and switch reconfiguration are executed. In some of
the more recent experiments, the header processing is executed optically, as mentioned
earlier for OLS. What has kept OPS in the laboratory until now is the lack of an optical
equivalent of the RAM.6 Instead, delays for purposes of packet queueing, synchroniza-
tion, and header processing are currently implemented using fiber delay lines (FDLs).
The use of a delay line for buffering means that optical packets are always in motion
even when they are buffered, as opposed to electronic buffering using RAM, which can
hold packets in place indefinitely. A fiber of a fixed length provides only a fixed delay for
a packet. By combining fibers with optical switches, the equivalent of variable delays
can be achieved, but these arrangements still do not achieve the functionality of a RAM,
because the packet stored in the fiber cannot be accessed at an arbitrary time, nor can
it be stored indefinitely because the bits are eventually mutilated in the delay process.7
The main advantage of optical, as compared with electronic packet switching is that the
payload remains in the optical domain throughout, which produces higher throughput
with less power consumption. These benefits are currently achieved most easily with
architectures where the header processing is realized in the electrical domain, the data
payload switching is done in the optical domain, and delays are realized using FDLs.
These will be the main focus of this chapter, although we will discuss some promising
systems involving all-optical header processing as well.
Given the difficulties of optical buffering, it is helpful to divide the discussion into
unbuffered and buffered networks.

10.1.1 Unbuffered Networks

The simplest type of OPS network has a transparent WDM architecture at the payload
level with unbuffered nodes, without wavelength conversion, where control information
6
At this writing an optical RAM is highly speculative, although progress is being made [Dorren+03, Liu+04].
7
The delay line approach to provide buffering is not new. Some of the earliest digital computers, built in the
late 1940s and early 1950s, used delay lines in which acoustic pulses traveling down a tube of mercury were
used to represent stored bits.
760 Multiwavelength Optical Networks

header
header Network

DEMUX
detector
detector Fast Memory
Forwarding Control
Table Interface

Switch Controller

fiber
delay Switching
Fabric

WDM Network Element

Figure 10.1 Optical packet-switching node.

in the headers is read but not written. In this case the optical switch would typically be a
high-speed wavelength-selective cross-connect (WSXC), shown as the WDM network
element in Figure 10.1.8 As indicated in the figure, an optical tap extracts a small
fraction of the optical power from each incoming packet for purposes of reading the
header and extracting switch timing and routing information. This information is fed to
the forwarding table, which, in conjunction with the Network Control Interface, provides
inputs to the switch controller. The controller then issues the required reconfiguration
commands to the switch at times chosen to accommodate the arriving packets. A short
optical packet delay is introduced using an FDL at each input port. This serves to
temporarily buffer the input packets while the headers are read, the next switch setting
is determined, and the switch is reconfigured.9 Because there is no writing of packet
headers, no buffering, and no wavelength conversion, the optical processing function is
very simple and efficient.
In the simplest form of packet routing, the path for each packet is predetermined,
based on its source and destination address and perhaps other information, such as class
of service. This constitutes a static routing approach with fixed routing information
stored in the forwarding tables in each node. More dynamic approaches can also be
used, where routes may be varied due to local conditions such as packet contention for
limited capacity on outbound links. (This is the case for deflection routing, described
below.) Because of the very rapid timescales over which link occupancy conditions
change, dynamic routing decisions are necessarily based on local information only; i.e.,
a node is myopic. It cannot use knowledge of conditions on links several hops away from

8
Network access connections are not shown in the figure, suggesting that the node performs optical transport
functions only. If it were to provide network access, say for a logical switch (e.g., IP router) or end system,
one or more of the input/output fiber pairs would serve as an access link from the WSXC to an NAS
providing the interface to the external equipment.
9
Although this constitutes optical buffering, it is of a fixed duration so that it cannot be used to queue packets
for arbitrary lengths of time to resolve contention for outbound links.
 Optical Packet-Switched Networks 761

Header
recognition/reading

Packet A Packet A Photonic Packet B

switch
Packet B Packet B Packet A

Add Drop

Figure 10.2 A generic OPS node architecture for an unslotted network. (From [Yao+00, Figure 6].
Copyright
c 2000 IEEE. Used by permission of The Institute of Electrical and Electronics
Engineers, Inc.)

it in making a forwarding decision. As will be seen, this can be a significant handicap in

attempting to keep packet losses low.
The performance of the basic unbuffered architecture can be improved by introducing
synchronized time slots, as used in slotted random access systems (Section 5.6.1).
Figure 10.2 shows a generic unslotted OPS node, including add/drop units for external
access. In the unslotted (asynchronous) case, the packets are of variable lengths, arrive
at the switch at arbitrary times as indicated by the positions of the packets in the figure,
and pass through the switch in the same relative positions in which they arrived (delayed
by a fixed time interval due to the FDLs).
In the slotted case, Figure 10.3, all packets are of a fixed length (as in ATM cells)
and are transmitted on each link within well-defined time slots. Because the time slots
on different input links are not aligned with each other, variable delays are required on
each switch input port to realign the boundaries of the packets within the switch and on

Switch control
unit

Packet A Sync. Packet A Photonic Packet B

stage switch
Packet B Sync. Packet B Packet A
stage

Add Drop

Figure 10.3 A generic OPS node architecture for a slotted network. (From [Yao+00, Figure 1].
Copyright
c 2000 IEEE. Used by permission of The Institute of Electrical and Electronics
Engineers, Inc.)
762 Multiwavelength Optical Networks

1/2 Packet 1/4 Packet 1/8 Packet 1/16 Packet

length length length length
delay delay delay delay

Input Output

Figure 10.4 An FDL-based synchronizer. (Adapted from [Yao+00, Figure 4]. Copyright

c 2000
IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

the outbound links. The figure shows a synchronization stage at each input port on the
switch (taking the place of the fixed FDL in the unslotted case of Figure 10.2), which
accomplishes this task.
Synchronization circuits have been realized using combinations of switches and FDLs,
which produce the effect of a controllable variable delay. Figure 10.4 is an example of
a synchronizer composed of a combination of FDLs and 2 × 2 switches. An FDL is
included or excluded in the total delay by setting the switch in front of it appropriately.
Four stages are shown giving a variable delay in increments of one-sixteenth of a packet
length, but it could be extended to n stages yielding increments of 2−n .
In slotted networks, the packets are placed in time slots that are a little larger than
the packet size, thereby inserting guard times between successive packets. This provides
tolerance for slight misalignments as well as time for clock recovery for header reading
on a packet-by-packet basis. An additional guard time may also be placed between
the header and the data payload. This might be necessary, for example, if headers are
extracted from incoming packets and written in front of outgoing packets, as in optical
label switching (see Section 10.5).
A generic optical packet format for a slotted architecture is shown in Figure 10.5.
Header reading circuits in the switch control unit are used to recognize the special
synchronization bit pattern at the beginning of the header and thereby locate the packet
boundaries, providing information necessary to adjust the synchronization stages and
determine switch reconfiguration times. The header may consist of several fields in
addition to synchronization bits, including source and destination labels, packet type

Switching time slot

Header Guard Payload Guard

band band

Sync SRC DSTN PT PN OAM HEC

label

Figure 10.5 A generic packet format for a slotted network. (From [El-Bawab+02, Figure 2]. Copy-
right

c 2002 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers,
Inc.)
 Optical Packet-Switched Networks 763

(for when priorities are used), sequence numbers, an operations, administration, and
maintenance (OAM) field, and header error correction field (HEC), among others.
In unslotted networks, the packets are not aligned prior to entering the switch, and
thus there is no need for synchronization stages. In general, slotted networks have a
performance advantage over unslotted networks, because fixed packet sizes and aligned
time slots reduce the probability of time overlap among packets contending for the same
outbound link at a switch. However, unslotted networks use simpler node architectures
and do not require packet segmentation and reassembly at the source and destination
nodes [Yao+00].
In the basic unbuffered version of optical packet switching the rule used by the switch
controller is as follows:

1. When a packet arrives at the switch, its header is read and its outbound port is
determined by the controller.
2. If the designated outbound link is free on the packet’s wavelength at the time the
header is read, the switch is set to route the packet to the corresponding port and
the switch setting is held for the duration of the packet transmission. Otherwise the
packet is lost.

The main problem encountered in OPS is the resolution of contention. An outbound

link is free on a packet’s wavelength only if there are no other arriving packets overlapping
that packet in time, which are destined for the same link on the same wavelength.10 If
these conditions are not met, contention for the outbound link leads to packet loss in the
simple bufferless, wavelength-continuous architecture described above.
An example of packet loss due to contention is shown in Figure 10.6, illustrating
what happens when several packets on the same wavelength are contending for the same
output port. The time intervals shown in the figure are defined as follows:

th = time to read header

tsw = time to configure switch
tg = inter-packet guard time > tsw
td = packet delay > th + tg . (10.1)

As shown in the figure a packet labeled (x, y) arrives on port x and is destined to
port y.11 The packet’s header is read and a scheduling decision is made after an interval
th , required to read the header. If the decision is to accept the packet, a path is created
through the switch from input x to output y on the packet’s wavelength. While the path
is being configured, the packet is propagating through a fiber delay td on its way to the
switch. In the scenario shown, the first packet on input port 1 is accepted and appears
on output port 1 with a delay td . (The delay within the switch is assumed negligible.)
A second packet arrives on input line 1 after an interval somewhat longer than the guard
10
In addition, for an outbound link to be free for a given packet on a given wavelength no other packet on
that wavelength (from another input port) can be scheduled ahead of it on that outbound link.
11
These labels are shown to simplify the illustration. In a real network, information in the packet header
would normally be translated into a destination port via routing tables stored in the controller.
764 Multiwavelength Optical Networks

th tsw

Input 1 (1, 1)
tg
X (1, 1)

Input 2 (2, 1)

Input 3 (3, 1) X

td td

Output 1 (1, 1) (2, 1)

Figure 10.6 Optical packet contention.

time tg . There are three packets that could be scheduled next: from input ports 3, 2,
and 1 in that order. When the header from the packet on input 3 is read, the channel
on output 1 is occupied, so that packet must be dropped. The packet from input 2 is
accepted because the output channel is free when header reading is completed. The
second packet from input 1 is dropped even though the channel is free when the header
is read, because the packet from input 2 was scheduled before it. (The blocked packets
are indicated with an X in the figure.) Note that the delays td are chosen to be sufficiently
long so that a guard time of at least tg will appear between successive packets on the
same wavelength, ensuring that there is enough time to set the switch before each packet
passes through it. This illustration was presented in the context of an unslotted system
wherein the packets are forwarded or dropped on a first-come-first-served (FCFS) basis.
In the slotted case, all packets are presented to the switch in the same time slot (after
slot synchronization), so it is not clear what order should be used in determining which
packets should be dropped in cases of contention. This provides an opportunity to use
criteria such as priorities or QoS considerations to make packet-loss decisions.
Without introducing special contention resolution techniques packet loss in unbuffered
systems can be very high. Procedures to ameliorate contention can be implemented in
the space domain (using deflection routing), in the time domain (using packet buffering),
and in the wavelength domain (using wavelength converters). These can be used alone
or in combination, as will be shown below.

10.1.2 Deflection Routing

Deflection routing (known as hot potato routing when it was proposed for the early
packet-switched networks [Baran64]) is the intentional misrouting of packets by the
switch when there is contention for a given outbound link. In the case of deflection
 Optical Packet-Switched Networks 765

routing, rather than using a single predefined path for each packet, normally the shortest
path from source to destination, a packet is allowed to be deflected from that path at
an intermediate node if it is blocked on the shortest path. In essence, in the case of
deflection routing the links on the network are a distributed storage medium, serving
as a global “virtual buffer” in the absence of local buffering at each node. Deflection
routing has been investigated thoroughly for regular topologies such as Shufflenets and
the Manhattan Street Network (MSN) [Maxemchuk89].
Typically, some constraints must be imposed on deflection routing to avoid the pos-
sibility of looping and to keep path lengths from growing too long. Overly long paths
introduce delays and cause congestion, decreasing the network throughput. In severe
cases congestion can lead to unstable network operation and ultimately network fail-
ure. For example, in MSNs it was shown that there is a 30–45% reduction in network
throughput when deflection routing is used rather than buffering the packets that are
in contention [Maxemchuk89]. This problem was also studied in terms of a deflection
index, defined as the maximum number of hops that one deflection adds to the shortest
path between a source and destination node in the network [Greenberg+93]. It was
shown that the deflection index in an arbitrary mesh network must be kept as small as
possible to avoid a large reduction in the network throughput when deflection routing is
utilized.
A simple way of imposing a deflection constraint is to incorporate a time-to-live field
in the packet header or to incorporate a limited set of allowable alternate paths in the
routing tables. This solution will also alleviate looping or livelock12 [Leonardi+96], a
term denoting the condition where packets are deflected endlessly without ever reaching
their destination. Standard approaches that have been used to eliminate looping in a
network with deflection routing may cause unnecessary complexity in the processing of
the header. This has led to the development of efficient algorithms adapted to OPS that
guarantee loopless deflection. A loopless deflection routing algorithm was developed in
[Jue02] and evaluated for a 15-node mesh network topology. Simulation results indicate
that the loopless deflection approach offers a reasonable trade-off between average hop
distance and packet losses compared to the unrestricted deflection approach.
As an illustrative example of deflection routing consider the Petersen network shown
in Figure 10.7. Using shortest path (SP) routing and A as the source node, we can
reach nodes B, F, and E in a single hop and the remaining nodes in two hops. Suppose
SP routing is used from node A to node D, together with a set of deflected routes
limited to distances not exceeding four hops. The following possible paths are obtained
(Figure 10.8): A-E-D (the shortest path), A-B-C-D and A-F-I-D (the paths deflected
from the shortest path if link A-E is completely occupied), and A-B-G-I-D and A-F-H-
C-D, the paths deflected from the intermediate nodes B and F, respectively. The different
deflection paths will generally have different propagation delays, so if a sequence of
packets enters the network at node A with destination node D, they may arrive at node D
out of order due to the fact that each packet may take a different path. Thus, deflection

12
Also termed deadlock [Greenberg+93].
766 Multiwavelength Optical Networks

J G B
E

I H

D C

Figure 10.7 Petersen network graph.

routing may require resequencing at the destination, which in turn means that packet
headers must carry sequence information.

10.1.3 Performance Analysis of Deflection Routing

We now present estimates of packet-loss probabilities in deflection routing as a function
of network load and other parameters. As will be shown, the wavelength-continuity
constraint plays a key role in exacerbating contention and hence packet loss. We therefore
consider the possibility of relaxing this constraint by using a WIXC as the switching
element in Figure 10.1. This means that packet contention in a network operating with
W wavelengths on each fiber will occur if and only if more than W packets destined
for the same outbound link (fiber) overlap in time. Their wavelengths on the inbound

B C
2

4
G

1 E
A D

H
5

3
F I

Figure 10.8 Paths from A to D.

 Optical Packet-Switched Networks 767

links are irrelevant because the WIXC will be set by the switch controller to make
the wavelength conversions required to avoid wavelength contention on the outbound
links.
Certain assumptions are helpful to keep the problem simple and tractable. We assume
that the links are carrying packets of variable lengths in a random (Poisson) stream at a
rate of λ packets/s with average packet length µ−1 . Furthermore,

r Each wavelength on each fiber in the network is equally loaded with traffic intensity
ρ = µλ and
r The occupancies of each wavelength on each fiber at a given point in time are mutually
independent random events.

The independence assumption above is only an approximation but is necessary for

tractability.
Suppose a WDM network with N nodes and B links operates on W wavelengths and
each node injects ρs units of traffic into the network. If the traffic is uniformly distributed
on all fibers and packets require an average of H̄ optical hops, then the load on each
wavelength on a fiber is

Nρs H̄
ρ= . (10.2)
2BW
In the case of the Petersen network with traffic from each source node uniformly dis-
tributed to the other nine nodes, using SP routing (measured in optical hops), one-third
of the traffic requires a single hop and two-thirds requires two hops for an average hop
count of H̄ = 35 . Furthermore, because the network is completely symmetric, the load
on each wavelength on a fiber is
5ρs
ρ= . (10.3)
9W

Packet Loss in SP Routing

Under the independence assumption stated above, packet-blocking probabilities are eas-
ily computed. A packet is blocked at a switch if all outbound fibers and wavelengths that
are allowable for that packet are occupied when the packet is presented for transmission.
Because there is no buffering in the network, all blocked packets are lost. In a network
with wavelength conversion (WC) the probability of a packet being blocked on an h-hop
path using the above assumptions is

PbW C = 1 − (1 − ρ W )h . (10.4)

Without WC, the blocking probability on an h-hop path on a specific wavelength

using the above assumptions is

Pb = 1 − (1 − ρ)h . (10.5)
768 Multiwavelength Optical Networks

Using Equation (10.4) for the one- and two-hop traffic on the Petersen network and
applying the appropriate weighting for each, the total blocking probability is found as
1 2 5
PSWPC = ρ W + [1 − (1 − ρ W )2 ] ≈ ρ W for small ρ W . (10.6)
3 3 3
In the Petersen network without WC, the probability of a packet being blocked
on a one-hop path using the above assumptions is ρ W and on a two-hop path it is
ρ W + (1 − ρ W )ρ. This gives
1 W 2 W 2
PS P = ρ + [ρ + (1 − ρ W )ρ] ≈ ρ for small ρ W . (10.7)
3 3 3
Comparing Equations (10.6) and (10.7) for some typical parameters illustrates the
importance of WC in this case. Taking ρ = 0.5 and W = 10 we have PSWPC ≈ 0.002
and PS P ≈ 0.33. Clearly, SP routing without WC is not a feasible option here while it
has fairly good performance with WC. The high blocking without WC is caused by the
fact that a wavelength must be assigned to the two-hop packet traffic without knowing
whether that wavelength will be available on the second hop. The deflection routing
protocol analyzed next attempts to circumvent this problem.

Packet Loss in Limited Deflection Routing

The discussion in this section is limited to the two-hop traffic in the Petersen net-
work; each network topology yields different results. The focus is on the traffic from
A to D with the alternate paths numbered from 1 to 5 as indicated in Figure 10.8.
With the assumption that the shorter path is used if possible at each node (with
node B used before node F), the following expression for blocking probability for
the two-hop traffic in the Petersen network using limited deflection routing (LDR)13 is
obtained:
5

PL D R = PCi PBi + PC0 , (10.8)
i=1

where PCi is the probability that path i is selected, i = 1, 2, · · ·5, PC0 is the probability
that no path is selected, and PBi is the probability that the packet is blocked on path
i given that path i is selected. In the case of networks with WC these probabilities
are
PC0 = ρ 3W
PC1 = 1 − ρ W
PC2 = ρ W (1 − ρ W )2 (10.9)
2W W 2
PC3 = ρ (1 − ρ )
PC4 = ρ 2W (1 − ρ W )
PC5 = ρ 3W (1 − ρ W )
13
LDR means allowable deflection paths are chosen so that there is a limit (in this case, four hops) on the
number of hops after a packet is deflected.
 Optical Packet-Switched Networks 769

and
PB1 = ρ W
PB2 = ρ W
PB3 = ρ W (10.10)
PB4 = 1 − (1 − ρ W )3
PB5 = 1 − (1 − ρ W )3 .
This gives PLWDCR ≈ ρ W for small ρ W . This is only an improvement of a factor of
approximately 2 over SP routing, assuming that ρ remains the same in both cases.
Moreover, it must be recalled that ρ is proportional to H̄ , which is larger in the case of
deflection routing than it is for SP routing. (The difference is small for small blocking
probabilities.) Thus, the improvement is somewhat less than a factor of 2.
In the case of networks without WC the quantities PCi and PBi given in
Equations (10.9) and (10.10) are modified as follows:
PC0 = ρ 3W
PC1 = 1 − ρ W
PC2 = ρ W (1 − ρ W )(1 − ρ)
PC3 = ρ 2W (1 − ρ W )(1 − ρ) (10.11)
W +1 W
PC4 = ρ (1 − ρ )
2W +1
PC5 = ρ (1 − ρ W )
and
PB1 = ρ
PB2 = ρ
PB3 = ρ (10.12)
3
PB4 = 1 − (1 − ρ)
PB5 = 1 − (1 − ρ)3 .
From Equations (10.11) and (10.12) we have PL D R ≈ ρ for small ρ W , which is about
the same as the blocking on two-hop paths using SP.14 Furthermore, as noted above,
LDR tends to increase H̄ which in turn increases ρ for a fixed offered traffic, ρs , so it
can actually increase the blocking probability. If losses are recovered by retransmission,
this can cause instability in the network. Thus, without WC, LDR is of little value
in this example and may even be detrimental. In larger networks it may have some
advantages, depending on the number of deflection alternatives (i.e., the degree of each
network node). In general the use of local information only in making deflection routing
decisions at a node is detrimental in networks without WC. This is due to the fact that
14
The reason for the lack of improvement using LDR in this case is that wavelength continuity forces the
blocking probability on the last link in each path to be equal to ρ no matter how many routing alternatives
there are.
770 Multiwavelength Optical Networks

wavelength continuity requires the chosen path to be free on a given wavelength all the
way to the final destination. Yet this information is not available to the node making the
routing decision.
Many studies of deflection routing with and without WC have appeared in the liter-
ature. For example, in [Castanon+97] it was shown that when wavelength conversion
is used, the probability of deflection in optical deflection-routed Manhattan Street and
ShuffleNet networks decreases and the propagation delay improves significantly. Specifi-
cally, it was shown that under uniform traffic, more than 60% of the propagation delay in-
crease using deflection routing as compared to store-and-forward routing was recovered
with the use of wavelength conversion when more than five wavelength channels were
used in the network. In addition to the network performance improvements, wavelength
conversion without deflection causes no extra packet jitter and no packet-reordering
problems.

10.1.4 Buffering: Time Domain Contention Resolution

The unbuffered architectures examined in the previous section typically have significant
packet-loss probabilities, which can be reduced by including packet buffering and wave-
length conversion as means of resolving contention in the time and wavelength domains.
The effectiveness of contention resolution greatly influences network performance in
terms of probability of packet loss, network utilization, average packet delay, and average
source-destination path lengths. In this section we describe some promising buffering
schemes for this purpose and then move on to systems with wavelength conversion in
the next section.

10.1.4.1 Buffered Nodes

In the absence of optical RAM, the fiber delay line, introduced earlier in this chapter, is
currently the medium of choice for optical packet buffering. The simplest FDL packet
buffer is made up of a length of fiber, typically several km long, accommodating several
packets queued in a FCFS manner. Because the length of fiber required to store one
packet is inversely proportional to the packet bit rate, high bit rates are advantageous
for FDL buffered systems. For example, a 5000-bit packet transmitted at 1 Gbps would
occupy about 1 km of fiber, whereas the same packet transmitted at 100 Gbps would
require only 10 m.
A number of node architectures can be employed that provide time domain contention
resolution based on FDLs. They vary in the location and techniques of utilization of the
FDLs. Optical node architectures are typically modeled after the structures commonly
used for buffering in electrical networks, including input, output, and feedback (recir-
culation or reentry) buffering. It is therefore worthwhile to recall briefly the electronic
approach.
Implementation of buffered packet switches in electronics is a straightforward task
using RAMs. In a simple configuration each incoming packet is read into a common
memory and then written out to its requested output line when that line is free. The
 Optical Packet-Switched Networks 771

Controller

Delay, synchronize,
1 read header 1

2 2

Optical Buffers (OB) R x R WSXC

R R

Figure 10.9 Input buffered optical packet switch with WSXC.

common memory queues contending packets while they are awaiting transmission on
their requested links. The read/write operations take the place of a reconfigurable switch
fabric, so a hardware arrangement with clearly defined sections for switching and buffer-
ing does not exist in this case. However, the common memory configuration with proper
packet scheduling is equivalent to a switch with output buffering. Provided that the
RAM is sufficiently large (to reduce probability of buffer overflow) and the read/write
speed is sufficiently high, an output buffered switch produces optimal performance
in terms of packet loss. Packet processing speed is the limiting factor in electronic
packet switches. In a slotted switch, for example, packets must be read into memory
from each input line and out of memory to each output line in a time interval of less
that one slot, placing a heavy burden on the electronics. Higher speeds are achieved
by implementing packet transfer using multistage fabrics similar to those discussed
in Section 2.3.2.1. Nevertheless, attempts to scale these switches to higher through-
puts and larger numbers of switch ports are eventually impeded by the speed limits
of the electronics – the electronic bottleneck – at which point OPS becomes a viable
alternative.
In OPS with input buffering, FDLs capable of storing multiple packets are placed at
the input ports of the switch and the arriving packets are queued in their respective FDLs.
Figure 10.9 shows a block diagram of an input buffered configuration based on a WSXC.
Another arrangement using WMUX/WDMUXs and multiple single wavelength space
switch planes is shown in Figure 10.10. The latter requires considerably more hardware.
The buffers for one input line in the system of Figure 10.9 could be realized in an
asynchronous system by a long FDL holding several packets or in a slotted system by
series of FDL buffers holding one packet per buffer as shown in Figure 10.11. Using the
configuration of Figure 10.9 each FDL carries wavelength multiplexed packets, whereas
in Figure 10.10 each FDL carries packets on a single wavelength only.
The main difficulty that arises in input-buffered FDL architectures is that delay lines
necessarily operate in FCFS order, which causes head-of-the-line (HOL) blocking (see
772 Multiwavelength Optical Networks

Controller

WDMUX WMUX
Delay, read header
Delay, read header
1 1
Delay, read header

2 2
Optical Buffers R x R OXC
Optical Buffers R x R OXC
Optical Buffers R x R Space Switch

R R

Figure 10.10 Input buffered optical packet switch using multiple space switch planes.

Section 5.6.3). This is a common problem in input-buffered switches no matter how

they are realized. HOL blocking arises when the packet at the head of a queue cannot be
switched to its corresponding outbound link because that link is occupied or reserved
for a packet contending for the same outbound link. This in turn means that the packets
behind the blocked packet in the same queue cannot be sent to their output links even
if those links are available. An example of HOL blocking is shown in Figure 10.12 for
an input-buffered queue with four buffers per link. In this case, the second packet in the
third queue is prevented from reaching output 3, even though that link is free, because
the packet ahead of it is experiencing contention for output 1. It has been shown that
HOL in input-buffered systems limits the attainable throughput in the switch to less than
two-thirds of the full switch capacity [Hui90].
In output buffering, buffers are associated with each of the outbound ports and packets
are held in these buffers while awaiting service on their outbound link. Output buffering
increases the complexity of the switch, because multiple contending packets must be able
to access the queue on the same outbound link almost simultaneously. This is especially
difficult in OPS. Nevertheless, architectures that at least approximate output buffering
are worth considering because of their superior packet-loss performance. The absence
of HOL blocking produces very low packet-loss probabilities.
Improved switch architectures can also be realized by interspersing FDLs and switches
in a multistage configuration. An example, known as a feed-forward or switched delay
line architecture, is shown in Figure 10.13 for a 2 × 2 switch. In this case, the switch is
implemented using three 2 × 2 switches surrounding FDL packet buffers, each holding a
single (fixed length) packet. (Note the similarity to the synchronizer of Figure 10.4.) The
switches enable packets from different input lines, contending for the same outbound

...
1 2 N
Figure 10.11 FDL input buffer.
 Optical Packet-Switched Networks 773

4 1 2

2 2 1

Space switch
4 3 1

2 1 4

Figure 10.12 Example of head-of-the-line (HOL) blocking. (From [Hunter+98b, Figure 5]. Copy-
right
c 1998 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers,
Inc.)

line to be delayed by different numbers of time slots. If incoming packets on the two
input lines, arriving in the same time slot are destined for the same output link, one is
forwarded immediately to that link, by setting the switches to bypass the FDLs. The
other is placed in an FDL. For example, if both packets are destined for the upper output
port the packet on the upper input port can be forwarded directly to the upper output port
by setting the first and third switches in the cross state and the second in the bar state.
This will place the second arriving packet in DL1. If DL2 is empty then the packet in
DL1 can be forwarded directly to the upper output port in the next time slot by setting
the second and third switch to the cross state. If DL2 is occupied the two switches are
set in the bar state and the packets in the two FDLs are delivered to the upper output
port in FCFS order. Clearly, this architecture allows each packet two opportunities to be
buffered before it reaches its destination. If contention persists, even after the second
buffering opportunity, then the packet will have to be blocked.
Generally, in input- or output-buffered systems packet-loss performance improves
with the number of buffers employed. However, large buffer sizes are difficult to im-
plement in OPS. A way of improving contention resolution with a limited number of
buffers is by using the feedback delay line architecture of Figure 10.14. In this case,
when there is contention between two incoming packets, one of them is directed to one
of the outbound links associated with the FDLs. By setting the switch to recirculate the
packet through the FDL, the buffered packet can keep circulating inside the feedback

delay line delay line

DL1 DL2

inputs outputs

2x 2 switch 2x 2 switch 2x 2 switch

Figure 10.13 Feed-forward delay line architecture. (From [Chlamtac+96, Figure 1]. Copyright
c
1996 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)
774 Multiwavelength Optical Networks

outputs
inputs

4x 4 switch

fiber

Figure 10.14 Feedback delay line architecture.

loop until the outbound link becomes available. This recirculation scheme in effect cre-
ates an “endless” delay – in theory. In practice, attenuation in the loop and/or ASE noise
introduced by amplification within the loop degrade the quality of the optical signal
every time the packet recirculates through the switch fabric. After passing through the
fabric a few times, the signal will be mutilated and the packet will have to be dropped.
Another feedback delay line node architecture with FDLs that has been investigated is
the partially shared buffering (PSB) architecture [Diao+99]. This is an output buffering
scheme in the sense that there is an optical buffer (called the prime buffer – essentially
a set of FDLs) for each output line, and in addition there is an output common buffer
(FDLs) that is shared among all the outputs. Overflow packets (arriving packets that
cannot be accommodated) that would otherwise be blocked are now sent to the shared
buffer for temporary storage. Simulation studies in [Diao+99] demonstrated that the
PSB architecture can achieve a higher throughput without significantly increasing the
size of each prime buffer or heavily utilizing wavelength converters. It is also found
that the increase in packet delay caused by the PSB is very minor, and the mean packet
delay approaches an upper bound (the mean packet delay of an M|D|1 queue15) when
the prime buffer size is large enough.

10.1.4.2 Dump-and-Insert Buffering

One way of emulating output buffering and thereby approaching its performance benefits
is to use a more sophisticated packet-scheduling scheme in an input-buffered system.
An example is shown for a slotted system in Figure 10.15, which we refer to as a
dump-and-insert (DI) buffer architecture. Using a structure similar to Figure 10.13, a
single N -stage cascade of FDL buffers, each holding a single packet, is inserted at each
input port. Each FDL is placed between a pair of 2 × 2 controllable directional couplers
operating as “dump-and-insert” switches (essentially add/drop switches). The switches
control the path taken by the packets. As indicated in Figure 10.15 the upper path between

15
See Equations (5.79) and (5.80).
 Optical Packet-Switched Networks 775

Controller

1 Delay, synchronize
read header
1

2 2
Dump
Optical Buffers R x R Space Switch

R R

Fiber delay

Controllable directional coupler

Figure 10.15 Dump-and-insert buffer architecture.

the switches interconnects the fiber delays, and the lower path detours around the delays.
The system shown in the figure is assumed to be operating on a single wavelength, so
it would represent one layer of optical buffers in the architecture of Figure 10.10, but it
could also operate on wavelength-multiplexed packets if it used wavelength selective DI
switches, in which case it would represent the optical buffer block in Figure 10.9.
In a slotted system, after passing through the delay, synchronization, and header
reading blocks, at most one packet will arrive at each input of the buffering system in
each time slot. It is the function of the controller to set the 2 × 2 switches either in
the bar or cross state to schedule the packets according to some desired optimization
criterion. This might mean maximizing throughput and/or minimizing packet loss due
to contention. To understand the role the switches play in packet scheduling, imagine
a packet arriving at an input port in a given time slot. If all buffer stages are currently
occupied on that port, and all switches are in the bar state, all packets are routed along
the upper path joining the buffer stages, passing successively through each buffer, so
the system acts as an ordinary FCFS queue. In each successive time slot, the arriving
packet is loaded into the buffer at the end of the queue, and all other packets move to
the right, with the packet at the head of the queue traversing the OXC and entering its
designated outbound line. Of course, the loading and shifting of packets are all occurring
continuously in time as the packets propagate through the fiber delays. However, it is
convenient to view the buffer states in terms of static “snapshots” at the beginning of
each time slot.
Now, imagine a different situation where all buffers are empty when the packet
arrives. Under the same conditions as described before (all switches in the bar state),
the packet would again be loaded into the left-hand buffer and would then experience an
(unnecessary) delay of N time slots before being transmitted. However, if the first and
776 Multiwavelength Optical Networks

Arrivals 1 2 3 4 Departures

1 4 2 4
3 Transmit
Dump
2 2 3 3 Transmit
Dump
1 3 3
3 4 Transmit
Dump
1 1 1 Transmit
4 3
Dump
Current slot

Arrivals 1 2 3 4 Departures

1 3 4 2
1 4
Dump
3 3
2 1 2
Dump
1 3 4
3 1 3
Dump
1 1
4 1
Dump
Next slot
Fiber delay

Controllable directional coupler

Figure 10.16 Typical packet sequence in DI buffers for a 4 × 4 optical switch.

last switches are placed in the cross state with the remaining switches in the bar state,
then the arriving packet detours along the lower path in the multistage buffer structure
bypassing all delay states, and immediately enters the OXC.
Many alternatives for packet scheduling within the multistage buffers are possible
because of the flexibility offered by the switches. We shall now describe a packet-
scheduling strategy designed to make the buffer arrays approximately emulate an output-
queued OXC. The basic idea is to attempt to insert each arriving packet as close to the
head of its queue as possible without causing contention with packets at other input ports.
This technique circumvents the problem of HOL blocking. The strategy is illustrated
using a 4 × 4 OXC preceded by four stages of buffering on each port, as shown in
Figure 10.16. The figure represents a snapshot of the buffer contents at the beginning of
one time slot. Each stage of the buffer array is indicated as a column, numbered from 1
to 4, with the number 4 at the head of the queue. A number inside a circle indicates the
output port to which the packet currently stored in that buffer is destined. The contents
of the fiber delays at the beginning of the current time slot are indicated in the upper
part of the figure, and the contents one time slot later are shown in the lower part. At the
end of the current time slot, the controller changes the OXC settings to route each packet
 Optical Packet-Switched Networks 777

in column 4 to its indicated output port. These packets begin their traversal of the OXC
at the beginning of the next time slot. (Guard times are required between successive
packets to allow time for switch reconfiguration.) To avoid contention for an outbound
link, all entries in each column must be distinct. Note that the buffer contents in this
example are collisionless and that there are several empty positions in the queues. In
each case any attempt to fill the empty buffers by moving packets forward in the queue
from their current positions will create contention.
A set of four packets labeled with their designated output ports, arriving in the
beginning of the current time slot, is shown at the entrance to the buffer array. The
headers of these packets were read when the packets passed through the input delays,
and this information is used to control the buffer switch settings as well as the OXC
settings for the packets reaching the head of the queue. To keep things simple, we assume
the following protocol for packet scheduling:
1. Starting from input port 1 and working down, an arriving packet is placed in the
feasible buffer position closest to the head of its queue. (A position is feasible if it
is vacant and it does not cause contention with packets in the same column at other
input ports.) If the packet is placed at the head of its queue it is transmitted in the
next time slot.
2. If there is no feasible position in the packet queue the arriving packet is dumped
(lost).
This protocol ensures that packets destined for the same output port are delivered
in correct order. However, packets with different destinations may be scheduled out of
order to increase throughput and reduce delays.
The method of switch setting used to adhere to the above scheduling rules is illustrated
in Figure 10.16. Switches set to the cross state are shown as “X”s, in the figure. All
others are set to the bar state. If all switches are held in the bar state, then the contents of
the current slot are simply shifted one position to the right in the next time slot, meaning
that the contents of column 4 are transmitted through the OXC in the next slot. All
arriving packets are placed in column number 1 (the end of the queue) in the beginning
of the next time slot. For example, at input port 1, the packet destined for output 3
appears in column 1 at the beginning of the next slot, as shown in the lower part of the
figure.
To place an arriving packet in a vacant buffer position farther forward in the queue,
the switch preceding the first buffer and the switch immediately following the targeted
vacant buffer are set in the cross state. The effect of this maneuver is to route the arriving
packet to the lower path, bypassing several buffer stages and then reentering the upper
path just ahead of the desired buffer. The arriving packet then appears in the targeted
buffer location at the beginning of the next time slot. (This location is one column to
the right of its position when the operation was initiated.) At the same time the original
contents of the desired buffer are detoured to the lower path and then dumped. Because
the desired buffer position was vacant in this case, nothing is lost in this maneuver. As
an example, the packet at input port 2, destined for output port 2, sees a vacant (and
778 Multiwavelength Optical Networks

feasible) buffer location in column 4. By setting the indicated switches in the cross state
the arriving packet is moved ahead to appear one column to the right of column 4 in
the next slot. In this case that means that the arriving packet is directly connected to the
input of the OXC at the beginning of the next time slot.
In a similar way, the packet arriving at input port 3 and destined for port 4 is detoured
to column 3 (which is feasible) and appears in column 4 at the beginning of the next time
slot. Finally, the packet arriving at input port 4 and destined for port 3 sees no feasible
buffer location. It cannot be appended to the end of the queue because it will collide with
the arrival on port 1, and it cannot be moved forward to the vacant position in column
1, because there is another packet with the same destination in that column. Thus the
arriving packet must be dropped, which is accomplished by placing the left-hand switch
in the cross state and dumping the packet from the lower path.
Note that when packet dropping is required there are various possible choices. In this
case, for example, if the packet arriving at port 4 was higher priority than that arriving
at port 1, the latter could be dropped rather than the former. In other cases it might be
feasible and desirable to dump some other packet already in the queue in favor of a
newly arriving packet in the same queue.
These operations tend to emulate output queueing by rescheduling packets contending
for the same output line. What makes this different from a true output-queued system is
that there are empty spaces in the input queues that cannot be filled by moving the packets
forward, because collisions will occur. However, by using the detouring function of the
DI switches, newly arriving packets can be inserted into the empty spaces provided that
they do not create contention. Thus, this protocol represents a middle ground between
an input- and output-buffered switch.

10.1.5 Buffering and Wavelength Conversion: Time/Wavelength Domain

Contention Resolution
Using buffering alone to achieve good OPS performance may require inordinately large
numbers of buffers. Thus, it is worthwhile to attempt to combine buffering with other
contention resolution techniques. In the case of deflection routing, we have already
seen that wavelength conversion can vastly improve the performance of unbuffered
architectures. Similarly, adding wavelength converters to buffered architectures can
produce still more performance improvement while at the same time reducing the number
of required FDL buffers. The ultimate architecture would clearly be one in which all three
contention resolution techniques, time domain, wavelength domain, and space domain,
are used together. Such a network node architecture was experimentally demonstrated
in [Xue+04].
The importance of wavelength conversion in contention resolution is evident. If two
packets on the same wavelength arrive at a switch at the same time on different input
links, and if they are both destined for the same outbound link, there will be a contention
problem. Placing one of the two in an input buffer will resolve the contention, but it can
also be resolved without buffering by changing the wavelength of one of them.
 Optical Packet-Switched Networks 779

Tunable wavelength
converters Buffer with
DMUX
B positions
λ1
λ1 … λN λ1 … λN
1 1
MN x M(B/N + 1)
Space
λN B/N + 1
Switch
λ1
λ1 … λN
λ1 … λN M
M

λN

Figure 10.17 Generic node architecture with TOWCs at the input lines. (From [Danielsen+98, Fig-
ure 6]. Copyright

c 1998 IEEE. Used by permission of The Institute of Electrical and Electronics
Engineers, Inc.)

Wavelength converters can be placed in the input section of a switch and can be com-
pletely or partially shared among the input lines, or they can be completely partitioned
[Callegati+02]. Clearly, depending on the architecture used, the ability of this technique
to offer contention resolution varies, and consequently the network performance will
vary as well. Some examples follow.

10.1.5.1 Switches with Tunable Wavelength Converters

Consider the generic node architecture of Figure 10.17 that employs a demultiplexer and
tunable optical wavelength converter modules (TOWCs), a nonblocking space switch
(a high speed OXC), and output buffering using fiber delay lines [Danielsen+98]. This
is a slotted system (synchronizers and header readers not shown.) The switch size
is M N × M(B/N + 1) where N is the number of wavelengths, M is the number of
node input and output links, and (B/N + 1) is the number of parallel buffering fibers
associated with each output link. The i-th buffering fiber on a link consists of i-l FDLs,
each one capable of holding up to N multiplexed packets in one time slot. Figure 10.18
shows the details of the output buffers on one link. Since the parallel fibers are combined
on the output link, only one packet on each wavelength can be allowed to reach the
output link in a given time slot to avoid contention. This in turn implies that no more
than B buffer positions can be occupied on each link at any given time.
The system operates as follows. Each packet arriving at the switch in a given time
slot, to be routed to a given output, is switched directly to its designated output if it
is available and there are no packets queued ahead of it on that line. Otherwise it is
switched to a free packet position in the buffer on that output line. The availability of the
line or the buffers depends on the number of contending packets. Up to M N arriving
packets could be contending for the same output line, which can accommodate a total
of N packets on distinct wavelengths in each time slot. For each wavelength the buffers
can accommodate B/N packets distributed among any of the B/N time slots. If there
780 Multiwavelength Optical Networks

B/N + 1

Combiner
Output buffers with B positions

Figure 10.18 Details of output buffers for TOWC switch.

are not enough available wavelength and time slot positions, some incoming packets are
lost. An incoming packet is directed to an available wavelength and time slot position
on a given output line by setting its TOWC to the available wavelength and routing the
wavelength-converted packet through the space switch to the available buffer on the
output line.
Traffic modeling and analysis for the case of random traffic in this switch clearly
demonstrates that the presence of tunable wavelength converters reduces the number
of delay lines required in the node by many orders of magnitude [Danielsen+97].
Figure 10.19 shows the packet-loss probability versus the number of delay lines, B/N ,
for a 16 × 16 switch with traffic intensity of 0.8 for each wavelength channel and for
each input link. For example, when eight wavelengths are used, six delay lines are needed
on each output line for a packet-loss probability of less than 10−11 . In the case of no
wavelength conversion, and extrapolating from the existing graph, the estimate is that
approximately 50 delay lines would be required. Similar results are also obtained for
bursty16 rather than random traffic [Danielsen+98].

Without TOWCs
(independent of N)
Packet-Loss Probability

N=4
With TOWCs

N=8

0 2 4 6 8 10 12 14 16
Fiber Delay Lines

Figure 10.19 Packet-loss probability versus number of FDLs with and without wavelength con-
version. (From [Danielsen+98, Figure 9]. Copyright
c 1998 IEEE. Used by permission of The
Institute of Electrical and Electronics Engineers, Inc.)

16
In this study, burstiness was increased over purely random slot occupancy by increasing the probability of
pairs of packet arrivals in successive time slots.
 Optical Packet-Switched Networks 781

DMUX MUX

λ1 λ1
λ1 … λN λ1 … λN
1 1
Space
Switch
λN λN
matrix
λ1 λ1
λ1 … λN λ1 … λN
M M

λN λN
Optical fiber

Tunable Optical Wavelength

Converter (TOWC)

Figure 10.20 Generic node architecture with TOWCs that are shared among input lines. (From
[Eramo+00, Figure 1]. Copyright
c 2000 IEEE. Used by permission of The Institute of Electrical
and Electronics Engineers, Inc.)

10.1.5.2 Shared Wavelength Converters

Because TOWCs are costly devices, a more economical TOWC-based architecture would
use wavelength converter sharing. An architecture similar to the one in Figure 10.17 that
shares the TOWCs among the input channels is shown in Figure 10.20.17 It has M input
and output fibers, each supporting N wavelengths. The switch architecture contains
r shareable TOWCs. On detection of contention between two incoming packets (by
the node controller – not shown in the figure), one of the two packets will have its
wavelength converted while the other will pass through the switch fabric “untouched.”
Packets requiring wavelength conversion will be directed to one of the r shared TOWCs,
and after being converted they will be directed to the designated output line. Figure 10.20
shows an example of this. Two packets on wavelength λ1 arrive at input lines 1 and M,
both destined for output line 1. While the packet on input line 1 is directly switched
to the output line, the packet on input line M is directed to the pool of TOWCs where
its wavelength is converted to λ N before it is also switched to output line 1. Sharing
of the TOWCs is feasible because typically not all the packets contending for the
same output line at any given time will be on the same wavelength. In cases where all

17
The figure shows a bufferless case, but output buffers could be included.
782 Multiwavelength Optical Networks

the TOWCs are utilized and all contentions have not been resolved then one or more
packets may be lost. Of course, if output buffering is included in the system as in
Figure 10.17, the burden of contention resolution is shared between the TOWCs and the
buffers.
Two issues arise in a shared TOWC architecture: (1) determination of the num-
ber of shared TOWCs required to obtain an acceptable packet loss probability and
(2) assignment of TOWCs and wavelengths to the contending packets. A simple control
algorithm was presented in [Eramo+00] to address the latter. The control unit randomly
selects an output fiber among those that require wavelength conversion and allocates
TOWCs and output wavelengths. This procedure is repeated on other output fibers until
there are no more available TOWCs or no more contentions to resolve.
Analytical work in [Eramo+00], together with simulations to verify the results, also
addressed the problem of minimizing the number of TOWCs needed to achieve a speci-
fied performance. Numerical results in [Eramo+00] showed that only a relatively small
number of TOWCs is enough to guarantee low packet-loss probability for both random
and bursty traffic. As an example, for N = M = 16, and with a total offered load on
each input line of 0.8, only 10 converters are required to keep the packet-loss proba-
bility at less than 10−10 as compared with 256 converters required for the generic node
architecture that uses one dedicated converter for each input wavelength and each input
port.

10.1.6 Comparison of Contention Resolution Techniques for Asynchronous

OPS Networks
So far we have examined contention resolution in the isolated context of a single node.
However, it is equally important to examine performance in a network context. A com-
prehensive study of contention resolution techniques employing various combinations
of buffering and wavelength conversion in asynchronous networks was presented in
[Yao+03]. Two cases of buffered systems were investigated with the buffers placed in a
feedback configuration: (a) a single-wavelength delay line (Figure 10.21[a]), where the
delay line can only accept one packet at a time, and (b) a multiwavelength delay line
(Figure 10.21[b]), where the delay line is equipped with a multiplexer and demultiplexer
and can accommodate multiple packets, provided that packets on the same wavelength
do not overlap in time. A third configuration involving wavelength conversion without
buffering was also examined (Figure 10.21[c]), where the input signals are demultiplexed
and routed through a space switch to TOWCs. A controller recognizes a contention con-
dition and selects a path and output wavelength on the desired outbound link. All switch
configurations have add/drop ports for access by external end systems. Finally, a node
architecture for the case of wavelength conversion with multiwavelength buffering was
studied, as shown in Figure 10.21(d). To complete the comparisons, deflection routing
(with no buffering and no wavelength conversion) was also included in the study. After
a contention condition has been recognized, the switching node selects an available out-
bound link, with a free wavelength the same as the incoming packet’s wavelength, that
 Optical Packet-Switched Networks 783

add drop
add drop
(a)

(b)

add drop
add drop

(c) (d)

Figure 10.21 Node architectures for different contention resolution schemes: (a) single-wavelength
delay line, (b) multiwavelength delay line, (c) wavelength conversion, and (d) wavelength conver-
sion with multiwavelength buffering. (From [Yao+03, Figure 3]. Copyright
c 2003 IEEE. Used
by permission of The Institute of Electrical and Electronics Engineers, Inc.)

leads to the second shortest path. If no such port is available at the time of contention,
the packet is dropped.
In the study a wide range of performance results were obtained via simulation for vari-
ous contention resolution schemes using the node architectures of Figure 10.21. These in-
cluded throughput, packet-loss ratio, average end-to-end delay, and average hop distance.
Not surprisingly, based on our earlier discussions, results showed that wavelength conver-
sion is the most beneficial, and the best approach is to provide a unified resolution strat-
egy that combines wavelength conversion with limited buffering and selective deflection
routing.
The possibilities provided by combinations of contention resolution techniques sug-
gest that support of differentiated services can be facilitated by proper use of the available
choices. A study in [Zhang+04], explored how to provide differentiated loss and delay
for QoS in OPS networks exploiting recirculation buffering and (loopless) deflection
routing schemes. For example, by varying the deflection routing method for different
classes of service, differentiated loss and delay were obtained. Results indicated that
combining these two techniques for contention resolution can provide sufficient options
for supporting differentiated service requirements.
784 Multiwavelength Optical Networks

To/from IP router

Buffers
Controller

Controller

In Out
1 1
Optical delay
2 In Out
2 header extraction
1 1
Delay, read
WSXC 2 header, escape 2
WSXC
R R
R R

(a) (b)

Figure 10.22 Switch architecture with electronic buffering and wavelength conversion.

10.1.7 Hybrid Electronic and Optical Buffering

In addition to the three contention resolution techniques presented above, there is a
fourth possibility, which is to use buffering and/or wavelength conversion implemented
in the electronic domain as a “backup” for all-optical switching. As opposed to optical
implementation, electronic implementations of buffering and wavelength conversion
improve signal quality rather than degrading it, because signals are regenerated during
the electronic operations.
Such a hybrid approach can potentially combine the desirable features of buffering and
wavelength conversion (low packet loss) in the electrical domain, with the low latency
and low processing load characteristic of purely optical packet switching while using
only minimal optical processing. The negative effects of electronic packet switching
(high processing load and large and variable delays) can be mitigated by aiming to keep
the bulk of the traffic in the optical domain.
Figure 10.22(a) shows an unslotted OPS node architecture without packet buffer-
ing and wavelength conversion (essentially the same functionality as Figure 10.2).
Figure 10.22(b) shows the switch of Figure 10.22(a) enhanced to provide electronic
buffering and wavelength conversion. Note that the optical delay and header extraction
box is essentially the same, except for one additional feature. Any packet that is not
accepted for transmission is diverted to the electronic domain and buffered rather than
being dropped. Based on control rules to be described subsequently, the electronically
buffered packets, henceforth called “E-packets,” are converted back to optical form and
transmitted on a designated wavelength and output port. Thus the electronic buffering
function carries with it an implicit wavelength conversion and regeneration function. In
the development that follows we assume an asynchronous architecture.
 Optical Packet-Switched Networks 785

The function of the controller has now become more complex: it must not only
schedule packets for transmission through the optical switch and give reconfiguration
commands to the WSXC as before but also schedule transmissions of the electronically
buffered packets. The latter function opens up a wide array of possibilities for perfor-
mance enhancement and added features. For the time being, we focus on the simplest
form of control. We assume that there is a queue of E-packets associated with each output
link; i.e., packets are queued separately according to their destination ports. This is the
equivalent of an output-buffered switch and thus avoids the problem of HOL blocking.
Also transmission of E-packets is assumed to be FCFS with each packet assigned any
free wavelength. Packets scheduled for transmission in the optical domain, henceforth
called O-packets, are given priority over E-packets, and all successive transmissions of
E-packets on a given wavelength are separated by at least one guard time tg as in
Figure 10.6. This ensures that they can be processed all-optically at the next
switch.
The effect of this approach is that under light loads the overall system operates
essentially as an all-optical packet switched system, with a small fraction of packets
being diverted to the electronic domain. In a dense WDM system, it is likely that a
free wavelength (different from the packet’s wavelength) will exist on the designated
outbound link. Thus a diverted packet will immediately be remodulated onto a different
wavelength and transmitted, with not much more delay than if it were transmitted
all-optically. In effect the system is acting as an electronic wavelength-interchanging
switch with electronic signal regeneration as a by-product. As the load increases, higher
contention in the optical domain will cause a larger fraction of the packets to escape to
the electronic domain. Moreover, increased queueing will tend to delay the packets for
variable lengths of time. Eventually, under high congestion some packets may be lost
due to finite buffer capacities. However, with reasonable buffer sizes and proper traffic
engineering, packet loss will be rare, as explained below. Furthermore, the control
flexibility inherent in this hybrid architecture offers a wide range of possibilities for
providing special QoS features for designated traffic classes.

Performance Analysis
In such a hybrid WDM network, end-to-end wavelength continuity is no longer required
for successful transmission of a packet. Thus, the basic performance parameters can be
determined by focusing on the behavior of one link. Then overall end-to-end performance
can be determined by extrapolating from a single link, utilizing suitable independence
assumptions (which are always approximations).
Assuming connectionless optical packet switching (where preserving the packet order
is not an issue), an arriving packet is scheduled for optical transmission if its wavelength
is free on its designated outbound link. Otherwise it is converted to an E-packet and
queued for later transmission. When at least one wavelength is free on an outbound
link the first packet in the queue is transmitted, with its wavelength chosen at random
from among the free wavelengths. (In this queueing system, a packet scheduled for
transmission or in the process of transmission is considered to be in the queue until
transmission is completed.) In scheduling a queued packet for transmission, a guard time
786 Multiwavelength Optical Networks

tg is inserted between the beginning of its transmission and the end of any transmission
ahead of it on the chosen wavelength. (This is equivalent to introducing extra overhead
in the form of an extended header.) There are two reasons for this. First, as in the case of
an all-optical switching path, the guard time ensures that there will be a sufficient time
interval to allow for switch reconfiguration at the next switch on the transmission path.
Second, if a newly arrived optical packet becomes a contender for the chosen output port
and wavelength during the interval tg , the O-packet can preempt the transmission of the
E-packet. If preempted, the E-packet is held in its buffer and another free wavelength is
tried.18
All quantities of interest in this system can be determined from a suitable queueing
model for the link. In keeping with the usual models for packet traffic, we assume that
the packet arrival process is Poisson with rate λ and packet lengths are random and
exponentially distributed with average transmission time µ−1 . This leads to modeling
a single link in a hybrid WDM network as an M|M|W |N queue, where the number of
servers, W , corresponds to the number of wavelengths on the link, and N is the size of
the buffer (the maximum number of packets that can be stored for transmission on one
link, including the packets in service).
Note that if traffic arriving at an R × R switch is uniformly distributed among input
and output ports, with total arrival rate λ on each port, then the arrival rate for each
queue is λ and the traffic intensity per wavelength for each queue is defined as ρ/W ,
where ρ = λ/µ. In the case of infinite buffer size, the queue becomes saturated and
unstable as ρ/W → 1.
Because buffering is electronic in this system it is relatively easy to size the buffers so
that packet loss is negligible. For example, a system using W = 120 wavelengths running
at a traffic intensity per wavelength of ρ/W = 0.8 and equipped with 240 buffers (two
per wavelength) has a packet loss probability of approximately 10−15 .
Because the hybrid system maintains very low packet loss by allowing contending
packets to “escape” to the electronic layer, it is of particular interest to examine the
fraction of packets f e that are diverted to the electronic domain. This fraction is relatively
independent of the wavelength conversion and buffering operations, because when a
packet arrives at a node it is optically switched only if its wavelength is free, and the
fraction a free wavelengths depends almost completely on ρ/W . For example, in a system
with W = 120 equipped with one to two buffers per wavelength, f e is approximately
equal to 1 − ρ/W for ρ/W ≤ 0.9. Given the model used here this is not surprising. This
is just the probability of an arriving packet seeing a busy server in an M|M|1 queue, and
this corresponds to the situation where an arriving optical packet sees its wavelength
occupied. Actually, the queueing model used here overestimates f e for small switch
sizes, because it does not take into account the fact that packets arriving on any one link
never overlap in time and therefore do not contend with each other. Also the priorities
given to O-packets over E-packets is ignored in the model, leading to an additional
overestimate of f e .

18
Although this gives priority to the optical packet, it is not a preemptive priority in the normal sense of the
term, since the preempted packet has not yet started its transmission.
 Optical Packet-Switched Networks 787

The packets diverted to the electronic domain are the ones that would be lost using all-
optical switching without buffering or wavelength conversion. Even though this fraction
may be substantial, its effect on overall network performance can be mitigated if different
traffic streams are designated for various types of special handling depending on their
QoS requirements. For example, real-time loss-tolerant traffic such as voice or video
might be designated for dropping in cases of contention rather than diversion to the
electronic layer.
The previous sections have presented a sampling of OPS network architectures, focus-
ing for the most part on proposed approaches rather than networks that have been realized
in the laboratory or in the field. To make the transition from theoretical approaches to
implementations we briefly discuss enabling technologies in the next section, which sets
the stage for Section 10.3 presenting some of the pioneering OPS testbed implemen-
tations. Most of the recent packet-switching testbed demonstrations are based on OLS
and are described in Section 10.5.3.2.

10.2 OPS Enabling Technologies

This section outlines briefly some of the operations that must be performed in an optical
packet switch, including the enabling technologies that are needed for implementation of
these operations. This is merely an overview, and the reader should consult the articles
referenced in this section, as well as Chapter 4 for additional information on these
technologies.
Referring to the generic OPS node architectures of Figures 10.2 and 10.3, when
packets enter the switch the header must be extracted and read. It can be transmitted
either serially on the same wavelength as the data payload (bit serial approach), with
a guard time between the header and the payload to allow for header extraction and
reinsertion, or multiplexed with the baseband payload, typically on a subcarrier, in
which case the header and payload are transmitted in parallel (see Section 5.2.2).
Because packets arriving at a given input port are separated by idle periods, (at least
of the length of a packet guard time in a slotted system), a packet delineation procedure
is necessary to determine the beginning and end of a packet. Rapid identification of
the beginning of the packet followed by fast clock recovery and bit synchronization
are important in OPS systems to extract and read the header so that the switch can be
reconfigured in time to route the packet to the proper outbound link. This is normally
performed by extracting a portion of the optical signal power, converting it to electrical
form and processing the electrical signal using a high-speed electronic clock recovery
and synchronization circuit.
While some processing operations that only involve packet headers can be relegated
to the electrical domain, in ideal OPS all processing operations should be handled in
the optical domain. There are several reasons for this. Although electronics is a ma-
ture and inexpensive technology at relatively low bit rates, the electronic bottleneck
comes into play at high speeds (tens of gigabits/s): electronics becomes expensive,
consumes an inordinate amount of power, and is difficult to implement. The principal
788 Multiwavelength Optical Networks

functions for which all-optical enabling technology is needed are outlined in the follow-
ing subsections.

10.2.1 Packet Synchronization

In a slotted system, as explained in Section 10.1.1, fixed length packets arrive at the
switch with random delays with respect to the local node’s time slots. This means that
adjustable delays must be inserted at the input ports to align each incoming packet to the
local time slots. This packet alignment process, executed in the synchronization stages
of Figure 10.3, can be implemented with switched FDLs, as shown in Figure 10.4.
Another method of achieving a controllable delay uses a highly dispersive fiber together
with a wavelength converter tuned to translate the packet wavelength to that required to
produce the desired group delay in the fiber.

10.2.2 All-Optical 2R or 3R Regeneration

As explained in Section 4.11, full 3R signal regeneration involves reamplification, re-
shaping, and retiming. These are all straightforward operations in the electronic domain
but all-optical approaches are considerably more difficult. This is an important func-
tion for OPS implementation, as it “cleans up” the signal at nodes along its path, for
further transmission and switching in the network. Recent work on all-optical regener-
ation has focused on semiconductor optical amplifier- (SOA) based approaches using
Mach–Zehnder interferometers (MZIs). Another promising approach is one that uses
the concept of soliton transmission together with in-line synchronous intensity/phase
modulation and optical narrowband filtering [Lavigne+00]. The all-optical approach
is still experimental, requiring more work before it is commercially viable. The most
commonly used methods of wavelength conversion typically provide 2R regeneration as
a by-product, and so tunable wavelength converters are often seen as useful components
of proposed optical packet switch architectures.

10.2.3 Optical Switching

Rapid switching is of great importance for OPS networks, as the switch fabric has to
be reconfigured on a packet-by-packet basis. As the separation between packets can
be extremely short at data rates of 10 s of Gbps, switching speeds in the nanosecond
range are essential. Electro-optic LiNbO3 switches have nanosecond switching times
but they are polarization sensitive, and they have relatively high insertion losses, which
prevent the implementation of large multistage switches. SOA-based switches also offer
nanosecond switching speeds, they can be integrated on a large scale (and thus are
scalable), and they can compensate for power losses due to their inherent amplification
characteristics. However, they are also polarization sensitive and they add noise to the
signal. Another approach to implementation of space switching is through a combination
 Optical Packet-Switched Networks 789

of wavelength converters and wavelength routers (e.g., AWGs). See Section 10.5.3.2 for
some recent examples and Section 4.10.4.4 for a discussion of AWG technology.

10.2.4 Wavelength Conversion

This is another important enabling technology, used both for contention resolution
and in wavelength routing switches. All-optical wavelength conversion is implemented
most commonly in OPS using SOAs with cross-gain or cross-phase modulation (see
Section 4.11). For both techniques wavelength conversion at bit rates up to 100 Gbps is
feasible. Wavelength mixing is another way to achieve all-optical wavelength conversion
(Section 4.11). These techniques are format and bit-rate transparent, but they require
polarization control. Comparisons of various techniques and their applications can be
found in [Elmirghani+00].

10.2.5 Optical Header Processing

A novel approach to all-optical header processing and packet delineation uses an op-
tical correlator for header reading and optical flip-flop memories for header storage
[Dorren+03]. A given header (and trailer) pattern is recognized using a two-pulse
correlation technique implemented in a SLALOM (semiconductor laser amplifier in a
loop optical mirror [Eiselt+95]) device. In this approach the entire packet is passed
through the processor, which recognizes the header and trailer patterns, delineating
the beginning and end of the packet by outputting two sharp correlation light pulses.
These are used to toggle the optical flip-flop memory. The flip-flop consists of a pair
of coupled laser diodes at two different wavelengths. It has two stable states, emitting
a continuous wave signal on wavelength λ1 in state 1 and λ2 in state 2. It is triggered
from one state to another by the output pulse of the header processor. This device
has the advantage of being bit-rate independent so it can operate on packets having
different bit rates in the header and the payload. By biasing the flip-flop asymmetri-
cally, its function can be changed to an optical threshold function (OTF), wherein the
presence of a packet at its input will trigger the system into state 2 irrespective of
the information in the packet header. (An application of the OTF to optical buffering
follows.)

10.2.6 Optical Buffering

As noted above, buffering is a key operation in packet switching, and the easiest way to
do this in the optical domain is with FDLs. However, FDLs are bulky and do not scale
well, so their numbers are limited to a few tens at best. Recirculating loops have been
proposed to reduce the amount of fiber used in FDLs [El-Bawab+02]. This approach
poses a number of problems, however, the most important of these being the need for
optical amplification and the buildup of ASE noise. In [Sakamoto+01] a variable-delay
790 Multiwavelength Optical Networks

λs
Packet 1
Packet 1

Optical Threshold
λs Function

Packet 2 λ1 or λ2 Pass-port
λ1
Wavelength FDL
Converter Packet 2
λs
λ2
Buffer-port
Demultiplexer

Figure 10.23 All-optical buffering and switching architecture. (From [Dorren+03, Figure 13].
Copyright
c 2003 IEEE. Used by permission of The Institute of Electrical and Electronics
Engineers, Inc.)

circuit is proposed based on wavelength converters and arrayed waveguide gratings

(AWGs) as a means of implementing a revolving type of delay.
Another approach to optical buffering and switching has been demonstrated by
combining all-optical header processing with SOA-based wavelength conversion
[Dorren+03, Liu+04] (see Figure 10.23). In the figure, packets arrive simultaneously
at two input ports of a slotted switch, both on wavelength λs . They are destined for
the same output port, so packet 2 (the lower priority packet) must be buffered to avoid
a collision at the output. This is accomplished by converting its wavelength to λ2 and
using a wavelength-routing switch, which then directs packet 2 via a DMUX to an FDL,
delaying it one slot. (Not shown in the figure are the arrangements for directing the
delayed packet to the desired output port and if required, converting its wavelength back
to λs ). The OTF realized by the header processor is now used to provide the logic for
determining whether buffering is needed; i.e., whether a higher priority packet is present
in the time slot. If it is, the OTF emits wavelength λ2 as an input signal to a nonlinear
wavelength converter, which converts packet 2 to that wavelength, after which it is routed
to the FDL buffer. (See Figure 4.71 for an example of a wavelength converter performing
this function.) This general approach has been demonstrated both for buffering and to
implement a 1 × 2 optical switch based on wavelength conversion. In the latter case only
one input is used and no FDL is needed.
This is not a wavelength-multiplexed device, so wavelength conversion is used only
for purposes of internally routing and buffering the packet. However, the OTF and
wavelength converter in Figure 10.23 could be used as an all-optical implementation of
a two-wavelength TOWC in the buffered wavelength-routed switch of Figure 10.17, and
three-wavelength devices have also been demonstrated [Liu+04].
There is another possibility on the horizon for optical buffering technology: slow
light. If the speed of light could be significantly slowed down in media such as fibers
or SOAs, buffering would become more practical. At this writing there have been a
number of attempts to achieve this effect, all of them based on producing a large and
 Optical Packet-Switched Networks 791

controllable decrease in group velocity using various nonlinear effects, including four-
wave mixing in SOAs [Chen+06] and stimulated Raman and Brillouin scattering in
fibers [Herraez+06, Zhu+06]. The basic problem in these attempts is that although the
group delay has been reduced by factors of 70 or more, the effect extends over a very
limited portion of the optical spectrum. This reduces its application to fairly narrow band
and hence wide optical pulses, which in turn implies low bit rates. Thus, the parameter
of importance in evaluating the usefulness of a slow light device for optical buffering
is delay × bit rate, which is a measure of how many bits can be stored in the buffer.
At this writing experimental results have produced typical values of about 0.5 for this
parameter (e.g., a 47-ps delay at 10 Gbps using 2 km of HNLF [Zhu+06]). Clearly,
several orders of magnitude improvement are needed for feasibility in packet buffering.
However, for applications to packet synchronization relatively small and controllable
delays are required, and small delays are sufficient for providing the time necessary for
header processing and switching in an OPS node, suggesting that these might be the first
areas where slow light could be used to advantage.

10.3 OPS Network Testbed Implementations

A number of groundbreaking testbeds for optical packet-switching networks were built

around the turn of the millennium. To provide some perspective on the origins of OPS,
this section discusses briefly a few of the early implementations. We summarize the
technologies developed and some of the important results and observations obtained
from them. This is by no means an exhaustive list. It is merely a sampling of some
of the larger projects so that the reader can gain some insight into the experimental
work surrounding the development and implementation of OPS when it was still just
a gleam in the eyes of some forward-looking engineers and scientists. To complete the
testbed picture, the reader is referred to Section 10.5.3 where some recent OLS network
implementations are discussed.

10.3.1 CORD Testbed

Contention Resolution by Delay Lines (CORD) was a 2.5-Gbps/wavelength WDM
packet-switched network testbed developed by Stanford University, the University of
Massachusetts, and GTE Laboratories [Chlamtac+96]. The goal of the project was
to develop contention resolution optics (CRO) consisting of optical switches and
fiber delay lines, multichannel subcarrier multiplexed signaling to transmit the header,
networkwide distributed slot synchronization techniques, signaling synchronization for
fast data recovery, and ultrafast clock recovery. The project culminated in the construc-
tion of a testbed network featuring all the aforementioned capabilities.
The contention resolution technique that was employed, using switched delay lines,
is the one described in Section 10.1.4 and shown in Figure 10.13. This approach shifts
the contending packets in time and in space to resolve the packet overlap. The actual
792 Multiwavelength Optical Networks

delay line CRO

switch
10/90 Rx
splitter
WDM
demux
Electronic Control Logic
λ1
1
λ22 λ1 λ2
Header
detector
t t

Figure 10.24 Physical implementation of the CRO device. (From [Chlamtac+96, Figure 8]. Copy-
right
c 1996 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers,
Inc.)

physical implementation of the CRO device is shown in Figure 10.24, where two packets
on different wavelengths overlapping in time arrive multiplexed on a fiber and must be
detected by a single receiver (Rx in the figure). This requires delaying one of the packets
to remove the overlap, which can be done by setting the switches appropriately, as
described in Section 10.1.4.
The CORD testbed topology is shown in Figure 10.25. It was essentially a passive star
network with two nodes. Each node had a fixed-wavelength transmitter and one receiver.
Each node transmitted fixed-size packets at 2.5 Gbps and these packets were destined
both for itself and the other node. This way, contention was created in the network,
which was resolved using the CRO device. To route the packets optically, the packet
destination was encoded in the header, the header was modulated onto a subcarrier on
a unique microwave frequency for each node (3 and 3.5 GHz, respectively, for nodes 1
and 2), and the subcarrier was multiplexed with the (baseband) high-speed data payload
[Hofmeister+97]. This was a slotted system, so synchronization was required both at
the packet level and the bit level (clock recovery). For distributed slot synchronization

TX
λ1 λ1 TX
Traffic Traffic
source source
& sink & sink
RX CRO RX Rx
λ1 & λ2 λ1 & λ2
3 dB
coupler
Node Node
Performance Performance
Monitor Monitor

Node 1 Node 2
Equipped with CRO No CRO

Figure 10.25 CORD testbed. (From [Chlamtac+96, Figure 7]. Copyright

c 1996 IEEE. Used by
permission of The Institute of Electrical and Electronics Engineers, Inc.)
 Optical Packet-Switched Networks 793

(DSS), one of the nodes was designated as a “master” node while the others were the
“slave” nodes. All nodes sent a three-bit slot marker (a “ping”) at the beginning of each
time slot on the subcarrier used for header transmission. The slave nodes listened to
both their own ping and the ping received from the master node and attempted to align
them at their receiver. This way, slot synchronization was achieved for all nodes in the
star network. For clock recovery in the header channel a digital oversampling technique
was developed that was able to recover the header channel clock in just four bits while
clock recovery of the data channel on a per packet basis was performed by extracting a
clock tone transmitted along with the payload [Chlamtac+96, Hofmeister+97].
Several optical and optoelectronic components were developed as a part of the CORD
project. These included an integrated version of the CRO device, integrated semiconduc-
tor optical amplifiers, and a high-performance digital optical switch in an X-configuration
for the switched delay line, that was polarization and wavelength independent, and ex-
hibited switching speeds of a few nanoseconds [Chlamtac+96].

10.3.2 KEOPS Testbed

The European Advanced Communications Technology and Service (ACTS) KEys to
Optical Packet Switching (KEOPS) project, a follow-on to the European ATM Optical
Switching project (ATMOS), was initiated in late 1995. Its goal was the analysis and
demonstration of optical bit-rate transparent packet switching within all-optical network
architectures [Gambini+98, Guillemot+98]. The focus was both on network architec-
tures and testbed implementation based on components developed for the project. Fixed-
duration packets were utilized in a slotted system, with both the header and payload using
the same wavelength carrier (the bit-serial approach). (The packet synchronizers for a
slotted system are not shown in the switch architectures described below.) The bit rate
of the header was low: 622 Mbps for easy processing, whereas the data payload was
carried on a user-defined bit rate (up to 10 Gbps). Figure 10.26 shows the packet format
used in the KEOPS project.
Two switch architectures were proposed. The first was a wavelength-routed switch,
where an incoming packet can be switched to a single outbound link (i.e., no multicasting/
broadcasting capability). This architecture is shown in Figure 10.27.19 Switching was
implemented by means of TOWCs, and delay lines were used for contention resolution.
The switch consisted of two stages: contention resolution and space switching. Header
information was read and then erased at the input ports, and a new header was written
in the second stage. In the contention resolution stage, each input port was routed to an
FDL by setting its TOWC to the appropriate wavelength. The routing decision chose
the shortest available delay and tended to minimize head-of-the-line blocking closely
19
The figure represents a system wherein there is no wavelength multiplexing on the input and output
ports. (The TOWCs cannot operate on wavelength-multiplexed packets.) In a WDM network, a three stage
wavelength-space switch structure of the general form of Figure 2.21 could be used with the packet switch
discussed here in the middle stage. This would require n-fold replication of the packet switch for an
n wavelength WDM network. Because packets may have different wavelengths leaving the middle stage
than when they entered, the WMUX stage of Figure 2.21 must be replaced by a passive combiner stage.
An illustration of this configuration for the WASPNET switch is shown in Figure 10.30(b).
794 Multiwavelength Optical Networks

Time slot T: 1.646 µs = 128 bytes (622 Mb/s)

Header Load
Synchronization bits Synchronization bits

guard guard guard

time Routing time
Tags time

Header Load

64.3 ns 180 ns 26 ns 64.3 ns

5 bytes 14 bytes 2 bytes 102 bytes 5 bytes time

Figure 10.26 Packet format for the KEOPS project. (From [Gambini+98, Figure 2]. Copyright
c
1998 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

emulating the performance of an output-queued switch. The second stage of the switch
was used for routing optical packets to the correct output ports using another array of
TOWCs and an arrangement of WDMUXs and WMUXs that acted as a static wavelength
router (see Figure 2.10). A packet arriving at the input to the second stage could be routed
to a given output port by tuning its TOWC to the wavelength corresponding to that port.
Combining wavelength converters and static wavelength routers is a common method
of constructing a high-speed switch fabric for optical packet switches.
The second node architecture shown in Figure 10.28 allowed for selective multicas-
ting and broadcasting of the incoming data packets using WDM in a broadcast star
configuration within the switch. It consisted of three sections: a wavelength encoding
section, a buffering section including multiwavelength gates, and a wavelength selector

First stage for contention resolution Second stage for switching packets

Input 1 TOWC TOWC Output 1

Input N Output N

Figure 10.27 Proposed unicast node architecture for the KEOPS project. (From [Xu+01, Figure 5].
Copyright
c 2001 IEEE. Used by permission of The Institute of Electrical and Electronics
Engineers, Inc.)
 Optical Packet-Switched Networks 795

Wavelength encoding Optical fiber delay-line buffer Wavelength selector

multi- single
wavelength wavelength gates wavelength gates
converter
λ1
λ1
1 1
0xT λM
PD

Electronic Control
PD
λ1
λM
M λM
M
RxT
WDM passive
coupler coupler

Figure 10.28 Proposed multicast/broadcast node architecture for the KEOPS project. (From
[Gambini+98, Figure 9]. Copyright
c 1998 IEEE. Used by permission of The Institute of
Electrical and Electronics Engineers, Inc.)

section. Packet detectors and an electronic control unit for extracting header information,
making buffering decisions, and setting the gates and wavelength switches are shown in
the first section. Every packet that arrived at the switching node passed through a wave-
length converter in the first section and was assigned a wavelength based on its input
port before it was forwarded to the buffer section. All packets were then multiplexed
and broadcast through a passive splitter to an array of FDLs with delays of 0 through
R slot times, so R + 1 copies of each packet were buffered with delays running from
0 to R slots. Optical multiwavelength gates sandwiched between passive splitters and
combiners were then used to select packets belonging to the appropriate time slots and
direct them to the desired output. At this point the packets on different wavelengths were
still multiplexed. Because all wavelengths were gated at the same time by these gates, a
fast wavelength selector was employed in the wavelength selector section to select the
packet with the desired wavelength; i.e., the packet coming from the desired input port.
The same wavelength could potentially be selected at some or all of the outputs in the
same time slot, thereby implementing the multicast function.
The KEOPS project utilized both optical buffering and wavelength conversion tech-
niques to alleviate the contention problem in OPS networks. An architecture of interest
that was developed for this project was the switch with large optical buffers (SLOB),
which cascaded many small switch elements, creating a larger switch with a greater
buffer depth (Figure 10.29) [Hunter+98a]. The SLOB architecture had m inputs and
outputs and the delay line lengths increased from left to right of the structure, so that in
the k-th stage each buffer had a depth of m k−1 packets. Thus, the buffer depth increased
exponentially with the number of stages. SLOB emulated an output-buffered switch
and essentially imitated electronics by implementing deep buffers. Experimental results
validated this architecture, as 40 such stages could be cascaded with the aid of all-optical
regeneration [Chiaroni+98].
The KEOPS project developed a host of network components that were used in the
testbed implementation including
r Polarization-insensitive clamped-gain SOAs based on bulk-tensile InGaAsP/InP struc-
tures with DBR reflectors for application as gates
796 Multiwavelength Optical Networks

1 1
1 m m2
2 2
PSE 2 PSE 2m PSE 2m2 PSE
Inputs 3 3
0 1 2 k unused

m m
(m-1)m0 (m-1)m1 (m-1)m2 (m-1)m
m 1 1

2
outputs

Figure 10.29 SLOB architecture. Each stage is a photonic switch element (PSE). (From
[Hunter+98a, Figure 3]. Copyright
c 1998 IEEE. Used by permission of The Institute of Elec-
trical and Electronics Engineers, Inc.)

r Polarization-independent InGaAsP/InP switch matrices with four 2×2 Mach–Zehnder

electro-optic switches that exhibited 5-dB losses and switching times below 200 ps
r Mach–Zehnder interferometric (MZI) devices based on cross-phase modulation in
SOAs for application as wavelength converters
r 2R all-optical regenerators based on a polarization-insensitive cross-gain modulation
(XGM) wavelength converter using SOAs and an active/passive MZI wavelength
converter
r 3R all-optical regenerators based on a wavelength converter using cross-gain mod-
ulation in an SOA and a wavelength converter using cross-gain modulation in an
SOA-based MZI structure
r Real-time header recovery circuits (GaAs gate array ASIC operating at the header bit
rate of 622 Mbps)

The reader is referred to [Guillemot+98] for additional information on the enabling

technologies developed for the KEOPS project.

10.3.3 WASPNET Testbed

The Wavelength Switched Packet Network (WASPNET) was a collaboration between
three British universities that had as a goal the investigation of packet-based optical
WDM transport networks. A time-slotted architecture was used in WASPNET as in the
KEOPS project. The switch architecture in this project used feedback delay lines along
with tunable wavelength converters and wavelength routers based on arrayed wave-
guide gratings (AWGs) as shown in Figure 10.30 [Hunter+99]. Figure 10.30(a) shows
one plane of the switch, accepting packets on a single wavelength, and Figure 10.30(b)
shows the full implementation of the switch in a WDM network. As in the WDM version
of the KEOPS wavelength-routed switch (Section 10.3.2), there are n planes, one for
 Optical Packet-Switched Networks 797

TOWCs TOWCs
1 1

2 2

N N

N+1 N+1
1 1
1
N+2 N+2
2 2
AWG AWG 2
2N 2N N
N N

(a) One plane of the WASPNET optical packet switch

n n
λ1, …, λn
λ1, …, λn
1
1

1
TOWCs TOWCs
1 1

2 2

N N

n 1 N+1 N+1
1
1
2 N+2 N+2
N
2
AWG 2
N AWG
N 2N 2N N
N Combiner
1

DMUXs
(b) Construction of a multiwavelength switch from multiple planes,
each operating at a single wavelength

Figure 10.30 WASPNET optical packet switch. (From [Hunter+99, Figure 3]. Copyright

c 1999
IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

each wavelength, and wavelength k on each input link is demultiplexed to plane k using
a wavelength DMUX. Although the inputs of each plane have the same wavelength, the
outputs of each plane do not. A passive combiner at the output of the switch is therefore
necessary to combine the outputs of each plane. The switch controller is responsible for
the wavelength assignment and routing decisions to ensure that there is no wavelength
conflict for packets arriving at the combiner from different planes.
As in KEOPS, combinations of TOWCs and wavelength routers (in this case AWGs)
were used to create high-speed switch fabrics, and FDLs were used as buffers for
contention resolution. In the WASPNET switch, the TOWCs on the input ports of the
first AWG were tuned to route the incoming packets either directly through to the
second TOWC/AWG combination, which switched the packets to the desired output
port, or to the TOWC/FDL combination for contention resolution. The feedback delay
798 Multiwavelength Optical Networks

line arrangement produced the effect of shared buffers, which improved the efficiency
of buffer utilization.
WASPNET used what is termed a scattered wavelength path, a procedure where the
wavelengths of the packets were intentionally changed along their end-to-end path. This
alleviated the packet contention problem, which in turn not only reduced the required
buffer capacity but eventually reduced packet-loss probability as well. Analytical studies
have indicated, for example, that for a packet-loss rate of 10−5 and a load of 0.9 this
approach reduced the required buffer size threefold compared to other methods. It was
further shown that packet losses as low as 10−12 were achievable at a load of 0.7 per
wavelength channel and a buffer depth of only 9 per wavelength [Hunter+99]. Another
advantage of the WASPNET switch was that it could support optical packet priorities.
For example, after leaving the delay line, an optical packet may be delayed again because
of preemption by a higher priority optical packet destined at the same time for the same
outbound link on the same wavelength.
As a part of the WASPNET project, a number of enabling technologies were developed,
including wavelength routers based on integrated grating spectrometers and AWGs, and
wavelength converters based on an integrated SOA/DFB laser, a multiwavelength grating
cavity laser, and SOA cross-gain modulation [Hunter+99]. An OPS testbed was also
implemented that examined the cascadability of the node architecture with distances
between nodes varying from 50 to 250 km. The testbed also examined contention
resolution and the physical implications associated with overcoming the contention
conditions.

10.4 Optical Burst Switching

Burst switching was first introduced in [Amstutz83] and [Kulzer+84] for transferring
voice and data over TDM links in traditional networks. Optical burst switching (OBS)
is a scheme that “falls between” circuit and packet switching in optical networks and
was reported initially in [Yoo+98] and [Turner99]. In OBS, packets that have the same
destination and some common attributes (e.g., quality of service) are assembled at a
source node into bursts (by using buffers on a spatial and/or temporal basis) and these
bursts of traffic are transmitted through the network by reserving network resources
only when data (bursts) are transmitted. This in turn improves the network throughput
and provides low latency compared to the circuit-switching approach. When the data
burst reaches the destination node, it is disassembled into the individual data packets
and these packets are forwarded to their final destinations. OBS tries to exploit the fact
that data traffic is typically bursty, unlike voice traffic, which resembles a continuous bit
stream.
The main difference between OPS and OBS is that in OBS the control packet (header)
is sent ahead of the payload to reserve the resources (bandwidth) and set the optical
switches, followed by the actual data after some offset time (Figure 10.31). (In the case
of OPS, the offset time between header and payload is zero.) OPS and OBS are compared
in Figure 10.31. In Figure 10.31(a), data are transmitted through a packet switch from
 Optical Packet-Switched Networks 799

Header recognition,
Payload processing, and generation
Header
A
Synchronizer Setup C
1 Switch
1 1
Incoming
fibers
2 2 2
Fixed length
(but unaligned) New
B headers D
(a)

A C
Control
wavelengths
2 2

OEO
1 1
Control Control packet processing
packets Offset time (setup/bandwidth reservation)
2 Switch 2

Data
wavelengths
1 1

B
Data bursts D
(b)

Figure 10.31 (a) Optical packet switching and (b) Optical burst switching. (After [Qiao00,
Figure 1]. Copyright
c 2000 IEEE. Used by permission of The Institute of Electrical and Elec-
tronics Engineers, Inc.)

fiber A to C and from B to D, using (electronic) header processing and optical payload
switching on a packet-by-packet basis. In the burst switching case of Figure 10.31(b),
the data are first assembled into bursts and transmitted along the same paths as in the
OPS case, but with a single header preceding the burst, transmitted on a separate control
channel, in this case a separate wavelength.
The main advantage of this scheme is that by offsetting the data from the control
packet, the data do not need to wait at intermediate nodes for the control packet to be
processed. Hence, no buffering (typically in FDLs) is necessary at the intermediate nodes
to delay the burst while the network node processes the control packet. This way, the
data burst cuts through the intermediate nodes (switches) without the buffering needed
in optical packet switching. The data need to be buffered only at the source, and this can
be implemented in the electronic domain at the edge of the network, with inexpensive,
high-capacity, flexible electronic memory. The absence of a buffering requirement in
the core network eliminates the need for optical buffering which is expensive and has
limited size and functionality. Figure 10.32 illustrates the basic concept for an optical
800 Multiwavelength Optical Networks

Header cell
Signal
Source
OBS node

Data Signal
data path Destination
Control
channel

Figure 10.32 OBS architecture concept.

burst-switching network. When an end system has a burst of data to send, it selects an
unused channel on the access link for the data burst, and before it transmits the data burst
into the network it first transmits a control packet (denoted in the figure as header cell)
on a control channel. The control packet contains information indicating the channel on
which the data burst is being transmitted, the data burst length, and the destination of the
burst. When the optical burst switch receives the control packet, it selects an outbound
link leading to the destination, reserves the appropriate time interval on an outbound
channel (wavelength), configures its switch for the incoming burst, and forwards the
control packet on the chosen outbound link to the next burst switch in the path. This
process is repeated at every burst switch in the path from source to destination. At the
end of the burst the path is released.
To implement OBS the leading edge of the time interval for an arriving burst must
be reserved at just the right time for it to coincide with the arrival time of the burst,
and this must be done at each node on the end-to-end path. This in turn means that
all the delays encountered in the network, including the propagation delays in the fiber
links and processing and switching delays in the network elements must be known and
accounted for. Thus, a global knowledge of the network is required, which is a downside
of OBS as compared to basic OPS.
A WDM burst-switched network may be transparent (no wavelength conversion) or
opaque (including wavelength converters). In the former case the nodes in Figure 10.32
are wavelength selective (WSXCs), whereas in the latter case they are wavelength
interchanging (WIXCs). The difference is crucial. In the transparent case, wavelength
continuity must be observed in reserving channels for a burst on each link on an end-to-
end path. Because this is done on the fly when the control packet arrives at each switch,
it is likely that a burst may advance part of the way to its destination and then be blocked
at an intermediate node because its wavelength is not available on the next hop. Not
only is the burst lost, but the bandwidth it occupied on the first part of its path has been
wasted. In most proposals for OBS, it is assumed that the network includes wavelength
 Optical Packet-Switched Networks 801

conversion, which means that a burst will not be blocked at an intermediate node unless
all wavelengths on the next hop are occupied. This corresponds to the context in which
burst switching was first introduced in traditional networks.
The effect of wavelength conversion on blocking in OBS can be understood by com-
paring it to the circuit-switched case in WDM networks. Successful transmission of a
burst is roughly the same as successful setup of a wavelength-routed path in a WDM
network. An approximate formula for (circuit-switched) blocking appears in Equa-
tion (6.46). The reader should observe the importance of a large number of wavelengths,
W , in reducing blocking in the opaque case. In the transparent case, the source node has
a choice of wavelength, so blocking at that node is governed by the Erlang-B formula
(Equation [5.91] and Figure 5.32). Beyond the first node, blocking increases rapidly with
path length, because the initial burst wavelength must be maintained on all subsequent
hops. Unless otherwise stated, we shall assume in the remainder of this section that the
networks under discussion provide wavelength conversion.
Routing of the bursts can be implemented either dynamically or based on precomputed
paths. In the former case, the control packet carries the network address of the destination
node, and the burst switch consults a routing database to choose dynamically one of its
outbound links toward that destination. In the latter case, burst transmission is preceded
by a selection of an end-to-end path similar to a virtual circuit in ATM networks (or a
tunnel in MPLS networks). An identifier (similar to the virtual circuit identifier in ATM)
can then be used to forward the burst through the core OBS network.
Compared to optical wavelength routing (circuit switching), OBS achieves better
bandwidth utilization (allows for statistical multiplexing of bursts on each wavelength)
and has a shorter end-to-end latency, as the offset time between the control and data
packets is much shorter than the time required to set up a wavelength-routed connection.
Furthermore, compared to optical packet switching, OBS has the advantage of lower
control overhead, it has less stringent synchronization requirements between bursts and
between a burst and its control packet, and its bursts can cut through the optical switches
without requiring delays for processing the packet headers. Because OBS processes only
one header for a burst of packets, it also relaxes the processing speed required at core
routers by at least one order of magnitude in the case of IP over WDM networks, thus
bringing the forwarding speed of the core routers within the multiple Tbps range.
A number of protocol variations have been proposed for optical burst switching. The
main ones, just enough time (JET) and just in time (JIT) switching, are outlined below.

10.4.1 Just Enough Time Protocol

In the just enough time (JET) approach, the control packet is sent ahead of the payload
by an offset time T that is greater than or equal to the delay encountered by the control
packet at each node times the number of hops from source to destination [Qiao+00]. JET
also uses delayed reservation (DR) where the channel carrying the burst on an output link
at intermediate node j is reserved starting from the time the burst is expected to arrive at
that node (and not from the time the node finished processing the control packet), until
the time the burst departs from the node. This implies that in JET the control packet
802 Multiwavelength Optical Networks

T(i)
S 1 2 D i
time
t’ t t+l
control

T δ (b)

burst δ
1st control packet 1st burst
δ
t’1 t1 t1 + l1
T(i)
case 2 case 1
i
time
time
t’2 t2
(a)
2nd control packet 2nd burst

(c)

Figure 10.33 Just enough time (JET) protocol. (From [Qiao+00, Figure 1]. Copyright

c 2000
SPIE. Used by permission of The International Society for Optical Engineering.)

has to include the burst length and the remaining offset time after the node has finished
processing the control packet. To achieve this, the control packet is stamped with its
arrival and departure time, and the remaining offset time is then the original offset
time minus the time it took for the control packet to be processed. Figure 10.33 shows
an example of the use of an offset time and delayed reservation in a JET scheme. In
Figure 10.33(a) the offset time is T ≥ 3δ where δ is the delay encountered by the control
packet at an intermediate node and the number of hops from source to destination is 3.20
In Figure 10.33(b) the channel on the outbound link of node i is reserved from time t,
the time the burst arrives (and not t ′ , the time the node finishes processing the control
packet), to t + l (for burst length l). Figure 10.33(c) demonstrates an example where the
use of DR reduces burst blocking even in the absence of buffers. In this case, when a
second control packet arrives at the node, if t2 > t1 + l1 or t1 > t2 + l2 , then the node
can reserve the channel for the second burst as well [Qiao+00].
Even though FDLs are not required for buffering in the JET protocol, they can be used
to our advantage to reduce the number of blocked bursts. Furthermore, if the control
delay is large compared to the average burst length, which implies that the JET protocol
can use the FDLs solely for the resolution of conflicts, JET will achieve lower burst
blocking probability than comparable OPS schemes, where some of the buffering must
be used for delaying a packet while the header is being processed.
A by-product of the delayed reservation technique employed in the JET protocol
is that the gaps between scheduled bursts that have not yet arrived can be used for
other bursts. A number of “void-filling” scheduling algorithms have been developed

20
H
In general T ≥ h=1
δ, where H is the total number of hops to the destination.
 Optical Packet-Switched Networks 803

that utilize the gaps between already scheduled data bursts to improve utilization. Al-
gorithms such as latest available unused-channel first with void filling (LAUF-VF) and
others that are presented in [Xiong+00] and [Tancevski+99] are examples of tech-
niques used to improve network utilization when the JET protocol is used in OBS
networks.
Extensions to the JET protocol to support priorities and minimize the probability of
dropping high priority bursts have also been examined in the literature. The priority
JET (pJET) protocol assigns an extra offset time to each high priority class burst and
still keeps the processing of the control messages on a first-come-first-served basis. By
doing this the control messages corresponding to the high priority bursts make channel
reservations considerably earlier than the low priority bursts, thus giving them a lower
blocking probability. Simulations in [Yoo+98] have shown that for an additional offset
time toffset = 3L 0 for the high priority bursts, where L 0 is the average length of the
low priority bursts (assumed to be exponentially distributed), 95% isolation between
the high- and low-priority classes is achieved. This means that the blocking probability
of the high-priority class becomes almost independent of the offered load in the low-
priority class. Furthermore, with this degree of isolation it was shown that the blocking
probability of high-priority traffic is several orders of magnitude lower than that of the
low-priority traffic (even though the overall blocking probability depends on the overall
traffic load and the number of wavelengths used in the network) [Qiao+00]. The reader is
referred to [Yoo+98] for additional simulation results and to [Yoo+00] for an additional
discussion of the pJET approach in the presence of FDLs, which further improves QoS
in optical burst-switching networks.

10.4.2 Just In Time Protocol

The just in time (JIT) protocol is similar to the JET protocol, the main difference
being how soon before a burst arrival and after a burst departure the optical switch is
made available to the incoming burst. The JIT protocol also includes some additional
signaling messages that are used to provide explicit feedback on delivery of data bursts.
Figure 10.34 shows an example of one variation of the JIT protocol. The source node
sends a JIT-SETUP message to the node to which it is attached in the OBS network, and
this node replies with a JIT-CALL-PROCEEDING message that includes the amount
of time the source node should wait before launching the burst into the network. As in
the case of JET, this offset time is calculated as the processing time in an intermediate
node times the number of hops from source to destination. The intermediate node, on
receiving the SETUP message, attempts to reserve a wavelength on the output port
toward the next node on the desired path and forwards the SETUP message to that node.
(The possibilities for available wavelengths depend on whether wavelength conversion
is permissible.)
The burst leaves the source as soon as the time specified by JIT-CALL-PROCEEDING
elapses. If no output wavelengths are available farther down the path at an intermediate
node, a JIT-BLOCKED message is sent to the source node. On the way to the source,
the JIT-BLOCKED message releases the reserved wavelengths on the outbound links
804 Multiwavelength Optical Networks

Source Intermediate Nodes Destination

tp Setup
Call tp
Setup
Proceeding
tc
tp + tb tp
Setup
tc
Start data transmission
tp
tp Release Setup
tc
tp
Release tp

tp Connect
Release tp

Connect tp
tp Release
tp
tp Connect

tp

Figure 10.34 Just in time (JIT) protocol. (From [Wei+00, Figure 4]. Copyright

c 2000 IEEE. Used
by permission of The Institute of Electrical and Electronics Engineers, Inc.)

from the intermediate nodes.21 If there are available resources in the entire path, the JIT-
SETUP message will eventually reach the destination node, and then a JIT-CONNECT
message is sent back to the source node. When the source node finishes transmitting
the data burst, it also sends a JIT-RELEASE message to release the reserved bandwidth
[Wei+99]. The reader is referred to [Baldine+02] for additional variations of the JIT
signaling protocol that differ slightly from the one described above.

10.4.2.1 JIT Signaling Performance

A comparison is made here between the JIT OBS scheme and the traditional circuit- and
packet-switching approaches. The following assumptions and definitions are used:
r The propagation delay is the same for all network links.
r The processing delays of all signaling messages are identical for all network nodes.
r t f : The propagation delay from a source to the node it accesses
r t p : Control message processing time
r tc : Cross-connect cut-through switching
r tl : The propagation delay between two nodes
r tb : Duration of the data burst
21
In this case the burst has already left the source, but because there was no available wavelength at some
intermediate node downstream, the call is blocked.
 Optical Packet-Switched Networks 805

To analyze the performance of the various switching schemes we focus on three

values: the source setup time,22 the channel hold time,23 and the end-to-end latency.24
The source setup time is calculated explicitly as it affects the end-to-end latency, and the
channel hold time provides information on the network’s usage of resources [Wei+00].
r Source setup time for a route of n + 1 hops
– Packet switching: t p
– Circuit switching: 4t f + 2(n − 1)tl + (2n + 3)t p + ntc
– JIT OBS : (n + 2)t p + tδ , where tδ ≥ tc − t p , tδ ≥ 0, is the extra delay required for
switching at the last optical switch
r End-to-end latency for a route of n + 1 hops
– Packet switching: 2t f + (n − 1)tl + (n + 2)t p + ntc + tb
– Circuit switching: 6t f + 3(n − 1)tl + (2n + 3)t p + ntc + tb
– JIT OBS : 2t f + (n − 1)tl + (n + 2)t p + tb + tδ
r Channel hold time for a route of n + 1 hops
– Packet switching: t p + tb + tc
– Circuit switching: 4t f + 2(n − 1)tl + (2n + 3)t p + ntc + tb
– JIT OBS : (n + 2)t p + tb + tδ

From the equations above, we note that JIT OBS has much lower setup time and
channel hold time than circuit-switching, which implies that it provides significant
advantages in terms of network efficiency and throughput, and it also has the lowest
end-to-end latency among all three schemes.
Comparing the JET and JIT protocols, we see that JET utilizes the duration of the burst
to schedule the cross-connect setting at the intermediate nodes, which in turn results
in better utilization than the JIT scheme. However, JET is more complicated than JIT
because of additional synchronization and scheduling required, and it cannot support
conventional circuit switching as is the case for JIT. (The set of signaling protocols in
JIT were designed to be adaptable to circuit switching as well as OBS.)
The JIT signaling protocol software was successfully implemented and tested under
the MONET program. Additional information on that implementation can be found in
[Wei+00].
A variation of the JIT signaling protocol presented in [Duser+02] requires an end-to-
end reservation that satisfies performance criteria such as latency and packet-loss rate
prior to the transmission of the data burst. This technique, which requires a two-way
connection establishment prior to the data being sent, is closer to a circuit-switched
wavelength-routed approach, and it is appropriately termed wavelength-routed optical
burst switching (WR-OBS). It is essentially a fast circuit-switching architecture in which
the optical connections are set up only for the duration of a burst and then are released

22
The source setup time is defined as the time duration from the instant that the data burst appears at the
source node until transmission of the burst begins.
23
The channel hold time is defined as the duration for which an outgoing channel is reserved by an optical
switch for a data burst.
24
The end-to-end latency is defined as the time duration from the instant that the data burst appears at the
source node until it is completely received at the destination node.
806 Multiwavelength Optical Networks

for subsequent requests. Simulation results show that this type of architecture can offer
increased optical connection utilization and wavelength channel reuse only when the
signaling round-trip time is shorter than the edge delay.25 This requirement in turn
implies that this architecture is most suitable to networks with small diameters, such as
metropolitan area networks.

10.4.3 Contention Resolution in OBS Networks

Contention in OBS networks occurs when multiple bursts compete for the same outbound
link in a burst-switching node. The contention problem is much more serious in OBS
than in OPS because of the long and highly variable burst sizes. Wavelength conversion
and deflection routing are two techniques that have already been described and can
minimize the blocking probability in packet switching, and we have already pointed out
the advantage of wavelength conversion in OBS. As these two techniques have been
analyzed in depth earlier they are not repeated here. Instead, the discussion here focuses
on additional schemes that may be useful to further reduce the blocking probability in
OBS networks.
In [Yoo+00] an offset scheme similar to the one described above for pJET is used to
isolate classes of bursts, so that the low-priority traffic does not contend with the high-
priority traffic. They demonstrate the effectiveness of wavelength conversion combined
with the offset scheme and FDL buffering in reducing blocking.
One of the problems that is encountered in OBS networks is that when contention
exists between two bursts, one of the bursts is completely dropped without examining
the amount of overlap between the two bursts (which could potentially be minimal). In
applications that are delay sensitive but loss tolerant, such as real-time voice or video,
it may be worthwhile to lose a few packets that overlap with another burst rather than
losing the entire burst and having to retransmit it at a later time. Burst segmentation, a
technique that was proposed in [Vokkarane+03], addresses this issue by dropping only
the packets that overlap between two contending bursts. To achieve this, it is necessary
to break up the burst into a number of segments, each consisting of a header and a
payload (Figure 10.35). Clearly, there are a number of ways to segment the burst. The
segments, for example, may be of fixed or variable size. Fixed-size segments allow for
easy synchronization at the receiver. However, variable-size segments can accommodate
variable-size packets more efficiently. Furthermore, the length of the segment is also
important. If the segments are long, more information will be lost during contention,
whereas if the segments are short, more overhead will be required (more segment headers
for the burst).
When there is contention between two bursts, only the segments of one burst that
overlap with the other are dropped. A decision has to be made whether to drop the
tail-end segments of the original burst (the burst that arrives at the switch first) or the
head-end segments of the contending burst (the burst that arrives at the switch at a later

25
The edge delay is defined as the time it takes from the instant the first bit of the first packet of the burst
enters the buffer in the source node until the entire burst is released into the network.
 Optical Packet-Switched Networks 807

Seg 1 Seg 2 Seg 3 Seg 4 Seg 5

Segment

Guard Payload Seg Segment Checksum

Bits Type ID Length

Segment Header

Figure 10.35 Segmentation of a burst. (From [Vokkarane+03, Figure 2]. Copyright

c 2003 SPIE.
Used by permission of The International Society for Optical Engineering.)

time). Tail-end segment dropping has the advantage that the dropped (and retransmitted)
segments will most likely reach the destination in-sequence compared to the case where
the head-end segments are dropped. The advantage of head-end segment dropping is
that it ensures that a burst that has reached a switching node with no contention will not
be segmented.
The segmentation approach can be taken a step further by combining segmentation
with deflection routing. To avoid dropping the segments of a burst that are contending
with another burst, these segments can be deflected to another outgoing link. A cap placed
on the burst hop count will avoid deflection routing problems such as looping, multiple
deflections that waste bandwidth, and long final routes that increase the total processing
time thus rendering the initial offset time insufficient. A number of approaches for
implementing segmentation with deflection routing were presented in [Vokkarane+03].
Simulation results compared a number of different policies for contention resolution
that included both segmentation and deflection techniques. The five different policies
that were examined were
1. Drop policy (DP): The entire contending burst is dropped. (Used as an upper bound
for comparison purposes.)
2. Deflect and drop (DDP): Deflect the entire burst to another port. If the port is not
available, drop the burst.
3. Segment and drop (SDP): Segment the original burst and drop its tail.
4. Segment, deflect, and drop (SDDP): Segment the original burst and deflect its tail if
the alternate port is free. If the port is not available, drop the tail.
5. Deflect, segment, and drop (DSDP): Deflect the contending burst to a free port if
available, otherwise segment the original burst and drop its tail.
Figure 10.36, showing the low load case, indicates that policies with segmentation
perform better than the policies with no segmentation and that the policies with deflection
perform better than the policies with no deflection. (At low loads spare capacity is
plentiful, which allows for the successful completion of deflection routing.) Additional
simulation results in [Vokkarane+03] demonstrate that at high loads deflection-routing
808 Multiwavelength Optical Networks

100

10−1
Packet Loss Probability

10−2

10−3

10−4
SDP
SDDP
DP
DDP
DSDP
10−5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Load (in Erlangs)

Figure 10.36 Packet-loss probability versus load for different contention resolution policies in
OBS. (From [Vokkarane+03, Figure 8]. Copyright
c 2003 SPIE. Used by permission of the
International Society for Optical Engineering).

policies do not fare as well in terms of blocking probability, as the network is already
heavily loaded and deflection routing may add to the load, thus increasing the probability
of packet loss.

10.5 Optical Label Switching

In optical label switching (OLS) individual packets are switched through an optical
network as in basic OPS, without signaling ahead of time to the network elements that a
packet is due to arrive and without converting the packet from optical to electrical form.
In addition, a label of local significance is used for packet forwarding in a manner similar
to MPLS packet forwarding described in Chapter 9. This means that OLS inherits the
attributes of MPLS, including the important advantage of scalability. Packets in OLS
(including any higher layer headers such as IP or MPLS) are encapsulated with an optical
label as they enter the network, and routing decisions are made at each intermediate node
based on the label, which is swapped; i.e., erased and rewritten at each node.
Figure 10.37 shows an example of an optical label switching WDM network. The OLS
nodes are assumed to have a bit-rate transparent wavelength conversion capability but no
packet buffering except that needed to allow for label processing and switch reconfigura-
tion delay. In the case shown, the header (label) is transmitted in a bit serial manner with
a guard time separating it from the payload. Typically the header would be transmitted
at a common fixed bit rate for all packets in the network, whereas the payload might run
at a higher speed user-defined bit rate. (This is the approach that was used in KEOPS
[Section 10.3.2].) The label precedes the payload at the OLS switch where, based on
 Optical Packet-Switched Networks 809

Signal OLS node

Source

payload Signal
data path
Destination
label l1
t
t
payload label l2

Figure 10.37 OLS network.

the label information, the signal is translated to an appropriate wavelength and switched
to the desired outbound link. Label swapping also takes place in the network (in Fig-
ure 10.37 label l1 is swapped with label l2 at the OLS node). The switching nodes inspect
the label, erase and rewrite it, and set the state of the optical switch just prior to the arrival
of the packet to an input port of the switch fabric. All of this requires the same hardware as
basic OPS with the addition of the signal-processing devices involved in label swapping.

10.5.1 All-Optical Label Swapping

Although label processing could be implemented in the electronic domain, optical-
processing techniques have now advanced to the point where all-optical label swapping
(AOLS) is feasible. For example, [Way+00] demonstrated a fast optical single-sideband
(OSSB) subcarrier header-erasing and replacement approach for swapping optical la-
bels.26 Normally, subcarrier techniques have limited reach in the presence of dispersion
(see Section 4.3.2.3). However, label swapping using OSSB can be repeated numer-
ous times over a long-distance network without dispersion compensation, because the
subcarrier is reset at each switching node and because of the inherent advantages in
transmission offered by OSSB modulation. Other optical label-swapping techniques
were demonstrated in [Blumenthal+00, Blumenthal+03] and [Jiang+98].
An example of implementation of AOLS from [Blumenthal+03] is shown in
Figure 10.38. The context is the WDM OLS network with bit serial header illustrated
in Figure 10.37. The processor in Figure 10.38 would be preceded by a WDMUX so
WDM packets would be demultiplexed before undergoing processing. A small portion
of the input packet’s power is tapped off to the control plane in the figure, where it is

26
In OSSB a double-sideband microwave subcarrier is electrically multiplexed with the baseband data payload,
and the multiplexed signal modulates an optical carrier. Then one sideband of the subcarrier is filtered out
optically.
810 Multiwavelength Optical Networks

Control
Control processor plane
Input ports Output ports

Line Switch Line

Interface fabric Interface
card card
Data
Plane
Line Line
Interface Interface
card card

Buffer
Control
plane
Scheduling

(a) Switched packet

with new label
Wavelength
Optical
switch
Optical label label
Optical erasure writing
delay
Optical
tap
Input packet
with optical
label

Routing
control
Photodetector and
label recovery
(b)

Figure 10.38 All-optical processor for OLS. (From [Blumenthal+03, Figure 3]. Copyright

c 2003
IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

detected and read (electronically) to extract control information for wavelength conver-
sion, space switching, and optical label writing. (All operations except header reading
and interpretation are done optically.) The bulk of the signal power passes through the
data plane, experiencing a short fiber delay to allow for control processing and switch
reconfiguration, its label is erased and swapped, and its wavelength is converted, with
bit regeneration as a by-product. Regeneration of the payload and rewriting of head-
ers at each node effectively eliminates the accumulation of transmission impairments
along each packet’s end-to-end path. More details of the processing technology appear
in [Blumenthal+03].
The system can be operated in either a slotted or asynchronous mode, and the wave-
length conversion capabilities of the optical processors make it suitable for use in a
variety of switching-node architectures. For example, an array of these processors might
 Optical Packet-Switched Networks 811

be followed by a space switch and WMUXs to route WDM packets to outbound lines
determined by control information in the labels. The wavelength conversion properties
of the processor could additionally be used to reduce losses due to contention on the
outbound links. The processors might also be used as the tunable wavelength convert-
ers in the more elaborate setting of Figure 10.17, which includes packet buffering for
contention resolution.
The enabling technologies for AOLS are the same as those for basic OPS with the
addition of special label-swapping devices, whose technology will depend on the type of
label being used; e.g. bit serial, subcarrier, etc. For examples of the underlying technology
see [Blumenthal+00, Blumenthal+03, Jiang+98, Way+00].

10.5.2 Contention Resolution Techniques

In OLS as in basic OPS, the limited amount of optical buffering in the network constrains
performance, so careful traffic planning and good contention resolution strategies must
be employed to keep packet-loss rates low. Contention resolution in OLS can use all
the techniques outlined in the OPS section: variable FDLs, deflection routing, and
wavelength interchange, in addition to packet preemption based on priorities. In a large
OLS network there may be a choice of paths to a given destination. The switch controller
can decide, based on the information in the header, whether it is acceptable to deflect the
packet to an alternate path. In some cases it may be better to drop the packet than have
it arrive late. As we have seen, wavelength conversion is often a component of OPS and
OLS node architectures in any case, so the use of this technique for contention resolution
comes with little extra cost. Contention resolution via wavelength conversion can be
considered as analogous to path deflection, with the added benefit that the “wavelength
deflected” path follows the same physical fibers as the optimal (original) path.

10.5.3 OLS Network Implementations

10.5.3.1 Early OLS Testbed Demonstrations
The Optical Packet Experimental Routing Architecture (OPERA) project was the first
to demonstrate a WDM network architecture designed to route optical packets us-
ing label swapping, packet routing and forwarding operations, and wavelength reuse
[Carena+98]. OPERA used a subcarrier multiplexed addressing format, and header
recovery was performed using simple microwave filtering techniques, independent of
the payload data rate. When SCM is used, the packet consists of a burst of optical
power that is modulated by the original payload data at baseband and an easily separated
header signal traveling along with the packet but shifted up to an electrical subcarrier
frequency. Figure 10.39 shows an example of a subcarrier transmission system that
can be utilized for such an application. The OPERA experiment demonstrated wave-
length conversion at the packet rate and routing between all-optical nodes. The routing
function was implemented by a combination of a wavelength converter and an AWG
wavelength router. The project experimentally demonstrated combined optical header
replacement and wavelength conversion using a two stage XGM and XPM SOA-based
812 Multiwavelength Optical Networks

Oscillator To Optical
Modulator
Optical Header Combiner

Data Payload (IP Packet)

Figure 10.39 OLS subcarrier transmission system. (From [Meagher+00, Figure 1]. Copyright
c
2000 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

wavelength converter with regeneration of header and payload operating at better than
10−9 BER.
The first IP over WDM optical label-switching testbed featuring computer-to-
computer communication is described in [Chang+00, Meagher+00], including the phys-
ical architecture, the electronic technologies used, and the physical network control in-
terfaces and applications that were incorporated in the early demonstration. The purpose
of the testbed was to implement a first proof-of-concept optical label-switching network
demonstration using four computer-based edge-routers and two optical cross-connects as
the core routers in the backbone and to provide a demonstration of computer-to-computer
communication using an optical label-switching mechanism. Data applications gener-
ated by computer hosts were transported over the network.
A block diagram of the network node used in that testbed appears in Figure 10.40. The
optical label switching was based on an SCM optical header multiplexed on the same
wavelength as the data packet, as illustrated in Figure 10.39.27 In testbed experiments it
was shown that it took less than 175 ns to perform all necessary decision and switching
functions for optical headers modulated at 250 Mbps and multiplexed at a 3-GHz
subcarrier frequency.
In other work performed at the University of California, Santa Barbara, as a part of
a Next Generation Internet (NGI) project, all optical label swapping was implemented
in a testbed using optical labels at 10 Gbps and variable-length packets at 80 Gbps
[Blumenthal+00, Blumenthal+03]. In Section 10.5.1 we briefly described a key element
of that project – the optical signal processor shown in Figure 10.38. Both bit serial
and subcarrier approaches for label swapping were demonstrated, and a wide range of
technology was developed for the all optical signal-processing operations.

27
Optical single-sideband modulation, as described in Section 10.5.1, was used for carrying the SCM label.
 Optical Packet-Switched Networks 813

Delay Line

Control Line

Optical Header Receiver 1

From Sender 1
To Receiver 1
Switch
Data Bus LiNbO3
(Electrical) Node
Controller Switch
Optical Header Receiver 2

From Sender 2 To Receiver 2

Delay Line

Figure 10.40 OLS network node. (From [Meagher+00, Figure 3]. Copyright

c 2000 IEEE. Used
by permission of The Institute of Electrical and Electronics Engineers, Inc.)

10.5.3.2 More Recent OLS Testbed Demonstrations

Most of the recent testbed implementation and experimental work on optical packet
switching focused on the specific case of OLS. Work performed at the University of
California, Davis, and described in [Xue+04] centered primarily on contention reso-
lution techniques. An OLS system was designed and experimentally demonstrated for
a variable-length optical packet-switching system, where unified contention resolution
(contention resolution in the time, space, and wavelength domains) was applied. This
testbed used subcarrier encoded optical labels for switching. The core node architecture
consisted of tunable- and fixed-wavelength converters and a shared pool of FDLs (see
Figure 10.41). The main component of the node architecture was a field programmable
gate array- (FPGA) based optical controller that processes the label and facilitates the
contention resolution schemes. If no contention is observed for a specific time slot, the
controller just forwards the packet to the desired output port. If two or more packets
contend for the same output port, the controller forwards the packets to output ports it
deems appropriate, using switch arbitration units. Contention may be resolved through
wavelength conversion, deflection routing, or delay. An optical time-to-live (OTTL)
function was also successfully demonstrated in this architecture (with the support of
an optical performance monitoring method). It was utilized to drop packets that were
looping aimlessly in the network.
The STOLAS (Switching Technologies for Optically Labeled Signals) project
[Vlachos+03], demonstrated an OLS testbed that concentrated on enabling technologies
for OLS. It utilized frequency-shift keying/intensity modulation (FSK/IM) or differential
phase-shift keying/intensity modulation (DPSK/IM) labeling schemes (that is, a combi-
nation of two orthogonal modulation dimensions) for the label (FSK or DPSK) and the
payload (IM) (Figure 10.42). This approach simplified label swapping and maintained
exact synchronization between the label and the payload without interference between
the payload and the label. Labeling was implemented on a packet burst rather than a
814 Multiwavelength Optical Networks

control
1 1
1 1
W W
. .
. .
. .
1 1

W AWGR W
NxN
1 1

. W W
.
. .
. 1 .
1
K K
W W

F: fiber delay lines

T-WC, tunable wavelength converter

F-WC, fixed wavelength converter

Figure 10.41 Network node architecture for an OLS testbed demonstration. (From [Xue+04, Fig-
ure 2]. Copyright

c 2004 IEEE. Used by permission of The Institute of Electrical and Electronics
Engineers, Inc.)

single packet basis.28 Packets were aggregated to bursts at an optical edge-router that was
used to set the wavelength and label the data. Label swapping was implemented using
an MZI that utilized SOAs in its two branches. The core node architecture consisted
of tunable wavelength converters at the input ports, with AWGs in the switch fabric.
An FDL feedback loop was utilized to buffer the data and for multicasting purposes.
Figure 10.43 shows the details of the optical router and the optical label swapper used in
the testbed. As indicated in the figure, incoming signals are demultiplexed and then pass
through a label swapper unit composed of an MZI that utilizes SOAs in its two branches

IM payload

FSK λ

label

Figure 10.42 FSK/IM orthogonal labeling scheme used in the STOLAS project. (From [Koonen+05,
Figure 2]. Copyright
c 2005 IEEE. Used by permission of The Institute of Electrical and Elec-
tronics Engineers, Inc.)

28
This procedure was termed LOBS (labeled optical burst switching).
 Optical Packet-Switched Networks 815

Electronic Label Processing

Routing table

M[N+2] control signals

Incoming link 1
1 SOA-MZI Variable

DEMUX
λ1 ... λ4 Delay lines
GCSR laser
Outgoing link 1
λ1 ... λ4

combiner
SOA-MZI Variable
Delay lines
GCSR laser
2

Arrayed Waveguide Router

SOA-MZI Variable
GCSR laser Delay lines

.
.
.
Incoming link 2 .
Label 0 λ1 ... λ4 1 SOA-MZI Variable

DEMUX
GCSR laser Delay lines
Outgoing link 2

combiner
label
processing or
2 λ1 ... λ4
SOA-MZI Variable
Label 0 new DPSK
GCSR laser Delay lines
label 2
Ein 1:10 λin
coupler delay SOA-MZI Variable
GCSR laser Delay lines

Label 0 Label 1
λout
SOA-MZI Variable
SOAs GCSR laser Delay lines

tunable phase Eout

new FSK laser
label
modulator
new λ
single InP chip IP router

Fiber splitter for

multicasting

Figure 10.43 Optical router and optical label swapper used in the STOLAS project. (From
[Vlachos+03, Figure 2]. Copyright
c 2003 IEEE. Used by permission of The Institute of Elec-
trical and Electronics Engineers, Inc.)

and a fast tunable laser based on a GCSR (grating-assisted codirectional coupler with
sampled grating reflector). The incoming signal is split (10% of it passes through label
processing while the rest is delayed in a fiber delay loop), and the label (e.g., label 0
in Figure 10.43) is processed. The outcome of label processing determines the new
wavelength that is used and the GCSR laser is set accordingly. The IM payload exits the
delay loop and it is wavelength converted to the new wavelength through cross-phase
modulation (XPM) in the MZI-SOA. The original label is erased in this process and a
new label is reinserted (shown in the figure as label 1). If an FSK label is used, it is
reinserted by modulating the phase current of the GCSR laser. DPSK label reinsertion is
achieved utilizing a phase modulator integrated with the MZI-SOA on a single InP chip.
After the payload is wavelength converted and its label is swapped, it passes through
variable fiber delay lines and an arrayed waveguide grating router before exiting the core
node at an appropriate outgoing link.
Experimental results demonstrated that the FSK/IM method performed better than
the DPSK/IM scheme, and verified the payload/label transmission and label swapping
operations at the 50- and 312-Mbps FSK label rates and 10-Gbps payload data rate over
50 km of dispersion-compensated single-mode fiber. The power penalty for single-hop
transmission and label swapping was shown to be less than 2 dB.
The OPSnet project [Klonidis+05a] developed a complete OLS testbed with edge-
and core-routers, as well as a distributed control plane. (A parallel project, OPORON
[Nejabati+06] was devoted to the development of edge-routers for OPS.) The ingress
816 Multiwavelength Optical Networks

Optical
Packets

header lookup
Control

λ lookup
Layer

Pre-processor
λ1
Packet aggregator
λ2

Network Interface

Optical Interface
Fiber 1
λ1

Fiber 2
Classification Electronic Buffers
IP Routers IP Packets
Payloads

Figure 10.44 Architecture of the edge-router in OPSnet. (From [Klonidis+05b, Figure 1a]. Copy-
right
c 2005 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers,
Inc.)

edge-router aggregates electronic IP packets presented to the network, maps the IP

traffic into a variable-length optical packet payload, assigns the optical label and the
wavelength for the optical packet, and launches it into the network. The core-router
performs wavelength conversion and packet switching (based on AWGs), and the egress
edge-router receives the asynchronous optical packets and extracts the data. The GMPLS
control plane provides information on the optical label and wavelength used for each
packet.
The architecture of the edge node is shown in Figure 10.44. Aggregation of the elec-
trical packets (shown in the figure as arriving from IP routers) to the optical packet
depends on a number of parameters such as QoS, CoS, and destination address. At
the network interface, information is collected on the IP traffic (destination address,
priorities, type of data, etc.) and the packets are classified and buffered according to
this information. The information is also passed on to a controller that decides on the
assembly strategy, the wavelength to be used, and the optical packet header. The aggre-
gate IP packets exiting the electronic buffers pass through an optical interface where
modulation on a specific wavelength occurs and an optical header is attached, thus
creating variable-length optical payloads. The structure of the core node is based on
the WASPNET architecture described in Section 10.3.3 and is shown in Figure 10.45.
The header/data modulation in this testbed is DPSK/IM, as in STOLAS, except that
the header and data are transmitted in series rather than in parallel. Tunable wave-
length converters at the inputs and outputs of an arrayed waveguide grating router,
in conjunction with a control unit that decodes and processes the header, are used to
determine the appropriate output port and wavelength for the incoming data. The ex-
traction of the header in this architecture is performed all-optically, whereas header
processing and the generation of the new header to be inserted is done in the electrical
domain.
 Optical Packet-Switched Networks 817

O->E Control Header Processing

λ lookup λ lookup
Header
decoder

λ1
TOWC 1 TOW C
λ2 Combiners
l1
TOWC AWG TOWC
nxn
λn
TOWC TOWC

λ1
TOWC 2 TOWC
l2 λ2
TOWC AWG TOWC
nxn
λn
TOWC TOWC

λ1
TOWC m TOWC
λ2
TOWC TOWC
lm AWG
nxn
λn
TOWC TOWC

Figure 10.45 Architecture of the core-router in OPSnet. (From [Klonidis+05a, Figure 2]. Copyright

c 2005 IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

Five key technologies were developed for this project

r Header decoding scheme using optical decoding of the DPSK-modulated header bits
with a passive delay interferometer (DI) (see Section 4.7.1)
r Optoelectronic control using fast FPGAs. After the header is processed, table lookup is
used for assignment of the new header and the new wavelength. The signal generated
after processing the header controls a tunable laser determining the wavelength of the
outgoing optical packet.
r Fast tunable laser based on a GCSR that exhibited less than 4-ns laser tuning times
r Wavelength conversion via dual-pump FWM [Morgan+99] utilizing an SOA
r Optical packet source generating variable-length packets with payload data rates be-
tween 10 and 160 Gbps and an optical header at a fixed bit rate

Experiments performed utilizing the OPSnet testbed demonstrated data transmission

and header processing for a 3.125-Gbps label and a 12.5-Gbps data payload in a network
with two edge nodes, one core-router, and GMPLS control [Klonidis+05b]. Additional
818 Multiwavelength Optical Networks

experiments in [Klonidis+05a] focused on the transmission and header processing of

a 3.125-Gbps header with higher payload rates (40-Gbps) through a core node. The
40-Gbps experiments determined that there was a 5-dB power penalty for the signal
after going through the two conversion stages in a core node.
The goal of the testbeds has always been to achieve progressively higher data rates
while keeping performance at acceptable levels. Experiments described in [Wang+05]
and [Wada+01] demonstrate 160-Gbps payload transmission through an OLS core
switch. High-speed all-optical label swapping together with very high speed data trans-
port was demonstrated for the first time in [Wang+05] with payload transmission at 160
Gbps using variable-length packets and 10-Gbps labels. Raman-enhanced fiber cross-
phase modulation (RE-XPM) was used for wavelength conversion. Experimental results
demonstrated that the optical label processing exhibits good BER performance and ex-
tinction ratio after the label is erased and rewritten, and the data payload experiences
only a small power penalty (2 dB) with 2R regeneration.
The experiments described in [Wada+01] utilize OLS based on optical code label
processing where each label is mapped onto an optical code. The optical code in this
case is a sequence of bipolar optical pulses (called optical chip pulses), and optical label
recognition utilizes optical correlation between the optical codes. These techniques
are similar to those used in optical CDMA (see Section 5.2.3). The reader is referred to
[Kitayama+02] for a complete description of optical-code based OLS and to [Wada+01]
and [Wada+04] for experiments on label processing and payload transmission using
optical-code based OLS, for 40 and 160 Gbps payloads, respectively.
An OLS testbed described in [Chang+05a] demonstrated multihop payload trans-
mission with optical label swapping and variable bit-rate payload. A 2.5-Gbps label
and 10- and 40-Gbps payload bit rates are used in this experiment with the label and
payload modulated onto separate optical carriers. The optical packets are transmit-
ted through three nodes and over 200 km with only a small power penalty (less than
1.5 dB).
The enabling technologies include a NOLM (see Section 4.11.2) for all-optical wave-
length conversion, and a dual-arm LiNbO3 modulator (LN-MOD) for generating label/
payload carriers on separate optical frequencies by a process of optical carrier suppres-
sion and separation (OCSS). Two options for optical label swapping were investigated.
In one, OOK-modulated payloads are used and label swapping is achieved using the
OCSS technique for generating carriers for the outgoing label and the wavelength-
converted outgoing packet. In a second, DPSK-modulated payloads are used, and there
is no wavelength conversion [Chang+05b].
The core node shown in Figure 10.46 is configured for the first case. In the figure, the
input channel at wavelength λi is demultiplexed using an AWG, the label and payload
(carried at optical frequencies separated by 2 f 0 ) are extracted by filtering, and the label
is processed electronically to obtain information for label swapping and switch setting.
If the payload is to be forwarded to another node, it is sent to a wavelength converter
and then combined with a new label, carried on an optical frequency at the same 2 f 0
spacing from the payload as was used on the incoming signal. The two new carriers
are generated by the dual-arm LN-MOD driven by a DFB laser at wavelength λ j .
 Optical Packet-Switched Networks 819

f0 Optical filter
λ1 DFB
Modulator

λ2 ωj(λj) ωj + f0
LN-MOD
Wavelength
conversion OC WDM Packets
WDM Packets
Inputs Output
ωj + f0

payload

λi(ωi) ωi + f0 Optical switching

Drop
Add
New label generation
λa Optical Label
filter processing MUX
AWG

Figure 10.46 Core node configuration for label swapping and packet switching. (From [Chang+05a,
Figure 1]. Copyright
c 2005 IEEE. Used by permission of The Institute of Electrical and Elec-
tronics Engineers, Inc.)

The operation of this device in a setting without a wavelength converter is shown in

Figure 10.47.
Figure 10.48 shows the complete testbed with a transmitting node (node 1), a core node
(node 2), and a receiving node (node 3). It shows the path of channel 4, carrying a 40-
Gbps payload, which is combined with seven other channels running at 10 Gbps. Channel
4 is separated from the other channels in the core node using an FBG, converted to a new
wavelength, and combined with a new label. (The other channels are demultiplexed and
multiplexed using AWGs.) A NOLM is used for wavelength conversion, and dual-arm

payload

2f0 λ0
Clock
Modulator 1 2f0
f0 λ0
λ0
Optical filter λ0
CW Optical coupler
Clock f0 label

Dual-arm MOD λ0

Modulator 2

Figure 10.47 OCSS technique. (From [Chang+05b, Figure 1]. Copyright

c 2005 IEEE. Used by
permission of The Institute of Electrical and Electronics Engineers, Inc.)
820 Multiwavelength Optical Networks

Payload
PC AWG
LN-MOD

New
HNL-fiber
label
FBG
DFB
NODE 2 LN-MOD5 25 GHz

DCF Old label

AWG Ch4
Cir
Ch1 Ch1
Ch2
Ch2
Ch3
0 dBm
AWG
Ch5 NODE 2
100 km
SMF
16 dBm Ch8 Ch8

EDFA 16 dBm
100 km
SMF
EAM
Packet envelope EDFA
generator
OC3

OC1 OC2
LN-MOD AWG
40 Gb/s 2.5 Gb/s LN-MOD
LN-MOD 20 GHZ
LN-MOD Ch1 Ch2 Ch8
2.5 Gb/s
EAM LN-MOD10 Gb/s 10 GHz
40 GHz EAM
AWG Interleaver
TOF

25 GHz LN-MOD LN-MOD

10 GHz
Ch4
Ch4 Label Label Payload
WDM source
Ch4 Payload
Ch1, 2, 3, 5, 6, 7, 8
NODE 1 NODE 3

Figure 10.48 Experimental setup for multihop packet transmission and multirate payload. (From
[Chang+05a, Figure 2]. Copyright

c 2005 IEEE. Used by permission of The Institute of Electrical
and Electronics Engineers, Inc.)

LN-MODs are used for generation of label/payload carriers (spaced at 50 GHz from
each other) in the transmitting and core nodes.
Further work using the OCSS technique with wavelength conversion, [Chang+05b]
experimentally demonstrated cascaded label swapping in a network with four nodes
and through a distance of 300 km for 40-Gbps payloads and 2.5-Gbps label bit rates.
That work also demonstrated a novel dynamically reconfigurable folded-path time-
delay buffer for delaying the optical packets (delay ranges from nanoseconds to mil-
liseconds) [Yeo+04]. For more details on the experimental setups and results for the
OCSS and optical label-swapping techniques, the reader is referred to [Chang+05b,
Chang+06].

10.6 Conclusions

Optical packet switching is a networking approach that looks increasingly promis-

ing from both the technology and commercial deployment points of view. With the
 Optical Packet-Switched Networks 821

exception of optical buffering, important advances have been made in the enabling
technology, from fast packet wavelength converters, regenerators, and all-optical signal
processors to high-speed switches and wavelength routers. The technological advances
in all-optical signal processing have made optical label switching, which requires more
elaborate header processing than basic OPS, a possible contender for future networks.
The absence of the optical RAM – the missing link for OPS – means that contention
resolution, which is handled by the “store-and-forward” function (and implemented us-
ing electronic RAM) in traditional packet-switched networks, becomes an overriding
concern in OPS.
The lack of optical buffering is currently circumvented by contention resolution
schemes involving deflection routing, wavelength conversion, use of a very limited
amount of FDL buffering, and procedures such as optical burst switching that operate
well in the absence of optical buffering. On the deployment side, current trends, including
the increasing need for merging the WDM optical layer with the (electronic) IP overlay,
the growth of fiber-based access networks, and the maturing of WDM technology for the
metro area, all suggest that a natural migration will take place from electronic overlays
to the optical layer, with OPS playing an increasing role.
Clearly, many of the enabling technologies for OPS are still in the experimental
stage. Nevertheless, several testbeds that were implemented in recent years have made
significant headway in moving key OPS/OBS/OLS technology toward commercial vi-
ability. Optical packet-switching testbeds began with basic OPS at the end of the 20th
century and continued into the new millennium with emphasis on OBS and OLS. The
importance of these testbeds is that they were able to produce significant results on a
network level using advanced enabling technology, often involving devices that had not
yet emerged from the laboratory. In that respect, they continue to follow the pattern set
by the pioneering WDM network testbeds of the last decade of the 20th century (see
Section 11.2).
As candidates for commercial deployment once the underlying technology is mastered,
basic OPS and OBS are simpler than the third alternative, optical label switching, which
requires more complex processing for label swapping. However, OLS offers better
scalability and more functionality than the other two techniques. Among the variants
of OPS, optical burst switching is in the lead for deployment, as it does not require
optical buffering, placing most of the control and buffering functions at the edge of the
network where electronics can be used to advantage. Thus OBS technology involves a
hybrid of electronic and optical implementation, the former at the edge and the latter
in the core. It therefore offers a natural transition from electronic packet switching in
the logical layer (e.g., IP networks) to purely optical packet switching in the WDM
infrastructure. At this writing a product line already exists in which OBS is used for
Ethernet access to WDM ring networks.29 This is an indication that we are already on
the path from traditional circuit-switched optical networks to optical packet-switched
networks.

29
Matisse Networks Ether BurstTM Optical Switch.
822 Multiwavelength Optical Networks

B C

A D

F E

Figure 10.49 K 3,3 network.

10.7 Problems

1 Explain the differences among optical packet, burst, and label switching. What are
the advantages and disadvantages of each technique?
2 Optical packet-switched networks have been discussed extensively in the last few
years but they have not been practically implemented yet. What are the issues that have
delayed this deployment? Describe possible techniques and algorithms that can counter
these problems.
3 Deflection routing is one of the methods used to resolve contention in optical packet-
switched networks. Describe a loopless deflection routing algorithm implemented in the
literature and propose modifications to that heuristic to achieve better performance.
What are the parameters of interest here?
4 For the K 3,3 network shown in Figure 10.49 and under the assumptions detailed in
Section 10.1.3, calculate PSWPC and PS P for ρ = 0.7 and W = 16.
5 For the K 3,3 network shown in Figure 10.49 and under the assumptions detailed in
Section 10.1.3 derive an expression for blocking probability for the two-hop traffic in
the K 3,3 network using limited deflection routing (LDR).
6 In Figure 10.15 determine an optimal assignment of directional coupler settings for
the next two time slots assuming that there are three buffer stages, packets arriving to
input ports 1 and 2 in the first slot are destined to output ports 3 and 4, respectively, and
packets arriving to input ports 3 and 4 in the first slot are destined to output ports 1 and
2. In the second slot packets arriving at input ports 2, 3, and 4 are all destined for output
port 4, and no packet arrives at port 1. Construct a snapshot of the buffer occupancies
at the beginning of each time slot and determine the fractional packet loss over all time
slots until the switch buffers are emptied.
 Optical Packet-Switched Networks 823

7 Propose a method of modifying the switch structure and/or changing the functioning
of the dump-and-insert switch to operate in an unslotted and variable packet-length
environment. (Assume that the value of the packet length is available in its header.)
8 A hybrid system was discussed in this chapter that allows contending packets to
“escape” to the electronic layer, where they are buffered and later on reinserted in the
optical transport network. Is this a practical technique? What are the advantages and
disadvantages of such an approach?

Bibliography

[Amstutz83] S. Amstutz. Burst switching – An introduction. IEEE Communications Mag., 21(8):

36–42, 1983.
[Arthurs+88] E. Arthurs, M. S. Goodman, H. Kobrinski, and M. P. Vecchi. HYPASS: An opto-
electronic packet switching system. IEEE J. Select. Areas Commun. 6(9):1500–1510, 1988.
[Baldine+02] I. Baldine, G. N. Rouskas, H. G. Perros, and D. Stevenson. JumpStart: A just-in-
time signaling architecture for WDM burst-switched networks. IEEE Communications Mag.,
40(2): 82–89, 2002.
[Baran64] P. Baran. On distributed communication networks. IEEE Trans. Commun., 12(1): 1–9,
1964.
[Blumenthal+00] D. J. Blumenthal, B.-E. Olsson, G. Rossi, T. E. Dimmick, L. Rau, M. Anovi,
O. Lavrova, R. Doshi, O. Jerphagnon, J. E. Bowers, V. Kaman, L. A. Coldren, and J. Bar-
ton. All optical label swapping networks and technologies. IEEE/OSA J. Lightwave Technol.,
18(12):2058–2075, 2000.
[Blumenthal+03] D. J. Blumenthal, J. E. Bowers, L. Rau, H. F. Chou, S. Rangarajan, W. Wang,
and H. N. Poulsen. Optical signal processing for optical packet switching networks. IEEE
Communications Mag., 41(2):23–29, 2003.
[Callegati+02] F. Callegati, G. Corazza, and C. Raffaelli. Exploitation of DWDM for optical
packet switching with quality of service guarantees. IEEE J. Select. Areas Commun., 20(1):190–
201, 2002.
[Carena+98] A. Carena, M. D. Vaughn, R. Gaudino, M. Shell, and D. J. Blumenthal. OPERA:
An optical packet experimental routing architecture with label swapping capability. IEEE/OSA
J. Lightwave Technol., 16(12):2135–2145, 1998.
[Castanon+97] G. Castanon, O. Tonguz, and A. Bononi. Performance of wavelength translation
in all-optical cross-connected networks. In Proceedings of the IEEE/LEOS Summer Topical
Meeting, Montreal, Canada, August 1997.
[Chang+00] G.-K. Chang, G. Ellinas, H. Dai, B. Meagher, W. Xin, T. F. Chen, A. Chowdhury,
M. Z. Iqbal, J. Young, S. J. B. Yoo, Y. J. Chen, C. Lee, and X. Yang. A proof-of-concept, ultra-low
latency optical label switching testbed demonstration for next generation Internet networks. In
Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Baltimore, MD, March 2000.
[Chang+05a] G.-K. Chang and J. Yu. Multirate payload switching using a swappable optical
carrier suppressed label in a packet-switched DWDM optical network. IEEE/OSA J. Lightwave
Technol., 23(1):196–202, 2005.
[Chang+05b] G.-K. Chang, J. Yu, A. Chowdhury, and Y.-K. Yeo. Optical carrier suppression and
separation label-switching techniques. IEEE/OSA J. Lightwave Technol., 23(10):3372–3387,
2005.
824 Multiwavelength Optical Networks

[Chang+06] G.-K. Chang, J. Yu, Y.-K. Yeo, A. Chowdhury, and Z. Jia. Enabling technolo-
gies for next-generation optical packet-switching networks. Proc. IEEE, 94(5):892–910,
2006.
[Chen+06] Z. Chen, B. Pesala, and C. Chang-Hasnain. Experimental demonstration of slow light
via four-wave mixing in semiconductor optical amplifiers. In Proceedings of the IEEE/OSA
Optical Fiber Commun. Conf., Paper OWS1, Anaheim CA, March 2006.
[Chiaroni+98] D. Chiaroni, B. Lavigne, L. Hamon, A. Jourdan, F. Dorgeuille, C. Janz, E. Grard,
M. Renaud, R. Bauknecht, C. Graf, H. P. Schneibel, and H. Melchior. Experimental validation
of an all-optical network based on 160 Gbit/s throughput packet switching nodes. In Pro-
ceedings of the European Conference on Optical Communications, Madrid, Spain, September
1998.
[Chlamtac+96] I. Chlamtac, A. Fumagalli, L. Kazovsky, P. Melman, W. H. Nelson, P. Poggiolini,
M. Cerisola, A. N. M. M. Choudhury, T. K. Fong, R. T. Hofmeister, C.-L. Lu, A. Mekkittikul,
D. J. M. Sabido IX, C.-J. Suh, and E. W. M. Wong. CORD: Contention resolution by delay lines.
IEEE J. Select. Areas Commun., 14(5):1014–1029, 1996.
[Danielsen+97] S. L. Danielsen, B. Mikkelsen, C. Jorgensen, T. Durhuus, and K. E. Stubkjaer.
WDM packet switch architectures and analysis of the influence of tunable wavelength converters
on the performance. IEEE/OSA J. Lightwave Technol., 15(2):219–227, 1997.
[Danielsen+98] S. L. Danielsen, P. B. Hansen, and K. E. Stubkjaer. Wavelength conversion in
optical packet switching. IEEE/OSA J. Lightwave Technol., 16(12):2095–2108, 1998.
[Diao+99] J. Diao and P. Chu. Analysis of partially shared buffering for WDM optical packet
switching. IEEE/OSA J. Lightwave Technol., 17(12):2461–2469, 1999.
[Dorren+03] H. J. S. Dorren, M. T. Hill, Y. Liu, N. Calabretta, A. Srivatsa, F. M. Huijskens,
H. de Waardt, and G. D. Khoe. Optical packet switching and buffering by using all-optical
signal processing methods. IEEE/OSA J. Lightwave Technol., 21(1):2–12, 2003.
[Duser+02] M. Duser and P. Bayvel. Analysis of dynamically wavelength-routed optical burst
switched network architecture. IEEE/OSA J. Lightwave Technol., 20(4):574–585, 2002.
[Eiselt+95] M. Eiselt, W. Pieper, and H. G. Weber. SLALOM: Semiconductor laser amplifier in
a loop mirror. IEEE/OSA J. Lightwave Technol., 13(10):2099–2112, 1995.
[El-Bawab+02] T. El-Bawab and J.-D. Shin. Optical packet switching in core networks: Between
vision and reality. IEEE Communications Mag., 40(9): 60–65, 2002.
[Elmirghani+00] J. Elmirghani and H. Mouftah. All-optical wavelength conversion: Technologies
and applications in DWDM networks. IEEE Communications Mag., 38(3): 86–92, 2000.
[Eramo+00] V. Eramo and M. Listanti. Packet loss in a bufferless optical WDM switch employing
shared tunable wavelength converters. IEEE/OSA J. Lightwave Technol., 18(12):1818–1833,
2000.
[Gambini+98] P. Gambini, M. Renaud, C. Guillemot, F. Callegati, I. Andonovic, B. Bostica,
D. Chiaroni, G. Corazza, S. L. Danielsen, P. Gravey, P. B. Hansen, M. Henry, C. Janz, A. Kloch,
R. Krahenbuhl, C. Raffaelli, M. Schilling, A. Talneau, and L. Zucchelli. Transparent optical
packet switching: Network architecture and demonstrators in the KEOPS Project. IEEE J.
Select. Areas Commun., 16(7):1245–1259, 1998.
[Greenberg+93] A. Greenberg and J. Goodman. Sharp approximate models of deflection routing
in mesh networks. IEEE Trans. Commun., 41(1):210–223, 1993.
[Guillemot+98] C. Guillemot, M. Renaud, P. Gambini, C. Janz, I. Andonovic, R. Bauknecht,
B. Bostica, M. Burzio, F. Callegati, M. Casoni, D. Chiaroni, F. Clerot, S. L. Danielsen,
F. Dorgeuille, A. Dupas, A. Franzen, P. B. Hansen, D. K. Hunter, A. Kloch, R. Krahenbuhl,
B. Lavigne, A. Le Corre, C. Raffaelli, M. Schilling, J.-C. Simon, and L. Zucchelli. Transparent
 Optical Packet-Switched Networks 825

optical packet switching: The European ACTS KEOPS Project approach. IEEE/OSA J. Light-
wave Technol., 16(12):2117–2134, 1998.
[Herraez+06] M. G. Herraez, K.-Y. Song, and L. Thevenaz. Broad-bandwidth Brillouin slow
light in optical fibers. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper
OTuA2, Anaheim, CA, March 2006.
[Hofmeister+97] R. T. Hofmeister, L. G. Kazovsky, C. L. Lu, P. Poggiolini, and F. Yang. CORD:
Optical packet-switched network testbed. Fiber Integra. Opt., 16(2):199–219, 1997.
[Hui90] J. Y. Hui. Switching and Traffic Theory for Integrated Broadband Networks. Norwell,
MA: Kluwer Academic 1990.
[Hunter+98a] D. K. Hunter, W. D. Cornwell, T. H. Gilfedder, A. Franzen, and I. Andonovic.
SLOB: A switch with large optical buffers for packet switching. IEEE/OSA J. Lightwave
Technol., 16(10):1725–1736, 1998.
[Hunter+98b] D. K. Hunter, M. C. Chia, and I. Andonovic. Buffering in optical packet switches.
IEEE/OSA J. Lightwave Technol., 16(12):2081–2094, 1998.
[Hunter+99] D. K. Hunter, M. H. M. Nizam, K. M. Guild, J. D. Bainbridge, M. C. Chia,
A. Tzanakaki, M. F. C. Stephens, R. V. Penty, M. J. O’Mahony, I. Andonovic, and I. H. White.
WASPNET: A wavelength switched packet network. IEEE Communications Mag., 37(3):120–
129, 1999.
[Jiang+98] X. Jiang, X. Chen, and A. Willner. All optical wavelength independent packet header
replacement using a long CW region generated directly from the packet flag. IEEE Photon.
Technol. Lett., 10(11):1638–1640, 1998.
[Jue02] J. Jue. An algorithm for loopless deflection in photonic packet switched networks. In
Proceedings of the IEEE Int’l Conf. Commun. (ICC), pp. 2776–2780, New York, NY, April
2002.
[Kitayama+02] K. Kitayama, K. Onohara, and M. Murata. Capability of optical code-based
MPLS (OC-MPLS). In Proceedings of the 6th International Federation for Information Pro-
cessing (IFIP) Working Conference on Optical Network Design and Modeling (ONDM), Torino,
Italy, February 2002.
[Klonidis+05a] D. Klonidis, C. T. Politi, R. Nejabati, M. J. O’Mahony, and D. Simeonidou.
OPSnet: Design and demonstration of an asynchronous high speed optical packet switch.
IEEE/OSA J. Lightwave Technol., 23(10):2914–2925, 2005.
[Klonidis+05b] D. Klonidis, C. T. Politi, R. Nejabati, D. Simeonidou, and M. O’Mahony. De-
signing a high speed optical packet switched network. In Proceedings of the IEEE International
Conference on Transparent Optical Networks (ICTON), Paper Tu.C1.2, Barcelona, Spain, July
2005.
[Koonen+05] T. Koonen, I. T. Monroy, J. van der Tol, E. Zouganeli, J. Cheyns, G. Morthier,
and J. Jennen. Optical label switched networks – The FP5-IST STOLAS project. In Proceed-
ings of the European Conference on Optical Communications, Glasgow, Scotland, September
2005.
[Kulzer+84] J. Kulzer and W. Montgomery. Statistical switching architectures for future services.
In Proceedings of the International Switching Symposium (ISS), Session 43A, Florence, Italy,
May 1984.
[Lavigne+00] B. Lavigne, O. Leclerc, P. Brindel, D. Chiaroni, B. Dany, P. Guerber, D. Rouvillain,
and A. J. Alcatel. All-optical 3R regeneration. In Proceedings of the Laser and Electro-Optics
Society Annual Meeting, Puerto Rico, November 2000.
[Leonardi+96] E. Leonardi, F. Neri, M. Gerla, and P. Palnati. Congestion control in asynchronous
high-speed wormhole routing networks. IEEE Communications Mag., 34(11): 58–69, 1996.
826 Multiwavelength Optical Networks

[Liu+04] Y. Liu, M. T. Hill, R. Geldenhuys, N. Calabretta, H. de Waardt, G. D. Khoe, and H. J. S.

Dorren. Demonstration of a variable optical delay for a recirculating buffer by using all-optical
signal processing. IEEE Photon. Technol. Lett., 16(7):1748–1750, 2004.
[Maxemchuk89] N. Maxemchuk. Comparison of deflection and store-and-forward techniques in
the Manhattan Street and shuffle-exchange networks. In Proceedings of the IEEE Conference
on Computer Communications (Infocom), Ottawa, Canada, April 1989.
[Meagher+00] B. Meagher, G.-K. Chang, G. Ellinas, Y. M. Lin, W. Xin, T. F. Chen, X. Yang,
A. Chowdhury, J. Young, S. J. Yoo, C. Lee, M. Z. Iqbal, T. Robe, H. Dai, Y. J. Chen, and
W. I. Way. Design and implementation of ultra-low latency optical label switching for packet-
switched WDM networks. IEEE/OSA J. Lightwave Technol., 18(12):1978–1987, 2000.
[Morgan+99] T. Morgan, R. S. Tucker, and J. P. R. Lacey. All-optical wavelength translation over
80 nm at 2.5 Gbps using four wave mixing in a semiconductor optical amplifier. IEEE Photon.
Technol. Lett., 11(8):982–984, 1999.
[Nejabati+06] R. Nejabati, D. Klonidis, G. Zervas, D. Simeonidou, and M. J. O’Mahony. Demon-
stration of a complete and fully functional end-to-end asynchronous optical packet switched
network. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper OWP5, Ana-
heim, CA, March 2006.
[Qiao00] C. Qiao. Labeled optical burst switching for IP-over-WDM integration. IEEE Commu-
nications Mag., 38(9):104–114, 2000.
[Qiao+00] C. Qiao and M . Yoo. Choices, features and issues in optical burst switching. SPIE
Optical Networks Mag., 1(2):36–44, 2000.
[Sakamoto+01] T. Sakamoto, K. Noguchi, R. Sato, A. Okada, Y. Sakai, and M. Matsuoka.
Variable optical delay circuit using wavelength converters. Electron. Lett., 37(7):454–455,
2001.
[Tancevski+99] L. Tancevski, A. Ge, G. Castanon, and L. Tamil. A new scheduling algorithm
for asynchronous variable length IP traffic incorporating void filling. In Proceedings of the
IEEE/OSA Optical Fiber Commun. Conf., Paper ThM7, San Diego, CA, February 1999.
[Turner99] J. Turner. Terabit burst switching. J. High Speed Networks., 8(1):3–16, 1999.
[Vlachos+03] K. G. Vlachos, I. T. Monroy, A. M. J. Koonen, C. Peucheret, and P. Jeppesen.
STOLAS: Switching technologies for optically labeled signals. IEEE Communications Mag.,
41(11):S9–S15, 2003.
[Vokkarane+03] V. Vokkarane and J. Jue. Burst segmentation: An approach for reducing packet
loss in optical burst-switched networks. SPIE Optical Networks Mag., 4(6):81–89, 2003.
[Wada+01] N. Wada, W. Chujo, and K.-I. Kitayama. 1.28 Tb/s (160 Gbps × 8 wavelengths)
throughput variable length packet switching using optical code based label switch. In Proceed-
ings of the European Conference on Optical Communications, vol. 6, pp. 62-63, Amsterdam,
Netherlands, September 2001.
[Wada+04] N. Wada, H. Harai, and F. Kubota. 40Gbit/s, multi-hop optical packet routing using
optical code label processing based packet switch prototype. In Proceedings of the IEEE/OSA
Optical Fiber Commun. Conf., Paper FO7, Los Angeles, CA, February 2004.
[Wang+05] W. Wang, L. Rau, and D. J. Blumenthal. All-optical label switching/swapping of
160 Gbps variable length packets and 10 Gbps labels using a WDM Raman enhanced-XPM
fiber wavelength converter with unicast/multicast operation. In Proceedings of the IEEE/OSA
Optical Fiber Commun. Conf., Paper PDP8, Anaheim, CA, March 2005.
[Way+00] W. I. Way, Y.-M. Lin, and G.-K. Chang. A novel optical label swapping technique using
erasable optical single-sideband subcarrier label. In Proceedings of the IEEE/OSA Optical Fiber
Commun. Conf., Paper WD6, Baltimore, MD, March 2000.
 Optical Packet-Switched Networks 827

[Wei+99] J. Y. Wei, J. L. Pastor, R. S. Ramamurthy, and Y. Tsai. Just-in-time optical burst

switching for multiwavelength networks. In Proceedings of the International Federation for In-
formation Processing (IFIP) Broadband Communications Conference, Hong Kong, November
1999.
[Wei+00] J. Y. Wei and R. McFarland. Just-in-time signaling for WDM optical burst switching
networks. IEEE/OSA J. Lightwave Technol., 18(12):2019–2037, 2000.
[Xiong+00] Y. Xiong, M. Vandenhoute, and H. Cankaya. Control architecture in optical burst
switched WDM networks. IEEE J. Select. Areas Commun., 18(10):1838–1851, 2000.
[Xu+01] L. Xu, H. G. Perros, and G. Rouskas. Techniques for optical packet switching and
optical burst switching. IEEE Communications Mag., 39(1): 136–142, 2001.
[Xue+04] F. Xue, Z. Pan, H. Yang, J. Yang, J. Cao, K. Okamoto, S. Kamei, V. Akella, and S. J. B.
Yoo. Design and experimental demonstration of a variable-length optical packet routing system
with unified contention resolution. IEEE/OSA J. Lightwave Technol., 22(11):2570–2581, 2004.
[Yao+00] S. Yao, B. Mukherjee, and S. Dixit. Advances in photonic packet switching: An
overview. IEEE Communications Mag., 38(2): 84–94, 2000.
[Yao+03] S. Yao, B. Mukherjee, S. J. B. Yoo, and S. Dixit. A unified study of contention-resolution
schemes in optical packet-switched networks. IEEE/OSA J. Lightwave Technol., 21(3):672–683,
2003.
[Yeo+04] Y.-K. Yeo, J. Yu, and G.-K. Chang. A dynamically reconfigurable folded-path time
delay buffer for optical packet switching. IEEE Photon. Technol. Lett., 16(11):2559–2561,
2004.
[Yoo+98] M. Yoo and C. Qiao. A new optical burst switching (OBS) protocol for supporting
quality of service. In Proceedings of the SPIE All-Optical Commun. Syst. Conf., vol. 3531, pp.
396–405, Boston, MA, November 1998.
[Yoo+00] M. Yoo, C. Qiao, and S. Dixit. QoS performance of optical burst switching in IP-over-
WDM networks. IEEE J. Select. Areas Commun. 18(10):2062–2071, 2000.
[Zhang+04] T. Zhang, K. Lu, and J. P. Jue. Differentiated contention resolution for QoS in
photonic packet-switched networks. IEEE/OSA J. Lightwave Technol., 22(11):2523–2535, 2004.
[Zhu+06] Z. Zhu, A. M. C. Dawes, and D. J. Gauthier. 12-GHz-bandwidth SBS slow light in
optical fibers. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper PDP1,
Anaheim, CA, March 2006.
11 Current Trends in Multiwavelength
Optical Networking
Co-authored by Neophytos Antoniades
The City University of New York/College of Staten Island

Despite the fact that optical fiber communications has been an active area of research
since the early 1970s and optical transmission facilities have been widely deployed since
the 1980s, serious activity in optical networking did not reach beyond the laboratory until
the 1990s. It was in the early 1990s that a number of ambitious optical network testbed
projects were initiated in the United States, Europe, and Japan. Although the testbeds
were largely government financed, they planted the seeds for subsequent commercial
developments, many of which were spin-offs of the testbed activities. The commercial
ventures benefited from the knowledge accumulated from the testbeds as well as from
the burgeoning worldwide demand for bandwidth. As a result, multiwavelength optical
networks are deployed today in metropolitan area as well as wide area applications with
increasing current activity in local access as well. In this chapter we give an overview
of current developments in metropolitan and wide area networks. Recent developments
in access networks were discussed in detail in Chapter 5. The chapter begins with a
brief discussion of the role of business drivers and relative costs in creating the current
trends. This is followed by a summary of the early testbed projects in the United States
and Europe, which provides the context for a description of current commercial activity
in multiwavelength metro and long-haul networks. We continue with a discussion of
new applications and services made possible by the unique features of intelligent optical
networks, and conclude with some thoughts for the future.

11.1 Business Drivers and Economics

Traffic growth in the Internet and other new data communications services and deregula-
tion of the telecommunications industry have resulted in new business opportunities and
challenges for telecommunications network operators. Deregulation across the industry
is currently resulting in the destruction of regulatory boundaries separating different
markets, with each carrier trying to provide end-to-end network services under a single
brand name. The increased demand and the competitive pressures of deregulation are
the main driving forces behind the need for low-cost increased bandwidth. WDM is a
proven method of increasing bandwidth by a factor of 30 at 50% of the cost of alternate
methods. These cost advantages are particularly significant in cases in which new fiber
builds are avoided by using WDM equipment. The saturation of fiber capacity, known
as fiber exhaust, is a serious problem for network operators. A study of some major
 Current Trends 829

intercity routes in the United States in 2001 showed that 70% of the lit fiber is in use,
which means that the network operators are reaching a point where increased capacity
is required. The decreasing number of spare fibers available in the cables of the long-
distance network carriers and the local exchange carriers (LECs) exacerbates the fiber
exhaust problem and has brought about a mass deployment of WDM into the network.
In particular, WDM optical networks allow for the following:
r Duct, fiber, and cable exhaust relief
r Reduction in the number of regenerators/optical amplifiers
r Equipment savings with advanced architectures
r Reduced network deployment cost achieved by introducing flexibility and configura-
bility into the optical layer of the network

Several studies have addressed the economics of multiwavelength optical network-

ing [Arijs+00a, Bala+96, Bala+97, Cardwell+00, Chen97, Melle+95, Parys+00,
Sasaki+99]. On the transmission level, not surprisingly, there is a clear advantage
with WDM, which has been shown to be a more cost-effective solution than laying
more fiber in the ground [Melle+95]. By replacing a single optical carrier on a fiber
with 80 or more signals on distinct λ-channels, the inherent bandwidth of the fiber is
utilized much more efficiently, and additional advantages are obtained through the use of
optical rather than electronic amplification on long transmission links (see Section 4.4).
However, WDM networking1 also requires routing (switching) that is implemented by
optical cross-connects of various types. Some have suggested that WDM switching
should be opaque [Bala+95], and others have recommended that WDM routing should
be performed using transparent optical approaches [Coathup+95].
In addition to the obvious capacity expansion advantages of WDM transmission, the
advantages of all-optical approaches going beyond transmission include transparency
to signal formats, upgradeability, and the ability to provide high-bandwidth “clear”
channels directly to the users’ end systems. As studied quantitatively in Section 4.9, a
significant disadvantage of transparency is the accumulation of transmission and switch-
ing impairments, such as noise, dispersion, nonlinear distortion, and switch cross-talk.
Among WDM equipment that is commercially available at this time, there are varying
degrees of transparency. Some examples, in decreasing order of transparency are
r Purely optical cross-connects
r Optical transmission line-terminating equipment performing WDM multiplexing/
demultiplexing functions as well as partial to full (electronic) regeneration (see
Section 2.5)
r Fully regenerative terminal and switching equipment, with regeneration and switching
performed electronically

In this section, we study some evolutionary trends and economic trade-offs for a range
of commercial products, applications, and architectures – from point-to-point connec-
tions to rings to cross-connected mesh networks using optical networking elements for
1
To conform to current usage in this section, we shall use the term WDM network for our generic WRN.
830 Multiwavelength Optical Networks

dynamic wavelength provisioning and protection. The reader should be aware that the
availability of these products as well as their costs are evolving rapidly, so the figures that
we use in our illustrations reflect costs at this writing only. Given the characteristics of
current equipment, the primary advantage of multiwavelength networking (as opposed to
transmission only) is the ability to route and switch signals at the line rate; for example,
STS-48 (2.5 Gbps), STS-192 (10 Gbps), or higher. Demultiplexing down to lower rates
(as is usually done in electronic switches) is costly and often unnecessary.
WDM is currently being widely deployed by telecommunications network operators
across the United States. The deployment has been led by long-distance or interexchange
network operators and the local exchange is following suit. The major deployments
were initially in the form of point-to-point WDM transmission systems, followed more
recently by the implementation of ring and mesh network deployments.
As mentioned previously, WDM is utilized to increase the network capacity and
subsequently solve the problem of capacity exhaust that some long-haul and local
exchange networks are nowadays faced with. The terms capacity exhaust, fiber exhaust,
conduit exhaust, and cable exhaust are used interchangeably in this book. In general
practice, each term has its own significance with regard to the utilization of the fiber
resources within a conduit – the physical structure in the ground carrying fiber cables.
There might be several cables (e.g., four) in a single conduit. Each cable typically
contains several fibers (e.g., 96). The term fiber exhaust implies that a single cable has
run out of fiber capacity. However, there might be room in the conduit for more cables
to be installed. The term conduit exhaust implies that there is no room in the physical
structure to carry more cables. This case has the greatest economic impact. It is both
costly and time-consuming to dig up the ground to build new conduit facilities.
In such a scenario, there are two ways to add capacity to the network: (1) dig up the
ground, install new fiber conduits, and continue to deploy traditional TDM systems (the
“backhoe solution”) and (2) upgrade the remaining fibers to higher capacity systems:
TDM or WDM.
Let us consider the trade-offs associated with these options using some simple case
studies. The first case examined involves upgrading via point-to-point WDM transmis-
sion. Next, a ring architecture upgraded via both TDM and WDM switching is treated,
and the final case studies the use of WDM cross-connects in an arbitrary mesh configu-
ration.
The terminology and symbols used in these examples are chosen to reflect current
usage in the telecommunications network community. It is worthwhile to point out the
correspondence between current usage and the generic terms used throughout this book.
Central Offices (COs) are typically large facilities that contain a variety of switching,
multiplexing, transmission, and end-system equipment. The switching and multiplex-
ing equipment is largely electronic at present, including SONET DCSs, ADMs, and
TDM multiplexers, but this is rapidly changing. These systems fit into our generic cat-
egory of logical switching nodes, with associated optical transceivers (see Section 2.6).
The WDM equipment currently being deployed includes WDM terminals, also called
wavelength terminal multiplexers (WTMs) or WDM transport systems, WADMs, and
optical cross-connects. The WDM terminals correspond to our generic network access
 Current Trends 831

stations, whereas the WADM (OADM or ROADM) and optical cross-connect (OXC)
exemplify generic optical network nodes (see Sections 2.3 and 2.4). If they are limited to
wavelength-selective permutation switching, they fit under our category of WSXCs, the
basic switching nodes in wavelength-routed networks. If they have multipoint switching
capability and/or switch on a waveband basis, they are examples of generic LDCs, which
are the switches in LLNs.
There are a number of studies comparing the cost of upgrading the capacity of a
transmission link either by installing additional fibers and terminating equipment or by
utilizing the WDM option. For example, [Melle+95] showed that for distances longer
than 50 km upgrading to WDM yields the best results in terms of cost. A second study
in [Bala+96] for a telephone company’s LEC network showed that the WDM option
produced significant cost savings (compared to the case of no WDM) for both the cases
of fiber exhaust (savings up to 36%) and conduit exhaust (savings up to 33%). A third
study in [Cardwell+00] also demonstrated that WDM should always be considered in
metropolitan areas under cable or conduit exhaust scenarios, with savings up to 25%
for high-demand areas. Detailed examples of potential savings in point-to-point, ring,
or mesh networks when WDM is introduced are presented in the next section.

11.1.1 Cost Issues for WDM Point-to-Point Systems

Figure 11.1 shows a network with six central offices. Two of the offices are major hub
switching offices. The figure shows a scenario in which the link between the hubs is
running at capacity exhaust, and the remaining links have large spare capacity.
Figure 11.2 shows a WDM point-to-point transmission system deployed between the
hubs to alleviate the capacity exhaust between these nodes. The trade-off between the
backhoe and WDM solution is as follows:
1. Laying new fiber in the ground costs about $60,000 per mile. For a distance of 100
miles between the central offices, the cost for installing new fiber is $6 million.

Large Spare
Capacity
H2 4
ha ty

La apac
s
Ex paci
t

rge ity
us

ile

C
0m
Ca

Sp
10

are

H1 3
cit re
La Cap

pa pa
rg ac

y
Ca e S
e
Sp ity

rg
La
ar

1 2
e

Large Spare
Capacity

Figure 11.1 Six Central Offices, including two hubs, with capacity exhaust.
832 Multiwavelength Optical Networks

l
Te DM
in a
W
rm
H2 4

l
Te DM
ina
W
rm

Capacity Exhaust
Relief
H1 3
16-fold Increase
in Capacity

1 2

Figure 11.2 Application of WDM point-to-point systems to alleviate capacity exhaust.

2. Installing a WDM transmission system that increases the capacity of each fiber
between the hubs by a factor of 16 or more is the other option in this case. The cost of
a WDM terminal with 16 wavelengths can be less than $1 million. The total system
cost (for two terminals in a point-to-point configuration) is less than $2 million.

In this type of capacity exhaust scenario, WDM increases the capacity of each fiber
between the hubs by a factor of 16 or more and is the clear winner between the two
alternatives.
Even in cases in which there is no fiber exhaust, WDM can be a very attractive
solution. This is especially true in the long-haul network scenario. For example, a long-
haul transmission link might be as long as 1000 km, with several electronic regenerators
en route spaced at about 40 km. As opposed to regenerators, optical amplifiers (e.g.,
EDFAs) currently can be spaced as far as 120 km. Furthermore, only one optical amplifier
is required to amplify simultaneously all wavelengths multiplexed on a fiber, as opposed
to one regenerator for each wavelength in the electronic case (together with wavelength
multiplexing and demultiplexing equipment).

11.1.2 Cost Issues for WDM Rings

Figure 11.3 shows a traffic scenario in which two hub nodes terminate an OC-48 (2.5-
Gbps) signal from each central office, as shown on the left side of the figure. There are
four links on the ring with capacity/cable exhaust, each with a length of 100 miles. Three
options for network deployment are considered:

1. Backhoe solution: Dig up the fiber routes and install new fiber. This is a costly option.
At $60,000 per mile for the four sections, this results in an overall cost of $24 million.
2. Upgrade to OC-192 (10-Gbps) TDM: In this case two OC-192 TDM rings are
required, which results in a total of 10 SONET ADMs. This results in a total cost of
$300,000, assuming that each OC-192 ADM costs $30,000.
 Current Trends 833

H1 H2 Large Spare
Capacity
H1 H2
100 miles

Ca haus
ha y
Ex acit

Ex
t
us

pa
1 2 3 4

city
Ca
Traffic:

t
One OC-48 between
Each Central Office
and the Two Hubs 1 4

cit re
pa pa
Ca xha

y
E

Ca e S
pa us
cit t

rg
y

La
2 3
Capacity
Exhaust

Figure 11.3 Six Central Offices, including two hubs, with capacity exhaust.

3. Upgrade to OC-48 2.5-Gbps WDM: In this case six WADMs using a 16-wavelength
system are required, for a total cost of $180,000 (assuming that each WADM costs
$30,000). In this case we assume that the WDM ring is path protected (see Chapter 8).
The WDM ring has 50% spare capacity for future upgrades.

As shown in Figure 11.4, the WDM solution is clearly superior. It is less expensive
on first-installed cost and has 50% spare capacity for upgrading network capacity.

11.1.3 Cost Issues for WDM Cross-Connect Networks

Figure 11.5 shows typical, traditional electronic switching and multiplexing equipment
based on DCSs in a central office using WDM transmission. The central office might
represent a node in a mesh network architecture. In this case, WDM transport systems
terminate the individual wavelengths at the central office. We compare an electronic
switching scenario with an optical cross-connect approach in Figure 11.6.
In the example shown, each transport system carries 16 wavelengths, each of which
terminates on a SONET TDM terminal (labeled TDM MUX in the figure). The inter-
connection between the TDM and WDM terminals is via a short-reach 1.3-µm optical
interface. Each TDM terminal demultiplexes the 2.5-Gbps (OC-48) signals into its con-
stituent DS3 tributaries (48 DS3s in an OC-48). In the traditional mode of operation,
the individual DS3s (running at 45 Mbps) are terminated onto a large digital cross-
connect, which is responsible for grooming, provisioning, and protecting the individual
DS3 channels (see Appendix F). This makes sense if there is a need to access these
individual low-speed bit streams; that is, if all of them terminate at the switching node
at DS3 speeds, and none are terminated at higher speeds or passed through. However,
in the emerging data-centric world, IP routers will be available with direct OC-48c and
OC-192c (concatenated OC-48 and OC-192), as well as DWDM interfaces. It is there-
fore unnecessary and costly to demultiplex these high-speed optical signals into their
low-speed components.
834 Multiwavelength Optical Networks

1. Backhoe Solution:
Very high cost - $24 Million
4 × OC-48
4 × OC-48 H1 H2

2. OC-192 TDM Solution:

0% Spare Capacity
In Network
10 ADMs 1 4
Cost = 10 × $300,000
= $3 Million OC-48
OC-48 OC-48
OC-48

2 OC-48 3 OC-48
OC-48 OC-48

OC-48 OC-48
OC-48 OC-48
3. OC-48 WDM OC-48
OC-48 H1 H2
16 λ Solution:
OC-48
50% Spare Capacity OC-48
In Network
6 WADMs
Cost = 6 × $375,000
= $2.25 Million
1 4

OC-48
OC-48
OC-48 OC-48

OC-48
2 OC-48 3
OC-48 OC-48

Figure 11.4 Economic case for WDM rings.

To compare the optical and electronic switching approaches, let us assume that the
network is dominated by data traffic that terminates on equipment (e.g., IP routers) with
OC-48c interfaces so demultiplexing to DS3 speeds at the cross-connect is unnecessary.
Let the cost of each DS3 termination on the digital cross-connect be $200, and let the cost
of the SONET TDM terminal (unprotected; protection provided at the cross-connect) be
$15,000.
In the electronic switching approach of Figure 11.5, a 2880 × 2880 DCS is required
to switch signals at the DS3 level with 576 DS3 local access ports, assuming that most
of the traffic is pass-through.
Now compare this with the case in which an optical cross-connect is installed at the
node to switch the high-speed optical signals without demultiplexing them to the DS3
level. The connections between the WDM terminals and the OXC are short lengths of
fiber operating at 1.3 µm, as in the case of the electronic switching approach. As shown
 Current Trends 835

16 λ

WDM
Transport

OC-48

TDM TDM
MUX 17 MUX 32

DS3

TDM TDM
MUX 1 MUX 33
2,880 × 2,880
DCS
16 λ WDM WDM 16 λ
Transport OC-48 DS3 DS3 OC-48 Transport

TDM TDM
MUX 16 MUX 48

576 DS3s for Local Access

Figure 11.5 Node in Central Office: Electronic cross-connect.

in Figure 11.6, the local traffic is now dropped at the node in the form of 12 OC-48s,
which are then demultiplexed into 576 DS3s. Replacing the DCS by the optical cross-
connect reduces the number of SONET TDM terminals from 48 to 12 and the size of the
low-speed digital cross-connect from 2880 × 2880 to 1152 × 1152. (The smaller DCS
is still needed for switching local traffic.) This can result in savings of about $800,000,
depending on the cost of the cross-connect. Note that the combination of WDM transport
systems and an optical cross-connect functions as a WIXC. The cross-connect alone is
essentially a space switch without wavelength selectivity, because wavelength selectivity
is handled in the WDM transport systems.

11.1.4 Open versus Closed WDM Installations

As described in this chapter, large numbers of point-to-point WDM links have been
deployed in various types of architectures to increase the capacity of the carriers’ existing
fiber plant without costly installation of new fibers. The capacity of point-to-point WDM
systems is increasing at a rapid pace. Whereas the first systems were deployed with eight
wavelengths per fiber, the number of wavelengths has increased to as many as 160
wavelengths per fiber. The total span of these systems varies from tens of kilometers
(local) to 1000 km (long haul) with optical amplifiers at intermediate locations for
boosting the signal level. Network operators are already expecting to increase the number
836 Multiwavelength Optical Networks

WDM
Transport

OC-48

Optical
Cross-Connect

WDM WDM
Transport Transport

TDM TDM TDM

MUX 1 MUX 11 MUX 12

1152 × 1152
DCS

576 DS3s for Local Access

Figure 11.6 Economic case for WDM optical cross-connect.

of wavelengths per fiber to even higher counts. Advances in optical amplifiers and lasers
suggest that there is still considerable room for growth in the number of wavelengths
carried on a single fiber.
An issue that arises in these large systems is whether the newly deployed equipment
is to be in the form of open or closed architectures. Figure 11.7 illustrates an open
architecture (resembling that of Figure 11.6) using standard short-reach SONET (TDM)
interfaces operating at 1.3 µm. In the figure, an optical cross-connect (OXC) is interfaced
to three point-to-point links terminated by WDM transport systems. These consist of
WMUX/WDMUXs and transponders converting optical signals between 1.3 µm and a
wavelength in the ITU grid. The presence of the transponders makes this an “opaque”
approach. The following are some advantages of this configuration:
r Multivendor interoperability using a standard 1.3-µm interface
r No cascading of physical impairments
r Wavelength conversion
 Current Trends 837

WDM
Transport

Provisioning
and
Restoration
Transponder

OXC

WDM WDM
Transport Transport

1.3-µm
short-reach interface

Figure 11.7 Open WDM network architecture: Opaque network.

In a different configuration, illustrated on a single link in Figure 11.8, some suppliers

are supporting an “integrated” or closed system by providing “compliant” ITU grid
wavelengths directly from their TDM (e.g., SONET terminal) equipment. There is a
debate within the industry on open versus integrated/closed WDM architectures. The
open systems have the advantage that they enable competition among the suppliers on
both TDM and WDM equipment. The open system allows network operators to obtain
competitive pricing for their equipment. For example, in Figure 11.7 two suppliers
compete to supply the TDM and the WDM equipment. The integrated approach generally
results in a single-supplier solution. Although the number of components in the integrated
system is lower (no transponders), the cost of this system can be higher due to the fact
that the network operator is now locked into a single vendor environment. It is difficult
to predict the direction that future network deployment will take. At this writing the
trend is in favor of open systems.

Compliant Signals Directly

from TDM Equipment
SONET SONET
Terminal Terminal

SONET MUX/ MUX/ SONET

Terminal DMUX DMUX Terminal

SONET SONET
Terminal Terminal

Figure 11.8 Integrated closed WDM network architecture.

838 Multiwavelength Optical Networks

11.2 Multiwavelength Optical Network Testbeds

A picture of current trends in optical networking would not be complete without

some historical perspective. We supply that now with a description of some of the
early wide area WDM network testbeds. These include the Optical Networks Technol-
ogy Consortium (ONTC) [Chang+96], Multiwavelength Transport Network (MWTN)
[Hill+93, Johansson96], RAINBOW [Hall+96], All Optical Network (AON) Consor-
tium [Alexander+93, Kaminow+96], MONET [Wagner+96], National Transparent Op-
tical Network Consortium (NTONC) [Wilt97], Optical Pan-European Network (OPEN)
[Jourdan+97], and PROMETEO [Merli+97]. We give brief overviews of a representa-
tive subset of these below. Other more recent testbeds that are focused on optical control
issues and optical packet-switching technologies are presented in Chapters 9 and 10,
respectively.

11.2.1 Optical Networks Technology Consortium

The mission of the Optical Networks Technology Consortium (ONTC) project was to
investigate and to demonstrate WDM networking in a testbed environment. Formed
in 1992 under partial sponsorship of DARPA in the United States, the consortium
members (including Bellcore, Case Western Reserve University, Columbia University,
Hughes Research Laboratories, Northern Telecom, Bell Northern Research, Lawrence
Livermore National Laboratory, Rockwell Science Center, United Technologies Pho-
tonics and United Technologies Research Center) designed and built a demonstration
testbed consisting of four network access nodes, distributed between two WDM optical
fiber rings joined by a 2 × 2 WDM cross-connect. Each ring had two WADMs and each
WADM had an access node attached to it. The WADM and the WDM cross-connects
were implemented using a hybrid optomechanical switch approach and an acousto-optic
tunable filter (AOTF) approach. The architectural approach was unique in that it used
a combination of ATM and WDM networking to provide a data communications net-
work for high-speed services. The fundamental architecture of the ONTC network was
layered, separating the optical (WDM) layer from the logical-switching (ATM) layer.
A number of different technologies were used and evaluated in the testbed, including
erbium-doped fiber amplifiers (EDFAs), multiwavelength laser arrays, multichannel in-
tegrated receivers, optical cross-connect switches using acousto-optic tunable filters,
and hybrid cross-connects utilizing a three-stage architecture of the type illustrated in
Figure 2.21, with multilayer dielectric thin-film interference filters for WMUX/
WDMUXs and discrete 2 × 2 optomechanical switching elements in the middle stage.
Signaling in this network was implemented using in-band isochronous ATM data cells.
Out-of-band optical layer signaling (at 1310 nm) and electrical overlay signaling were
also tested in the network. The work culminated in a successful demonstration of a
four-wavelength WDM network at the Optical Fiber Communications Conference in
1995 [Chang+96]. This program was later extended to an eight-wavelength network
(the NTONC project, discussed in Section 11.2.5).
 Current Trends 839

11.2.2 All-Optical Network Consortium

The AON Consortium was formed with DARPA funding in 1993 and included AT&T,
Bell Laboratories, Digital Equipment Corporations, and the Massachusetts Institute of
Technology. Its objective was to develop architectures and technologies for high-speed
WDM and TDM networks. A testbed was set up in the Boston metropolitan area with a
20-channel WDM system running at a data rate per wavelength ranging from 10 Mbps
to 10 Gbps. The network was all-optical in that the transmitted data within the network
did not undergo any optoelectronic and electro-optic conversions. Several technologies
were demonstrated by the AON Consortium, such as silica-on-silicon passive devices,2
rapidly tunable distributed Bragg reflector (DBR) lasers, tunable receivers based on fiber
Fabry–Perot filters, and various types of frequency changers. Control and management
of the network and of individual devices in an optical terminal was available through a
graphic user interface (GUI) and a command-level scripting interface, providing detailed
telemetry and state information.

11.2.3 European Multiwavelength Optical Network Trials

Several testbed efforts were mounted in Europe as a part of the RACE and the ACTS pro-
grams set up by the European Community Commission. The Multiwavelength Transport
Network (MWTN) project was one of the first testbeds to demonstrate successfully the
concepts of wavelength routing in optical networks [Hill+93, Johansson96]. The RACE–
MWTN consortium was set up to develop and demonstrate the concept of WDM optical
networking. Two network elements, the optical cross-connect and the optical ADM,
were the building blocks of the network architecture. Several network nodes were de-
veloped based on photonic technologies, involving space switches, tunable filters and
lasers, and wavelength-flattened EDFAs. The optical ADM nodes were presented at the
European Conference on Optical Communications in 1995 in a network demonstration.
The final demonstration of the MWTN was carried out in a real network environment
in the Stockholm area. The fiber network formed a part of the local operator’s network
of about 130 km of standard fiber. The management system of the MWTN demonstra-
tor was developed in accordance with the Telecommunications Management Network
(TMN) standard. This provided the necessary management functions and offered com-
munication between an operations system and different network elements. A logically
separate data communications transport network supported data transfer for network
management, and a local management console provided a GUI for manipulating the
managed objects.
Other European trials included the Optical Pan-European Network that demonstrated
the operation of a wavelength cross-connect mesh network in a field environment
spanning several countries, including Norway, Denmark, France, and Belgium, the

2
An efficient silica-on-silicon multiplexer based on an arrayed waveguide grating, a demultiplexer based on
an optical phased array dispersive element, and a waveguide grating router were fabricated and demonstrated
by the consortium.
840 Multiwavelength Optical Networks

Pan-European Photonic Transport Overlay Network, another field testbed with WDM
using 16 wavelengths over 500 km at 10 Gbps and also spanning several countries,
including Germany and Austria, and the Management of Photonic Systems and Net-
works (MEPHISTO) project that applied the principles of network management (TMN)
to advanced optical networks that utilize WDM and optical switching.

11.2.4 Multiwavelength Optical Network

The Multiwavelength Optical Network (MONET) project was initiated in 1995 as an
industrywide consortium of companies and government agencies (including NASA,
DARPA, NSA, NRL, DIA, DISA, Bellcore, Lucent Technologies, AT&T, Bell Atlantic,
BellSouth, Pacific Telesis, and Southwestern Bell) collaborating to deploy a multiwave-
length optical network in a field trial. The program was established to define, demonstrate,
and help form an industry consensus on how best to achieve multiwavelength optical net-
working on a national scale that serves both commercial and government applications.
MONET concentrated on the local exchange and long-distance network and was built
on a hierarchical network architecture to achieve scalability to national size. MONET
developed and combined a wide array of network elements into a scalable network archi-
tecture, including wavelength add/drop multiplexers, WDM cross-connects, wavelength
routing stars, four- and eight-wavelength 2.5-Gbps InP laser arrays, InP photodetector/
preamplifier arrays, wavelength-selectable laser sources, LiNbO3 polarization-
independent switch arrays, InP switch arrays, wavelength interchangers, gain-flattened
EDFAs, and silica-on-silicon multiplexers and routers. A field trial was conducted in the
Washington, DC area in 1999, using two interconnected eight-wavelength WDM rings
that were provided by two different vendors with full interoperability at their intercon-
nection points. The diameter of each ring was somewhat less than 100 km and each ring
executed a UPPR protocol for protection switching.

11.2.5 National Transparent Optical Networks Consortium

This project utilized WDM technology that evolved from the devices developed previ-
ously by ONTC to produce a switched multiwavelength MAN using an existing fiber
infrastructure in the San Francisco Bay area. The 10+ Gbps, 400-km network demon-
strated high-bandwidth applications and tested emerging optical technologies–hardware,
software, and protocols. Various services were carried, including SONET OC-48, sub-
carrier modulated analog (FM) video, and digital (BPSK) data.

11.2.6 The Importance of the Testbeds in Driving the Telecommunications Infrastructure

The optical network testbeds deployed in the last few years of the 20th century pro-
vided worthwhile field trial experience for the carriers, the equipment manufacturers,
and some users of bandwidth-intensive applications. These trials were invaluable in
moving ideas from the laboratory toward viable commercial products and network
 Current Trends 841

architectures. The first operational optical networks of the 21st century were built
around these products and architectures. They enabled a host of new services that
dramatically lowered costs in the telecommunications infrastructure. This has in turn
accelerated demand for more bandwidth, which should stimulate further development
of next-generation networks.
More recent networking developments that build on the early testbed experience are
discussed next: metropolitan area networks and long-haul network deployments utilizing
intelligent optical components (in Sections 11.3 and 11.4, respectively).

11.3 Metropolitan Area Networks

11.3.1 Metro Network Unique Characteristics

Since the adoption of the SONET/SDH standard, the architectures of optical metropolitan
area networks (MANs) have constantly been evolving. Scaling such network architec-
tures has provided a particular deployment challenge for incumbent local exchange
carriers (ILECs) and competitive local exchange carriers (CLECs) alike. The deploy-
ment costs of a typical metro networking infrastructure have two significant components:
the costs associated with fiber deployment and the costs associated with the electrical
and optical components, including customer-premise equipment and customer-located
equipment that enable services across the metro fiber infrastructure. As fiber deploy-
ments reach out further to lower density applications and closer to the customer, the
above costs are shared among a significantly smaller revenue base compared to, for
example, networks in the long-haul domain. Clearly, new metropolitan area network-
ing equipment must offer significantly increased functionality and performance at a
lower cost. In addition, metropolitan area networks are more sensitive to space and
power consumption of the network equipment involved as ILECs and CLECs intro-
duce WDM and try to pack more channels at higher bit rates and closer in wavelength
spacing to respond to the increasing bandwidth demands. Moreover, the presence of
multiple customers having diverse bandwidth needs creates traffic requirements that
are constantly changing with a surprising 80% year-over-year growth [Elby03] so it
is of utmost importance to have the right technology to deal with possible network
reconfigurations and varying network loads. Finally, another significant challenge fac-
ing today’s carriers is the optimum balancing between deploying new-build (“Green-
field”) networks and using existing legacy infrastructure that is SONET/SDH based
[Wright+04].
Introduction of all-optical networking in metro was suggested by the early testbed
efforts in the field described in Section 11.2 mainly as a solution that provided complete
flexibility to bit rate, modulation format, and wavelength spacing due to the elimi-
nation of the optical-to-electrical-to-optical (OEO) conversions that were the basics
of the legacy SONET/SDH backbone infrastructure in metro [Brackett96]. As opti-
cal networking technology matured over the past few years it became evident that
842 Multiwavelength Optical Networks

it has the potential of providing a strong economic alternative to the costly O-E-O
scenario especially in the cost-sensitive metropolitan networking arena. The savings are
basically due to the elimination or at least the extensive reduction of expensive O-E-O
regenerators [Saleh00, Saleh02]. The introduction of optical reconfigurable devices like
wavelength add/drop multiplexers (WADMs) to address the unique changing traffic
requirements in metro provided an additional factor in the cost-versus-performance
equation that moved the balance toward all-optical WDM networking in the metro en-
vironment as presented in [Saleh03]. WDM rings used in metro networks can also have
architectural configurations similar to TDM SONET rings, offering enhanced surviv-
ability at the optical layer by reserving bandwidth (wavelengths and fiber) for protection
purposes.

11.3.2 Defining the Metropolitan Networking Domain

Traditional architectures in metro consist of interconnected fiber SONET/SDH
rings as shown in Figure 11.9 [Ghani+02]. For example, LECs have installed

ADM ADM

ADM HUB
HUB DCS
SONET/SDH protected IOF ring SONET/SDH protected IOF ring
(example 80–200 km, OC-48/192) (example 80–200 km, OC-48/192)
ADM
DCS
ADM
ADM ADM
DCS
Backbone interconnection
(long-haul, regional etc.)
ADM

ADM ADM

SONET/SDH protected Edge ring

(example 10–50 km, OC-12/48)

ADM ADM

DS-n/OC-n
Electronically aggregated services

Figure 11.9 Current legacy SONET/SDH design in U.S. metropolitan regions. (After [Ghani+02,
Figure 8.2]. Copyright

c 2002 Academic Press. Used with permission of Elsevier.)
 Current Trends 843

both two-fiber unidirectional path-switched rings (UPSRs) and two-fiber bi-

directional line-switched rings (BLSRs), as described in Chapter 8. On a physical
level, SONET/SDH is a time-division multiplexing (TDM) frame format that offers
bit-rate multiplexing capabilities in DS-n or OC-n hierarchical fashion [Goralski02]. In
Figure 11.9 three levels of a SONET/SDH ring hierarchy are presented.

r Level 1: The edge rings span distances from a few kilometers to a few tens of kilometers
and are composed of add/drop multiplexers (ADMs) that electronically aggregate
traffic onto the fiber. Connectivity on these rings is mostly hubbed to a Central Office
location (CO) denoted by a combination of ADMs and digital cross-connect switches
(DCSs) in the figure.
r Level 2: DCSs interconnect SONET/SDH edge rings to the larger interoffice (IOF)
rings that span possibly tens to hundreds of kilometers.
r Level 3: IOF rings can also be interconnected using DCSs forming mesh connectivity
optical networks or larger SONET/SDH rings as shown in the figure.

Capacity expansion of legacy metro networks like the one in Figure 11.9 can be
accomplished in three ways: (1) the introduction of fiber overlays, which can be very
costly in the metropolitan environment (new fiber ducts have high costs as discussed in
Section 11.1.1), (2) the introduction of DWDM, offering the desired capacity upgrades
without the need for more fiber, or (3) the introduction of a combination of fiber overlays
and DWDM.
Initial efforts on investigating the feasibility of WDM and optical transparency in
metropolitan optical networks focused on the transformation of the networks of Fig-
ure 11.9 into a form of interconnected all-optical rings as shown in Figure 4.94 and
discussed in the simulation case study of Section 4.13.2. This type of metro network
is composed of the feeder, referred to also as IOF or core sections, and the distribu-
tion (referred to also as collector or edge) sections as first defined in [Saleh+99]. The
feeder sections aggregate traffic from distribution sections at the so-called hub or su-
perhub nodes and redirect them to the backbone long-haul network. The feeder (IOF)
sections can vary in size. Typically they are of the order of 100–200 km in circumfer-
ence with an average number of nodes between five and eight. Extended versions of
such IOF rings that could reach circumferences of 400–600 km have been suggested
[Humblet01] and are commonly referred to as regional metro rings. The distribution
(edge) sections interface with customer premises, delivering and collecting traffic. They
can have a variety of topologies, e.g., tree, bus, and single- and double-homed rings, can
employ coarse WDM or no WDM, and can be totally passive, composed of low-cost,
low-maintenance components. Their typical circumference is 10–50 km with an average
number of nodes between three and six. Today’s trend is for metro DWDM to regularly
extend beyond city limits (at specific nodes called superhubs as outlined below) and con-
nect cities together. Depending on the density and distance, metro DWDM is even being
stretched to cover entire states or more because the cost is lower than long-haul DWDM
[Nachum+05].
844 Multiwavelength Optical Networks

11.3.3 Metro Network Evolution

11.3.3.1 The Initial Vision
The infancy of WDM metropolitan area optical networking in North America can
be traced back to the early to mid-1990s when the various DARPA-funded consor-
tia [Alexander+93, Brackett+93, Wagner+96] were well under way with the first of
their kind research studies. Their purpose was to integrate next-generation network
architectures, advanced technology, and business drivers to achieve high-capacity, high-
performance, cost-effective, reliable, transparent multiwavelength optical networks that
met the government demands at the time. The technology applications were intended
initially for networks of national scale [Wagner+96] but eventually were scaled back to
metropolitan area-type of designs when certain transparency restrictions became obvi-
ous [Vodhanel+97]. The above-mentioned consortium efforts clearly helped shape and
influence the initial industry attempts to understand the limits of transparency, determine
its feasibility in the metropolitan environment, quantify its benefits, and, finally, attempt
to design and engineer networks in its framework. Figure 11.10 presents one of the early
attempts to theoretically study transparency in networks of the metropolitan regional
size [Antoniades+00a]. The architecture was in line with the theoretical research at the
time and prototypes such as the multiwavelength optical network (MONET) testbed that
was being constructed in Washington, DC [Anderson00]. The idea was to investigate
transparency employing two interconnected WDM rings with dedicated (1+1) protec-
tion, using 32 channels in the range 1536.609–1561.419 nm with a channel spacing of
100 GHz. The network of Figure 11.10 was not based on any realistic network traffic
scenario but rather on the initial thoughts that optical nodes need to be connected in
rings (following the traditional SONET paradigm) and that possibly different rings could
be designed by different vendors and for different application domains but be intercon-
nected in a transparent way. The network nodes of such a design were assumed to be

5 6 8 9
4
7
10
Protection path
WSXC Working path

WADM
14
3
11
1 13
2 12

Figure 11.10 Metro WDM interconnected-ring simulation case study. (From [Antoniades+00a, Fig-
ure 1]. Copyright
c 2000 IEEE. Used by permission of The Institute of Electrical and Electronics
Engineers, Inc.)
 Current Trends 845

Dark Boxes: Superhubs

White Boxes: Hubs

Gray Boxes: Central Offices

Figure 11.11 Typical ILEC metro network in the 2004 time frame. (From [Elby02, Slide 8]. Copy-
right
c 2002 Verizon. Used by permission of Verizon.)

WADMs or WSXCs following traditional parallel architectures of the time (see Fig-
ure 4.82). The study of [Antoniades+00a] showed that trying to transparently engineer
the network of Figure 11.10 for the worst-case path (bold line) was challenging consider-
ing the fact that the all-optical nature of the network ensured the presence of a number of
physical-layer transport effects (i.e., fiber nonlinearities, optical cross-talk, etc). Similar
studies in the area [Antoniades+00b, Chen+99, Farjady+03, Gerstel+99, Humblet01,
Wong00] helped stimulate commercial interest in the subject and the first metro network
element prototypes started being developed in the commercial sector’s optical network-
ing laboratories and deployed in field-trial implementations [Kim+02, Noguchi+03,
Noirie+02, Reichmann+01, Richter+02, Takachio+02, Toliver+03, Tomkos+02a] in
the late 1990s to early 2000.

11.3.3.2 What Has Happened Instead

The telecommunications market slowdown of the early 2000s inevitably affected the
equipment and DWDM technology deployment scenarios in the metro area. The view of
Figure 4.94 with transparent interconnections of IOF and edge rings has not yet material-
ized. Figure 11.11 shows the deployment of a typical network in a highly populated metro
region that is served by an ILEC during the 2004 time frame [Elby02]. All rings shown
in Figure 11.11 were still based on SONET/SDH equipment and the move to DWDM
involved only the superhubs (dark boxes), where stacked SONET OC-48/OC-192 rings
run over point-to-point DWDM between the superhubs. In typical densely populated
metropolitan areas like New York City, Philadelphia, or Atlanta, superhubs will number
only a few and will typically be spaced 30 to 50 km apart with the possibility of much
shorter distances.
Superhubs generically all looked the same and a typical view is presented in Fig-
ure 11.12.
846 Multiwavelength Optical Networks

UNI NNI

Local IOF

Superhub

R-DWDM

R-DWDM
IAD
DCS

ADM
SONET
ATM
IAD SONET
ADM

IAD
ATM/SONET

R-DWDM: Ring DWDM

DCS: Digital Cross-Connect System
ADM: Add/Drop Multiplexer
IAD: Integrated Access Device

Figure 11.12 Typical view of a superhub in the ILEC metro network of Figure 11.11 (2004 time
frame). (From [Elby02, Slide 13]. Copyright

c 2002 Verizon. Used by permission of Verizon.)

DWDM was introduced only on selective edge rings mainly in a hubbed config-
uration (ring DWDM) and as point-to-point fiber relief among the superhubs. Elec-
tronic add/drop multiplexers (ADMs) and digital cross-connect systems (DCS) were still
at the core of these locations as shown in Figure 11.12. What has transpired since then
is the introduction of reconfigurable optical add/drop multiplexer (ROADM) modules at
the superhub nodes and a transition to ring and mesh optical networking among these
locations as shown in Figure 4.98 and described in the case study of Section 4.13.2. Sys-
tem integrators are currently evaluating the cost-performance trade-offs associated with
the introduction of the latest products such as wavelength cross-connects (WXCs) (see
Chapter 4) or multiservice provisioning platforms (MSPPs) (described in more detail
below). The former represents a current trend in metro systems and networks, namely,
integration of transmission and switching technologies as described in [Elby03]. Fig-
ure 11.13 presents the view of a currently evolving superhub in a typical metro network
of an ILEC [Elby02].
Potential integration of products such as the optical cross-connect with its 1310-nm/
1550-nm transponders and electronic switching fabric, with the emerging all-optical
WXC products, is at the heart of the near-future implementation of these superhubs.
In addition, current trends also include the expansion of DWDM and reconfigurable
OADMs into the hub and Central Office locations of Figure 11.11 as demand for single-
wavelength services continues to grow.
 Current Trends 847

UNI NNI
IOF
Access
Local
Potential Integration

WXC
Superhub

ROADM OXC DWDM

R-DWDM

MSPP
NG-SONET MSPP

MSPP ATM/SONET
IAD

MSPP: MultiService Provisioning Platform

ROADM: Reconfigurable Optical Add Drop Multiplexer
OXC: Optical Cross-Connect
WXC: Wavelength Cross-Connect

Figure 11.13 Typical view of a current superhub in an ILEC metro network. (From [Elby02,
Slide 14]. Copyright

c 2002 Verizon. Used by permission of Verizon.)

11.3.4 Metro Networking State of the Art

11.3.4.1 Emerging Technologies and Architectures
Gigabit Ethernet
Ethernet, the most popular technology today for local area networking has seen multiple
improvements since the early 1980s, going from 10 to 100 Mbps (fast Ethernet) and
then to 1 Gbps (GbE) so it can expand to the metropolitan environment. This technology
is fully backward compatible with the existing Ethernet protocol and increases speed
tenfold over fast Ethernet. The Gigabit Ethernet protocol, ratified by IEEE under the
802.3-2000 standard, encompasses the previous 802.3z and 802.3ab standards. The orig-
inal Ethernet specification was defined by the frame format and support for CSMA/CD
protocol, full duplex operation, flow control, and management objects as defined by the
IEEE 802.3 standard. Gigabit Ethernet looks identical to Ethernet from the data link
layer upward and implements all these functions. The most important changes from fast
Ethernet to Gigabit Ethernet include the data rates and the additional support of full du-
plex operation in the MAC sublayer, whereas fast Ethernet supported only CSMA/CD.
At the physical layer, Gigabit Ethernet supports fiber optic media to be able to deliver 1
Gbps data rates to the fiber.

10-Gigabit Ethernet
10-Gigabit Ethernet (10 GbE) has been developed to meet the increasing bandwidth
needs of the LAN market and to take Ethernet out into the storage, metro, and wide area
848 Multiwavelength Optical Networks

Edge Edge
(collector)
(collector)
10 GbE
10 GbE
IOF rings
Metro/DWDM
OC-48/OC-192

Central office
(Hub or Superhub) Central office
(Hub or Superhub)

Long Haul/DWDM
OC-192/STM-64

Figure 11.14 Typical deployment of 10 GbE technology in the metro environment.

networks. The migration to 10 Gbps Ethernet is an economically viable path because it

uses the same frame format as 10M/100M/1G Ethernet. This allows the leveraging of
existing Ethernet installations, the Ethernet management tools, and the people already
trained in Ethernet. Ultimately 10 Gbps transmission provides a solution with easy
integration and low cost of ownership. The 10-gigabit Ethernet 802.3ae standard has a
number of features such as
r Preserves the 802.3 Ethernet frame format
r Preserves the minimum and maximum frame size of the current Ethernet standard
r Supports full duplex operation only
r Supports a speed of 10 Gbps at the MAC interface

In contrast to LANs, where Ethernet has been the technology of choice over the
years, metro networks impose fundamentally different demands as they are expected to
carry more than just best-effort data and quality-of-service (QoS) guarantees. Network
resilience, operations, administration, and maintenance (OAM), and scalability impose
new requirements on the 10 GbE technology. In [Zapata+04] the above requirements
are discussed from the point of view of high-speed metro Ethernet networks, and typical
ring architectures are analyzed. These architectures essentially involve optical packet
switching (OPS) and optical burst switching (OBS) technologies that are addressed in
detail in Chapter 10. Clearly, the developing 10 GbE technology in the metro domain is
focusing on the edge (collector) ring with target maximum distances on standard single
mode fiber of about 30 to 40 km. Figure 11.14 shows a metro network implementation
using 10 GbE running at the edge of the network with Central Offices or hubs as
aggregation points. The above technology is also penetrating the core (IOF) where the
aggregation points of Figure 11.14 might be the superhubs of Figure 11.11.
 Current Trends 849

Next-Generation Access
Traditional passive optical networks (PONs) were focused on serving the access network,
which centered around the residential domain. As described in Section 5.7, PONs are
composed of an optical line terminal (OLT), normally in a Central Office, which is con-
nected by passive components to various optical networking units (ONUs) residing on
customer premises. With the establishment of Gigabit-capable PON (GPON), extensions
of Ethernet-based PON (EPON) to the 10 gigabit domain, and the emergence of support-
ing technologies such as Gigabit Ethernet and high-performance cost-effective optical
components, there is a move toward next-generation access networks [Davey+06a].
These networks use PONs along with WDM and 10-Gbps transmission, and they have
expanded reach with increased loss budgets (higher power splits). (See Section 5.7.5.)
As a result, network operators are able to use GPON and EPON over an extended reach
that essentially blends access into the metro domain [Davey+06b].

Resilient Packet Ring

Resilient Packet Ring (RPR) is an emerging network architecture and technology de-
signed to meet the requirements of a packet-based metropolitan area network. The
problem of effectively managing a shared resource (in this case the fiber ring is the
resource that needs to be shared across thousands of subscribers in a metro area) is most
efficiently solved at the MAC layer of the protocol stack. RPR (IEEE 802.17) is a media
access control (MAC) protocol designed to optimize bandwidth utilization and facilitate
services over a ring network while satisfying carrier-class requirements like 50-ms recov-
ery under any single network fault (i.e., failure of a fiber span or node). As a result RPR
is designed for metro fiber ring networks. In [Simard+05] the interplay between RPR
and next-generation SONET/SDH in their ability to support metro Ethernet is analyzed.
It is shown that RPR enables more efficient statistical multiplexing and finer grooming
thus increasing transport efficiency compared to traditional SONET/SDH technology
that has been used to map Ethernet services onto circuits.

11.3.4.2 Emerging Products

Multiservice Provisioning Platform
New technologies and equipment allow service providers to rapidly respond to traffic
and service needs in the metro environment while reducing power consumption, space
requirements, and overall cost. With the evolving edge, new equipment that can provide
the above service while collapsing multiple technologies in one box is the market trend.
Several leading equipment vendors have announced multiservice provisioning platform
(MSPP) products that provide the functions of multiple network elements in a single
platform. These usually provide TDM solutions with interfaces such as DS-1, DS-3,
EC-1, and data solutions with 10/100/1000 Mbps Ethernet and with OC-3 to
OC-192 optical transport bit rates, including integrated DWDM wavelengths. In addi-
tion typically they support aggregation and transport of services from DS-1 to OC-192,
switched 10/100/1000 Mbps Ethernet for improved bandwidth utilization, and line-rates
of 10 Mbps, 100 Mbps, and Gigabit Ethernet transport.
850 Multiwavelength Optical Networks

Small Form Factor Pluggable Modules

In metro networks a new generation of pluggable-hosting line cards are being introduced
that allow the line card and the pluggable optics to remain unchanged in cases where the
bit rates must increase or the protocol is changed (no powering-down of the line card).
With small form factor pluggable modules (SFPs) the above tasks are achieved using
software and as a result the costly power-down of line cards is avoided. A technoeconomic
study of the effect of SFPs on the provisioning of services in metro DWDM networks is
presented in [Esmacher05].

11.3.4.3 Network Elements

Following the evolutionary scenario presented above, system vendors develop all-optical
network elements based on the anticipated utilization rates and capacity planning cycles
in the metro space. In the early 2000s, capacity utilization rates were fairly high, typically
averaging above 80%. Utilization rates of 70% and above are tipping points for adding
new equipment, and capacity planning cycles typically produce plans for one to two
quarters of demand ahead of the current demand. Existing equipment is rarely replaced
and is not turned off-line for upgrading as new systems are added to legacy systems. The
new scenario sets fresh standards on the compatibility and upgradeability requirements
for new equipment. Figure 11.15 portrays typical architectures of metro networks (based
on the scenario of Section 11.3.3 above) representing the current application space for
new network elements.
Figure 11.15(a) presents the logical connectivity pattern for a typical edge (or col-
lector) metro ring composed of four nodes. These nodes represent the Central Offices
(or end offices) and, provided that they are outfitted with WDM, these are usually static
add/drops with specifically assigned wavelength(s). Their logical connectivity pattern is
“hubbed” to a hub or superhub location. Figure 11.15(b) represents the logical connec-
tivity pattern of a typical IOF (or core) metro ring network with six nodes that are either
hubs or superhubs. Until recently the mesh connectivity pattern presented in the figure
was achieved using fixed optical add/drop multiplexers. This is currently changing with
the introduction of reconfigurable optical add/drop multiplexers. In addition, in some

hub/
superhub

(a) (b)

Figure 11.15 Typical metro network ring architectures.

 Current Trends 851

Band DMUX Band MUX

pre-amp post-amp
Optional switching
functionality
Working Path
West East
Pass-through bands Optional switching
functionality
Protection Path

post-amp
pre-amp
Channel Channel
DMUX MUX

Protection
switch

Rx Tx

Figure 11.16 Typical vendor generic WADM node architecture for the metro network application
space of Figure 11.15. (From [Antoniades+05, Figure 4]. Copyright

c 2005 OSA. Used by
permission of the Optical Society of America.)

of the nodes of Figure 11.15(b) wavelength selective cross-connect switches can be

installed to enable dynamic reconfigurability of wavelength services between IOF rings
in the backbone.
Figure 11.16 presents a vendor “generic” WADM node implementation for the ap-
plication space described above. The design is based on the concept of wavelengths
and wavebands described in Section 4.12.2 and generally presented in Figure 4.83.
The implementation can be static (OADM) or dynamic (ROADM) depending on
the presence of switching elements. As mentioned, edge ring implementations of the
type shown in Figure 11.15(a) are usually static, whereas IOF ring implementations can
be dynamic, in which case optical switching elements are needed in Figure 11.16. (Other
forms of ROADMs were presented in Figures 4.84 and 4.85 and described in detail in
Section 4.12.2.) The latter is an emerging configuration and can be used to switch in-
dividual wavelengths between fibers in an all-optical manner effectively functioning as
a WSXC. As of this writing, a variety of companies (for example, Capella Photonics,
JDSU, CoAdna, and Xtellus) manufacture the WSXC modules for system vendors such
as Alcatel, NEC, ADVA, and Ciena.

11.3.4.4 Metro WDM Network Deployments

In addition to the gradual WDM upgrades on the LEC metro networks, there has
been significant recent activity on WDM metro network deployment demonstrations
of next-generation technologies. In [Toliver+03] the concept of wavebands and wave-
band switching was demonstrated on the Advanced Technology Demonstration Network
(ATDnet), an optical network testbed that links a number of government agency labo-
ratories in the Washington, DC metropolitan area. ATDnet is a WDM metro network
built as a continuation to the MONET testbed [Anderson00] that was designed to
852 Multiwavelength Optical Networks

support eight 2.5 Gbps wavelengths per fiber. In the ATDnet the East ring uses MEMS-
and LiNbO3 -based transparent WSXCs. The fact that these nodes are transparent en-
abled the designers to perform an upgrade to the network to eight wavebands on a
200-GHz grid. In [Toliver+03], transparent waveband switching on one of these bands
is described. Four channels at OC-48 and OC-192 bit rates were spaced 25 GHz apart
within the band and transmitted through 100 km and two WSXCs with satisfactory BER
performance. The demonstrated transparent path was similar to those in the IOF ring
designs described above. The impact of polarization effects (PDL and PMD) as well as
dispersion and filter concatenation was examined.
In [Noguchi+03] the first field trial of a wavelength routing full-mesh WDM metro
network in the northern part of Japan is described. The network involved a 32 × 32
arrayed waveguide grating (AWG) router as the central node in a downtown location of
the city of Chitose as well as five WDM nodes strategically located throughout the down-
town inner area spaced 5 km from each other. Typical fixed optical add/drop multiplexer
equipment (similar to that of Figure 11.16) is used at each one of the five peripheral
WDM nodes and a 5 × 5 full-mesh logical connectivity is realized among these nodes.
Several experiments were carried on this network involving SONET/SDH, GbE, and
high-definition serial digital interface (HD-SDI). Several private networks running dif-
ferent protocols, bit rates, and levels of management policy were also demonstrated. The
network is currently open to the citizens of Chitose and is used for the educational and
administrative needs of that community.
A field trial of a 10-Gbps Resilient Packet Ring in downtown Tokyo, Japan, is reported
in [Tsurusawa+05]. It included a next-generation multiservice architecture platform that
supported GbE interfaces for data transfer and SONET/SDH interfaces for voice services
and dedicated private line services, including 3G cell phone service. The configuration
of the field trial was based on three nodes at the major operation centers of a Japanese
CLEC. The total circumference of the RPR was about 32 km. Various transport perfor-
mance measurements were taken on differentiated traffic carried on Gigabit Ethernet. In
addition, the feasibility of IP-based video multicast service using virtual private LAN
services was investigated.
In [Zeng+05] an all-optical metro WDM network interconnecting four research fa-
cilities within the city of Ottawa, Canada, was reported. The four network nodes were
based on MEMS switches, and the feasibility of dynamically provisioning transparent
optical paths (of the order of 60 km) among the four nodes was demonstrated.
Recent work has shown that the physical design of a transparent WDM metro network
should be tailored to both the network topology as well as the network traffic demands
[Madamopoulos+02, Peloso+04]. In [Peloso+04] a network featuring optical nodes at
each of 20 European capitals composing a 32-link topology was exhaustively tested us-
ing an experimental simulation technique. This was based on using a recirculating loop
experiment and testing all possible transparent path combinations among the 20 optical
nodes. The network topology used was based on a realistic traffic matrix that showed
that 85% of the traffic required links shorter than 2000 km. The established network
is an expanded metropolitan design, more like a regional network, where technologies
like enhanced forward error correction (EFEC), high-end metro amplifiers (total output
 Current Trends 853

50 km Z1
Z12 Z2
Z11 B

AA
X Y Z
W
60 km
Z3
Z10 Max length = 465 km
V
T
120 km S
Max number of hops = 12
5 km P
R
Z9 5 km
L Q O Z4
K
M H
I 5 km
J
80 km
G
E F
Z8 140 km ILEC
C Z5
Peripheral nodes
C
Z7
Z6 Distances are not drawn in scale

Figure 11.17 Metro network case study. (From [Madamopoulos+02, Figure 1]. Copyright
c 2002
IEEE. Used by permission of The Institute of Electrical and Electronics Engineers, Inc.)

power more than 17 dBm), and externally modulated transmitters were required. It was
shown that with the help of efficient routing algorithms, 99% of the connections could
be established without regeneration. A set of basic chromatic dispersion management
and amplification rules was applied. In [Madamopoulos+02] an effective top-down
simulation approach was presented that enabled the design of a 29-node, 11-Tbps U.S.
mesh metro network in a transparent way based on realistic network traffic requirements.
Projected traffic demands based on current known values and existing Central Office
locations in a major U.S. city were used to design network architectures for a number
of different U.S. carriers for year 2005. A physical-layer topology, based on network
dimensioning, was then superimposed on the network architectures. The network was
then engineered for 100% of the required optical connections using the three-step sim-
ulation methodology that was presented in Section 4.13.1. This involves engineering
paths with the longest length as well as the ones with the maximum number of optical
hops. Figure 11.17 presents the network topology for an ILEC (connected nodes). Its
maximum length path is shown as a gray line, and its maximum hop path appears as a
dotted line. Also shown in the figure are the nodes of a CLEC (diamonds in the figure),
which are left unconnected.
The optical network architecture described above is derived from traffic modeling of
a major U.S. metropolitan area with different market penetrations (low, moderate-, and
high-demand scenarios) of existing carriers, both ILECs and CLECs, as presented in
[Madamopoulos+02]. The traffic projections are based on projected voice, transaction
data, and Internet traffic demands, taking into account knowledge of current demands,
854 Multiwavelength Optical Networks

the change of the population, the change of the nonproduction employees, and the
Internet hosts [Madamopoulos+01]. The model has been described in [Dwivedi+00,
Vaughn+00] and uses publicly available data sources. The projected traffic demands and
CO locations are used to derive individual traffic matrices for the different carriers. The
traffic demand of the core network (which excludes the peripheral nodes) is the driver of
the assumed connectivity (selection of links connecting metro nodes in Figure 11.17).
After having selected a specific connectivity for the core, the projected traffic from the
peripheral nodes is also considered and the connectivity is expanded further to include
all possible connections between any two nodes in the network. A commercial network
dimensioning and routing tool (WDMNetDesign [Arijs+00b]) was then used to generate
required network parameters, such as connection lengths, add/drop percentages, and
number of fibers at each node, and to assign wavelengths for each connection. For the
network case study of Figure 11.17, 40 wavelengths per fiber and (1 + 1) protection were
considered. In the network engineering part of [Madamopoulos+02] it is shown that
transparent paths with maximum lengths of 465 km and paths with maximum number
of hops equal to 12 were possible, thus enabling all other paths in the 29-node mesh
WDM metro optical network of Figure 11.17.

11.4 Long-Haul and Ultra Long-Haul Networks

The widespread deployment of point-to-point WDM systems has resulted in large num-
bers of optical channels terminating at the Central Offices. This in turn has created a
pressing need to manage channels at the optical level without subrate demultiplexing
of the individual connections, making optical switching more attractive economically.3
Optical cross-connects can be deployed to interconnect the WDM transport systems,
as in Figure 11.7, provisioning and protecting wavelengths directly in the optical layer.
The changing economics of switching has introduced new perspectives on wide area
network architectures. We examine some of these now, followed by a look at some recent
deployments.

11.4.1 Current Considerations in Wide Area Network Architectures

Some network operators who have deployed open WDM transport systems with 1.3-µm
short-reach SONET interfaces have considered the deployment of these cross-connects
in an opaque architecture [Bala+95, Tkach+98, Young+98], whereas others have con-
sidered the transparent approach [Davis+98]. The opaque architecture has the disad-
vantage that it has extra components (transponders) but it has several advantages. It
clearly separates switching from transport and eliminates any cascaded impairments
that accumulate during transmission, by providing signal regeneration. Furthermore, it
enables wavelength interchange functionality between WDM transport systems by using
3
This is analogous to the situation that existed in the 1970s, when the widespread deployment of digital
transmission systems in the telephone industry (e.g., the T1 carrier) led to the development of digital
time-space division switching equipment (e.g., AT&T’s No. 4 ESS).
 Current Trends 855

the 1.3-µm interface as a common intermediate frequency between two WDM systems.
In this fashion, it enables multivendor interoperability and allows different wavelength
sets from different vendors to be interconnected by bringing them all to a common de-
nominator (i.e., the 1.3-µm standard interface). Whether opaque or transparent, the need
for optical cross-connects is clear, and they must be equipped for efficient provisioning
and fault recovery.
Separate WDM transport systems and optical cross-connects can also be integrated
to form wavelength-selective (WSXCs) or wavelength-interchanging cross-connects
(WIXCs) with both the transport and the switching subsystems integrated into one
network element [Jackel+96, Jourdan+97]. This can lower the cost of the cross-connect
by eliminating duplicate components. Later in this section we will discuss in more detail
the deployment of mesh networks utilizing opaque or transparent optical cross-connect
architectures.
Architectures for long-haul networks can be either ring based or mesh based. As
discussed in Chapter 8, the former have well-developed and rapid (e.g., 50 ms) protection-
switching protocols, while the latter generally have much higher network utilization than
rings, but without the simplicity and speed of ring-based protection schemes.4
Mesh architectures are the prevalent ones for backbone networks today, designed
as opaque or transparent, depending on the technologies used. This section focuses on
designs where the optical switches are completely transparent (or have optical bypass ca-
pabilities [Saleh00, Saleh03, Wagner+01]) and transmission is either regeneration-free,
or regeneration5 is limited as much as possible. Optical reach,6 physical impairments,
and the performance engineering of reconfigurable wavelength-routed networks, issues
that are of no consequence in networks deploying opaque switches, become the focal
points in transparent networks. For example, longer optical reach means that the optical
signal will require a smaller number of regenerations, which in turn will lead to lower
capital and operational costs. Of course, one has to take into account the fact that ultra
long-haul (ULH) technology generally requires more expensive transmitters and optical
amplifiers, so the question that one should ask is what is the correct point of operation?
The analysis of four representative North American long-haul networks performed in
[Simmons05] showed that an optical reach in the 2500- to 3500-km range yielded the
minimum total capital cost, and it was the point where there was a balance between the
decreasing operating costs and the increasing initial network cost.
Transparent networking in backbone networks requires the development of high-
capacity transport technologies capable of regeneration-free transmission over several
thousands of kilometers as well as the development of cost-effective all-optical cross-
connects [Saleh00].

4
Mesh restoration algorithms in the SONET layer can take several seconds to minutes to restore service
mainly because the restoration is done at a subrate (e.g., DS1 or DS3) level. WDM meshes that are protected
at the optical connection level have much fewer connections to reroute, resulting in an accelerated protection
time, which can approach that in rings (see Section 8.4.1).
5
All-optical regeneration is not yet commercially available.
6
Defined as the distance that an optical signal can travel before the signal quality decreases to the point
where regeneration is required.
856 Multiwavelength Optical Networks

Ultra long-haul systems are usually systems where reach extends to distances longer
than 1500 km [Wallace01]. Even though the enabling technologies for point-to-point
ULH transmission have been widely reported in the literature [Naito+99, Tomkos+02b],
work on ULH transmission in mesh network architectures with fully reconfigurable opti-
cal cross-connects is more recent. [Pratt+03], for example, reports on ULH transmission
of 40 × 10.7 Gbps channels over 1700 km in mesh network architectures with mixed
fiber types and with fully reconfigurable optical cross-connects and dynamic gain equal-
ization.
When some regeneration is required, [Saleh98] advocates the use of islands of trans-
parency, where each island is fully transparent internally but requires standard electronic
interfaces to access it. This approach allows the island size to match the reach that is
achievable and in this way solves the problem of expensive regenerators. Another ap-
proach is the managed reach solution [Wagner00]. In that case, regenerators are placed
only at distances that are consistent with the reach length of the technology, and routing is
performed taking into account the reach limitations and the locations of the regenerators.
A study in [Parys+01] tries to quantify the benefits of selective wavelength regeneration
in ULH networks. It is shown that IP traffic, which is distance independent and has an
average connection distance much longer than the distance for voice traffic, requires
an increase in the optical reach to decrease the regeneration cost and the transmission
cost per bit. A similar study in [Wagner+01] also looks at the characteristics of Internet
traffic and the reach length needed for Internet backbone applications. Internet traffic
requires a reach length of approximately 3000 km to satisfy 60% of the connections. It
is shown that for Internet traffic, when a managed reach approach is utilized, with reach
lengths of 2500 km, many regenerators are no longer needed and can be eliminated,
which in turn will significantly lower the network cost in terms of equipment.
The next-generation core backbone network that will encompass the long-haul and
ultra long-haul components discussed above will require major breakthroughs in net-
work resiliency, capacity, and transmission technology. In [Saleh06] an evolutionary
path toward the new backbone network is presented. This includes scaling up today’s
technology to meet the future network requirements and identifying areas where inno-
vative solutions will be needed. Among the objectives for the next generation are the
ability to support traffic of the order of 100 Tb/s at demand end points, a greater degree
of configurability than current networks, rapid reconfiguration capability, and the ability
to survive multiple failures.

11.4.2 Some Recent Commercial Network Deployments

This section describes three commercial deployments of mesh optical networks in the
core that were implemented in the past few years.
The first case was that of Dynegy Global Communications, which deployed a nation-
wide, optically switched, mesh network spanning 16,000 route miles and reaching 44
of the largest U.S. cities. The backbone used Tellium’s optical switches and Fujitsu’s
DWDM platform and offered 2.5- and 10-Gbps services. Dynegy’s network supported
end-to-end point-and-click provisioning and implemented shared mesh protection, as
 Current Trends 857

described in Chapter 8. Diversity of routes in Dynegy’s optical network was defined

using SRGs, which were configured by Dynegy’s network operators with the knowledge
of the physical fiber plant of the optical network. The maximum recovery times observed
in this network were less than 200 ms in the worst case and corresponded to the last
lightpath recovered as a result of a conduit failure (indicating the end of the protection
process) [Charalambous+03]. Dynegy’s network also supported mesh reprovisioning
in the case of multiple failures or when recovery of the original backup path did not
succeed, and mesh reoptimization during off-peak hours of operation. Reoptimization
by rerouting the backup paths gave Dynegy’s network operators the opportunity to regain
some of the network bandwidth that was currently in use. In a specific case, when the
backup paths in Dynegy’s live network containing 70 shared mesh protected demands
were rerouted during a maintenance window, this procedure saved 31% of the protection
ports, which in turn translated to 20% savings in the total number of ports. Also, the av-
erage length of the backup paths decreased from 5.87 to 4.88 hops, which in turn reduced
the backup latency (recovery time) by 25.61% (calculated using the average length of
the backup path in miles before and after the reoptimization). The complete backup
path reoptimization procedure for that network, including testing, took approximately
5 hours [Charalambous+03].
Another commercial deployment was made by AT&T using intelligent optical
switches (IOSs) and multiservice platforms (MSPs) in its nationwide network in a mesh
configuration. This new network was built on AT&T’s nationwide fiber-optic network
that by 2002 included 63,500 route miles of fiber-optic cable, as well as DWDM systems
capable of carrying 160-wavelengths running at 10 Gbps each. In the 2002 time frame,
the network was already connecting 40 U.S. cities and carrying live traffic, with plans
to expand the network significantly in the next few years. The optical switches deployed
were Ciena’s CoreDirector opaque optical switches, which acted as wavelength switches
but were also capable of grooming subwavelength circuits [Cortez02, Ranganathan+02].
These switches carry traffic from 45 Mbps to 10 Gbps and they use real-time signaling
and routing algorithms to make decisions without the need for manual intervention. The
MSPs are used to aggregate lower rate customer traffic into high-speed pipes (e.g., OC-48
and OC-192) for routing across the network using the optical cross-connects. The AT&T
network also featured point-and-click provisioning, survivability capabilities, and classes
of service with different recovery characteristics. Dynegy’s and AT&T’s deployments
were the first comprehensive commercial deployments of nationwide multiwavelength
optical networks by a carrier utilizing intelligent optical components. They showed the
need for the evolution of the carriers’ infrastructures to state-of-the-art technology. Such
a step, even though expensive in terms of capital expenditure, has proven to be invaluable
to carriers that want to create a network that supports simplified operations, improved
performance and reliability, and the ability to support new services and applications.
Finally, Broadwing Communications Services, Inc., deployed Corvis’s all-optical
switch (a transparent WSXC) in their network, marking the first time that a carrier
installed an all-optical switch for commercial use in a nationwide network. Networks
deploying transparent optical switches, such as the Broadwing/Corvis deployment, have
on the surface significant advantages over networks deploying opaque switches, such
858 Multiwavelength Optical Networks

as the Dynegy/Tellium and the AT&T/Ciena deployments. The end-to-end bit-rate and
data format transparency as well as significant footprint and power savings suggest cost
savings. However, the absence of field-deployable commercial wavelength conversion
technology, the interoperability requirements among different equipment vendors, the
absence of adequate control and management, and the inflexibility of configurations
due to the nonlinear limitations that appear in large-scale transparent optical networks,
suggest that completely transparent networks, even though desirable, have not matured
enough for commercial deployment [Ellinas+03]. The absence of wavelength conver-
sion can be dealt with in regeneration sites. ([Simmons02] has shown that sufficient
sparse wavelength conversion in all-optical express networks is achieved by utilizing the
wavelength conversion that is already available at regeneration sites.) However, the main
problem of providing the control and management functions is still present, because
access to the electrical signals is not possible in transparent networks, whereas these
signals are readily available in opaque networks.

11.5 New Applications and Services

11.5.1 Wavelength-on-Demand
Intelligent optical cross-connects in a mesh network can be used to request wavelengths
on a per-need basis, at a specified time of the day (scheduled service) or as a short-term
leasing service. The point-and-click provisioning capability of the network through a
wavelength management system can be used to request wavelengths on a per-need basis.
Alternatively, the client network elements can have direct access to the resources of the
network, requesting wavelengths on a per-need basis through an optical User-Network
Interface (UNI). This capability is particularly useful for emergency disaster recovery,
where additional wavelengths may be required to ensure that the network is running
normally after a disaster event. It can also be used to connect video points-of-presence
on demand. Wavelengths can also be scheduled for specific hours of the day, used for a
short time and then released at a predetermined time. This service is particularly useful
for applications that use bandwidth in a predictable way. Such examples include, among
others, storage area networks (SANs) where daily backups are performed at specific
times or Web cache and mirror site daily updates. In these applications wavelengths
can be billed by peak and off-peak hour usage. Finally, short-term wavelength leasing
services are useful for applications such as special event hosting (e.g., sports events,
concerts, and Webcasts).

11.5.2 Virtual Private Optical Networks

Networks utilizing optical cross-connects in conjunction with their management software
can also allow service providers to create virtual private optical networks (VPONs) for
their customers. In this way a defined group of customers can have exclusive access to
a specific portion of the network; including the ability to monitor performance and be
 Current Trends 859

aware of faults. VPON services can be offered to any entity that wants to manage and
monitor its transport networks. In conjunction with wavelength-on-demand services,
customers in VPONs can provision or release wavelengths and reconfigure their portion
of the network according to their needs at any particular moment. They can also mix and
match different QoS services, because they have access to the individual wavelengths in
their network.

11.5.3 Bandwidth Trading

Bandwidth trading has attracted considerable attention from the telecommunications in-
dustry as it is a way to maximize profits by matching a carrier holding surplus resources
with a carrier with a deficit of resources [Fuller01, Kohut01]. From a trading perspective,
resource surplus is considered to be the unused bandwidth for which the fiber infras-
tructure (including DWDM) is already installed. It is more economical for one carrier
to buy excess capacity from another carrier rather than to build its own resources – an
extremely expensive alternative. Networks that utilize optical cross-connect switches
together with a wavelength management system are ideal candidates for providing ser-
vices like wavelength or bandwidth trading as such a system requires near real-time
bandwidth delivery.
Trading or brokering can be achieved either through multilateral agreements between
the carriers, through multilateral operations interfaces (distributed approach), or through
a third-party bandwidth or wavelength broker with a clearinghouse operations model
(centralized approach). If the clearinghouse model is used, the clearinghouse receives
all bandwidth quotes and maintains prices per bandwidth type (OC-48, OC-192) de-
pending on the time of day the request is made, duration of the request, as well as the
QoS parameters of the request (protection, delay, etc.). The service providers then query
the clearinghouse for quotes and the buyers may bid for bandwidth or they may just
request the lowest quote. Once the partner service provider is identified, a request is
placed. Note that the clearinghouse is not involved in the signaling or physical channel
setup procedure. It does maintain pricing information, however, and it may be involved
in the exchange of billing records. If the distributed approach is used, each network
management system (NMS) for each carrier has operations support system (OSS) inter-
faces with all other NMSs. These could be either TeleManagement Forum-based or they
could be based on some other proprietary interface. An NMS then places bandwidth
queries with other NMSs and after accepting a quote, it places a provisioning request
either through the other provider’s NMS or using the optical UNI. Once the provisioning
request is accepted and the setup is complete, the service provider that is providing the
bandwidth starts the billing process. The billing terminates when the connection is finally
released.
Figure 11.18 shows a functional block diagram of a wavelength-brokering opera-
tional model using OXCs, an OXC wavelength management system, and the associated
brokering applications. The optical cross-connect switches work as pooling points con-
necting the buyers and the sellers. The wavelength management system is responsible for
cross-connect provisioning, performance monitoring, and fault management. Through a
860 Multiwavelength Optical Networks

Connection Performance Billing

Manager Manager
Interface Manager

OXC Wavelength Management System

Carrier A

Carrier B
Local Local
OXC OXC
network network
Carrier C
Clients
Clients
Pooling Point Pooling Point
Carrier D

Figure 11.18 Wavelength-brokering operational model.

“northbound” interface (from the client to the management system), third-party vendors
can develop customized applications such as SLA management and billing.
Other applications of bandwidth trading include: (1) infrastructure swapping (be-
fore the swap providers use the resources of their own network only, whereas after
the swap providers use resources in partner networks for enhanced network flexibility
[Figure 11.19]) and (2) multicarrier recovery (where the wavelengths that are used for
protection purposes may be owned by multiple carriers [Figure 11.20]).
Beyond maximizing profits, bandwidth trading may play an important role in breaking
down some of the barriers to the development of large WANs. In the early days of optical
networking, it appeared that it would be difficult to extend these networks over wide
areas because that would require using the resources of multiple smaller networks owned
by different carriers. Bandwidth trading is a step toward this goal.

Carrier A Carrier B

Carrier A Carrier A Carrier C

Carrier A

Carrier C Carrier C

Carrier C

Carrier C Carrier B

Carrier B Carrier A Carrier C

Carrier B

Carrier B Carrier B

Before the swap After the swap

Figure 11.19 Infrastructure swapping.

 Current Trends 861

Broker

Carrier A

Carrier B

Local Local
network network
OXC Carrier C
OXC
Clients Clients

Carrier D
Pooling Point Pooling Point

Figure 11.20 Multicarrier recovery.

11.6 Conclusions

In this concluding chapter, we attempted to provide a snapshot of current trends in

multiwavelength networks, their roots in the recent past, and some indication of how
they may evolve in the future. We began with the business drivers and economics
behind the use of WDM technology. Given the increasing demand for capacity, the
need for the low-cost bandwidth provided by WDM is more pressing than ever. In
addition to the economies achieved through the use of WDM equipment, an “intelligent”
reconfigurable optical network layer offers further efficiencies because of its flexibility
and manageability. Through some examples we provided quantitative cost comparisons
for the cases of point-to-point and ring networks showing clearly the potential cost
savings accruing from WDM. The cost issues in the case of networks utilizing optical
versus electronic cross-connects were also examined and compared, as well as the
properties of open and closed WDM equipment architectures.
The chapter continued with a brief description of the early WDM network testbeds
and their importance in the subsequent commercialization of multiwavelength optical
networks. The testbeds produced a large number of advanced optical components and
systems, many of which form the foundation of currently deployed networks. They
were also the driving force behind the development of intelligent optical networking
equipment that allows for point-and-click provisioning, efficient failure recovery, re-
optimization, grooming, and a host of other features. Thus, the testbeds served as
a proof-of-concept for the carriers who were interested in networks with improved
performance, reliability, and the ability to support new services and applications.
At this writing the metropolitan area is arguably the arena in which commercial
networking activity has been most intense. We presented an extensive discussion of
current metropolitan area network architectures and their unique characteristics in terms
of functionality, performance, and cost. The current goal for MANs is a deployment
with increased functionality and performance and with decreased cost, space, and power
862 Multiwavelength Optical Networks

requirements. Whereas the initial vision of WDM metropolitan area optical network-
ing focused on transparent IOF and edge ring connections, the current view favors a
translucent approach with islands of transparency. Recent metropolitan WDM network
deployment demonstrations in North America, Europe, and Japan include high per-
formance architectures (e.g., Gigabit and 10-Gigabit Ethernet, Resilient Packet Rings,
and Gigabit-capable PONs), based on new products (e.g., multiservice provisioning
platforms), and new implementations of network elements (e.g., WADMs). These net-
works deliver high data rates to the optical layer, expanded reach toward the customer
sites, optimized bandwidth utilization, improved reconfigurability, and efficient recovery
strategies.
Moving on to a discussion of long-haul and ultra long-haul networks, we described
some current directions in wide area network architectures, considering relative advan-
tages of transparent and opaque networks as they are currently constituted. We pointed
out that the increase in the numbers of optical channels terminating at Central Offices
has made optical switching a viable if not imperative alternative to electronics. We
then reviewed some recent network deployments implemented with optical switch based
architectures.
Finally, we examined several emerging applications for multiwavelength networks,
including wavelength-on-demand, VPONs, and bandwidth trading. Wavelength-on-
demand offers the capability for requesting wavelengths on a per-need basis, a feature
that is very useful in disaster recovery, storage networks, and special event hosting.
The VPON allows a specific group of customers access to a circumscribed portion of
the network, while bandwidth trading enables carriers to trade surplus resources, thus
maximizing revenue.
Summing up our journey from Chapter 1, the outlook for multiwavelength optical
networks is brighter than ever. In a time span of about one decade, they have moved
out of the laboratory and beyond government-sponsored experimental field trials, to
commercially sustainable deployments. This road has not been an easy one. Companies
involved in the development of optical networking technologies had to weather the tech-
nology bubble at the end of the 20th century, adjusting their activities to economically
viable and reasoned approaches. The latest progress in the area points in a number of
directions. New developments in photonic and electro-optic technology, the deployment
of cost-effective optical access and metropolitan area networks, new and efficient sur-
vivability techniques, and recent work on a control plane for optical networks – all of
these provide a foundation for significant network expansion to meet the current demand
pull. Furthermore, cutting-edge research on purely optical packet processing illustrates
the fact that the technology push is alive and well. Both are good omens for the future
of optical networking.

Bibliography

[Alexander+93] S. B. Alexander, R. S. Bondurant, D. Byrne, V. W. S. Chan, S. G. Finn,

R. Gallager, B. S. Robert, H. A. Haus, P. A. Humblet, R. Jain, I. P. Kaminow, M. Karol,
 Current Trends 863

R. S. Kennedy, A. Kirby, H. Q. Le, A. A. Saleh, and B. A. Schofield. A precompetitive con-

sortium on wide-band all optical networks. IEEE/OSA J. Lightwave Technol., 11(5/6):714–735,
1993.
[Anderson00] W. T. Anderson. The MONET Project final report. In Proceedings of the IEEE/OSA
Optical Fiber Commun. Conf., Paper WJ1-1, pp. 148–149, Baltimore, MD, February 2000.
[Antoniades+00a] N. Antoniades, M. Yadlowsky, and V. L. daSilva. Computer simulation of a
metro WDM ring network. IEEE Photon. Technol. Lett., 12(11):1576–1578, 2000.
[Antoniades+00b] N. Antoniades, A. Boskovic, J. Downie, N. Madamopoulos, D. Pastel,
J.-K. Rhee, and M. Yadlowsky. Engineering the performance of DWDM metro networks.
In Proceedings of the National Fiber Optics Engineers Conference (NFOEC), Session B2,
pp. 204–211, Denver, CO, August 2000.
[Antoniades+05] N. Antoniades, N. Menon, N. Chbouki, and G. Ellinas. Value proposition for
amplets as banded amplification solutions in evolving WDM metro network architectures. OSA
J. Opt. Network., 4(3):101–112, 2005.
[Arijs+00a] P. Arijs, W. Van Parys, and P. Demeester. Cost and availability comparison of WDM
mesh and ring network architectures. In Proceedings of the 2nd International Workshop on the
Design of Reliable Communication Networks (DRCN), pp. 33–38, Munich, Germany, April,
2000.
[Arijs+00b] P. Arijs, B. Van Caenegem, P. Demeester, P. Lagasse, W. Van Parys, and P. Achten.
Design of ring- and mesh-based WDM transport networks. SPIE Optical Networks Mag.,
1(3):25–40, 2000.
[Bala+95] K. Bala, R. R. Cordell, and E. L. Goldstein. The case for opaque multiwavelength
optical networks. In Proceedings of the IEEE/LEOS Summer Topical Meeting on Optical
Networks, Keystone, CO, August 1995.
[Bala+96] K. Bala, R. H. Cardwell, D. Feddor, K. Goel, T. Hodgson, H. Kobrinski, M. Mihail,
G. Pearson, O. J. Wasem, and M. Wilder. WDM network economics. In Proceedings of the
National Fiber Optics Engineers Conference (NFOEC), Denver, CO, September 1996.
[Bala+97] K. Bala, R. Cardwell, H. Kobrinski, and O. Wasem. WDM network economics sen-
sitivities. In Proceedings of the National Fiber Optics Engineers Conference (NFOEC), pp.
105–116, San Diego, CA, September 1997.
[Brackett+93] C. A. Brackett, A. Acampora, J. Sweitzer, G. Tangonan, M. Smith, W. Lennon,
K. Wang, and R. Hobbs. A scalable multiwavelength multihop optical network: A proposal for
research on all-optical networks. IEEE/OSA J. Lightwave Technol., 11(5/6):736–753, 1993.
[Brackett96] C. A. Brackett. Is there an emerging consensus on WDM networking? IEEE/OSA J.
Lightwave Technol., 14(6):936–941, 1996.
[Cardwell+00] R. H. Cardwell, O. J. Wasem, and H. Kobrinski. WDM architectures and eco-
nomics in metropolitan areas. SPIE Optical Networks Mag., 1(3): 41–50, 2000.
[Chang+96] G. K. Chang, G. Ellinas, J. K. Gamelin, M. Z. Iqbal, and C. A. Brackett. Multiwave-
length reconfigurable WDM/ATM/SONET network testbed. IEEE/OSA J. Lightwave Technol.,
14(6):1320–1340, 1996.
[Charalambous+03] P. Charalambous, G. Ellinas, C. Dennis, E. Bouillet, J.-F. Labourdette,
A. Akyamac, S. Chaudhuri, M. Morokhovich, and D. Shales. A national mesh network
using optical cross-connect switches. In Proceedings of the IEEE/OSA Optical Fiber Commun.
Conf., Paper FR6, pp. 789–790, Atlanta, GA, March 2003.
[Chen97] Y. J. Chen. Low cost WDM technology for emerging telecommunications applications.
In Proceedings of the IEEE/LEOS Annual Meeting, pp. 390–391, San Francisco, CA, November,
1997.
864 Multiwavelength Optical Networks

[Chen+99] Y. Chen, M. T. Fatehi, H. J. La Roche, J. Z. Larsen, and B. L. Nelson. Metro optical

networking. Bell Labs Technical J., 4(1):163–186, 1999.
[Coathup+95] L. Coathup, C. Y. Lu, and P. Roorda. Economical applications of all-optical net-
works. In Proceedings of the IEEE/LEOS Summer Topical Meeting on Optical Networks,
Keystone, CO, August 1995.
[Cortez02] B. Cortez. The emerging intelligent network: Now a reality. In Proceedings of the
IEEE/OSA Optical Fiber Commun. Conf., Paper WH1, pp. 224–225, Anaheim, CA, March
2002.
[Davey+06a] R. Davey, J. Kani, F. Bourgart, and K. McCammon. Options for future optical
access networks. IEEE Communications Mag., 44(10): 50–56, 2006.
[Davey+06b] R. Davey, P. Haley, I. Hope, D. B. Payne, O. Marmur, J. Ruhmann, and Y. Zuiderveld.
DWDM reach extension of a GPON to 135km. IEEE/OSA J. Lightwave Technol., 24(1):29–31,
2006.
[Davis+98] G. B. Davis, N. Robinson, S. K. Liu, J. Fee, and D. Way. Optical cross-connect system
technology trial. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper TuE3,
pp. 27–28, San Jose, CA, February 1998.
[Dwivedi+00] A. Dwivedi and R. Wagner. Traffic model for the USA long-distance optical
network. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper TuK1, pp.
156–158, Baltimore, MD, February 2000.
[Elby02] S. Elby. Public network reliability and resiliency: Presentation at Columbia Uni-
versity, http://www2.cvn.columbia.edu/course/elene6901/session/S.ElbyNetworkReliability
TalkPT1. ppt, October 30, 2002.
[Elby03] S. Elby. An ILEC’s view of next generation networking. In Proceedings of the IEEE/OSA
Optical Fiber Commun. Conf., Paper ThAA1, pp. 606–607, Atlanta, GA, March 2003.
[Ellinas+03] G. Ellinas, J. Labourdette, J. Walker, S. Chaudhuri, L. Lin, E. Goldstein, and K. Bala.
Transparent optical switches: Technology issues and challenges. IEC Ann. Rev. of Commun.,
56, 2003.
[Esmacher05] R. Esmacher. Optimizing metro networks using pluggable optics. In Proceedings
of the IEEE/OSA Optical Fiber Commun. Conf./National Fiber Optics Engineers Conference,
Paper NThB4, Anaheim, CA, March 2005.
[Farjady+03] F. Farjady, N. Antoniades, and M. J. Yadlowsky. Value of fiber overlays in WDM
metro networks. IEEE Photon. Technol. Lett., 15(2):329–333, 2003.
[Fuller01] M. Fuller. Bandwidth trading no longer a case of “if ” but “when.” Lightwave Mag.,
June 2001.
[Gerstel+99] O. Gerstel and R. Ramaswami. Upgrading SONET rings with WDM instead of
TDM: An economic analysis. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf.,
Paper ThE4, pp. 75–77, San Diego, CA, February 1999.
[Ghani+02] N. Ghani, J. Y. Pan, and X. Cheng. Metropolitan Optical Networks. In Optical Fiber
Telecommunications vol. IV-B, I. Kaminow and T. Li (Editors), San Diego: Academic Press,
2002.
[Goralski02] W. J. Goralski. SONET/SDH. Third Edition. New York: McGraw-Hill 2002.
[Hall+96] E. Hall, J. Kravitz, R. Ramaswami, M. Halvorson, S. Tenbrink, and R. Thomsen.
The RAINBOW-II gigabit optical network. IEEE J. Select. Areas Commun., 14(5):814–823,
1996.
[Hill+93] P. J. Chidgey, F. Kaufhold, T. Lynch, O. Sahlen, M. Gustavsson, M. Janson, B. Lager-
strom, G. Grasso, F. Meli, S. Johansson, J. Ingers, L. Fernandez, S. Rotolo, A. Antonielli,
S. Tebaldini, E. Vezzoni, R. Caddedu, N. Capanio, F. Testa, A. Scavennec, M. J. O’Mahony,
 Current Trends 865

J. Zhou, A. Yu, W. Sohler, U. Rust, and H. Herrmann. Multi-wavelength transport network: A

transport network layer based on optical network elements. IEEE/OSA J. Lightwave Technol.,
11(5/6):667–679, 1993.
[Humblet01] P. Humblet. The direction of optical technology in the metro area. In Proceedings
of the IEEE/OSA Optical Fiber Commun. Conf., pp. WBB1–1-WBB1-3, Anaheim, CA, March
2001.
[Jackel+96] J. L. Jackel, M. S. Goodman, J. E. Baran, W. J. Tomlinson, G.-K. Chang, M. Z. Iqbal,
G. H. Song, K. Bala, C. A. Brackett, D. A. Smith, R. S. Chakravarthy, R. H. Hobbs, D. J. Fritz,
R. W. Ade, and K. M. Kissa. Acousto optic tunable filters for multiwavelength optical networks:
Cross-talk considerations. IEEE/OSA J. Lightwave Technol., 14(6):1056–1066, 1996.
[Johansson96] S. Johansson. A transport network involving a reconfigurable WDM network
layer – A European demonstration. IEEE/OSA J. Lightwave Technol., 14(6):1341–1348, 1996.
[Jourdan+97] A. Jourdan, L. Berthelon, P. Bonno, F. Bruyere, M. Chbat, C. Coeurjolly,
J. Y. Emery, P. Gavignet, E. Grard, C. Janz, A. Noury, G. Soulage, and T. Zami. Fully re-
configurable WDM optical cross-connect: Feasibility validation and preparation of prototype
cross-connect for ACTS open field trials. In Proceedings of the European Conference on Optical
Communications, pp. 3.55–3.57, Edinburgh,UK, September 1997.
[Kaminow+96] I. P. Kaminow, C. R. Doerr, C. Dragone, T. Koch, U. Koren, A. A. M. Saleh,
A. J. Kirby, C. M. Ozveren, B. Schofield, R. E. Thomas, R. A. Barry, D. M. Castagnozzi,
V. W. S. Chan, B. R. Hemenway, Jr., D. Marquis, S. A. Parikh, M. L. Stevens, E. A. Swanson,
S. G. Finn, and R. G. Gallager. A wideband all-optical WDM network. IEEE J. Select. Areas
Commun., 14(5):780–799, 1996.
[Kim+02] H. D. Kim, J.-H. Shin, C.-G. Lim, S.-M. Lee, Y. Cho, C.-H. Lee, and Y. J. Oh.
Demonstration of bi-directional WDM Network. In Proceedings of the IEEE/OSA Optical
Fiber Commun. Conf., Paper WW5, pp. 353–355, Anaheim, CA, March 2002.
[Kohut01] P. V. Kohut. Understanding the need for bandwidth trading. Capacity, 2001.
[Madamopoulos+01] N. Madamopoulos, M. D. Vaughn, L. Nederlof, and R. E. Wagner. Metro
network architecture scenarios, equipment requirements and implications for carriers. In Pro-
ceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper WL2, Anaheim, CA, March
2001.
[Madamopoulos+02] N. Madamopoulos, N. Antoniades, I. Roudas, M. Vaughn, and R. E. Wag-
ner. Design, transport performance study and engineering of a 11Tb/s US metro network.
In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper ThH6, pp. 444–445,
Anaheim, CA, March 2002.
[Melle+95] S. Melle, C. P. Pfistner, and F. Diner. Amplifier and multiplexing technologies expand
network capacity. In Lightwave Mag., pp. 42–46, December 1995.
[Merli+97] S. Merli, F. Arecco, A. Mariconda, F. Pozzi, F. Veghini, and E. Iannone. The
PROMETEO testbed: A unidirectional WDM transparent self-healing ring in a field envi-
ronment. In Proceedings of the European Conference on Optical Communications, Edinburgh,
UK, September 1997.
[Nachum+05] E. Nachum and M. Schneider. Metropolitan WDM: Fact or fallacy? The require-
ments for regional reach in metro applications. In Proceedings of the IEEE/OSA Optical Fiber
Commun. Conf./ National Fiber Optics Engineers Conference, Paper NTuE1, Anaheim, CA,
March 2005.
[Naito+99] T. Naito, N. Shimojoh, T. Tanaka, H. Nakamoto, M. Doi, T. Ueki, and M. Suyama.
1 terabit/s WDM transmission over 10,000 km. In Proceedings of the European Conference on
Optical Communications, Paper PD2-1, Nice, France, September 1999.
866 Multiwavelength Optical Networks

[Noguchi+03] K. Noguchi, Y. Koike, H. Tanobe, and M. Matsuoka. The first field trial of a
wavelength routing WDM full-mesh network system (AWG-STAR) in a metropolitan/local
area. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper THAA5, pp.
611–613, Atlanta, GA, March 2003.
[Noirie+02] L. Noirie, F. O. Dorgeuillem, and A. Bisson. 32x10Gbit/s DWDM metropolitan
network demonstration with 10 waveband-ADMs and 155 km TeraLightTM metro fiber. In Pro-
ceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper TH4, pp. 442–443, Anaheim,
CA, March 2002.
[Parys+00] W. Van Parys, P. Arijs, and P. Demeester. Cost boundaries for economical OXCs.
In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper ThO5, Baltimore, MD,
March 2000.
[Parys+01] W. Van Parys, P. Arijs, O. Antonis, and P. Demeester. Quantifying the benefits of
selective wavelength regeneration in ultra long-haul WDM networks. In Proceedings of the
IEEE/OSA Optical Fiber Commun. Conf., Paper TuT4, Anaheim, CA, March 2001.
[Peloso+04] P. Peloso, D. Penninckx, M. Prunaire, and L. Noirie. Optical transparency of a het-
erogeneous Pan-European network. IEEE/OSA J. Lightwave Technol., 22(11):242–248, 2004.
[Pratt+03] A. Pratt, B. Charbonnier, P. Harper, D. Nesset, B. K. Nayar, and N.J. Doran. 40x10.7
Gb/s DWDM transmission over a meshed ULH network with dynamically reconfigurable optical
cross-connects. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper PD09,
Atlanta, GA, March 2003.
[Ranganathan+02] R. Ranganathan, L. Blair, and J. Berthold. Architectural implications of core
grooming in a 46-node USA optical network. In Proceedings of the IEEE/OSA Optical Fiber
Commun. Conf., Paper ThP2, pp. 498–499, Anaheim, CA, March 2002.
[Reichmann+01] K. C. Reichmann, P. P. Iannone, M. Birk, N. J. Frigo, D. Barbier, C. Cas-
sagnettes, T. Garret, A. Velucco, S. Perrier, and J. Philipsen. An eight-wavelength 160-km
transparent metro WDM ring network featuring cascaded erbium-doped waveguide amplifiers.
IEEE Photon. Technol. Lett., 13(10):1130–1132, 2001.
[Richter+02] A. Richter, H. Bock, W. Fischler, P. Leisching, P. Krummrich, A. Mayer,
R. E. Neuhauser, J. P. Elbers, and C. Glingener. Germany-wide DWDM field trial: Trans-
parent connection of a long-haul link and a multi-channel metro network. In Proceedings of
the IEEE/OSA Optical Fiber Commun. Conf., Paper TH4, pp. 442–443, Anaheim, CA, March
2002.
[Saleh98] A. A. M. Saleh. Islands of transparency – An emerging reality in multiwavelength
optical networking. In Proceedings of the IEEE/LEOS Summer Topical Meeting on Broadband
Optical Networks and Technologies, p. 36, Monterey, CA, July 1998.
[Saleh+99] A. A. M. Saleh and J. M. Simmons. Architectural principles of optical regional and
metropolitan access networks. IEEE/OSA J. Lightwave Technol., 17(12):2431–2448, 1999.
[Saleh00] A. A. M. Saleh. Transparent optical networking in backbone networks. In Proceedings
of the IEEE/OSA Optical Fiber Commun. Conf., pp. 62–64, Baltimore, MD, February 2000.
[Saleh02] A. A. M. Saleh. All-optical networking in metro, regional, and backbone networks.
In Proceedings of the IEEE/LEOS Summer Topical Meeting, p. 15, Mont Tremblant, Quebec,
Canada, July 2002.
[Saleh03] A. A. M. Saleh. Defining all-optical networking and assessing its benefits in metro,
regional and backbone networks. In Proceedings of the IEEE/OSA Optical Fiber Commun.
Conf., Paper WQ1, pp. 410–411, Atlanta, GA, March 2003.
[Saleh06] A. A. M. Saleh and J. M. Simmons. Evolution toward the next-generation core optical
network. IEEE/OSA J. Lightwave Technol., 24(9):3303–3321, 2006.
 Current Trends 867

[Sasaki+99] G. Sasaki and T. Lin. A minimal cost WDM network for incremental traffic. In Pro-
ceedings of the Information Theory and Communications Workshop, pp. 5–7, Kruger National
Park, South Africa, June 1999.
[Simard+05] L. Simard and I. M. White. Impact of RPR and next generation SONET features on
metro Ethernet service delivery costs. In Proceedings of the IEEE/OSA Optical Fiber Commun.
Conf./ National Fiber Optics Engineers Conference, Paper NtuC3, Anaheim, CA, March 2005.
[Simmons02] J. M. Simmons. Analysis of wavelength conversion in all-optical express back-
bone networks. In Proceedings of the IEEE/OSA Optical Fiber Commun Conf., Paper TuG2,
pp. 34–36, Anaheim, CA, March 2002.
[Simmons05] J. M. Simmons. On determining the optimal optical reach for a long-haul network.
IEEE/OSA J. Lightwave Technol., 23(3):1039–1048, 2005.
[Takachio+02] N. Takachio, H. Suzuki, M. Fujiwara, J. Kani, M. Teshima, K. Iwatsuki, and
T. Kitoh. 12.5-GHz-spaced super dense WDM ring network handling 256 wavelengths with
tapped-type OADMs. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper
WW2, pp. 349–350, Anaheim, CA, March 2002.
[Tkach+98] R. W. Tkach, E. L. Goldstein, J. A. Nagel, and J. L. Strand. Fundamental limits of
optical transparency. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper
WJ1, pp. 161–162, San Jose, CA, February 1998.
[Toliver+03] P. Toliver, R. Runser, J. Young, and J. Jackel. Experimental field trial of waveband
switching and transmission in a transparent reconfigurable optical network. In Proceedings of
the IEEE/OSA Optical Fiber Commun. Conf., Paper FR1, pp. 783–784, Atlanta, GA, March
2003.
[Tomkos+02a] I. Tomkos, R. Hesse, N. Madamopoulos, C. Friedman, N. Antoniades, B. Hal-
lock, R. Vodhanel, and A. Boskovic. Transport performance of an 80-Gb/s WDM regional
area transparent network utilizing directly modulated lasers. IEEE/OSA J. Lightwave Technol.,
20(4):562–573, 2002.
[Tomkos+02b] I. Tomkos, M. Vasilyev, J.-K. Rhee, B. Hallock, B. Szalabofka, M. Williams,
S. Tsuda, and M. Sharma. 80x10.7Gb/s ultra long-haul (+4200 km) DWDM network with
reconfigurable broadcast and select OADMs. In Proceedings of the IEEE/OSA Optical Fiber
Commun. Conf., Paper FC1, Anaheim, CA, March 2002.
[Tsurusawa+05] M. Tsurusawa, J. Matsuda, T. Otani, M. Daikoku, S. Aruga, and H. Tanaka.
Field trial of 10-Gb/s resilient packet ring for a next-generation multi-service metropolitan area
network. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf./ National Fiber Optics
Engineers Conference, Paper NThI2, Anaheim, CA, March 2005.
[Vaughn+00] M. D. Vaughn and R. E. Wagner. Metropolitan network traffic demand study. In
Proceedings of the IEEE/LEOS Annual Meeting, Paper MK4, Piscataway, NJ, November 2000.
[Vodhanel+97] R. Vodhanel, L. Garrett, S. Patel, W. Kraeft, J. Chiao, J. Gamelin, R. Wagner,
M. Maeda, J. Pastor, C. Allyn, A. Chraplyvy, Y. Sun, A. Chester, N. Jackman, and G. Newsome.
National scale WDM networking demonstration by the MONET Consortium. In Proceedings
of the IEEE/OSA Optical Fiber Commun. Conf., Paper PD 27-1, Dallas, TX, February 1997.
[Wagner+96] R. E. Wagner, R. C. Alferness, A. A. M. Saleh, and M. S. Goodman. MONET:
Multiwavelength optical networking. IEEE/OSA J. Lightwave Technol., 14(6):1349–1355, 1996.
[Wagner00] R. E. Wagner. Managed reach optical networks. In Proceedings of the IEEE/LEOS
Summer Topical Meetings, pp. IV33–34, Aventura, FL, July 2000.
[Wagner+01] R. E. Wagner, L. Nederlof, and S. De Maesschalck. The potential of optical layer
networks. In Proceedings of the IEEE/OSA Optical Fiber Commun. Conf., Paper TuT3, Ana-
heim, CA, March 2001.
868 Multiwavelength Optical Networks

[Wallace01] A. F. Wallace. Ultra long-haul DWDM: Network economics. In Proceedings of the

IEEE/OSA Optical Fiber Commun. Conf., Paper TuT1, Anaheim, CA, March 2001.
[Wilt97] D. Wilt. Optical networks: Wave of the future. Science Technol. Mag., April 1997.
[Wong00] P. Wong. Network architectures in the metro DWDM environment. In Proceedings of
the National Fiber Optics Engineers Conference (NFOEC), Session D4, pp. 588–596, Denver,
CO, August 2000.
[Wright+04] S. Wright, H. Fahmy, and A. J. Vernon. Deployment challenges for access/metro
optical networks and services. IEEE/OSA J. Lightwave Technol., 22(11):2606–2616, 2004.
[Young+98] M. Young, H. Laor, and E. J. Fontenot. First in-service network application of optical
cross-connects. Opt. Photon. News, 9(4): 48–49, 1998.
[Zapata+04] A. Zapata, M. Duser, J. Spencer, P. Bayvel, I. de Miguel, D. Breuer, N. Hanik, and
A. Gladisch. Next generation 100-gigabit metro Ethernet (100 GbME) using multiwavelength
optical rings. IEEE/OSA J. Lightwave Technol., 22(11):2420–2434, 2004.
[Zeng+05] H. Zeng, C. Huang, A. Vukovic, and J. M. Savoie. Transport performance of an
all-optical metro WDM network based on dynamic all-optical switches. In Proceedings of the
IEEE/OSA Optical Fiber Commun. Conf./ National Fiber Optics Engineers Conference, Paper
NTuE3, Anaheim, CA, March 2005.
A Graph Theory

Graph and hypergraph terminology has evolved over the years. The following definitions
are adapted from [Berge89, Bermond+97, Chartrand+96]. Some of the material in this
appendix is found in other parts of the book. It is repeated here for convenience.

A.1 Graphs

A graph G consists of a set of vertices V (G) and a set of edges E(G), where each edge
e is a pair of distinct vertices (u, v). (If the two vertices are the same, then the edge is
a loop. We rule out these cases.) A graph with vertex set V and edge set E is typically
denoted by G(V, E). If e = (u, v), then u and v are adjacent vertices and e is incident
on u and v. Two edges are adjacent if they are incident on the same vertex. Nonadjacent
edges or nonadjacent vertices are called independent. A set of pairwise independent
vertices of a graph G, which is of maximal cardinality, is called a maximal independent
set. Figure A.1 shows an example of a maximal independent set of vertices (outlined in
dashed circles).
A graph in which every two vertices are adjacent is called a complete or fully connected
graph. The complete graph with n vertices is denoted by K n . Figure A.2 shows K 5 .
A graph G is called bipartite if its vertices can be partitioned into two subsets, V1 and
V2 , (called partite sets) such that every edge of G joins a vertex in V1 to one in V2 . A
complete bipartite graph is one in which an edge exists between every pair of vertices
u ∈ V1 and v ∈ V2 . The complete bipartite graph is denoted by K p,q , where p and q are
the cardinalities of V1 and V2 , respectively. Figure A.3 shows K 3,3 .
Graphs that may contain more than one edge between the same pair of vertices as well
as loops are called multigraphs. (We use the term multigraph here to include graphs as
special cases. But unless otherwise stated, we exclude multigraphs from our discussion.)
The number of vertices in G, |V (G)|, is called the order of G, typically denoted by
n(G). The number of edges in G, |E(G)|, is called the size of G, typically denoted by
m(G).
The degree of a vertex v in a graph G is the number of edges incident on v. A graph
is regular if all of its vertices are of the same degree.
A graph S is a subgraph of a graph G if V (S) ⊆ V (G) and E(S) ⊆ E(G). A subgraph,
S of G, is said to be induced by a vertex set U ⊆ V (G), if V (S) = U and E(S) consists
870 Multiwavelength Optical Networks

Figure A.1 A maximal independent set.

of all edges incident on two vertices in U . A spanning subgraph of a graph G is a

subgraph containing all vertices of G.
A clique of a graph G is a complete subgraph of G.
A u–v walk of a graph is an alternating sequence of vertices and edges: u =
u 1 , e1 , u 2 , . . . , u i , ei , u i + 1 , . . . vk = v, beginning on vertex u and ending on vertex v,
where ei = (u i , u i + 1 ). Its length is defined as the number of edges it contains. A walk
is called a trail if no edge is repeated and is a path if no vertex is repeated. It is open

Figure A.2 The complete graph K 5 .

 Appendix A 871

V1 V2

Figure A.3 The complete bipartite graph K 3,3 .

if u = v and is closed otherwise. (All paths of a graph are trails, but all trails are not
paths.) A closed trail is called a circuit, and a circuit with no repeated vertices is called
a cycle. A graph is called acyclic if it contains no cycles.
The distance, d(u, v), between two vertices in a connected graph G is the minimum
of the lengths of all u–v paths in G. The diameter of a graph is the largest distance
between pairs of its vertices.
A graph is connected if it contains a path between every pair of its vertices. Otherwise,
it is disconnected. A graph G that is disconnected is made up of a number of connected
components, each of which is a subgraph of G.
A graph is k-edge connected if the removal of less than k edges never results in a
disconnected graph. A k-node connected graph is defined in a similar way.
A tree is an acyclic connected graph, and a forest is an acyclic graph with more than
one component. A spanning tree of a connected graph G is an acyclic spanning subgraph
of G. The chords of a spanning tree T of G are the edges of G not included in T . Every
spanning tree T of a connected graph G of order n and size m has n − 1 edges and
m − n + 1 chords. Each chord added to T creates a unique cycle, called a fundamental
cycle.
A bridge of a connected graph is an edge whose removal results in a disconnected
graph. Thus, a graph with bridges is one-edge connected. A cut of a connected graph
is a set of edges whose removal disconnects the graph. We reserve the term cut in this
book to mean a minimal cut, or cutset, which is a set of edges whose removal produces
two components and no subset of which has this property.
872 Multiwavelength Optical Networks

1 4

2 5

3 6

Figure A.4 Orientable cycle double cover for K 3,3 .

A.1.1 Cycle Double Covers

A cycle double cover (CDC) of a connected graph G is a collection C of cycles of G
such that every edge of G is in exactly two of the cycles in C. It has been conjectured
that every bridgeless connected graph has a CDC [Szekeres73].
An orientable cycle double cover (OCDC) of a connected graph G is a CDC with the
property that it is possible to choose a circular orientation for each cycle so that each edge
is taken in opposite directions in its two incident cycles [Alspach+85]. The existence
of OCDCs for all bridgeless connected graphs has also been conjectured. Figure A.4
shows an OCDC for the graph K 3,3 , consisting of three cycles.

A.1.2 Eulerian Graphs

An Eulerian circuit is a circuit that contains all of the edges and vertices of a graph or
multigraph. (In discussing Eulerian properties we include multigraphs.) An Eulerian trail
on a multigraph G is a trail containing all of the edges and vertices of G. A multigraph
is called Eulerian if it has a closed Eulerian trail. A connected multigraph is Eulerian if
and only if all of its vertices are of even degree, and it contains an open Eulerian trail if
and only if all but two of its vertices are of even degree.
Figure A.5(a) is the most famous example of a non-Eulerian multigraph. It is the
model for the Konigsberg Bridge problem mentioned in Section 6.3.5. Each vertex
represents a land mass or island in the city of Konigsberg, and each edge represents
one of the seven bridges in the city. The problem posed to Euler was to find a way to
 Appendix A 873

(a) (b) (c)

Figure A.5 Multigraphs: Non-Eulerian (a, b) and Eulerian (c).

traverse all of the bridges exactly once, returning to the point of departure (i.e., to find
a closed Eulerian trail). Euler proved that this was impossible in Konigsberg and that in
fact closed Eulerian trails exist in a connected multigraph G if and only if all vertices of
G are of even degree. Figure A.5(b) shows a graph containing an open Eulerian trail, and
Figure A.5(c) shows an Eulerian graph formed by adding a vertex and two edges to Figure
A.5(b). Algorithms that can be used to find Eulerian trails are found in [Fleischner90].

A.1.3 Planar Graphs

A graph is planar if it can be drawn on a plane (or, equivalently, on a sphere) so that no
two edges intersect except at their end points (vertices). When drawn in this way it is said
to be embedded on the plane, and the resultant embedding is called a plane graph. (A
given planar graph has many plane embeddings.) Each region of a plane graph bounded
by a set of edges is called a face, including the outer unbounded region, which is called
the outer face. (When embedded on a sphere, a plane graph has no outer face.)
For a connected plane graph G, the number of vertices n, edges m, and faces f are
related by

f = m − n + 2, (A.1)

where f is called Euler’s number.

A necessary and sufficient condition for a graph to be nonplanar is that it contain as
a subgraph either K 5 or K 3,3 , which are known as the Kuratowski graphs.

A.1.4 Matchings in Graphs

A set of pairwise independent edges of a graph (or multigraph) is called a matching.
Of particular interest here (see, for example, Appendix B) is a maximum matching of
a bipartite multigraph. A maximum matching is a matching of maximum cardinality.
874 Multiwavelength Optical Networks

V1

V2

Figure A.6 Maximum matching of a bipartite multigraph.

Figure A.6 illustrates a bipartite multigraph with maximum matching shown as the
dashed edges.

A.1.5 Graph Coloring

A vertex coloring of a graph G is an assignment of colors to the vertices of G so that no
two adjacent vertices have the same color. A minimal vertex coloring accomplishes this
with a minimum number of colors.1 This number is called the graph’s vertex chromatic
number. The vertex chromatic number of a graph is equal to or greater than the order of
its largest clique. Graphs in which the two numbers are equal are called perfect graphs.
An edge coloring of a graph G is an assignment of colors to the edges of G so that
no two adjacent edges have the same color. A minimal edge coloring accomplishes this
with a minimum number of colors. This number is called the graph’s edge chromatic
number.
Minimal coloring and determination of the chromatic number are both NP-complete
problems.2 However, the Dsatur algorithm given in [Brelaz79] is a simple and effective
vertex-coloring heuristic. It is based on saturation degree. The saturation degree of a
vertex is defined as the number of different colored vertices to which it is adjacent. The
steps in the algorithm are as follows:
1
One way of vertex coloring a graph is to partition the vertices into independent sets and to assign the same
color to all vertices in the same set. This, however, is not a practical way of finding a good (close to minimal)
vertex coloring of a large graph.
2
An informal definition of an NP-complete problem (NP meaning nondeterministic polynomial) is that it is
a problem that cannot be solved in polynomial time. That is, the computational complexity of the problem
grows faster than any power of the size of the problem (typically growing exponentially). In the coloring
problem, for example, size might refer to the number of vertices or edges in the graph to be colored.
 Appendix A 875

1. List the vertices in decreasing order of degree.

2. Assign color one to a vertex of maximal degree.
3. Choose an uncolored vertex of maximal saturation degree. If there is more than one,
choose any vertex of maximal degree in the uncolored subgraph.
4. Color the selected vertex with the lowest numbered permissible color.
5. If all vertices are colored, stop; otherwise, go to Step 3.

A.1.6 Digraphs
A digraph D(V, A), or directed graph, is a graph in which the edges A(D) (now called
arcs) are ordered pairs of distinct vertices. (We rule out loops.) Thus, each arc has a
direction. An arc (u, v) is incident from u and incident to v. Similarly, vertex u is adjacent
to v, and v is adjacent from u. The out-degree of a vertex u is the number of arcs adjacent
from u, and the in-degree of a vertex v is the number of arcs adjacent to v. A digraph
is regular if all vertices have the same in- and out-degrees. Walks, paths, distances, and
diameter are defined for digraphs in a fashion similar to graphs, except that directions
must be taken into account. Thus, the distance from a vertex u to a vertex v is the length
of the shortest u − v path, which is generally not the same as the distance from v to u.
The diameter is the maximum of the distances between all ordered pairs of vertices.
Let the reversal of the arc set A be defined as A = {(i, j)|( j, i) ∈ A}. Then given a
digraph D(V, A), its conjugate is defined as D = (V, A).
A subdigraph S of a digraph D is a digraph for which V (S) ⊆ V (D) and A(S) ⊆ A(D).
A bipartite digraph D is a digraph with vertices that can be partitioned into two partite
sets, V1 and V2 , so that each arc of D is adjacent from V1 and adjacent to V2. A complete
bipartite digraph is one in which an arc exists from every vertex in V1 to every vertex in
V2 .
A subdigraph of a digraph D is called a diclique of D if it is a complete bipartite
digraph. A diclique is indicated in Figure A.7 with dashed lines.

A.1.7 Moore Bounds

The Moore bound for graphs gives the maximum possible order NMoore (, D) of a graph
of maximum degree  and diameter D:
D−1

NMoore (, D) = 1 +  ( − 1)i
i =0
( − 1) D − 2
=  > 2. (A.2)
−2
d
The Moore bound for digraphs gives the maximum possible order NMoore (δ, D) for a
digraph of maximum out-degree δ > 1 and diameter D:

d δ D+1 − 1
NMoore (δ, D) = 1 + δ + δ 2 + · · · + δ D = . (A.3)
δ−1
876 Multiwavelength Optical Networks

Figure A.7 A diclique.

A.1.8 Max Flow–Min Cut

Consider a digraph D in which each arc is associated with a capacity or maximum flow.
We choose any vertex s as a source, and any other vertex d as a destination, and wish to
find the maximum possible flow between s and d given the capacity constraints.
Partitioning the vertices of D into two partite sets, V1 and V2 , we call the set of all arcs
(u, v), with u ∈ V1 and v ∈ V2 the cut induced by that partition.3 If s ∈ V1 and d ∈ V2 , the
cut is called an s–d cut. The capacity of an s–d cut is defined as the sum of the capacities
of its arcs, and a minimal s–d cut is an s–d cut with least capacity. Figure A.8 illustrates
these concepts. The cut {(a, c), (b, c), (b, e)} separates the partite sets V1 = {s, a, b} and
V2 = {c, e, d}. The numbers on the arcs indicate capacities. It can be seen that this is a
minimal s–d cut of capacity 11.
The Max Flow–Min Cut Theorem [Ford+64] states that:

The maximum possible flow between a pair of vertices s and d is given by the capacity
of the minimal cut.

A related result is that if the vertices of a digraph D are partitioned into two partite sets
V1 and V2 , then the maximum possible flow from all vertices in V1 to all vertices in V2
is given by the capacity of the cut between V1 and V2 .

3
The concept of a cut induced by a partition is equally valid for graphs, in which arcs are replaced by edges.
 Appendix A 877

Cut

a 4 c

8 5

s 4 3 d
2 6

6 9

b 5 e

Figure A.8 A cut.

A.2 Hypergraphs

Hypergraphs are direct generalizations of graphs or digraphs to the case in which an

edge (now, a hyperedge or hyperarc) may contain more than two vertices.

A.2.1 Undirected Hypergraphs

An undirected hypergraph H consists of a set of vertices ν(H ) and a set of hyperedges
ε(H ), where each hyperedge E ∈ ε(H ) is a subset of ν(H ). We typically denote a
hypergraph by H (ν, ε). The number of vertices in the hypergraph (its order) is n = |ν(H )|
and the number of hyperedges is m = |ε(H )|.
The degree of a vertex is the number of hyperedges containing that vertex. The size
of a hyperedge E ∈ ε(H ) is its cardinality |E|. (We rule out hyperedges of size less than
2. A hypergraph, all of whose hyperedges are of size 2, is a graph.) The rank of H is the
size of its largest hyperedge. A path connecting two vertices u and v in a hypergraph is
an alternating sequence of vertices and hyperedges u = v0 , E 1 , v1 , . . . , E k , vk = v with
{vi − 1 , vi } ⊆ E i for all 1 ≤ i ≤ k. The length of a path is the number of hyperedges in it.
The distance d H (u, v) between two different vertices u and v is the length of a shortest
path connecting them in H . The diameter of H is the maximum of the distances between
all pairs of vertices in H .
An undirected hypergraph of diameter D, maximum vertex degree , and rank r
is called a (, D, r )-hypergraph, and its order is denoted by N (, D, r ). The Moore
bound for a (, D, r )-hypergraph is given by

D−l

N (, D, r ) ≤ 1 + (r − 1) ( − 1)i (r − 1)i . (A.4)
i =0
878 Multiwavelength Optical Networks

A.2.2 Directed Hypergraphs

A directed hypergraph H consisting of a set of vertices ν(H ) and hyperarcs ε(H ) will
be denoted by H (ν, ε).
A hyperarc E ∈ ε(H ) is an ordered pair (E − , E + ) of nonempty subsets of ν(H ).4 The
vertex sets E − and E + are called the in-set and out-set of E respectively. (The in- and
out-sets need not be disjoint.) The quantities |E − | and |E + | are, respectively, called the
in-size and out-size of E. (If the in-size and out-size of all hyperarcs is one, then the
hypergraph is a digraph.)
Let v be a vertex in ν(H ). Its in-degree, d − (v), is the number of hyperarcs that contain
v in their out-set. Similarly its out-degree, d + (v), is the number of hyperarcs that contain
v in their in-set.
A walk in H from vertex u to vertex v is an alternating sequence of vertices and
hyperarcs u = v0 , E 1 , v1 , E 2 , v2 , . . . , E k , vk = v such that vi − 1 ∈ E i− and vi ∈ E i+ for
each 1 ≤ i ≤ k. The length of a walk is equal to the number of hyperarcs in it. The
distance and diameter are defined in the same way as in the undirected case.
A directed hypergraph of diameter D, maximum vertex out-degree d, and maximum
out-size s is called a (d, D, s)-directed hypergraph, and its order is denoted by n(d, D, s).
The Moore bound for a (d, D, s)-directed hypergraph is given by
D

n(d, D, s) ≤ (ds)i . (A.5)
i =0

4
This is not the only way to define a directed hypergraph. See [Ausiello+86] for another approach.
Note: The bibliography for Appendices A through G follows Appendix G.
B Fixed Scheduling Algorithm

The following algorithm for scheduling traffic optimally in a fixed frame is adapted from
[Gopal82]. The terminology used here is defined in Section 5.4.1.
The objective is to find a channel allocation schedule (CAS) that schedules all traf-
fic in a given normalized traffic matrix T in a number of time slots L min , given in
Equation (5.54). Finding an optimal CAS is equivalent to decomposing T into a sum
of matrices C (s) , with elements that are nonnegative integers, satisfying the system
constraints:
 (s)
ci j ≤ αi i = 1, 2, . . . , M (B.1)
j
 (s)
ci j ≤ β j j = 1, 2, . . . , N (B.2)
i
 (s)
ci j ≤ C (B.3)
ij

and the traffic constraints

L min

C (s) = T. (B.4)
s=1

First, some definitions:

r A matrix with nonnegative integer entries that satisfies Equations (B.1) and (B.2) is
called a transmission matrix (TM), and a TM satisfying those equations with equality
is called a complete TM.
r An N × N normalized traffic matrix for which all row and column sums equal some
positive integer B is called a quasidoubly stochastic (QDS) matrix, and B is called
the line sum.
r An M × N normalized traffic matrix for which R′ /β = T ′ /α = B, for all i and j,
j j i i
where B is a positive integer, is called a normalized quasidoubly stochastic (NQDS)
matrix. The quantity B is called the normalized line sum.

The steps in the algorithm are as follows. [These steps apply for the case C <
M N
min i=1 αi , j=1 β j . See comments at the end of the algorithm for the remaining
cases.]
880 Multiwavelength Optical Networks

1. Extend the M × N matrix T into an (M + 1) × (N + 1) NQDS matrix Q as follows.

M
(a) Corresponding to the extra column, define β N +1 = i=1 αi − C, and corre-
N
sponding to the extra row, define α M+1 = j=1 β j − C. Denote the row sums
q q q
and the column sums of Q by ri and c j , respectively. Initially, set r M+1 =
q q q
c N +1 = 0. (For the remaining row and column sums, ri = T i′ and c j = R′j ,
respectively.)
(b) Add entries to row M + 1 of Q as follows. To entry q M+1, j , 1 ≤ j ≤ N , add
q q
x M+1, j = min α N +1 L min − r M+1 , β j L min − c j , (B.5)
q q
updating r M+1 and c j with each addition. The process terminates when no further
q
additions can be made, in which case r M+1 = α N +1 L min .
(c) Add entries to column N + 1 of Q as follows. To entry qi,N +1 , 1 ≤ i ≤ M, add
q q
xi,N +1 = min β M+1 L min − c N +1 , αi L min − ri , (B.6)
q q
updating c N +1 and ri with each addition. The process terminates when no further
q
additions can be made, in which case c N +1 = β N +1 L min .
′
(d) Consider the submatrix Q of Q defined by the first M rows and first N columns
of Q. To entry qi j in this submatrix, add
q q
xi j = min β j L min − c j , αi L min − ri , (B.7)
q q
updating and with each addition. (The matrix Q ′ is just the matrix T with
cj ri
some “dummy” traffic added.)
(e) Now decompose the matrix Q into L min complete TMs in the following steps.
2. Use the column/row–expansion algorithm (see Section B.1) to expand Q to a Q DS
matrix, A.
3. Decompose A into a sum of permutation matrices A = a(1) + a(2) + · · · + a(L min )
(see Section B.2).
4. Use the column/row–compression algorithm (see Section B.3) to reduce the a(i)’s to
q(1), q(2), . . . , q(L min ), where the q(i)’s are complete TMs.
5. Remove the last column and row of each matrix q(i) to form an M × N matrix q ′ (i).
6. Remove dummy traffic from each q ′ (i) to convert it to C (i) by letting C (1) =
min(T, q ′ [1]), and C (i) = min T − i−1 ( j) ′

j=1 C , q [i] for i = 2, 3, . . . , L min .

In the case when C = Nj=1 β j , Q is constructed as an M × (N + 1) matrix. Entries



are added as indicated in Step 1, first to the (N + 1)th column, and then to the submatrix
M
Q ′ to make Q into an NQDS matrix. In the case when C = i=1 αi , Q is constructed
as an (M + 1) × N matrix in an analogous fashion. The various procedures used here
are described next in detail.

B.1 Column/Row–Expansion Algorithm

In Step 2, the matrix Q was expanded to an H × H QDS matrix, A , with line sum L min ,
where H = Nj=1
 +1  M+1
β j = i=1 αi . First, expand Q into an (M + 1) × H matrix Q 1 by
 Appendix B 881

Q
2 3
2 3 7 = (2, 2, 1) = (2, 3) Lmin = 5
2 2 8
1 5 0

B1 B2
3 0 5 2 0 Column-Expansion
2 0 0 3 5
Algorithm
0 5 0 0 0

Row-Expansion
Concatenate Algorithm

3 0 5 2 0
Q1 2 0 0 3 5
0 5 0 0 0 3 0 2 0 0 Concatenate A QDS
A1 0 0 3 2 0 Matrix
3 0 2 0 0
0 0 3 2 0
2 0 0 3 0
A2 2 0 0 3 0 H H
0 0 0 0 5
0 0 0 0 5
0 5 0 0 0
A3 0 5 0 0 0

Decomposition of A into 5 Permutation Matrices

1 1 1 1 1
1 1 1 1 1
a(i) 1 1 1 1 1
1 1 1 1 1
1 1 1 1 1

Row Compression

1 1 1 1 1 1 1 1 1 1
a (i ) 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1

Column Compression

1 1 1 1 1 1 2 2
q(i ) 2 2 2 1 1 1 1
1 1 1 1 1

Figure B.1 Example of decomposition of an NQDS matrix. (From [Gopal82, Figure 2.7]. Used by
permission of the author.)

replacing each column of Q by an (M + 1) × β j matrix B j . The following procedure is

used. Let cla be the column sum of the lth column in B j and rka the row sum of the kth
row of B j . Starting with B j = 0, add to the (k, l) element of B j , the quantity

xkl = min L min − cla , tk j − rka , (B.8)

882 Multiwavelength Optical Networks

= (2, 2, 2) M= 3
2 2
= (2, 3) N= 2
T 3 1 C=4
Lmin = 5
5 0

Add Row 4 and Column 3

2 2 3
3 1 2 ′ = (2, 2, 2, 1)
Q
5 0 5 ′ = (2, 3, 2)

0 5 0

Add to Submatrix Q′

2 5 3
3 5 2
Q′
5 0 5
0 5 0

Decomposition into TMs

1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1
q(i )
1 1 1 1 1 1 1 1 1 1
1 1 1 1 1

1 1 1 1 1 1 1
q′(i) 1 1 1 1 1 1 1 1
1 1 1 1 1

Removal of Dummy Traffic

1 1 1 1
C (i) 1 1 1 1
1 1 1 1 1

Figure B.2 Example of fixed-frame scheduling. (From [Gopal82, Figure 2.8]. Used by permission
of the author.)
 Appendix B 883

executing this operation for each element of B j and updating cla and rka each time. The
process terminates when no further additions can be made.
A similar operation is executed on the rows of Q 1 to give the H × H matrix A . Row
i of Q 1 is replaced by an αi × H matrix A i . Let clb be the l th column sum of A i and rkb
the k th row sum. Then add to the (k, l) element of A i , the quantity
xkl = min L min − rkb , tk′ j − clb , (B.9)
executing this operation for each element of A i and updating clb and rkb each time. The
process terminates when no further additions can be made. The resultant matrix A is an
H × H QDS matrix of line sum L min .

B.2 Decomposition into Permutation Matrices

In Step 3 the QDS expanded traffic matrix A is decomposed into a sum of permutation
matrices. First A is represented by a bipartite multigraph with vertices on the left (right)
representing rows (columns), and with an edge from left to right representing each unit
of traffic in A . Then, a complete maximum cardinality matching is found, defining a
permutation matrix a(1). (See [Lawler76] for an efficient matching algorithm.) This
is subtracted from A and the process is repeated until a decomposition A = a(1) +
a(2) + · · · + a(L min ) is determined.

B.3 Column/Row–Compression Algorithm

In Step 4 the H × H permutation matrices a(i) are compressed into (M + 1) × (N + 1)

matrices q(i). First, a(i) is compressed along the rows to obtain an (M + 1) × H matrix
a ′ (i) as follows. The kth row of a ′ (i) is constructed by taking the sum of the rows of a(i)
corresponding to αk . Similarly, an (M + 1) × (N + 1) matrix q(i) is formed with an lth
column that is the sum of the columns of a ′ (i) corresponding to βl . Each of the matrices
q(i) is a complete TM.
Examples of these procedures appear in Figures B.1 and B.2. Given an NQDS matrix
Q, the expansion, decomposition, and compression steps are illustrated in Figure B.1.
A 3 × 2 NQDS matrix Q is expanded to a 5 × 5 QDS matrix A (Step 2), which is
decomposed into five permutation matrices a(i) (Step 3). These are compressed in turn
into five 3 × 2 matrices q(i).
In Figure B.2, a 3 × 2 traffic matrix T is extended to a 4 × 3 NQDS matrix Q (Step 1),
which is decomposed into five permutation matrices (not shown) via the expansion and
decomposition steps illustrated in Figure B.1. These are compressed in turn into five
4 × 3 matrices q(i), which are reduced to five 3 × 2 matrices q ′ (i) and then to the final
form C (i) (Steps 4, 5, and 6).
C Markov Chains and Queues

At various points in the book, we use stochastic traffic and queueing models to represent
the behavior of a network under conditions of random demand. These are based on
Markov processes as well as some more general queueing models, which are summarized
in this appendix. A readable and comprehensive treatment of these models may be found
in [Kleinrock75].

C.1 Random Processes

Random processes, such as connection requests, contents of packet queues, and so

forth, can be described as sequences of random variables, often called the states of the
process, with state transitions occurring at successive (isolated) time points. (Between
state transitions, the state remains constant.) In discrete state processes, the states take on
discrete (typically integer) values, whereas in continuous state processes the states take
on a continuum of values. For example, a discrete state process might be the length of a
packet queue, whereas a continuous state process might be the random noise generated
in an electrical circuit. In discrete time processes, the transitions are spaced regularly
in time so that a complete description of the process is given by the state sequence
alone. In continuous time processes, the transitions may occur randomly, at any point in
time.
A realization of a random process is a specific sequence. In the case of discrete time
processes, a realization is completely specified as a sequence of states. In continuous
time processes, the transition times must also be specified. For example, if the process
consists of the state of a queue, a realization would consist of the sequence of queue
lengths and the state transition times corresponding to arrivals and departures from the
queue.
In many applications, only the transition (event) times themselves are of interest.
For example, the random process that consists of connection requests that arrive at a
network controller is described completely by the times at which these events occur.
Such processes are called point processes.
An important special case of a point process is the renewal process. A renewal process
is defined in terms of a sequence of event times, ti , where the times between successive
events τi = ti − ti−1 are independent identically distributed random variables. Thus a
renewal process is characterized completely by the PDF of its interevent times, f τ (x).
 Appendix C 885

An important special case of a renewal process is the Poisson process. A Poisson

process is a renewal process with exponentially distributed interevent times:

f τ (x) = λǫ −λx .

The positive parameter λ is the rate of the process.

Another equivalent way of defining a Poisson process is in terms of a set of postulates.
An informal statement of the Poisson postulates is as follows:
r The probability of an event occurring in a very small time interval of width t is λt.
r The probability of more than one event occurring in a very small time interval is
negligible.
r Events occurring in different intervals are statistically independent.

The latter property implies that a Poisson process is “memoryless.” That is, the
occurrence of an event at one point in time has no relation to the occurrence of an event
at any other point in time.1
From these postulates, it is possible to derive the Poisson distribution for the number
of events k occurring in a Poisson process of rate λ in a time interval T :
(λT )k −λT
Prob[k events in time T ] = ǫ . (C.1)
k!
From Equation (C.1) it can be deduced that the interevent times of the Poisson process
are distributed exponentially, as indicated previously. For this reason, the exponential
distribution is often called a memoryless distribution.

C.2 Markov Processes

One of the most useful mathematical models for analyzing a wide variety of stochastic
processes is the Markov process. In general random processes, statistical dependencies
typically exist among various combinations of successive states in the process. A Markov
process may be defined informally as a random process in which any future state depends
only on its current state. That is, the current state embodies the “probabilistic memory” of
the process back to the infinite past. As in general random processes, Markov processes
(or Markov chains) can be defined with discrete or continuous states, and in discrete or
continuous time.2 We focus here on discrete-state continuous-time chains. The simplest
example is a two-state chain, with a state transition diagram shown in Figure C.1.
This might represent a source generating call requests, which is alternately in an active
or idle state, indicated as 1 and 0, respectively. (The active state would correspond to
a call connection.) The state transition rate from 0 to 1 is indicated as r01 and from 1
to 0 as r10 . The significance of a state transition rate in a Markov chain is as follows. If
the chain is in a state i at time t, the probability that it will change to state j in a small
1
Addicts of slot machines do not believe in Poisson processes!
2
The term Markov chain is sometimes reserved for a discrete-time Markov process. We shall use it to mean
any Markov process.
886 Multiwavelength Optical Networks

r01

0 1

r10

Figure C.1 Two-state chain.

time interval, (t, t + t), is given by the state transition rate ri j times the length of the
interval, t.3 Thus, if we denote the time-dependent state probabilities for this chain as
p1 (t) and p0 (t) and use the rates just defined, the evolution of these probabilities in time
is governed by the forward differential equations
dp0
= −r01 p0 + r10 p1 (C.2)
dt
dp1
= −r10 p1 + r01 p0 . (C.3)
dt
A chain is said to be ergodic if, as t → ∞, its state probabilities tend toward a
unique set of positive equilibrium or ergodic probabilities independent of their initial
values. These correspond to a statistical steady state. For example, letting π0 and π1
be the equilibrium values of p0 and p1 , respectively, for the two-state chain, positive
equilibrium probabilities exist for this chain as long as both transition rates are nonzero.
(If one rate is zero, one of the equilibrium probabilities is zero, and if both rates are
zero, the chain remains in its initial state forever.) If the chain is ergodic, the equilibrium
probabilities can be found by setting the derivatives to zero in the forward differential

equations and using the normalization relation i πi = 1. In the case of the two-state
chain, this gives π0 = r10 /(r10 + r01 ).
This discussion generalizes in a natural way to chains with any finite number of states.4
An example is the (C + 1)-state chain shown in Figure 5.31(b), which represents the
behavior of a population of M users competing for C < M channels in a circuit-switched
system.
Markov chains may also be defined with an infinite number of states. An important
example of such a chain is the infinite birth–death process shown in Figure C.2(a).
As indicated in the figure, the “birth rate,” λn , is the transition rate from state n to
state n + 1, and the “death rate,” µn , is the transition rate from state n to state n − 1.
Ergodicity of infinite chains typically depends on certain “stability” conditions related
to the underlying physical phenomena they represent.
The relations among the equilibrium probabilities, πn , for an ergodic birth–death
process are deduced easily by noting that in the steady state a “balance condition” must

3
Because of this property, the time that a Markov chain remains in any given state is distributed exponentially,
with a mean value equal to the reciprocal of the sum of the transition rates from that state.
4
Certain easily verifiable conditions on the state transition diagram are required to ensure ergodicity in the
general case.
 Appendix C 887

0 n

0 1 n n +1

1 n n +1
(a
a)

0 1 n n+1

(b
b)

Figure C.2 Birth–death process.

exist among the average “flows” between states. The balance condition requires that
πn λn = πn+1 µn+1 , n = 1, 2, . . . . Using this relation recursively gives
4 n 6
5 λk−1
πn = π0 , n = 1, 2, . . . , (C.4)
k=1
µk

where
 ∞
4 n 6−1
 5 λk−1
π0 = 1 + . (C.5)
n=1 k=1
µk

The Erlang traffic model described in Section 5.5.1 is an example of a finite birth–
death process with (C + 1) states. It represents an infinite population that generates
connection requests in a Poisson fashion at a rate λ, with exponentially distributed
holding times of average duration µ−1 . The birth rates in this case are λn = λ, and the
death rates are µn = nµ, n = 1, 2, . . . , C.

C.3 Queues

Queues are simply storage devices (buffers or “waiting rooms”) that are useful in any
system in which a resource (e.g., a communication link or bank teller) is being shared
among a number of users (e.g., packets or bank customers) and for which there is random
demand on the resource. For example, in a packet switch, randomness exists because the
packet arrival process is random, and the packet service and/or processing time may be
random as well. In a circuit-switched system, randomness exists because the connection
request process is random and holding times are typically random as well.
A queue is described by its arrival process, its service time distribution, and other
factors such as its service discipline (e.g., FCFS or priorities), the number of servers,
and the buffer size (finite or infinite). To fix the terminology, let us think in terms of a
packet queue. The key parameters defining its operation are its average arrival rate, λ
(packets per second); average service rate, µ (packets per second); and traffic intensity,
888 Multiwavelength Optical Networks

K
n
Server
Arrivals

Figure C.3 A queue.

ρ = λ/µ. The state, n, of the queue is defined as the number of packets it contains. The
performance parameters of interest in a queue are average queue length, average time
spent in the queue, and possibly other quantities such as average packet loss, complete
queue length distributions, and so forth. To illustrate, Figure C.3 shows a single-server
queue in state n, with a finite buffer holding a maximum of K packets.

C.3.1 The M|M|1 Queue

For purposes of analysis, the M|M|1 queue is the simplest. The three arguments in the
queue notation refer to the arrival process (the first M for Markovian), the service process
(the second M, again Markovian), and the number of servers (one). A Markovian arrival
process is a Poisson process. A Markovian service process corresponds to exponentially
distributed service times. For example, in the case of a packet queue, the service time
refers to the time to transmit a packet at some fixed bit rate. It is, approximately,
distributed exponentially if packets have random lengths (in bits) drawn from a geometric
probability distribution.
The significance of the fact that both the arrival and service processes are Marko-
vian is that the queue can be modeled as an infinite birth–death process, as shown in
Figure C.2(b). The chain has an infinite number of states corresponding to an infinite
number of packet buffers. (A more realistic model would be truncated to a finite number
of states reflecting finite buffer capacity.) The birth rates λ represent the average packet
arrival rate to the queue, and the death rates µ represent the average packet service rate.
Taking the queue’s traffic intensity as ρ = λ/µ, we find that the queue is stable (i.e.,
ergodic) if and only if ρ < 1. For all larger values of ρ (including ρ = 1), the average
queue length grows without bound so that all state probabilities tend toward zero as
t → ∞. In the stable case, Equations (C.4) and (C.5) reduce to the geometric queue
length distribution πn = ρ n (1 − ρ), with an average queue length of N̄ = ρ/(1 − ρ).

C.3.2 The M|G|1 Queue

Many typical queues are not Markovian. A commonly encountered class is the M|G|1
queue, where G denotes a general service process. These queues have Poisson arrivals
and a service time τ , with a general probability density function f τ (x). A very simple
example is the M|D|1 queue (D for deterministic service time) with f τ (x) = δ(x −
t0 ), with a constant service time t0 . Markov chains cannot be used for analysis of
M|G|1 queues. Although the details of performance, such as complete queue length
 Appendix C 889

distributions, are not computed easily, the average queue length has the following closed-
form expression:
ρ  ρ 
N̄ = 1 − (1 − µ2 σ 2 ) , (C.6)
1−ρ 2
where µ−1 and σ 2 denote the mean and the variance of the service time distribution,
respectively. Equation (C.6) is known as the Pollaczek–Khinchine formula. Note that in
the case of the M|D|1 queue Equation (C.6) reduces to N̄ = 1 −ρ ρ 1 − ρ2 .

C.3.3 Little’s Formula

There is a very simple and general relationship between the average time spent in a queue
and the average queue occupancy, known as Little’s formula. Denoting the average time
in the queue (including time being served) by T̄ and the average queue occupancy
(including the packet in service) as N̄ , Little’s formula is
N̄ = λT̄ , (C.7)
where λ is the average arrival rate to the queue. For example, for an M|M|1 queue,
Equation (C.7) gives T̄ = 1/µ(1 − ρ).
D A Limiting-Cut Heuristic

The concept of the limiting cut, introduced in Section 6.3.1.2, stems from the Min Cut–
Max Flow relation in multicommodity flow problems. We first give a brief summary of
this problem and then present a heuristic for finding limiting cuts.

D.1 The Multicommodity Flow Problem and Limiting Cuts

In the most common version of the multicommodity flow problem, a set of demands are
prescribed between source-destination node pairs in a network with a given topology
and link capacities. (Each source-destination demand is known as a commodity, and the
network can be anything—gas pipelines, airline routes, highways, and so on.) The basic
issue is whether the prescribed demands can be satisfied within the capacity constraints;
that is, whether all commodities can be routed through the network (in a bifurcated
manner if necessary) so that the total flow of all commodities on each link does not
exceed its capacity. If so, the demands are said to be feasible.
In wavelength-routed networks (WRNs), the commodities (demands) are LCs, each
requiring one λ-channel, so the capacity of a cut Ci is Fi W, where Fi is the number of
fiber pairs in the cut and W is the number of available wavelengths. Because a channel
in a WRN is a single point-to-point entity, bifurcated routing is not permitted in a WRN.
(An exception would be a case in which several λ-channels are required to carry the
flow on one LC.)
The relations between cut capacities and feasible demands were stated in Section A.1.8
for the single-commodity case. In the multicommodity case, which is of interest here,
the relations are considerably more complex. However, it is clear that a prescribed set of
demands is not feasible if a “bad cut” exists with a capacity that is less than the total flow
required from all sources on one side of the cut to all destinations on the other side. In the
context of WRNs, the capacity of a cut is proportional to W. Thus, the capacity of all bad
cuts can be scaled up by increasing W so that all bad cuts can be eliminated by making
W sufficiently large. (However, the nonexistence of bad cuts does not guarantee that the
prescribed connections can be accommodated on the network.) This leads directly to the
expression for WLimcut given in Equation (6.14), which is a necessary (but not sufficient)
condition on the number of available wavelengths required to support a given set of LCs.
Even though W Limcut is only a lower bound on the actual number of required wave-
lengths, experiments with realistic network topologies indicate that the bound is often
 Appendix D 891

tight. Thus it is important to be able to compute it (or something close to it), especially
in cases of large networks in which heuristics must be relied on for routing and channel
assignment.

D.2 A Heuristic

Because the complexity of computing W Limcut grows exponentially with the number
of network nodes N , it is out of the question to compute it for large N . Instead we
present a heuristic for determining a lower bound, W Limcut , on W Limcut for a prescribed
set of M LCs on a network with a given topology. The algorithm is based on examining
randomly selected cuts of the network rather than all cuts and has a computational
complexity O(N 3 ).
Denote the set of network nodes by V , where N = |V |. Let C x y be a cut induced by
partitioning V into two sets X and Y . Let Mx y be the number of prescribed LCs from X
to Y and Fx y be the total number of fiber pairs in all links comprising the cut C x y . Two
procedures are used: Swap and Limcut.

D.2.1 Swap (X, Y )

This procedure determines a lower bound, W0 , on the number of wavelengths required
to support LCs across cuts between subsets of nodes X and Y of cardinality k and
N − k, respectively, and also determines a node partition (X 0 , Y0 ) with connections
across the cut that require W0 wavelengths.

procedure Swap ( X, Y )

Initially All nodes are unmarked.

let W0 = W = Mxy /Fxy
let X0 = X and Y0 = Y
repeat
M(X−i+j),(Y−j+i)
Select two unmarked nodes i ∈ X and j ∈ Y so that
F(X−i+j),(Y−j+i)
is maximum.
Swap i and j so that X ← X + j - i and Y ← Y + i - j.
Update W and mark i and j.
if W ≥ W0 then
W0 ← W, X0 ← X and Y0 ← Y.
end if
until all nodes in X or Y are marked
return ( X0 , Y0 , W0 )
end Swap
892 Multiwavelength Optical Networks

D.2.2 Limcut
The procedure Limcut determines W Limcut , a lower bound on W Limcut , by applying the
Swap procedure to partitions of V containing k nodes, with k ranging from 1 to N /2.

procedure Limcut

WLimcut = 0

repeat q times

for k = 1 to N/2

Randomly partition V into a set Xk , containing k nodes,

and Yk , containing N − k nodes.

let Wk − MXk Yk /FXk Yk

repeat

( X0 , Y0 , W0 ) = Swap ( Xk , Yk )

if W0 > Wk then

Wk ← W0

Xk ← X0

Yk ← Y0

end if

until no increase is observed after 2 consecutive swaps

end for

if maxk Wk > WLimcut then

WLimcut ← maxk Wk

end if

end repeat

return WLimcut

end Limcut

The parameter q is chosen sufficiently large (e.g., q = 200) to ensure that W Limcut ≈
W Limcut .
E An Algorithm for
Minimum-Interference
Routing in Linear
Lightwave Networks
Given a source s and destination d for a point-to-point connection on a selected waveband
in an LLN, the Min-Int algorithm presented here attempts to find a minimum-interference
optical path p = s, d for that connection on the given waveband. The exact sense in
which interference is minimized requires some explanation and is defined in Section E.3.

E.1 The Image Network

The approach used to find a path that minimizes interference is based on shortest path
calculations, where the path “length” takes into account weights or “lengths” representing
currently active interfering signals. These weights are associated with nodes rather than
links. A useful way of visualizing the node-weighting procedure is shown in the image
network of Figure E.1.1 In the network shown in the figure, each node of the original
network is “blown up” to create additional intranodal links between each input/output
port pair. This is nothing more than a representation of the internal structure of the
LDC on the chosen waveband (see Figure 2.19[b]). The image network of Figure E.1
corresponds to the state of activity in the network of Figure 6.55. Two optical connections,
(1, 1∗ ) and (2, 2∗ ), are active, with signal S 1 transmitted from station 1 to 1∗ and signal
S 2 transmitted from station 2 to 2∗ .
We shall denote an internodal link from node i to node j by (i, j) and assign it a
positive weight d(i, j). An intranodal link inside node k joining inbound link (i, k)
to outbound link (k, j) will be assigned a weight Ik (i, j), representing the additional
interference accumulated by inseparability as a signal traverses node k on a path i–k– j.
The transnodal weight Ik (i, j) is computed as follows. Let G(i, k) be the set of signals
carried on link (i, k) on the chosen waveband, and let M be a large number (greater than
the sum of all internodal link weights). Then

Ik (i, j) = M|G(k, j) − G(i, k)|. (E.1)

For example, in Figure E.1 we have G(B, D) = {S 1 , S 2 }, G(A, B) = {S 1 }, and

G(C, B) = {S 2 }. Thus, I B (A, D) = I B (C, D) = M. The transnodal weights for the cur-
rent state of the network (two connections in progress) are shown in the figure.

1
This is shown for illustrative purposes only. The algorithm does not require the construction of an image
network.
894 Multiwavelength Optical Networks

1*

S1 + S2

1
B
S1
A D F
0
S1
M S1 + S2 0 S1 + S2 0 S1 + S2
0 G
0 0
M 0
2M 2M
0
S2 0
C
S1 + S2
M
0 0 E
0 3* 2*
0
S2

2 3

Figure E.1 Image network.

Now a new connection from station 3 to 3∗ is to be routed on a minimum-interference

path (i.e., one for which the number of interfering signals accumulated along the way is
minimum). The algorithm attempts to find such a path by choosing a path p = 3, 3∗ ,
which is shortest according to the assigned internodal and transnodal weights. In the
case at hand, if all internodal links have weight one, the shortest path for this connection
is 3–C–D–E–F–3∗ , for which the total “length” is five, which is the total number of
hops (including the access links). Note that the minimum-hop path 3–C–D–F–3∗ is not
chosen. It has a total length 4 + 2M, indicating that there are two interfering signals on
this path due to inseparability. The five-hop path detours around the link carrying these
signals and thereby avoids interference. In general, using this approach, the effect of the
transnodal weights outweighs all link weights, so paths are found based on interference,
with ties broken by the internodal link weights. If W is the total length of a path based on
the assigned weights, then ⌊W/M⌋ is the number of signals interfering with the desired
signal on the path in question.

E.2 The Min-Int Algorithm

The Min-Int algorithm for finding minimum-interference shortest paths is presented

here. It is based on Dijkstra’s shortest path algorithm [Tarjan83], modified to include the
transnodal weights.
 Appendix E 895

Min-Int Algorithm

Finds the shortest paths from node 1 to all other nodes in a network
based on internodal weights d(i, j) and transnodal weights Ik (i, j) (as
defined earlier).

INITIALIZATION

To each node, i, assign a label, li , a weight, wi (denoting distance

from node 1 to i), and a predecessor, pi (denoting the node preceding
i on the shortest path from 1 to i). The label li = 0 (li = 1) indicates
that the weight and predecessor are temporary (permanent). Set
w1 = p1 = 0,
and l1 = 1. For all other nodes, set wi = ∞, pi = 0, and li = 0.

BEGIN

STEP 1: Find a node i such that li = 0 and wi is minimum among all

nodes
labeled temporary (lj = 0). Set li = 1.

STEP 2: For every node k adjacent to i:

IF wi + Ii (pi , k) + d(i, k) < wk

THEN Set wk = wi + Ii (pi , k) + d(i, k). Set pk = i.

STEP 3: IF all nodes are labeled permanent, then the algorithm

terminates and the minimum-interference path to any node t can be
obtained by tracing back the predecessors from t to 1.

ELSE GOTO STEP 1.

END

E.3 Minimum Interference

The paths found by Min-Int are not necessarily feasible because they have not been
checked for DSC violations, color clashes, or channel availability. However, they have
the following minimum-interference property:
Theorem: If the DSC condition is satisfied on a path p = s, d , found by the Min-Int
algorithm, then it is the path along which the maximum incremental interference is
minimized from among all paths between s and d that satisfy the DSC constraint on
the chosen waveband.2
In this theorem the incremental interference associated with a path p is defined as the
additional interference caused at a receiver on the chosen waveband due to the activation
of a connection on path p. The maximization is over all stations that receive some
additional interference due to the activation of the new connection.

2
See [Bala92] for proof.
F Synopsis of the SONET Standard

SONET, the acronym for synchronous optical network, is currently the prevailing stan-
dard for high-speed digital transmission in North America. Introduced in the 1980s,
it replaced an earlier standard, the plesiochronous digital hierarchy (PDH), which had
been in place for more than two decades prior to the introduction of the SONET standard
[Ballart+89]. The most frequently used lower levels of the PDH system are the DS1
(1.544 Mbps, designed to carry 24 64-Kbps digitized voice signals plus synchronizing
overhead) and DS3, running at 44.736 Mbps. An architecture similar to SONET, the
synchronous digital hierarchy (SDH), is currently used in Europe and Japan, replacing
an earlier European hierarchy similar to the PDH system. SONET can carry PDH bit
streams as well as many other types of digital traffic (e.g., ATM cells) as part of its pay-
load. One of the most important features of SONET is its highly organized protection
capability [Wu92].
The basic building block (i.e., the first level) of the SONET signal hierarchy is called
the synchronous transport signal-level 1 (STS-1). STS-1 has a bit rate of 51.84 Mbps
and is divided into two portions, transport overhead and information payload, and the
transport overhead is divided further into line and section overheads. (A line is composed
of one or more sections in series, separated by electronic regenerators.) The line overhead
is terminated at SONET terminals and add/drop multiplexers (ADMs), and the section
overhead is terminated at regenerators. The STS-1 frame consists of 90 columns and
nine rows of eight-bit bytes as shown in Figure F.1. The transmission order of the bytes
is row by row, from left to right, with one entire frame being transmitted every 125 µs.
The first three columns contain transport overhead, and the remaining 87 columns and
nine rows (a total of 783 bytes) carry the STS-1 synchronous payload envelope (SPE).
As an example, a DS3 signal can be carried within an STS-1 SPE.
A SONET path is the basic end-to-end connection entity. The SPE contains a nine-byte
path overhead and is used for end-to-end service performance monitoring.
Optical carrier level 1 (OC-1) is the lowest level optical signal used at equipment
and network interfaces. The OC-1 signal is obtained from an STS-1 bit stream after
scrambling and electrical-to-optical conversion.
The overhead bytes are used in SONET for network control and management. This
includes a variety of functions such as performance monitoring and protection switching.
As an example, the K1 and K2 bytes form a two-byte APS message channel.
Higher rate optical signals are formed by byte interleaving an integral number of STS-
1s. Figure F.2 shows an example of how an OC-N signal is formed in the multiplexing
 Appendix F 897

90 Columns
3 Columns 87 Columns

Framing Framing STS-1 ID

87 Columns
A1 A2 C1
Section BIP.8 Orderwire User Trace
Overhead B1 E1 F1 J1
Datacom Datacom Datacom BIP 8
D1 D2 D3 B3
Pointer Pointer Point Signal
H1 H2 Action H3 Label C2
9 BIP.8 APS APS
Rows B2 K1 K2 Path
Status G1 9
Datacom Datacom Datacom Rows
D4 D5 D6 User
Line
Channel F2
Overhead Datacom Datacom Datacom
D7 D8 D9 Multiframe
H4
Datacom Datacom Datacom
D10 D11 D12 Growth
Z3
Growth Growth Orderwire
Z1 Z2 E2 Growth
Z4
Transport Overhead Growth
Z5
Path
STS-1 Payload Capacity
Overhead
SPE

Figure F.1 SONET STS-1 frame and overhead channels. (From [Wu92, Figure 2.7]. Copyright

1992
c Artech House, Inc. Reproduced by permission of Artech House, Inc.)

side of a SONET TDM terminal. Services like DS1 and DS3 are mapped into an SPE.
A number of STS-1s are then byte interleaved and multiplexed to form an STS-N. The
frame is scrambled after section overhead (except framing and STS-N ID) is added.
Framing and STS-N ID are then added into the section overhead of the scrambled
STS-N, and finally the STS-N signal is converted into the optical signal OC-N. For
example, an OC-48 signal might carry 48 DS3s. The reverse operations are executed on
the demultiplexing side.
The OC-N line rate is exactly N times that of OC-1, and all higher rate standard
interface signals are readily defined in terms of STS-1. An OC-N signal may be formed
from various combinations of lower speed signals. For example, an OC-12 might be
composed of 12 OC-1s, four OC-3s, or any other combination resulting in the correct
aggregate bit rate.
The STS-1 SPE can carry one DS3 signal (called a tributary) or a mix of sub-DS3
signals. The DS1 is a typical sub-DS3 signal that can be mapped into a SONET unit
called the virtual tributary-1.5 (VT-1.5). An STS-1 SPE can carry as many as 28 VT-1.5s.
VT-1.5 has an actual bit rate of 1.728 Mbps.
Broadband services requiring more than the capacity of one STS-1 are transported by
concatenated STS-1s. For example, high-definition television signals requiring 135 Mbps
can be carried by three concatenated STS-1s, denoted by STS-3c, which has transport
overheads and payload envelopes that are aligned.
 L
SPE O
H

SPE Synchronization
(Frame Alignment)

SPE

Service Service P L STS-N L L

O SPE O 1:N O .... O
Payload Payload H H MUX SPE H H

P L
Service O O
H H

L
SPE O
H

Byte-Interleaved
Multiplexing

Scramble
STS-N

T T STS-N T T STS-N
OC-N O/E
o e O ... O O ... O
H H SPE H H SPE

SOH
Framing and STS-N ID except
Framing and STS-N ID
LOH = Line overhead; POH = Path overhead;
SOH = Section overhead; TOH = Transport overhead

Figure F.2 Creating an OC-N signal. (From [Wu92, Figure 2.8]. Copyright
1992
c Artech House,
Inc. Reproduced by permission of Artech House, Inc.)

TOH
(9 Columns) 3 × 87 Columns
A1 A1 A1 A2 A2 A2 C1 C1 C1
SOH B1 E1 F1 3 × 87 Columns
D1 D2 D3 J1
H1 H1* H1* H2 H2* H2* H3 H3 H3 B3
B2 B2 B2 K1 K2 C2
D4 D5 D6 G1
LOH 9
D7 D8 D9 F2 STS-3c Payload Capacity Rows
D10 D11 D12 H4
Z1 Z1 Z1 Z2 Z2 Z2 E2 Z3
Z4
Z5

POH
(1 Column)
* Concatenation Indication
Unused Overhead Byte

Figure F.3 Structure of a concatenated SONET frame. (From [Wu92, Figure 2.9]. Copyright
1992
c
Artech House, Inc. Reproduced by permission of Artech House, Inc.)
 Appendix F 899

Figure F.3 shows the frame structure of a SONET STS-3c signal. In the format shown
in the figure, the first of three H1 and H2 bytes contains a valid SPE pointer, whereas
the second and third H1 and H2 bytes contain a concatenation indicator that prevents
the STS-3c signals from being demultiplexed. Concatenation specifies that these signals
are considered as a single unit and are to be transferred as such through the network.
When two or more SONET lines are incident on a network node, they are typically
switched via a SONET ADM or DCS, which in current systems operates on the signals
at the DS3 level. Thus, as shown in Figure 11.5, for example, the SONET terminals
demultiplex the OC-N signals down to the DS3 tributaries, and the DCS is responsible
for routing the DS3s to the correct outbound lines. In addition, the DCS implements
protection and grooming functions.1
1
Grooming in SONET packs tributary signals efficiently into higher rate optical carriers. For example,
grooming assigns DS3s to an OC-48 signal such that their paths all terminate at the same Central Office
without requiring any demultiplexing along the way. Efficient grooming reduces the need to gain access to
tributary signals at intermediate network nodes before they reach their destination.
G A Looping Algorithm

We present a looping algorithm that can be used to establish connections through the
Benes switch described in Section 2.3.2.1 [Benes65, Hui90]. Because this is a rearrange-
ably nonblocking switch, the algorithm begins with a complete list of all connections
requested and with no switches set. It finds all required switch settings recursively, be-
ginning with the three-stage factorization of the n × n switch shown in Figure 2.16. The
following procedure executes one recursion, in which first- (input) stage and third- (out-
put) stage 2 × 2 switch settings are determined for all connections, so as to route them
either via the upper second-stage switch (U ) or the lower switch (L) (see Figure G.1).
Subsequent recursions are required to determine the switch settings within U and L.

Looping Algorithm

INITIALIZATION

Choose any input switch and denote it by I.

BEGIN

STEP 1: (Forward loop) Set I to route an unconnected input via U to

its output stage switch and denote that switch by O.

STEP 2: (Backward loop) Route the connection terminating on the

unused output of O backward through L, to its input switch, and
denote that switch by I.

STEP 3: (Loop termination) If I has no unused input GO TO STEP 4.

ELSE GO TO STEP 1.

STEP 4: (New loop) Choose any unused input switch, denote it by I

and GO TO STEP 1. If there are no unused switches the algorithm
terminates.

END

One execution of the looping algorithm described above establishes complete sets
of required connections through the two second-stage switches U and L. Unless
they each consist of a single 2 × 2 switch (in which case the procedure terminates),
these switches can themselves be factored into three stage networks of the form of
 Appendix G 901

1 1
2 2
U
3 3
4 4

5 5
6 6

L
7 7
8 8

Figure G.1 Looping, first step.

Figure 2.16. The looping algorithm is then applied to these smaller switches, and the
recursion is continued in this manner until all 2 × 2 switch settings are found.
As an example, Figure G.1 shows the first step in applying the algorithm to an 8 × 8
switch to make the connections, 1 → 5, 2 → 4, 3 → 1, 4 → 2, 5 → 6, 6 → 3, 7 → 7,
8 → 8. A forward loop (shown as a solid line) is made from input 1 through U to output
5, and then looped back from output 6 to input 5, then looped forward from input 6 to
output 3, and then looped back from output 4 to input 2, where it terminates. Then a new
forward loop (shown dashed in the figure) is started on input 3, to output 1, and then
looped back from output 2 terminating on input 4. A final loop is shown as a dash-dot
line, involving inputs and outputs 7 and 8. A second recursion determines the final switch
settings as shown in Figure G.2.

1 1
2 2

3 3
4 4

5 5
6 6

7 7
8 8

Unmarked switches are in bar state.

Those marked with an X are in cross state.

Figure G.2 Final switch settings.

902 Multiwavelength Optical Networks

Appendices A–G Bibliography

[Alspach+85] B. R. Alspach and C. D. Godsil. Cycles in Graphs, Annals of Discrete

Mathematics, vol. 115. Amsterdam: Elsevier Science 1985.
[Ausiello+86] G. Ausiello, A. D’Atri, and D. Sacca. Minimal representation of directed
hypergraphs. SIAM J. Comput., 15(2):418–431, 1986.
[Bala92] K. Bala. Routing in Linear Lightwave Networks. Ph.D. thesis. Columbia
University, New York, 1992.
[Ballart+89] R. Ballart and Y.-C. Ching. SONET: Now it’s the standard optical network.
IEEE Communications Mag., 27(3):8–15, 1989.
[Benes65] V. E. Benes. Mathematical Theory of Connecting Networks and Telephone
Traffic. New York: Academic Press 1965.
[Berge89] C. Berge. Hypergraphs, North-Holland Mathematical Library, vol. 45.
Amsterdam: Elsevier Science 1989.
[Bermond+97] J.-C. Bermond, R. Dawes, and F. O. Ergincan. deBruijn and Kautz bus
networks. Networks, 30(3):205–218, 1997.
[Brelaz79] D. Brelaz. New methods to color the vertices of a graph. Commun. ACM,
22(2):251–256, 1979.
[Chartrand+96] G. Chartrand and L. Lesniak-Foster. Graphs and Digraphs. New York:
Chapman & Hall 1996.
[Fleischner90] H. Fleischner. Eulerian Graphs and Related Topics, Part I: Vol. 1. Annals
of Discrete Mathematics, 45. Amsterdam: Elsevier Science 1990.
[Ford+64] L. Ford and D. Fulkerson. Flows in Networks. Princeton, NJ: Princeton
University Press 1964.
[Gopal82] I. S. Gopal. Scheduling Algorithms for Multibeam Communications Satel-
lites. Ph.D. thesis. Columbia University, New York, 1982.
[Hui90] J. Y. Hui. Switching and Traffic Theory for Integrated Broadband Networks.
Norwell, MA: Kluwer Academic 1990.
[Kleinrock75] L. Kleinrock. Queueing Systems, vol. 1. New York: Wiley 1975.
[Lawler76] E. Lawler. Combinatorial Optimization: Networks and Matroids. Philadel-
phia: Saunders College 1976.
[Szekeres73] G. Szekeres. Polyhedral decomposition of cubic graphs. Bull. Austral.
Math. Soc., 8:367–387, 1973.
[Tarjan83] R. E. Tarjan. Data Structures and Network Algorithms. Philadelphia: Society
for Industrial and Applied Mathematics 1983.
[Wu92] T.-H. Wu. Fiber Network Service Survivability. Norwood, MA: Artech House
1992.
Acronyms

1R Reamplification
2R Reamplification and reshaping
3R Regeneration with retiming and reshaping
ACTS Advanced communications technology and service
ADH Average distance heuristic
ADM Add/drop multiplexer
AIP American Institute of Physics
AIS Alarm indication signal
AM Amplitude modulation
AOLS All-optical label swapping
AON All Optical Network
AOTF Acousto-optic tunable filter
AP Access point
APD Avalanche photodiode
APON ATM-based passive optical network
APS Automatic protection switching
ARPA Advanced Research Projects Agency
ARQ Automatic repeat request
AS Autonomous system
ASE Amplified spontaneous emission
ASIC Application-specific integrated circuit
ASK Amplitude shift keying
ASON Automatically Switched Optical Network
ATDnet Advanced Technology Demonstration Network
ATM Asynchronous transfer mode
ATMOS ATM Optical Switching
AWG Arrayed waveguide grating
BER Bit error rate
BERT Bit error rate tester
BGP Border Gateway Protocol
BH Buried heterostructure
904 Multiwavelength Optical Networks

B-ISDN Broadband integrated services digital network

BLSR Bidirectional line-switched ring
BLT Balanced light-tree
BP Blocking probability
BPF Bandpass filter
BPON Broadband passive optical network
BPSK Binary phase shift keying
BS Base station
C3 Cleaved–coupled cavity
CAS Channel allocation schedule
CATV Cable television
CCR Counterclockwise ring
CDC Cycle double cover
CDMA Code division multiple access
CE Core exchange
CH Contraction heuristic
CIR Carrier-to-interference ratio
CLEC Competitive local exchange carrier
CLNP Connectionless Network Protocol
CNR Carrier-to-noise ratio
CO Central Office
CORD Contention Resolution by Delay Lines
CoS Class of service
CR Clockwise ring
CRC Cyclic redundancy check
CR-LDP Constraint-based Label Distribution Protocol
CRO Contention resolution optics
CRZ Chirped return-to-zero
CS Central station
CSMA/CD Carrier-sense multiple access with collision detection
CSO Composite second order
CSRZ Carrier suppressed return-to-zero
CTB Composite triple beat (interference)
CWDM Coarse wavelength division multiplexing
DA Destination address
DAH Dual ascent heuristic
DARPA Defense Advanced Research Projects Agency
DBR Distributed Bragg reflector
DCA Distinct channel assignment
DCF Dispersion-compensating fiber
 Acronyms 905

DCN Data Communications Network

DCS Digital cross-connect system
DDP Deflect and drop
DFB Distributed-feedback
DFC Difference frequency converter
DGD Differential group delay
DI Delay interferometer; Dump-and-insert
DLC Data link control
DML Directly modulated laser
DMUX Demultiplexer
DNH Distance network heuristic
DP Drop policy
DPRING Dedicated protection ring
DPSK Differential phase-shift keying
DPSK/IM Differential phase-shift keying/intensity modulation
DQDB Distributed queue dual bus
DR Delayed reservation
DRA Distributed Raman amplifier
D-RZ Duobinary return-to-zero
DSB Double sideband
DSC Distinct source combining
DS-CDMA Direct sequence code division multiple access
DSDP Deflect, segment, and drop
DSE Dynamic spectrum equalizer
DSF Dispersion-shifted fiber
DSL Digital subscriber line
DSn Digital signal-level n
DSS Distributed slot synchronization
DWDM Dense wavelength division multiplexing
EA Electro-absorption
EA-DFB Electro-absorption distributed-feedback
EAM Electro-absorption modulator
ECOC European Conference on Optical Communications
ED Electrical detector
EDC Electronic dispersion compensation; Electronic dispersion compensator
EDF Erbium-doped fiber
EDFA Erbium-doped fiber amplifier
EDWA Erbium-doped waveguide amplifier
EFEC Enhanced forward error correction
EFM Ethernet in the First Mile
906 Multiwavelength Optical Networks

EGP Exterior Gateway Protocol

EML Externally modulated laser
E-NNI Exterior Network-Network Interface
EON European Optical Network
EPON Ethernet-based passive optical network
ETDM Electrical time division multiplexing
FBG Fiber Bragg grating
FCFS First-come-first-served
FDDI Fiber distributed data interface
FDL Fiber delay line
FEC Forward error correction; Forwarding Equivalence Class
FFH-CDMA Fast frequency hopping code division multiple access
FITNESS Failure Immunization Technology for Network Service
Survivability
FM Frequency modulation
FP Fabry–Perot
FPGA Field programmable gate array
FR Frame relay
FSAN Full services access network
FSK Frequency-shift keying
FSK/IM Frequency-shift keying/intensity modulation
FSR Free spectral range
FTTB Fiber to the building
FTTC Fiber to the curb
FTTcab Fiber to the cabinet
FTTH Fiber to the home
FTTN Fiber to the node
FT-TR Fixed transmitters with tunable receivers
FWHM Full width at half maximum
FWM Four-wave mixing
GbE Gigabit Ethernet
GCSR Grating-assisted codirectional coupler with sampled grating reflector
GMPLS Generalized Multiprotocol Label Switching
GPON Gigabit-capable passive optical network
GRIN Graded refractive index
GUI Graphic user interface
GVD Group velocity dispersion
HD-SDI High definition serial digital interface
HDTV High-definition television
HEC Header error correction
 Acronyms 907

HFC Hybrid fiber-coaxial

HNLF Highly nonlinear fiber
HOL Head-of-the-line (blocking)
IAD Integrated access device
ID Identifier
IDT Interdigital transducer
IETF Internet Engineering Task Force
IF Intermediate frequency
IGP Interior Gateway Protocol
ILEC Incumbent local exchange carrier
ILP Integer linear program
IM/DD Intensity-modulated direct detection
I-NNI Interior Network–Network Interface
IOF Interoffice
IOS Intelligent optical switch
IP Internet Protocol
IRP Island Restoration Protocol
IPTV Internet Protocol Television
ISI Intersymbol interference
IS-IS Intermediate System to Intermediate System
ISO International Standards Organization
ISIS-TE Intermediate System to Intermediate System Protocol with Traffic
Engineering extensions
ISP Internet service provider
ITU International Telecommunications Union
ITU-T International Telecommunications Union–Telecommunication
Standardization Sector
JET Just enough time
JIT Just in time
KEOPS KEys to Optical Packet Switching
k-SP k shortest path
λSXC λ-channel-selective cross-connect
LAN Local area network
LAS Local access subnet
LAUF-VF Latest available unused-channel first with void filling
LC Logical connection; Liquid crystal
LCC Lost calls cleared
LCG Logical connection graph
LCH Logical connection hypergraph
LD Laser diode
908 Multiwavelength Optical Networks

LDC Linear divider-combiner

LDP Label Distribution Protocol
LDR Limited deflection routing
LE Local exchange
LEC Local exchange carrier
LED Light-emitting diode
LER Label edge router
LIB Label Information Base
LL Logical layer
LLC Logical link control
LLID Logical link identifier
LLN Linear lightwave network
LMP Link Management Protocol
LN Logical network
LN-MOD Lithium niobate modulator
LN-MZM Lithium niobate Mach–Zehnder modulator
LO Local oscillator
LOA Linear optical amplifier
LOBS Labeled optical burst switching
LP Logical path
LPF Low-pass filter
LRN Logically routed network
LS Logical switch
LSA Link State Advertisement
LSN Logical switching node
LSP Label-switched path
LSR Label-switching router
LT Logical topology
LTE Line terminating element
LTN Logical terminal node
MAC Media access control
MAI Multiple access interference
MAN Metropolitan area network
MBFS-d Modified breadth-first search with parameter d
MC-CR Minimum-cost collapsed ring
MCH Minimum-cost heuristic
MC-LGS Multicast-capable logical-grooming switch
MC-RCA Multicast routing and channel assignment
MC-WRN Multicast-capable wavelength-routed network
MD-RZ Modified duobinary return-to-zero
 Acronyms 909

MEMS Microelectromechanical system

MFC Maximum free capacity
MFL Multifrequency laser source
MI Multilayer interference
MIP Mixed integer program
ML Maximum likelihood
ML-LD Mode-locked laser diode
MONET Multiwavelength Optical Network
MPCP Multi-point Control Protocol
MPLS Multiprotocol Label Switching
MPS Multipoint subnet
MQW Multiple quantum well
MRU Maximizing resource utilization
MSN Manhattan Street Network
MSP Multiservice platform
MSPP Multiservice provisioning platform
MST Maximizing single-hop traffic
MUX Multiplexer
MWNA Multiwavelength network architecture
MWS Multiwavelength switch; Multiwaveband switch
MWTN Multiwavelength Transport Network
MZ Mach–Zehnder
MZI Mach–Zehnder interferometer
NA Numerical aperture
NAS Network access station
NC Network connection
NDF Nonzero dispersion fiber
NETRATS Network Restoration Algorithm for Telecommunications Systems
NGI Next Generation Internet
NL Nonlinear (device)
NMS Network management system
NNI Network–Network Interface
NOLM Nonlinear optical loop mirror
NP Nondeterministic polynomial
NQDS Normalized quasidoubly stochastic
NRZ Non-return-to-zero
NTONC National Transparent Optical Network Consortium
OA Optical amplifier
OADM Optical add/drop multiplexer
OAM Operations, administration, and maintenance
910 Multiwavelength Optical Networks

OAM&P Operations, administration, maintenance, and provisioning

OBI Optical beat interference
OBPF Optical bandpass filter
OBS Optical burst switching
OC Optical connection
OCDC Orientable cycle double cover
OCDMA Optical code division multiple access
OCG Optical connection graph
OCh Optical channel
OCH Optical connection hypergraph
OChS Optical channel-section
OC-n Optical carrier-level n
OCSS Optical carrier suppression and separation
OCT Optical coherence tomography
OEIC Optoelectronic integrated circuit
O-E-O Optical-electrical-optical (switch, cross-connect)
OEO Optical-to-electrical-to-optical (conversion)
OF Optical filter
OIC Optical integrated circuit
OIF Optical Internetworking Forum
OLP Overlay processor
OLS Optical label switching
OLT Optical line terminal
OMS Optical multiplex section
ONC Optical node controller
ONM Optical network manager
ONN Optical network node
ONTC Optical Networks Technology Consortium
ONU Optical network unit
OOC Orthogonal optical code
OOK On–off keying
O-O-O Optical-optical-optical
OP Optical path
OPEN Optical Pan-European Network
OPERA Optical Packet Experimental Routing Architecture
OPS Optical packet switching
OR Optical receiver
OSI Open Systems Interconnection
OSNR Optical signal-to-noise ratio
OSPF Open Shortest Path First
 Acronyms 911

OSPF-TE Open Shortest Path First Protocol with Traffic Engineering extensions
OSS Operations support system
OSSB Optical single-sideband
OT Optical transmitter
OTF Optical threshold function
OTS Optical transmission section
OTTL Optical time-to-live
OXC Optical cross-connect
PBG Photonic band gap
PBS Polarization beam splitter
PC Polarization controller
PCF Photonic crystal fiber
PD Photodetector
PDF Probability density function
PDG Polarization-dependent gain
PDH Plesiochronous digital hierarchy
PDL Polarization-dependent loss
PHASAR Phased array
PIN p-type, intrinsic, n-type
pJET Priority just enough time
PLC Planar lightwave circuit
PLOAM Physical Layer Operations Administration and Maintenance
PM Physical media (layer)
PMD Polarization mode dispersion
PNNI Private Network-to-Network Interface
POH Path overhead
PON Passive optical network
POTS Plain old telephone service
PPWDM Point-to-point wavelength division multiplexing
PS Packet switch
PSB Partially shared buffering
PSD Power spectral density
PSE Photonic switch element
PSK Phase-shift keying
PSP Principal state of polarization
PT Physical topology
PXC Photonic cross-connect
QD Quantum dot
QDS Quasidoubly stochastic
QoS Quality of service
912 Multiwavelength Optical Networks

QPSK Quadrature phase-shift keying

RA Raman amplifier
RAM Random access memory
RBF Red/blue filter
RBS Rayleigh back-scattering
RCA Routing and channel assignment
RCN Randomly connected network
RE-XPM Raman-enhanced fiber cross-phase modulation
RF Radio frequency
RGC Raman gain coefficient
RIN Relative intensity noise
RIP Routing Information Protocol
RN Routing node; Remote node
ROADM Reconfigurable optical add/drop multiplexer
ROF Radio-over-fiber
RP Reception processor
RPR Resilient packet ring
RS Reed–Solomon (code)
R-SOA Reflective semiconductor optical amplifier
RSVP Resource Reservation Protocol
RSVP-TE Resource Reservation Protocol with Traffic Engineering extensions
RWA Routing and waveband assignment
RZ Return-to-zero
SA Source address; Saturable absorber
SaD Split-and-deliver
SAN Storage area network
SBS Stimulated Brillouin scattering
SCDMOD Subcarrier demodulator
SCH Set covering heuristic
SCM Subcarrier multiplexing
SCMA Subcarrier multiple access
SCMOD Subcarrier modulator
SCS Single-crystal silicon
SDDP Segment, deflect, and drop
SDH Synchronous digital hierarchy
SDP Segment and drop
SFP Small form factor pluggable (module)
SHR Self-healing ring
SLA Service-level agreement
SLALOM Semiconductor laser amplifier in a loop optical mirror
 Acronyms 913

SLM Spatial light modulator

SLOB Switch with large optical buffers
SLSP Short leap shared protection
SMF Single-mode fiber
SNA Systems Network Architecture
SNR Signal-to-noise ratio
SOA Semiconductor optical amplifier
SOH Section overhead
SOI Silicon-on-insulator
SONET Synchronous optical network
SP Shortest path
SPC Soft permanent circuit
SPD Shortest path with deletions
SPE Synchronous payload envelope
SPM Self-phase modulation
SPRING Shared-protection ring
SPT Shortest path tree
SP-WRN Splitter placement in wavelength-routed networks
SRG Shared risk group
SRLG Shared risk link group
SRS Stimulated Raman scattering
SSFBG Superstructure fiber Bragg grating
STOLAS Switching Technologies for Optically Labeled Signals
STS Synchronous transport signal
STS-n Synchronous transport signal, level n
TCP Transmission Control Protocol; Terminal connection point
TDM Time-division multiplexing
TDMA Time division multiple access
TE Transverse electric; Traffic engineering
TEM Transverse electromagnetic
T-F-CDMA Time-frequency code division multiple access
TM Transverse magnetic; Transmission matrix
TMF TeleManagement Forum
TMN Telecommunications Management Network
TOH Transport overhead
TOWC Tunable optical wavelength converter
TP Transmission processor
T-SCMA Time-subcarrier multiple access
TSI Time slot interchange
TT Trail termination
914 Multiwavelength Optical Networks

TT-FR Tunable transmitters with fixed receivers

TTL Time-to-live
TT-TR Tunable transmitters with tunable receivers
TW Transmission window
T-WDMA Time-wavelength division multiple access
UDP User Datagram Protocol
ULH Ultra long-haul
UNI User-Network Interface
UPPR Unidirectional path-protected ring
UPSR Unidirectional path-switched ring
UV Ultraviolet
VC Virtual channel; Virtual connection
VCI Virtual channel identifier
VCSEL Vertical-cavity surface-emitting laser
VOA Variable optical attenuator
VoIP Voice-over-Internet Protocol
VP Virtual path
VPI Virtual path identifier
VPN Virtual private network
VPON Virtual private optical network
VT Virtual tributary; Virtual topology
WADM Wavelength add/drop multiplexer
WAN Wide area network
WASPNET Wavelength Switched Packet Network
WC Wavelength conversion
WDF Wavelength-dropping filter
WDM Wavelength division multiplexing
WDMA Wavelength division multiple access
WDMUX Wavelength demultiplexer
WIXC Wavelength-interchanging cross-connect
WMUX Wavelength multiplexer
WRN Wavelength- or waveband-routed network
WR-OBS Wavelength-routed optical burst switching
WSC Wavelength-selective coupler
WSS Wavelength-selective switch
WSXC Wavelength- or waveband-selective cross-connect
WTM Wavelength terminal multiplexer
WXC Wavelength cross-connect
XGM Cross-gain modulation
XPM Cross-phase modulation
Index

2D MEMS switches, 250–251 Air-clad core PCF, 189

2R regeneration, 788 Algorithms. See also Heuristic algorithms;
3D CDMA system, 363 Integer linear program
3D MEMS switch, 252–253 adaptive routing, 485–486
3R regeneration, 788 alternate routing, 485
arc-disjoint, 703
Acceptable label set object, 749 auxiliary graph as framework for, 606
Acceptance angle, 170 Bellman-Ford, 734–735
Access fibers, 30, 132 connection allocation, 545–546
Access link, 30 constrained v. unconstrained, 486
Access station. See Network access station for DCS restoration techniques, 664,
(NAS) 666–668
Acousto-optic tunable filter (AOTF) Djikstra’s, 734–735, 738, 888
all-fiber, 268–269 Dsatur, 872
fabrication of, 268 fixed routing, 485
in ONTC testbed, 838 fixed scheduling, 875–879
polarization flipping in, 267–268 grooming, 584, 603–606
ACTS. See European Advanced Haberman’s, 505
Communications Technology and Service for hypernet design, 632–633
Acyclic graph, 871 for Kautz hypernets, 623–625
Adaptive routing algorithm, 485–486 k-SP, 487, 547–550, 553–555
Add/drop multiplexer (ADM) least loaded routing, 489
failure at, 656 link state, 735
as LS, 79 looping, 51, 765, 897–898
minimization of, 600–601 minimum cost routing, 637
in SONET, 79–81, 598–599 Min-Int, 547–550, 553–555, 888–892
Add/drop traffic, 78–79 multicast connection allocation, 558–568
Address shift routing optimization, 465–468
in directed Kautz hypergraph, 623–625 planarity-testing, 684
rule for, 591 for point-to-point connections, 545–549,
in undirected Kautz hypergraph, 625 551–557
ADH. See Average distance heuristic for RCA, 51–52, 127, 139, 455–468,
Adiabatic chirp, 213 474–591, 623–625, 637
Adjacency, in optical control plane, 719, 731, for ring covers, 664–665
738 SPD, 487–488
ADM. See Add/drop multiplexer void-filling scheduling, 802–803
Administrative status object, 749–750 for WDM mesh networks, 603–606
Advanced Technology Demonstration for WDM rings, 600–602
Network (ATDnet), 851–852 for wide-sense nonblocking switches,
Aggregate network capacity bound, 445–446 51–52
916 Index

All-optical label swapping (AOLS) channel selector and, 272–273

enabling technologies for, 811 as optical filtering technology, 265–266
example of, 809–811 routing properties of, 265
processor for, 810 as single thermo-optically controlled MZ
All-Optical Network (AON) Consortium, 839 switch, 271–272
All-optical 2R or 3R regeneration, 788 switches based on, 271–273
All-optical wavelength conversion, 275–278 versatility of, 266
ALOHA, 401–403 WADM based on, 271–272
Alternate routing algorithm, 485 in WDM-based PONS, 417–418
Amplets, 204–205 AS. See Autonomous Systems
Amplified spontaneous emission (ASE) noise, ASE noise. See Amplified spontaneous
195–197, 200, 227, 241 emission noise
Amplifier chains, in EDFAs, 197 Asynchronous OPS networks, 782–783
Amplifiers, 190–205, 257. See also Asynchronous transfer mode (ATM), 4, 6, 8,
Erbium-doped fiber amplifier; Raman 19
amplifier; Semiconductor optical APON based on, 412–414
amplifier ATMOS and, 793
amplets, 204–205 broadband services supported by, 19, 81–82
EDFA, 191–197 cell format of, 82
EDWA, 204 logical topology in, 82–83
front-end, 219–221 networks of, 81–85, 97
line, 190 Asynchronous transfer mode (ATM) network
linear optical, 205 ATM-over-SONET structure in, 82–83
LOA, 205 connection management in, 97
power, 190 as LRN, 81–85
preamplifiers and, 190 LS in, 81
rare-earth-doped fiber, 191 LT in, 82–83
self-saturation of, 195 NAS in, 83
SLALOM, 280, 789 point-to-point connections in, 82
SOA, 201–203, 205, 423, 788, 790, ATDnet. See Advanced Technology
814–815 Demonstration Network
spacing of, 197 ATM. See Asynchronous transfer mode
transimpedance, 219 ATM-based passive optical network (APON),
in WANs, 190–191 412–414
Analog modulation, 71, 181, 214 ATM network. See Asynchronous transfer
AOLS. See All-optical label swapping mode network
AON Consortium. See All-Optical Network ATM Optical Switching Project (ATMOS),
Consortium 793
AOTF. See Acousto-optic tunable filter ATMOS. See ATM Optical Switching Project
APD. See Avalanche photodiode Attenuation
APON. See ATM-based passive optical material absorption and, 175
network OAs and, 30
APS. See Automatic protection switching power losses from, 505
Arbitrary mesh architectures, 663–669 Rayleigh scattering and, 176
centralized restoration techniques in, as transmission impairment, 175–177
665–666 VOAs for, 241
distributed restoration techniques in, waveguide imperfections and, 176
666–669 Autocorrelation sequence, 357
ring covers for survivability in, 664–665 Automated neighbor discovery, 731
Arc-disjoint algorithm, 703 Automatic protection switching (APS),
Arrayed waveband grating (AWG), 259, 790, 453–454
796 in four-fiber BLSR, 657
 Index 917

for link failures, using protection cycles, UPSR v., 655–657

686–687 Bidirectional pumping EDFA, 192–193
one-for-N, 650, 653–654, 671–672 Bidirectional ring, 600. See also Bidirectional
one-for-one, 650, 653, 670–671 line-switched ring
one-plus-one, 650–653 connectivity implemented on, 143, 145–147
redundant trees for, 697 dynamic RCA case study of, 491–494
SONET linear, 650, 653 with full wavelength interchange, 493–494
Autonomous Systems (AS), 734 physical topology of, 93–94
Auxiliary graph RCA in, 129–132
access layer in, 605 ShuffleNet as, 155–156
as algorithm framework, 606 single fiber pair access in, 129–130
edges in, 606 spectrum reuse in, 93, 150–151
for grooming, 605–607 two fiber pair access in, 129–132
Auxiliary hypergraph, 636–637 Binary block code, 232
Avalanche photodiode (APD), 218–220, Bipartite graph, 869, 871
422 Birth-death process, 882–883
Average distance heuristic (ADH), 502 B-ISDN. See Broadband integrated services
AWG. See Arrayed waveband grating digital network
Bit error rate (BER)
Backup route computation, 678–679 CIR influencing, 345
Backward pumping EDFA, 192–193 cross-talk related to, 239
Balanced light-tree (BLT), 505 digital signal detection and, 221–226
Band gap, 201 of FFH-CDMA using FBGs, 362
Bandpass filter (BPF), 73–74 ISI increasing, 177
Bandwidth in OLP regeneration, 76
channelization of, 10 performance measured in, 167
OR constraints of, 31 in signal regeneration, 274–275
deregulation and, 828 Blocking. See also Blocking probability
OT constraints of, 31 color clashes causing, 546–547
of single fiber, 10 HOL, 406, 771–773, 776, 785
trading of, 3, 859–861 in networks with multifiber links, 554–555
wavelength-brokering operational model in networks with multiple wavebands,
and, 859–860 555–557
Banyan structure, 43, 50 permutation switches and, 50–52
Base station (BS), 420 in simulated multicast scenario, 567–568
Bellman-Ford algorithm, 734–735 wavelength interchange improving, 497
Benes switch, 49–51 Blocking probability, 390
cross-talk of, 286–288 admission control and, 397–398
implementation of, 285–286 in bidirectional ring dynamic RCA case
looping algorithm for, 51 study, 492
twisted, 66 framed system, 397
BGP. See Border Gateway Protocol matching time slots in, 396
BH. See Buried heterostructure (BH) in mesh networks, 494–495
Bidirectional line-switched ring (BLSR), for multicast connections, 567–568
656–657 multiple wavebands and, 556
BLSR-BLSR interconnected architectures in multipoint logical topologies, 638
of, 659, 663 rearrangeability and, 398–399
BLSR-UPSR interconnected architectures as RWA performance measure, 545
of, 659–663 in TDM/T-WDMA networks, 395–399
four-fiber, 657 in WDMA networks, 390–395
two-fiber, 656–657, 676 BLSR. See Bidirectional line-switched ring
918 Index

BLT. See Balanced light-tree packet switching in, 112–113, 117

Border Gateway Protocol (BGP), 726, 734 reflecting, 523
Bottleneck root node of, 140, 145
access, 2, 132, 409 as shared medium, 103, 117, 324–327
connectivity, 11, 20, 576–577 transceivers in, 112
distribution network as, 2 tree emulating, 145–146
electronic, 11, 15, 59 T-WDMA scheme for, 109–110
at limiting cut, 447–448 BS. See Base station
WIXC removing, 274 B&S OADM. See Broadcast and Select optical
Bounds. See Flow bounds add/drop multiplexer
BPON. See Broadband passive optical network Bubble switch, 249
Bragg grating. See Fiber Bragg grating Buffered nodes, 770–774
Bridged ring overlay, 461–462 Buffering
Bridgeless graph, 460, 685 contention resolution by, 770–787
Bridges dump-and-insert, 774–778
of connected graphs, 871 electronic, 770–771
failures in, 681 as enabling technology, 789–791
as LSs, 461–462 FDLs as medium of choice in, 770–771
as OLPs, 461–462 hybrid electric and optical, 784–787
in wavelength interchange, 460–462 input, 771–772
Broadband integrated services digital network of nodes, 770–774
(B-ISDN), 19 in OBS, 799–800
Broadband passive optical network (BPON) in OPS networks, 61, 759, 770–791,
MAC protocol for, 413–414 799–800, 821
OLTs in, 412–414 output, 772, 780
ONUs in, 412–414 virtual, by deflection routing, 765
transmission scenario in, 413 Buried heterostructure (BH), 207
Broadband services. See also Digital Burstiness, 111, 387, 389
subscriber line Burst segmentation, 806–808
ATM supporting, 19, 81–82 Burst switching. See Optical burst switching
availability of, 2–3, 19 Business drivers and economics, in
cable modems as, 3, 19 multiwavelength optical networking,
DSL as, 3, 19 828–837, 861
Broadcast-and-select operating method. See cross-connect network cost issues as,
also Static networks 833–836
DCA constraint of, 103, 127–128, 131, open v. closed WDM installations as,
138–139 835–837
large networks and, 122 point-to-point system cost issues as,
of shared medium networks, 103 831–832
Broadcast and Select (B&S) optical add/drop ring cost issues as, 832–834
multiplexer (OADM), 291–293 Busy destinations, 391
Broadcast star network
dynamic capacity assignment in, 113 Cable exhaust. See Fiber exhaust
embedded, 140–142, 146–147, 158, 160, Cable television (CATV) distribution
327, 522–523 networks, 227, 336
examples of, 325–326 Cantor network, 51
maximum number of stations in, 122 Capacity allocation. See also Dynamic
multipoint connections in, 103–106, capacity allocation; Fixed capacity
324–327 allocation
NAS role in, 40–41, 324–237 for dedicated connections, 371–389
nonreflecting, 522–523 for demand-assigned connections, 389–399
 Index 919

dynamic, 346–347, 383, 399–400, 406, 558 of λ-channels, 516–518, 528–531

fixed, 346–347, 383–389, 400 constraints on, 103, 125–128, 131, 138–139
Capacity exhaust. See Fiber exhaust in fixed-frame scheduling, 373–376
Carried traffic, 387, 389, 391. See also in LLNs, 528–540
Throughput network performance comparisons of,
in demand-assigned LCs, 539–540 535–538
maximization of, 471–474 routing compared with, 124
Carrier-sense multiple access with collision in RWA, 516–518
detection (CSMA/CD) transmission, 532–535
collision events in, 115 Channel reservation. See also Resource
end-to-end propagation time in, 115–116 Reservation Protocol
LLC sublayer of, 121 delayed, 801–802
as MAC protocol, 113–114, 120 in packet switching in optical layer, 406,
throughput efficiency of, 116 411–412
Carrier suppressed return-to-zero (CSRZ) in PONs, 411–412
keying, 231 slot-by-slot, 406
Carrier-to-interference ratio (CIR) Channel selector, 272–273
BER influenced by, 345 Channel sharing
OBI performance measured by, 342–345 advantages of, 151
Carrier-to-noise ratio (CNR), 227 collisions in, 115, 118–119, 401–402
CAS. See Channel allocation schedule λ-channels. See also Channel assignment;
Cascaded Mach-Zehnder (MZ) switch, 247 Routing and channel assignment
CATV distribution networks. See Cable assignment of, 516–518, 528–531
television distribution networks MPS and, 142
C-band, 176–177 ONNs switching, 30
CDC. See Cycle double cover shared, 532–535
CDMA. See Code division multiple access in single waveband, 60
Centralized backup route computation, spacing of, 35–37, 150, 186–187
678–679 wavelength continuity and, 63
Centralized control paradigm, in optical in wavelength-routed networks, 122
control plane, 722 Chirp, 212, 216
Centralized restoration, 665–666 adiabatic, 213
Central Office (CO), 830–831, 835 dispersion interaction with, 240
Central station (CS), 420–422 in FBG, 265
CH. See Contraction heuristic intensity-modulated analog systems, 214
Channel allocation penalty induced by, 240
by Max Reuse, 550–551, 552–553 transient, 213
by Min Reuse, 550–553 Chirped return-to-zero (CRZ) keying, 231
multiple waveband, 555–557 Chirp-induced penalty, 240
in RWA, 550–551 Chooser node, 667–668
single waveband, 552–555 Chords, 871
on trees, 566–567 Chromatic dispersion, 178–179
Channel allocation schedule (CAS), 119 CIR. See Carrier-to-interference ratio
examples of, 329–331 Circuits, 871
in fixed-frame scheduling, 374, 376 Circuit-switched operation
with full complement of channels, 376 of ONNs, 14, 60
information loss avoided in, 329 packet switching v., 756
logical multicast, 382 TDM and, 106–107
for TDM, 328–332 in telephone networks, 4
Channel assignment. See also Routing and Circulator, 257
channel assignment Class of service (CoS), 728–729
920 Index

Clear channels. See Demand-assigned Column/row-compression algorithm, 879

wavelengths Column/row-expansion algorithm, 876–879
CLEC. See Competitive local exchange carrier Competitive local exchange carrier (CLEC),
Client service associations, 32–33 841, 853
Clique, 870 Complete graph, 869–870
Closed walk, 871 Conduit exhaust. See Fiber exhaust
Closed WDM installations, 835–837 Conflict-free receiver tuning schedules, 111
Clos switch, 48–49, 51 Connected graph, 871–872
CNR. See Carrier-to-noise ratio Connection allocation algorithm, 545–546
CO. See Central Office Connection granularity. See Granularity
Coarse wavelength division multiplexing Connection interference graph, 529–531
(CWDM), 17, 177, 205 Connection management
Cochannel heterodyne cross-talk, 62, 236–237 in ATM network, 97
Cochannel multipath cross-talk, 62, 236–239 connection establishment phase of, 96, 100
Code division multiple access (CDMA), connection release phase of, 96–98, 100
352–367 control plane in, 98
coherent, 354, 360–363 information transfer phase of, 96–98
DS, 353–354, 360–363, 365–366 of LCs, 100–102
electronic processing in IM/DD systems in optical networks, 3, 9, 17–18, 20, 39,
and, 354–359 96–102
using FBG, 361–362 in static networks, 102
FFH, 353, 361–362 system of, 99–100
improvement of, 358–359 Connection rearrangement, 124–126, 139
MAI in, 353, 358 Connectivity bottleneck, 11, 20, 576–577
noncoherent, 354, 363–366 Constraint-based Routing Label Distribution
with optical processing, 352–353, 359–366 Protocol (CR-LDP), 726, 730
packet switching using, 366–367 Constraints
parallel transceiver structure of, 358–359 bandwidth, 31
self-contained uncoordinated manner of, channel assignment, 103, 125–128, 131,
366 138–139, 528–529, 559
as spread-spectrum technique, 352 DCA, 103, 127–128, 131, 138–139,
using SSFBG, 365–366 528–529, 559
T-F, 353 on deflection routing, 765
3D, 363 DSC, 135–139, 559
waveforms for, 356 inseparability, 135–136, 138–139, 546–550,
Coherent converters, 63 559
Coherent optical systems interference, 449–452
CDMA using SSFBG, 365–366 in LLNs, 509, 528–529
as enabling technology, 234–235 optical layer, 504–505
heterodyne OR in, 234–235 path length, 467
optical processing in, 363–366 physical layer, 548–549
phase masks in, 363–365 routing, 135–140, 509, 548–549
photocurrent in, 234 on static multipoint networks, 424
Coherent receiver. See Heterodyne optical traffic, 367–371
receiver transmission, 333–335
Collisions, 115, 118–119, 401–402 on waveband-routed networks, 38
Color clashes wavelength continuity, 127, 139, 450
in LLNs, 138–139, 546–547 Contending burst, 806
Max Reuse causing, 563 Contention resolution
Min Reuse minimizing, 551–553 for asynchronous OPS networks, 782–783
SPT causing, 562–564 by buffering, 770–787
 Index 921

by burst segmentation, 806–808 Couplers. See also Directional couplers; Star

comparison of techniques for, 782–783 couplers
CRO for, 791–793 cross-talk in, 243–244
by deflection routing, 764–770 GCSR, 815
MAC protocol and, 117 insertion loss in, 243–244, 250–251, 255
node architectures for, 783 as optical and photonic technology, 242–255
OBS and, 806–808 wavelength-selective, 191
in OLS, 811 CRC. See Cyclic redundancy check
in OPS, 763–764 CR-LDP. See Constraint-based Routing Label
testbed implementations and, 791–798 Distribution Protocol
time domain, 770–778 CRO. See Contention resolution optics
timewavelength domain, 778–782 Crossbar switch, 52, 55
by wavelength conversion, 778–783 implementation of, 281–284
Contention Resolution by Delay Lines optical, 282–284
(CORD) testbed, 791–793 path-independent loss, 282
Contention resolution optics (CRO), as permutation switch, 48
791–793 as space switch, 281–284
Continuous state process, 880 Cross-phase modulation, 185
Continuous time process, 880 Cross-points, 48, 51–52
Contraction heuristic (CH), 502 Cross-talk, 137
Control agents of Benes switches, 286–288
DCN used by, 98, 719 BER related to, 239
ONC as, 99–100 cochannel heterodyne, 62, 236–237
Control algorithms, 22 cochannel multipath, 62, 236–239
Controllability, 242 in cost-performance trade-off, 62–63
Controllable directional coupler in couplers, 243–244
as guided-wave switch, 244–246 dynamic nodes and, 62–63
semiconductor technology and, 245–246 in EFDAs, 195
Controllable dynamic couplers. See Dynamic FWM and, 276
directional couplers interchannel, 62, 236–237, 258
Controller. See Optical node controller; from MZ switches, 247
Switch controller order of, 286–288
Control paradigms, in optical control plane, overlapping signal spectra causing, 34–35
722 as performance impairment, 235–239
Control plane. See also Optical control plane power penalty for, 238–239
agents in, 98–99 reduction of, by switch fabrics, 52, 62
DCN and, 98, 719 of routers, 286–288
management plane v., 714 space dilation reducing, 288
Converter edge, 606 stimulated Raman scattering introducing, 62
Converters switch, 52, 62, 247, 286–288
coherent, 63 in WDM physical-layer simulation,
DFC, 64–65, 276–277, 281 306–307
wavelength, 63–67, 275–277, 281, CRZ keying. See Chirped return-to-zero
779–782, 793–794 keying
CORD testbed. See Contention Resolution by CS. See Central station
Delay Lines testbed CSMA/CD. See Carrier-sense multiple access
Core node, 818–819 with collision detection
Core router, 817–818 CSRZ keying. See Carrier suppressed
Correlation sequence, 357 return-to-zero keying
CoS. See Class of service Cube graph. See n-cube graph
Cost-performance trade-offs, 22, 62–63 Cutoff wavelength, 218
922 Index

Cuts capacity allocation for, 371–389

of connected graphs, 871 rearrangement of, 125
fiber, 647–648, 655–656 switched connections v., 124
limiting, 446–558, 886–889 Dedicated protection ring (DPRING), 672
in max flow-min cut theorem, 446, 872–873 Dedicated path-based protection, 692–693,
Cutset, 871 695
CWDM. See Coarse wavelength division Deflect and Drop Policy (DDP), 807
multiplexing Deflection index, 765
Cycle decomposition, 664–665 Deflection probability, 770
Cycle double cover (CDC) Deflection routing
in bridgeless graphs, 685 burst segmentation combined with, 807–808
of connected graphs, 872 constraints on, 765
directed, 683–684 contention resolution by, 764–770
in mesh networks, 460, 680–687 limited, 768–770
nonplanar case of, 683, 685–686 performance analysis of, 766–770
orientable, 460, 685–686 virtual buffering by, 765
planar case of, 683–684 Deflect, Segment, and Drop Policy (DSDP),
redundancy in, 680 807
shared optical protection by, 680–687 Delayed reservation (DR), 801–802
Cyclic redundancy check (CRC), 114 Delay interferometer (DI), 230
Demand-assigned logical connections (LCs).
DA. See Destination address See also Switched connections
DAH. See Dual ascent heuristic blocking calculations in T-WDMA
Dark current, 217 networks, 395–399
Data Communications Network (DCN) blocking calculations in WDMA networks
control agents using, 98, 719 under, 390–395
required for signaling protocol, 742 capacity allocation for, 389–399
Data link control (DLC) layer, 84, 120 carried traffic in, 539–540
Data networks in LASs, 538–540
history of, 18–19 Markov chain model for, 391
packet-switched mode of, 4 packet switching compared with, 399
DBR laser. See Distributed Bragg reflector Demand-assigned wavelengths, 7, 11, 13, 96
(DBR) laser Demultiplexing
DCA constraint. See Distinct channel by ADM, 80
assignment constraint by DCS, 80
DCF. See Dispersion-compensating fiber in RPs, 70
DCN. See Data Communications Network Dense hypernets, 613–615
DCS. See Digital cross-connect system Dense logical topologies (LTs)
DDP. See Deflect and Drop Policy desirable results of, 589
deBruijn digraph families of, 589–591
density of, 589–590 Dense wavelength division multiplexing
maximum throughput of, 590–591 (DWDM), 13, 19, 35, 60, 177
deBruijn graph DWDM/TDM PON and, 422–424
multihop networks and, 586 in MANs, 845–846
orders of, 436 modulation formats in, 230
deBruijn network, 436–437 physical-layer simulation of,
Decomposition 308–311
cycle, 664–665 size of, 251
ring, 458–462 Dense wavelength division multiplexing/time
tree, 560–566 division multiplexing passive optical
Dedicated connections, 96, 127 network (DWDM/TDM PON), 422–424
 Index 923

Deregulation, 3, 828 address shift routing in, 623–625

Design, of hypernets, 632–641 converted to undirected, 623
logical-layer, 632–633 duality construction of, 617, 619
physical-layer, 634–637 edge grouping construction of, 620–621
suboptimal approach to, 633 Directed Kautz hypernets, 630–631, 641
Destination address (DA), 114 Directed star network, 45, 103n7
DFB laser. See Distributed-feedback laser Directional couplers
DFC. See Difference frequency converter combining losses of, 44
DGD. See Differential group delay dynamic, 40–44
DI. See Delay interferometer lossless, 42, 54–55
Dicliques, 614–615, 621 multiwaveband, 57
Difference frequency converter (DFC), 64–65, power transfer, 42–44
276–277, 281 static, 42–44
Differential group delay (DGD), 241 waveband-selective, 57
Differential phase shift keying (DPSK), 813, Directly modulated laser (DML), 212–214
815–816 Direct modulation, 70, 212–214
Differentiated services, 5 Direct sequence code division multiple access
Digital cross-connect system (DCS) (DS-CDMA), 353, 354, 360–363,
algorithms for, 664, 666–668 365–366
centralized restoration techniques of, Discrete-state continuous-time Markov chains,
665–666 881
distributed restoration techniques of, Discrete state process, 880
666–669 Discrete time process, 880
as LS, 79 Dispersion, 177–183, 227
in SONET, 79–81, 582–583 analog modulation and subcarriers in, 181
Digital modulation, 71 chirp interaction with, 240
Digital signal detection chromatic, 178–179
ideal, 224–225 compensation for, 181–183
ML, 225 flattened, 179
noise, interference, and BER in, 221–226 intermodal, 170, 177–178
Digital subscriber line (DSL) management of, 181–183
as half-measure, 3 maximum bit rates influenced by, 180–181
origin of, 19 PMD, 179–180, 241
Digital switch. See Y-branch switch shifted, 179, 185
Digraphs. See also Kautz digraphs slope, 183
deBruijn, 589–591 Dispersion-compensating fiber (DCF), 182
generalized loopbacks creating, 690–692 Dispersion-flattened fibers, 179
Kautz, 615, 617–618, 622–623, 630–631, Dispersion relation, 173
639–640 Dispersion-shifted fibers (DSF), 179, 185
for logical topology, 589–591 Distance network graph, 502
max flow-min cut theorem for, 446, Distance network heuristic (DNH), 502
872–873 Distance vector protocols, 735–736
maximum throughput of, 590–591 Distinct channel assignment (DCA)
Moore bound for, 614, 872 constraint, 136, 148, 454
Direct detection optical receiver (OR), 71–74, of broadcast-and-select operating method,
233n26 103, 127–128, 131, 138–139
Direct detection OR, 233n26 channel assignment constrained by,
Directed cycle double cover (CDC), 683–684 127–128, 131, 559
Directed hypergraphs, 614–615, 617, color clashes violating, 138–139,
619–621, 623–625, 639, 874 546–547
Directed hypernets, 608, 630–632, 640 in LLNs, 528–529
Directed Kautz hypergraph, 615–619 spectrum reuse prevented by, 103
924 Index

Distinct source combining (DSC) of hypergraphs, 614–619

in LLNs, 135–139, 512–513 labeling convention of, 617–619
violations of, 137–138, 512–513, 546–547, limitations of, 620
559 Dump-and-insert buffering
Distortion-induced penalty, 240–241 architecture of, 774–775
Distributed backup route computation, contention resolution and, 774–778
678–679 in slotted system, 775
Distributed Bragg reflector (DBR) laser, typical packet sequence in, 776
207–208, 839 Duobinary return-to-zero (D-RZ) keying, 231
Distributed control paradigm, in optical DWDM. See Dense wavelength division
control plane, 722 multiplexing
Distributed-feedback (DFB) laser, 207–208, DWDM/TDM PON. See Dense wavelength
215–217 division multiplexing/time division
Distributed queue dual bus (DQDB), 458 multiplexing passive optical network
Distributed Raman amplifier (DRA), 199–200 Dynamic capacity allocation, 108
Distributed restoration advantages of, 112
DCS-based, 666–669 in broadcast star networks, 113
FITNESS protocol in, 667 contention arbitration and, 117
NETRATS protocol in, 668 in fixed-frame scheduling for packet traffic,
RREACT protocol in, 667–668 383
SelfHealing Network in, 666–667 fixed v., 408–409
trade-offs of, 668–669 logical topology and, 113
Distributed routing approaches, 737 packet switching in optical layer and, 112,
Djikstra’s algorithm, 734–735, 738, 888 399–400
DLC layer. See Data link control layer scheduling of, 117
DML. See Directly modulated laser (DML) simulation evaluating performance of,
Double fiber pair access, 456 567–569
Double-heterostructure geometry, 218 in TDM/T-SCMA, 346–347
DP. See Drop Policy Dynamic directional couplers, 40–44
DPRING. See Dedicated protection ring Dynamic nodes, 40, 46–48, 60–62
DPSK. See Differential phase shift keying connection states and, 46
DQDB. See Distributed queue dual bus cross-talk and, 62–63
DR. See Delayed reservation generalized, 46–48, 52–53
DRA. See Distributed Raman amplifier LDCs and, 47, 53–56
Drop-and-continue function, 658, 659, 662 loopback connections of, 55
Drop filter, 264 permutation, 45–46, 48–52, 54
Drop Policy (DP), 807 waveband-space switches and, 56–60
D-RZ keying. See Duobinary return-to-zero Dynamic restoration, 705
keying Dynamic routing and channel assignment
Dsatur algorithm, 872 (RCA)
DSC. See Distinct source combining algorithms for, 51–52, 127, 139, 455–458,
DS-CDMA. See Direct sequence code division 474–507
multiple access bidirectional ring case study of, 491–494
DSDP. See Deflect, Segment, and Drop Policy characterization of, 485–487
DSF. See Dispersion-shifted fibers fairness in, 489–491, 498
DSL. See Digital subscriber line grooming in, 601–602, 605
Dual-access ring interconnection interconnected ring case study of, 495–496
configurations, 657–663 LLN rules for, 544–568
Dual ascent heuristic (DAH), 502 in mesh networks, 494–495
Dual honing, 30n2 of multicast connections, 497–507
Duality construction, of hypergraphs for point-to-point connections, 544–558
 Index 925

static RCA v., 557–558 ORs, 217–227

in wavelength-routed networks, 484–507 OTs, 205–217
overview of optical connection, 167–168
EA-DFB. See Electro-absorption packet synchronization, 788
distributed-feedback laser performance evaluation of, 297–311
EDC. See Electronic dispersion compensator performance impairments and, 235–241
EDFA. See Erbium-doped fiber amplifier signal regeneration, 274–281, 788
Edge-disjoint tree, 560–561 transmission and switching technology
Edge grouping construction evolution, 166–167
of directed Kautz hypergraph, 620–621 wavelength conversion, 274–281, 789
of ShuffleNet, 614–615 End node, 442
Edge router End systems
LER, 723–725 ATM interfaced to, 82
in OPSnet, 816 full connectivity of, 91–92
EDWA. See Erbium-doped waveguide in PON, 410
amplifier in ShuffleNet, 588
EFEC. See Enhanced forward error correction End-to-end transmission channel
Effective-index PCF, 188–189 as enabling technology, 228–234
Effective interaction length, 185–186 equalization in, 233–234
Effective path length, 489 FEC and, 228, 231–232
Effective service time, 385 modulation formats and, 229–231
EFM Task Force. See Ethernet in the First processing operations in, 228
Mile Task Force Engset model, 393
EGP. See Exterior gateway protocol Enhanced forward error correction (EFEC),
Electro-absorption distributed-feedback laser 852
(EA-DFB), 215–217 E-NNI. See Exterior Network-Network
Electron-hole recombination, 201–202 Interface
Electronic bottleneck, 11, 15, 59 E-packets, 784–786
Electronic buffering, 770–771 EPON. See Ethernet-based passive optical
Electronic dispersion compensator (EDC), 233 network
Electronic equalizer. See Electronic dispersion Equalization, 183
compensator in end-to-end transmission channels,
Elementary station, 448–449, 510, 521, 233–234
527–528, 636 FEC combined with, 234
Embedded broadcast star network, 140–142, in WDM systems, 233–234
146–147, 158, 160, 327, 522–523 Equalizer. See Electronic dispersion
EML. See Externally modulated laser compensator
Enabling technology Equipment failure, 648
all-optical 2R or 3R regeneration, 788 Erbium-doped fiber amplifier (EDFA),
amplifiers, 190–205 191–197, 241
for AOLS, 811 as amplet, 204
buffering, 789–791 amplifier chains in, 197
coherent optical systems, 234–235 cross-talk in, 195
developed in WASPNET, 798 distributed, 197n12
end-to-end transmission channel, 228–234 drawbacks of, 191
for OPS, 757, 787–791, 821 emergence of, 18, 182
optical and photonic device, 241–274 energy levels in, 190–193
optical fibers, 168–190 gain profile of, 193–194
optical header processing, 789 gain saturation of, 194–195
optical switch architecture, 281–297, noise in, 195–197, 200
788–789 population inversion in, 193
926 Index

Erbium-doped fiber amplifier (EDFA) (cont.) at network nodes, 685

pump wavelengths in, 193 node, 648, 650, 656–657, 675, 685
RAs compared with, 197–199, 201 path, 651
structures of, 192–193 recovery from, 580, 647, 649, 651–652,
Erbium-doped waveguide amplifier (EDWA), 688
204 of underground telecommunication cables,
Ergodic chain, 882 647
Erlang model, 391–393 Failure Immunization Technology for Network
Etalon. See Fabry-Perot filter Service Survivability (FITNESS)
Ethernet, 402 protocol, 667
EPON, 414–416, 849 Fairness
gigabit, 847 in dynamic RCA, 489–491, 498
OBS and, 821 ratio of, 490, 498
ten-gigabit, 847–848 unfairness factor and, 490
Ethernet-based passive optical network wavelength interchange improving, 491,
(EPON), 849 494, 498–499
EFM Task Force developing, 414 Faraday rotator, 257
MAC protocol for, 415–416 FASTAR system
MPCP in, 414 restoration process of, 665–666
OLTs in, 415 speed of, 666
ONUs in, 415–416 Fast frequency hopping code division multiple
Ethernet in the First Mile (EFM) Task Force, access (FFH-CDMA), 353, 361–362
414 Fault isolation, 733
Eulerian network, 459–460 Fault protection. See Protection
European Advanced Communications Fault recovery. See Recovery
Technology and Service (ACTS), 793 Fault restoration. See Restoration
European multiwavelength optical network FBG. See Fiber Bragg grating
trials, 839–840 FCFS basis. See First-come-first-served basis
Explicit Route object, 748 FDDI. See Fiber distributed data interface
Exterior gateway protocol (EGP), 734 FDL. See Fiber delay line
Exterior Network-Network Interface (E-NNI), Feasibility, 550, 566
720, 744 of multivendor networks, 19
Externally modulated laser (EML), 214–215 of OPS networks, 757
External modulation, 70, 214–215 path, 487–488, 511, 546, 559
Eye diagram, 223–224 SPD algorithm accounting for, 487–488
Feasible paths, 511, 546, 559
Fabry-Perot (FP) filter, 259 FEC. See Forward error correction;
finesse of, 261 Forwarding equivalence class
MI filter compared with, 262 Feed-forward line delay architecture, 772–774,
spectral response of, 260 791–793
tunability of, 261 FFH-CDMA. See Fast frequency hopping
Fabry–Perot (FP) laser, 206–207 code division multiple access
Face traversal, 684 Fiber Bragg grating (FBG), 182
Failure chirped, 265
at ADM, 656 as drop filter, 264
bridge, 681 FFH-CDMA using, 361–362
communication, 647 manufacturing of, 263–264
equipment, 648 SSFBG, 365–366
from fiber cuts, 647–648 tunability of, 264
FITNESS protocol for, 667 written into planar waveguides, 264–265
link, 652, 674, 681–683, 685–687, Fiber cables
690–692, 750–751 ribbon, 174
 Index 927

typical designs of, 174–175 Field programmable gate array (FPGA), 813
underground, 647 Filters. See Optical filtering technology
Fiber cuts First-come-first-served (FCFS) basis, 764
causes of, 647 FITNESS protocol. See Failure Immunization
failure from, 647–648 Technology for Network Service
UPSRs and, 655–656 Survivability protocol
Fiber delay line (FDL), 759 Fixed capacity allocation, 108, 111, 116, 400
as buffering medium of choice, 770–771 dynamic v., 408–409
in feed-forward delay line architecture, in fixed-frame scheduling for packet traffic,
773–774 383–389
in JET protocol, 802 in TDM/T-SCMA, 346–347
packet loss probability v., 780–781 Fixed frame passive optical network (PON),
synchronizer based on, 762 412–414
Fiber-disjoint tree, 634–636 Fixed frame PON. See Fixed frame passive
Fiber dispersion. See Dispersion optical network
Fiber distributed data interface (FDDI), 458 Fixed-frame scheduling
Fiber exhaust, 828–829 algorithm for, 875–879
in WDM point-to-point systems, 831–832 channel assignments in, 373–376
in WDM rings, 832–833 dynamic capacity allocation in, 383
Fiber grating, 182, 259, 263–265, 361–362, examples of, 375–379, 878
365–366 heterogeneous, 379
Fiber links, 2, 10, 31–32, 126, 411, 554–555 for LCs, 534
Fiber-optic transmission systems, 647 multicast LCs and, 380–383
Fibers for packet traffic, 383–389
access, 30, 132 parameters of, 371–372
bandwidth of, 10 queues in, 384–388
bidirectional pairs of, 34 for stream traffic, 371–383
congestion of, 475 traffic matrix scaling in, 379–380, 593
dispersion influencing, 179, 182, 185 tunability in, 373–376
efficiency of, 536 Fixed routing algorithm, 485
as enabling technology, 168–190 Fixed scheduling algorithm
geometry of, 174–175 column/row-compression, 879
glass, 1–2 column/row-expansion, 876–879
graded-index, 169–171, 177–178 decomposition into permutation matrices,
gratings and, 182, 259, 263–265, 361–362, 877, 879
365–366 terminology of, 875–879
large effective area, 186–187 Fixed transmitters with tunable receivers
low-loss, 1, 18 (FT-TR), 109–110, 117–119, 147, 161,
optical, 168–190 331–332, 346, 353–354, 394–395, 404
PCFs as, 21, 188–190 Flow bounds, 444–448, 452
polarization-maintaining, 174 aggregate network capacity, 445–446
resources of, 34 fluid, 470–471
step-index, 168–170, 177–178 limiting cut, 446–448, 886–889
technological advances in, 20 Flow conservation equation, 466–467
wavelength requirements reduced by, Fluid flow bounds, 470–471
476–477 Folded bus topology, 41, 155
working and protection, 291 Forests
Fiber to the building (FTTB), 20, 409 definition of, 671
Fiber to the cabinet (FTTCab), 20, 409 of multicast connections, 505
Fiber to the curb (FTTC), 20, 409 Steiner, 635
Fiber to the home (FTTH), 3, 20, 409 Fortuitous destinations, 136, 513, 515,
Fiber to the node (FTTN), 20 528
928 Index

Forward error correction (FEC) G.872 networking standard, 32–33

encoding/decoding functions of, 232 Gain profile, of EDFAs, 194
end-to-end transmission channel and, 228, Gain saturation, of EDFAs, 194–195
231–232 Garbling, in routing, 137
equalization combined with, 234 Gate arrays
Forwarding equivalence class (FEC), 723 FPGA, 813
Forward pumping EDFA, 192–193 as guided-wave switches, 248–249
Four-fiber bidirectional line-switched ring Gateway interconnection example, 14–16,
(BLSR), 657 578–580
Four-fiber BLSR. See Four-fiber bidirectional Gateway node, 442
line-switched ring Gaussian approximation, 226, 439
Four-fiber shared-protection ring GCSR. See Grating-assisted codirectional
(SPRING) coupler with sampled grating reflector
link failure survived by, 674 Generalized dynamic nodes, 46–48, 52–53
node failure survived by, 675 Generalized Kautz hypergraph, 543–544, 615
WDM, 453–458, 673–675 as functionally equivalent to Kautz digraph,
Four-fiber SPRING. See Four-fiber 622–623
shared-protection ring symmetric case of, 621
Four-wave mixing (FWM), 182, 184–185, tripartite representation of, 621–622
276, 281 Generalized label object, 749
FP filter. See Fabry-Perot filter Generalized label request object, 749
FPGA. See Field programmable gate array Generalized loopbacks, 690–692
FP laser. See Fabry-Perot laser Generalized Multiprotocol Label Switching
FP resonator. See Fabry-Perot filter (GMPLS), 717
Frame schedule, for time division techniques, generality of, 729–730, 752–753
108 lightpath establishment and tear-down in,
Free spectral range (FSR), 260 750
Frequency shift keying (FSK), 67, 813–815 link management in, 730–733
Front-end amplifiers, 219–221 LSP hierarchy in, 741
FSK. See Frequency shift keying in optical control plane, 729–751
FSR. See Free spectral range OSPF-TE supporting, 738–742
FTTB. See Fiber to the building overview of, 729–751
FTTC. See Fiber to the curb path protection in, 736–737
FTTCab. See Fiber to the cabinet provisioning in, 744
FTTH. See Fiber to the home routing in, 730, 734–742
FTTN. See Fiber to the node RSVP-TE supporting, 749–750
FT-TR. See Fixed transmitters with tunable signaling in, 730, 742–751
receivers Generalized switches
Full grooming, 597 cross-points in, 52
Full wavelength interchange, 495 power transfer relations of, 53
bidirectional rings with, 493–494 Geometric optics, 168–171
fluid flow bound and, 470–471 Gigabit Ethernet, 847
RCA as optimization problem and, Gigabit passive optical network (GPON), 414,
468–469, 470–471, 493–494 849
Full width at half maximum (FWHM), Gimbaled mirror, 252–253
260–261 Glass fiber, 1–2
Fully connected graph. See Complete graph Global communications infrastructure,
Fundamental cycles, 664, 871 demand for, 1–2
Fundamental soliton solution, 188 GMPLS. See Generalized Multiprotocol Label
FWHM. See Full width at half maximum Switching
FWM. See Four-wave mixing GPON. See Gigabit passive optical network
 Index 929

Graded-index fiber Gratings

intermodal dispersion in, 170, 177–178 AWG, 259, 265–266, 271–273, 417–418,
rays in, 169–171 790, 796
Graded refractive index (GRIN) lens, 263 fiber, 182, 259, 263–265, 361–362,
Granularity 365–366
fluid flow bound and, 471 as optical filtering technology, 263–265
grooming resolving, 581 SSFBG, 365–366
in hierarchical networks, 16, 18 waveguide, 259, 263–265
label stacking influencing, 727 GRIN lens. See Graded refractive index lens
in optical networks, 12, 16 Grooming, 637–639
optical switches and, 13–14, 78 algorithms for, 584, 603–606
performance influenced by, 538–540 auxiliary graph for, 605–607
refinement of, 396–397, 539–540 benefits of, 597–598, 601
in SONET, 81 in dynamic RCA, 601–602, 605
Graph coloring edge, 606
brute force approach in, 451–452 full, 597
complexity of, 452 granularity resolved by, 581
Dsatur algorithm for, 872 in hypernets, 637–639
minimal edge, 636 in logical topology, 597–607
minimal vertex, 450–451 in LRNs, 581–585, 597–607, 642
in OCGs, 131 of multihop connections, 637
in optimization, 467–468 in multipoint logical topologies, 637–639
static RCA and, 449–452 of multitier networks, 581–585
Graphs. See also Digraphs; Graph coloring; node for, with optical bypass, 583–584
Hypergraphs; Logical connection graph optimal, 599–600
auxiliary, 605–607 in point-to-point logical topologies,
bipartite, 869, 871 597–607
bridgeless, 460, 685 policies for, 607
circuits in, 871 in SONET, 896n1
complete, 869–870 source, 597
connected, 871–872 in static RCA, 600–602
connection interference, 529–531 subrate functions of, 583
deBruijn, 436, 586 in WDM mesh networks, 602–607
distance network, 502 in WDM rings, 598–602
Hamilton, 690 Group velocity dispersion (GVD), 178–179
maximal, 436 Guard bands, 36, 259
maximal independent set in, 869–870 Guard times, 108
Moore, 434–437, 452, 468, 508–509 Guided wave optical transmission, 18, 21
Moore bound for, 434, 543, 586, Guided wave propagation
872 modes of, 171–174
multigraphs as, 560–561, 869 principles of, 168–174
n-cube, 436 rays and geometric optics in, 168–171
OCG, 131 Guided-wave switches
path interference, 449–452, 467–468 bubble, 249
subgraphs of, 869–870 controllable directional coupler as,
terminology of, 869–873 244–246
theory of, 869–874 fabrication of, 244
walks of, 870–871, 874 gate array, 248–249
Grating-assisted codirectional coupler MZ, 246–247
with sampled grating reflector (GCSR), Y-branch, 248, 284
815 GVD. See Group velocity dispersion
930 Index

Haberman’s algorithm, 505 Hybrid distributed-discrete amplification,

Hairpinning, 584–585, 597 199–200
Hamilton graph, 690 Hybrid electric and optical buffering
Header error correction (HEC) field, 763 contention resolution by, 784–787
Header processing, 789–790 packet loss by, 786
Header routing, 624 performance analysis of, 785–787
Head-of-the-line (HOL) blocking, 406, queuing model for, 786
771–773, 776, 785 Hybrid routing model, 718–719, 751–752
HEC field. See Header error correction field Hybrid switches, 296–297
Helmholtz equation, 172 Hyperarc, 541–543, 621
Heterodyne cross-talk. See Cochannel Hyperedge
heterodyne cross-talk full connectivity in, 634
Heterodyne optical receiver (OR), 73–74, as hyperarc, 541–543
234–235 in LCHs, 540–541
Heterogeneous traffic scheduling, 379 load of, 626–628
Heuristic algorithms MPS as, 147, 149
ADH, 502 Hypergraphs. See also Logical connection
CH, 502 hypergraph
DAH, 502 auxiliary, 636–637
DNH, 502 density of, 613–615, 623
limiting-cut, 886–889 directed, 614–615, 617, 619–621, 623–625,
MBFS-d, 561–563, 567 639, 874
MCH, 703–704 duality constructing, 614–615, 616–619
minimum-hop, 474–479 edge grouping constructing, 614–615,
Monte Carlo approach, 479–484 620–621
SCH, 502 fan-out in, 613
SPT, 501–502, 561–564 Kautz, 543–544, 615–625, 629–630,
for static RCA, 474–486 639–641
for traffic in WDM mesh networks, 604–605 Moore, 613–614
for traffic in WDM rings, 600–602 OCH, 517–518, 533–535
Hierarchical networks terminology of, 873–874
granularity in, 16, 18 underlying graph of, 610–611
long-haul networks as, 16–17 undirected, 613, 623, 625, 628–630,
manageability of, 18 639–640, 873–874
MANs in, 16–18 Hypernets. See also ShuffleNet
ONNs in, 17 algorithms for, 632–633
rationale for, 16–18 applications of, 608
Hitless reconfiguration, 596 dense, 613–615, 623
HOL. See Head-of-the-line blocking design of, 85, 87, 95, 156–161, 632–641
Holey fibers. See Photonic crystal fibers of diameter 1, 540–541
Holographic liquid crystal switch architecture, directed, 608, 630–632, 640
254–255 embedding of, 158, 160, 634–637
Homodyne cross-talk. See Cochannel grooming in, 637–639
multipath cross-talk Kautz, 615–631, 641
Hot potato routing. See Deflection routing layered view of, 607–608
Hybrid approach, to optical networks, 13–16, LCH for, 608–609, 632–633
20, 60, 87. See also Hybrid electric and LLN realization via, 147–149
optical buffering; Logically routed logical layer of, 632–633
networks logical topology building, 85, 87, 145,
Hybrid control paradigm, in optical control 147–149, 156–161, 607–632
plane, 722 LRNs as, 87, 95, 156–161
 Index 931

LSNs in, 157–161 Intensity-modulated direct-detection (IM/DD)

LTNs in, 157–161 systems, 168
MPSs in, 143, 158–161, 608–612, 631–632, CDMA and electronic processing in,
634–637 354–359
multicast virtual connections in, CNR and, 227
631–632 digital signal detection in, 221–226
multihop networks v., 628–631 ORs in, 217–227
multipoint connections in, 156–161 typical waveforms in, 222
as multipoint logical topologies, Intensity modulation, 168, 214, 217–227, 341,
607–632 354–361
multistar, 149, 540–544, 639–641 Interchannel cross-talk, 62, 236–237, 258
physical layer of, 634–637 Interconnected rings
physical topology of, 158 BLSR-BLSR, 659, 663
reduced to point-to-point networks, 612 BLSR-UPSR, 659–663
tree topology in, 158, 634–636 dual-access SHR, 657–663
undirected, 608–610, 635–636 dynamic RCA case study of, 495–496
waveband assignment in, 635–637 ripple effect in, 663
simulation case study of, in MANs,
IETF. See Internet Engineering Task Force 844–845
IGP. See Interior Gateway Protocol single-access SHR, 657–663
ILEC. See Incumbent local exchange carrier UPSR-UPSR, 659, 663
ILP. See Integer linear program WDM composed of, 302–308
Image network, 890–891 Interface identification object, 749–750
IM/DD systems. See Intensity-modulated Interfaces, of optical control plane
direct-detection systems illustration of, 721
Impact-ionization effect, 219 information abstraction in, 719–720
Impairments. See Performance impairments, Interface Switching Capability Descriptor, 740
in network environment Interference. See also Min-Int algorithm;
Incumbent local exchange carrier (ILEC), Multilayer interference filter; Optical beat
841, 845–847, 853 interference
Information abstraction, 719–720 in channel assignment, 528–531
Infrastructure swapping, 860 in CIR, 342–345
Injection electroluminescence, 202 color clash caused by, 546–547
Injection laser. See Semiconductor laser diode connection interference graph for, 529–531
I-NNI. See Interior Network-Network constraints from, 449–452
Interface digital signal detection and, 221–226
Input buffering, 771–772 ISI, 177, 180
Inseparability MAI, 353, 358, 366
illustration of, 548 path coalescence causing, 546–547
in LLNs, 135–136, 138–139, 546–550 path interference graph for, 449–452,
path coalescing caused by, 559 467–468
Insertion loss, in couplers, 243–244, 250–251, Interferometers
255 DI, 230
Integer linear program (ILP), 449 MZ, 214–215, 230, 246–247, 279–280, 788,
for design of optimum light-trees, 637–638 814–815
model, 503–504 Interferometric cross-talk. See Cochannel
in path-based protection, 694–697 multipath cross-talk
for p-cycles, 689–690 Interior Gateway Protocol (IGP), 734, 737
for SRGs, 700 Interior label-switching router (LSR), 723
for WDM mesh network, 603–604 Interior Network-Network Interface (I-NNI),
Intelligent optical switch (IOS), 857 720, 744
932 Index

Intermediate System to Intermediate System Just enough time (JET) protocol

(IS-IS) protocol, 726 FDLs in, 802
Intermediate System to Intermediate System JIT protocol v., 805
protocol with Traffic Engineering in OBS, 801–803
extensions (ISIS-TE), 730 priority, 803
Intermodal dispersion, 170, 177–178 void-filling scheduling algorithms for,
International Telecommunication Union (ITU) 802–803
networking standards Just in time (JIT) protocol
G.872, 32–33 JET protocol v., 805
transmission bands standardized by, 35 messages in, 803–804
wavelength bands of, 176 in OBS, 803–806
wavelength grid standardized by, 35 signaling performance of, 804–806
Internet Engineering Task Force (IETF), 717 variations of, 805–806
Internet Protocol (IP)
evolution of, 19 Kautz digraphs, 615
MPLS applications for, 729 density of, 589–590
routers for, 4, 83–84 directed Kautz hypernets compared with,
Internet Protocol (IP) network 630–631
IP routers in, 4, 83–84 generalized, 589
as LRN, 83–84, 152–153 generalized Kautz hypergraph as functional
MPLS in, 722–729 equivalent of, 622–623
Internet Protocol (IP) routers, 4, 83–84 Kautz hypergraph duality relations to,
Internet service providers (ISPs) 617–618
as key networking players, 19 maximum throughput of, 590–591
SLAs with, 5–6 ShuffleNet compared with, 630–631
Internodal distances undirected Kautz hypergraph constructed
as optical hops, 31–32, 438–439 from, 639–640
in random networks, 438–440 Kautz hypergraphs, 641
in ShuffleNet, 439–440 directed, 615–621, 623–625
Internodal network links, 30 duality construction of, 616–619
Interoffice (IOF), 304 duality relations of, 617–618
Intersymbol interference (ISI), 177, generalized, 543–544, 615, 621–623
180 undirected, 623, 625, 628–630, 639–640
IOF. See Interoffice Kautz hypernets, 615–628, 641
IOS. See Intelligent optical switch directed, 641
IP. See Internet Protocol growth of, 616, 623
IP network. See Internet Protocol network parameters of, 615–616
IRP. See Island Restoration Protocol performance comparisons of, 626–631
ISI. See Intersymbol interference routing algorithms for, 623–625
IS-IS protocol. See Intermediate System to useful properties of, 615
Intermediate System protocol Kautz network, 436–437
ISIS-TE. See Intermediate System to k-edge connected graph, 871
Intermediate System protocol with KEOPS testbed. See KEys to Optical Packet
Traffic Engineering extensions Switching testbed
Island protection. See Segment protection Keying
Island Restoration Protocol (IRP), 700 CRZ, 231
Isolator, 257 CSRZ, 231
ISPs. See Internet service providers DPSK, 813, 815–816
ITU networking standards. See International D-RZ, 231
Telecommunication Union networking FSK, 67, 813–815
standards NRZ, 229–231
 Index 933

OOK, 67, 221, 228–230 EA-DFB, 215–217

PSK, 67, 229–230 EML, 214–215
RZ, 229–231 FP, 206–207
KEys to Optical Packet Switching (KEOPS) MFL for, 418
testbed mode-locked, 208–209
network components in, 795–796 modulation technology for, 212–217
node architecture for, 793–795 origin of, 18
packet format for, 794 in OTs, 68, 70–71, 205–211
k-node connected graph, 871 semiconductor, 18, 205–207
k shortest path (k-SP) algorithm, 487 single-frequency, 181, 207–211
Min-Int algorithm vs., 549–550, 553–554 supercontinuum generation and, 208–209
for point-to-point connections, 547–549 tunability of, 210–212
for single waveband, 553–554 VCSEL, 205, 211–212
as static, 555 Latest available unused-channel first with void
steps of, 548 filling (LAUF-VF) algorithm, 803
k-SP algorithm. See k shortest path algorithm Latin routers, 45, 265
LAUF-VF algorithm. See Latest available
Label Distribution Protocol (LDP), 726, 730 unused-channel first with void filling
Label edge router (LER), 723–725 algorithm
Label Information Base (LIB), 723 Layer-2 protocols, 729
Label object, 748 Layers. See Logical layers; Multilayered
Label Request object, 748 networks; Optical-layer protection;
Labels. See also Optical label switching Packet switching, in optical layer;
protocols for assignment of, 725–726 Physical layers
reuse of, 724 LC. See Logical connection
signaling for distribution of, 745 LCC traffic model. See Lost calls cleared
stacking of, 726–727 traffic model
swapping of, 723–724, 809–811 LCG. See Logical connection graph
Label set object, 749 LCH. See Logical connection hypergraph
Label-switched path (LSP), 723 LC switch architecture. See Liquid crystal
control driven, 725 switch architecture
data driven, 725 LDC. See Linear divider-combiner
hierarchy of, in GMPLS, 741 LDP. See Label Distribution Protocol
in MPLS, 717, 723, 725 LDR. See Limited deflection routing
nested, 741–742 Least loaded routing algorithm, 489
tunnel establishment of, 748–749 LEC. See Local Exchange Carrier
Label switching. See Optical label switching LER. See Label edge router
Label-switching router (LSR), 85, 723–724, LIB. See Label Information Base
727 Lightpath, 7, 443, 603–604
LAMBDANET, 20 edges, in WDM mesh networks, 606
LANs. See Local area networks (LANs) establishment and tear-down of, 750
Large effective area fibers, 186–187 minimization of, 607
Large-scale free-space integrated switch Light-tree, 497–499
fabrics, 250–253 balanced, 505
LARNet, 417 ILP for optimum design of, 637–638
LAS. See Local access subnets list of, 499
Lasers MC-WRN and, 501–504
arrays of, 210–211 source-destination, 498
DBR, 207–208, 839 Lightwave networks. See Optical networks
DFB, 207–208, 215–217 Limited-connection case, 397
DML, 212–214 Limited deflection routing (LDR), 768–770
934 Index

Limiting cut bound, 446–448 spectrum partitioning in, 557–558

Limiting cuts, 446–558 wavelength requirements of, 433
heuristic for, 886–889 wavelength-routed networks compared
multicommodity flow problem and, 886 with, 150–151, 527
Line amplifiers, 190 Linear networks. See Linear lightwave
Linear divider-combiner (LDC), 47, networks
141–142 Linear optical amplifier (LOA), 205
combining ratios of, 54 Line-based survivability schemes
σ -δ, 141–142 path-based v., 678, 705
dividing ratios of, 54 protection in, 651–652, 657, 674, 679–692
dynamic nodes and, 47, 53–56 Line protection. See Shared optical protection
in LLNs, 134 Line sum, 875
permutation switches and, 54 Line terminating element (LTE), 653
power transfer relations of, 55 Linewidth enhancement factor, 212
in rooted routing, 564–565 Link bandwidth parameter, 740
routing rules and, 566 Link bundling, 733, 739–740
as star coupler, 54 Link capacity, 2, 10, 592
tree decomposition and, 566 Link cost metric, 740
Linear lightwave networks (LLNs), 23, 94–95, Link encoding type, 740
133–151, 162–163 Link failure, 652, 674
advantages of, 140 APS using protection cycles against,
channel assignment in, 528–540 686–687
color clashes in, 138–139, 546–547 CDC protecting against, 681–683
connection rearrangeability in, 139 generalized loopbacks protecting against,
constraints in, 509, 528–529 690–692
directed trees in, 513–515 in planar networks, 686–687
DSC in, 135–139, 512–513 protection against, 681–683, 685–687,
full connectivity in, 140, 145–147 690–692, 750–751
geographical reach limitation in, 577–578 Link functionality, 22
hypernet realization of, 147–149 Link management
inseparability in, 135–136, 138–139, in GMPLS, 730–733
546–550 LMP in, 719, 732–733
LCHs for, 541 neighbor discovery in, 731–732
LDCs in, 134 Link Management Protocol (LMP), 719
local access to, 519–521 features of, 732
Min-Int algorithm for routing in, 890–892 in GMPLS, 732–733
MPSs in, 140–145, 508, 540–544 link bundling in, 733
multipoint connections in, 134 in MPLS, 719
multistar, 149, 540–544 Link p-cycles, 688
network connections in, 94–95, 133–151, Link protection type, 740
558–567 Link State Advertisement (LSA), 738–741
nonblocking stations in, 510, 518–519, Link state algorithms, 735
527–528 Link state information dissemination,
ONN functionality in, 133 739–741
performance parameters of, 536–537 Link state protocols, 735–736. See also Open
properties of, 95 Shortest Path First protocol
purely optical approaches in, 576–578 Liquid crystal switch architecture
routing in, 135–140, 507–568, 890–892 advantages of, 254
as single-wavelength-per-waveband holographic, 254–255
networks, 134 multiwavelength, 269–270
special cases of, 134 Little’s formula, 885
 Index 935

Livelock, 765 for multistar LLNs, 541

LLC frame. See Logical link control frame for undirected hypernets, 608–609
LLN. See Linear lightwave networks Logical layers, 6–8, 13, 29
LMP. See Link Management Protocol adjacency in, 731
LOA. See Linear optical amplifier control techniques for, 21
Local access subnets (LASs), 91, 326 current protection and restoration
demand-assigned LCs in, 538–540 techniques in, 650–669
as elementary access station, 521 design of, 591–593, 632–633, 642
on Petersen network, 519–521 in hypernets, 632–633
Local area networks (LANs) logical topology of, 12, 77, 82, 97
early structures proposed for, 13, 20 in multihop networks, 591–593
local v. global, 114n14 of multilayered networks, 10, 12–13, 21, 29
Local carriers opaqueness of, 8, 10, 12
competition from, 3 physical layer interface with, 30
ILEC, 845–847 reconfigurable, 16, 79–80, 82, 94
LEC, 831, 841, 845–847, 853 Logical link control (LLC) frame, 114–116,
Local exchange carrier (LEC), 831, 841, 120–121
845–847, 853 Logically routed networks (LRNs), 94–95,
Logical connection (LC), 432, 578. See also 151–163. See also Hypernets;
Demand-assigned logical connections; Synchronous optical network
Logical connection graph; Logical ATM network as, 81–85
connection hypergraph characteristics of, 579–580
fixed-frame scheduling for, 534 design challenges of, 153, 641–642
in LRNs, 581–582 generic arrangement for, 77–78
management of, 100–102 grooming in, 581–585, 597–607, 642
many-to-one, 102, 162 hop count v. network capacity in, 153
multicast, 380–383 hypernet design of, 85, 87, 95, 156–161,
multiplexing of data streams from, 68 632–641
multipoint, 102 as hypernets, 87, 95, 156–161, 607–641
NAS as interface for, 67 IP network as, 83–84, 152–153
notation for, 103 LCG for, 585
ONCs and, 101 LCs in, 581–582
one-to-many, 102, 104, 162 logical topologies of, 145, 153–156, 580,
one-to-one, 104, 110, 162 585–593, 607–632
packet switching and, 111, 121–122, LPs in, 578
399 LSNs in, 151–161
transmission channel and, 31 MPLS as, 84–85
Logical connection graph (LCG), 140–141, multihop design of, 433, 591–607
442 multipoint, 607–632, 641
embedding of, on physical topology, physical topologies of, 580–581, 585
634–637 physical topology of, 436–437, 580–581,
in logical-layer design, 591–593 585
logical topology in, 91–93 point-to-point, 153–156, 578–579,
for LRNs, 585 585–591, 641
ShuffleNet, 593 rationale for, 576–581
structural bounds on, 586 recovery in, 580
Logical connection hypergraph (LCH), regularity in, 586
147–148, 382 SONET as, 79–81
diameter of, 158 STSs in, 580
hyperedges in, 540–541 versatility of, 641
for hypernets, 608–609, 632–633 wavelength requirements of, 433
936 Index

Logical networks, 6–7. See also Logically physical topology independent of, 79, 94
routed networks physical topology matched with, in static
connections organized by, 12 RCA, 444–448
hybrid architecture and, 16 point-to-point, 585–591, 597–607
LSNs in, 12, 77 survivable, 649
as middlemen, 12–13 wavelength requirements of, 444
overlays, 77–85 Long-haul networks. See also Ultra long-haul
Logical path (LP) networks
in LRNs, 578 commercial network deployments and,
SONET path v., 651 856–858
Logical signal, conversion of, 31, current considerations in WADM
67–68 architectures and, 854–856, 862
Logical switch (LS) hierarchical, 16–17
ADM as, 79 mesh architectures for, 855
in ATM network, 81 Long period fiber grating, 265
bridges as, 461–462 Loopback connections
DCS as, 79 for control purposes, 141
functions of, 77–79, 85–86 of dynamic nodes, 55
NAS as interface for, 78 generalized, 690–692
ONN interfaced with, 77 Looping algorithm, 51, 765, 897–898
Logical switching nodes (LSNs), 8, Loss. See also Packet loss
93–94 from attenuation, 505
degrees of, 153, 157 of directional couplers, 44
in hybrid architecture, 15–16 information, 329
in hypernets, 147–148, 157–161 insertion, 243–244, 251, 255
in IP networks, 83 from material absorption, 175
in logical networks, 12, 77 path-independent, 282
in LRNs, 151–161 power, 504–505
multiplexing and, 94 from Rayleigh scattering, 175
in optical networks, 152–153 splitting, 504–505
in ShuffleNet, 157 throughput, 36–38
Logical terminal node (LTN), 157–161 traffic, 673
Logical topology (LT), 12, 77, 97, 150. See from waveguide imperfections, 176
also Multipoint logical topology; Lossless directional couplers, 42, 54–55
Point-to-point topology Lossless scheduling, 405–407
in ATM networks, 82–83 Lost calls cleared (LCC) traffic model, 391
cyclic changes in, 596 Low-loss fibers, 1, 18
dense, 589–591 LRN. See Logically routed networks
digraphs for, 589–591 LS. See Logical switch
dynamic capacity assignment and, 113 LSA. See Link State Advertisement
embedded in physical topology, 91, LSNs. See Logical switching nodes
594–596, 634–637, 642 LSP. See Label-switched path
grooming in, 597–607 LSR. See Label-switching router
hypernet, 85, 87, 145, 147–149, 156–161, LT. See Logical topology
607–632 LTE. See Line terminating element
LCGs showing, 91–93 LTN. See Logical terminal node
of logical layers, 12, 77, 82, 97
of LRNs, 145, 153–156, 580, 585–593, Mach-Zehnder interferometer (MZI),
607–632 214–215, 230
in multihop networks, 585–591, 594–596 drawbacks of, 246
multipoint, 607–632 fabrication of, 246–247
 Index 937

regeneration by, 279–280 Max Reuse. See Maximum reuse

SOAs using, 788, 814–815 Maxwell’s equations, 171, 255
Mach-Zehnder (MZ) switch MBFS-d algorithm. See Modified breadth-first
cascaded, 247 search with parameter d algorithm
cross-talk from, 247 MC-CR. See Minimum-cost collapsed-ring
guided-wave, 246–247 MCH. See Minimum-cost heuristic
modification of, to WADM, 264–265 MC-LGS. See Multicast-capable logical
single thermo-optically controlled, grooming switch
271–272 MC-WRN. See Multicast-capable
MAC protocol. See Media access control wavelength-routed network
protocol M|D|1 queue, 884–885
MAI. See Multiple access interference Media access control (MAC) protocol,
Managed reach solution, 856 86
Management of Photonic Systems and ALOHA, 401–403
Networks (MEPHISTO), 840 characteristics of, 115, 120
Management plane contention arbitrated by, 117
control plane v., 714 CSMA/CD as, 113–114, 120
resources in, 721 Ethernet, 402
Manhattan Street Network (MSN), 765, 770 lossless scheduling, 405–407
MANs. See Metropolitan area networks for optical packet switching, 111–121,
Many-to-many optical connection (OC), 399–408
533–535 perfect scheduling, 407–408
Many-to-one logical connection (LC), 102, performance-to-cost ratio for, 120
162 for PONs, 412–416
Many-to-one optical connection (OC), 533 position of, in MWNA, 120–121
Markov chains role of, 102, 113, 400
birth-death process as, 882–883 sublayers of, 111–121
for demand-assigned LCs, 391 tell-and-go, 403–405
discrete-state continuous-time, 881 uncontrolled scheduling, 401–403
ergodic, 882 Memory. See Random access memory
Markov processes and, 881–883 Memoryless distribution, 881
queues and, 883–885 MEMS. See Microelectromechanical system
random processes and, 880–883 MEPHISTO. See Management of Photonic
Markov processes, 388, 881–883 Systems and Networks
Matched filter, 356 Mesh architectures
Material absorption, 175 arbitrary, 663–669
Matrix scaling, of traffic, 379–380, 593 centralized restoration techniques in,
Maxband, 545–546 665–666
Max flow-min cut theorem, 446, 872–873 distributed restoration techniques in,
Maximal graph, 436 666–669
Maximal independent set, 869–870 for long-haul networks, 855
Maximizing resource utilization (MRU), 605 optical-layer protection in, 677–703
Maximizing single-hop traffic (MST), ring covers in, 664–665, 669
604–605 shared optical protection in, 679–692
Maximum free capacity (MFC), 489 spare capacity of, 669
Maximum likelihood (ML) detection, 225 ultralong-haul networks and, 856
Maximum reuse (Max Reuse) Mesh network. See also Wavelength
channels allocated by, 550–551, 552–553 division-multiplexed mesh network
color clashes caused by, 563 access-fiber bottleneck in, 132
Min Reuse v., 552–553, 567 algorithms for, 603–606
wavelength list ordered by, 486–487, 499 blocking probability in, 494–495
938 Index

Mesh network. (cont.) M|G|1 queue, 884–885

CDCs in, 460, 680–687 Microelectromechanical system (MEMS), 20,
commercial deployments of, 856–858 242
connectivity implemented on, 143, 145, design challenges of, 252
147 electrostatic v. magnetic, 253
Eulerian, 459–460 in large-scale free-space integrated switch
full wavelength interchange in, 495 fabrics, 250–253
lightpath edges in, 606 mirrors in, 271
RCA in, 131–133, 494–495 scalability of, 250–251
ring decomposition of, 458–462 3D, 252–253
single fiber pair access in, 131–132 2D, 250–251
sparse wavelength interchange in, 460–461 WADM based on, 270–271
spectrum reuse in, 133, 150–151 Microstructured fibers. See Photonic crystal
transport network in, 715 fibers
tree embedded in, 147–149 Microwave subcarriers, 227
tree topology embedded in, 140–142, Microwave wireless systems, 227
147–149 Midsystem spectral inversion, 182
Metro networking state of the art MI filter. See Multilayer interference filter
emerging technologies and architectures in, Minband, 545–546
847–850 Minimal cut, 871
metro WDM network deployments, Minimal edge coloring, 636
851–854 Minimal vertex coloring, 450–451
network elements in, 850–851 Minimum-cost collapsed-ring (MC-CR),
Metropolitan area networks (MANs) 703–704
amplets in, 204–205 Minimum-cost heuristic (MCH), 703–704
capacity expansion of, 843 Minimum cost routing algorithms, 637
challenges of, 841 Minimum-hop heuristic, 474–479
CSMA/CD impossible in, 116 Minimum interference, 892
current state of, 845–847 Minimum reuse (Min Reuse)
deployment case study of, 853–854 channels allocated by, 550–553
domain of, 842–843 color clashes minimized by, 551–553
DWDM in, 845–846 Max Reuse v., 552–553, 567
early structures proposed for, 13, 20 wavelength list ordered by, 486–487
evolution of, 844–847 Min-Int algorithm
goal of, 861–862 as adaptive, 555
in hierarchical networks, 16–18 illustration of, 550, 888–889
ILEC, 845–847 image network and, 890–891
initial vision of, 844–845 k-SP algorithm v., 549–550, 553–554
interconnected-ring simulation case study for point-to-point connections, 547–549
of, 844–845 for routing in LLNs, 890–892
legacy architectures in, 842–843 for single waveband, 553–554
in multiwavelength optical networking, Minority carrier injection, 202
841–854, 861–862 Min Reuse. See Minimum reuse
networking state of the art, 847–854 MIP. See Mixed integer program
transparency in, 844–845 Mirror frequency, 64–65
typical ring architectures in, 850–851 Mirrors
unique characteristics of, 841–842 gimbaled, 252–253
WADMs in, 842, 851 in MEMS-based WADM, 271
WDM in, 35–36, 843–847, 851–854 in signal regeneration devices, 280,
MFC. See Maximum free capacity 789
MFL. See Multifrequency laser source in SLALOM, 789
 Index 939

Mixed integer program (MIP), 449 Moore hypergraph, 613–614

in hypernet design, 632–633 MPCP. See Multi-point Control Protocol
in multihop network design, 592 MPLS. See Multiprotocol Label Switching
in static RCA optimization, 464–465 MPS. See Multipoint subnet
ML detection. See Maximum likelihood MRU. See Maximizing resource utilization
detection MSN. See Manhattan Street Network
M|M|1 queue, 884 MSP. See Multiservice platform
Mode coupling, 177 MSPP. See Multiservice provisioning
Mode-locked laser, 208–209 platform
Modified breadth-first search with parameter d MST. See Maximizing single-hop traffic
(MBFS-d) algorithm Multicarrier recovery, 860–861
definition of, 561 Multicast-capable logical grooming switch
size of trees generated with, 567 (MC-LGS), 637–639
tree decomposition using, 562–563 Multicast-capable wavelength-routed network
Modulation (MC-WRN), 498
analog, 71, 181, 214 light-tree approach to, 501–504
cross-phase, 185 power loss in, 504–505
digital, 71 Multicast connection allocation
formats of, 229–231 algorithm for, 558–568
index, increase in, 345 completion of, 566–568
intensity, 168, 214, 217–227, 341, Multicast connections
354–361 allocation algorithm for, 558–568
in NAS, 70–71 blocking probability for, 567–568
overmodulation and, 345 forest of, 505
in SCMA, 71, 337–339 LC, 380–383
self-phase, 185 routing of, 497–507, 558–567
technology for, 212–217 survivability techniques for, 702–703
MONET. See Multiwavelength Optical threefold, 531
Network project in transparent networks, 500
Monochromatic signal, 172 virtual, in hypernets, 631–632
Monte Carlo approach Multicasting
deterministic approaches compared with, MPLS applications for, 729
479–480 with optical layer constraints, 504–505
flow chart of, 481 optical splitters and, 500, 503, 505–507
rules of, 481–482 sparse, 505–507
as stochastic approach, 479 Multicast optical connection (OC), 532–533
time trace of, 482–483 Multicast optical path (OP), 100, 106, 531
wavelength interchange in, 484 multiplexing and, 105
Monte Carlo simulation, 239 point-to-point connections compared with,
Moore bound 106
for digraphs, 614, 872 possible implementations of, 383
for graphs, 434, 543, 586, 872 RWA with, 522–525
for undirected hypergraphs, 613 Multicast path. See Multicast optical path
Moore graph Multicast tree, 100, 559–560, 631–632
orders of, 436 Multicommodity flow problem, 470, 886
Petersen, 434, 437, 440–442, 468, 508–509, Multidimensional codes, 360–361
766 Multifiber links, 126, 554–555
rings as, 452 Multifrequency laser source (MFL), 418
tessellations of plane in, 435–436 Multigraph, 560–561, 869
38-vertex, 435 Multihop electronic overlay, 94
vertices in, 434–435 Multihop groomed connection, 637
940 Index

Multihop networks. See also ShuffleNet SCM, 348–352, 811–812

deBruijn graphs and, 586 space division, 34
design of, 591–607 time division, 36
as hybrid approach, 20, 87 waveband, 37–38
hypernets v., 628–631 Multipoint connections, 85. See also Static
logical layers in, 591–593 multipoint networks
logical topology in, 585–591, 594–596 in broadcast star networks, 103–106,
LRNs as, 433, 591–607 324–327
optical hops in, 31–32, 438–439 in hypernets, 156–161
origin of, 585–586 LC, 102
physical layer in, 594–596, 642 in LLNs, 134
point-to-point topology and, 585–591 management of, 102
routing in, 591–593 operations for realizing, 109
Multihop packet transmission and multirate in static networks, 371–399
payload testbed, 819–820 terminology of, 31–32
Multilayered networks. See also Physical time division techniques for, 106–109
layers time/wavelength division techniques for,
case for, 12–16 109–111
challenges of, 20 in wavelength-routed networks, 122–123,
file-sharing services in, 6, 8 122n16
logical layers of, 10, 12–13, 21, 29 Multi-point Control Protocol (MPCP), 414
OCh layer of, 32–33 Multipoint logical topology (LT)
OMS layer of, 32 blocking probability in, 638
OTS layer of, 32–33 fully shared, 145–147
physical layers of, 7n4, 10–11, 11n6, 13, grooming in, 637–639
21–22 hypernets as, 607–632
PM layer of, 32 MPS capacity in, 611–612
SONET in, 6–8 Multipoint LT. See Multipoint logical topology
Multilayer interference (MI) filter, 259 Multipoint networks. See Static multipoint
FP filter compared with, 262 networks
integrated array of, 263 Multipoint subnet (MPS)
operation of, 261–262 capacity of, 611–612
switches based on, 271–272 channel-sharing parameters for, 611
tunability of, 262 λ-channels and, 142
Multipath cross-talk. See Cochannel multipath control of, 142–143
cross-talk creation of, 142
Multipath dispersion, 170 full connectivity realized on, 140
Multiple access interference (MAI), 353, 358, as hyperedge, 147, 149
366 in hypernets, 143, 158–161, 608–612,
Multiple waveband routing, 555–557 631–632, 634–637
Multiplexing. See also Add/drop multiplexer; implementation of, 143, 634
Demultiplexing; Dense wavelength in LLNs, 140–145, 508, 540–544
division multiplexing; Time division NASs in, 95, 143
multiplexing; Wavelength division in Petersen networks, 522
multiplexing Multiprotocol Label Switching (MPLS). See
CWDM, 17, 177, 205 also Generalized Multiprotocol Label
of LC data streams, 68 Switching
LSNs and, 94 applications of, 722, 728–729
multicast paths and, 105 flows managed by, 85
multiple access techniques for, 36, 327–367 header format of, 723–724
representative, 327–367 in IP networks, 722–729
 Index 941

LMP in, 719 taxonomy of, 85–87

as LRN, 84–85 in WANs, 29–30
LSPs in, 717, 723, 725 Multiwavelength optical networking. See also
LSRs in, 723–724 Multiwavelength network architecture
in optical control plane, 717, 719, 722–729 advantages of, 830
packet transport through, 722–727 business drivers and economics in,
protocol stack of, 727–728 828–837, 861
tunnels formed by, 726–727, 748–749 current trends in, 828–862
VPNs allowed by, 722, 728 long-haul and ultralong-haul, 854–858, 862
Multiservice platform (MSP), 857 MANs in, 841–854, 861–862
Multiservice provisioning platform (MSPP), new applications of, 858–861
849 outlook for, 862
Multistar network testbeds in, 828, 838–841, 861
connectivity implemented on, 143, 145 Multiwavelength Optical Network (MONET)
fiber length in, 464 project, 840–841
hypernets as, 149, 540–544, 639–641 Multiwavelength switch (MWS)
LLN, 149, 540–544 AOTF, 267–269, 838
optical spectrum efficiently utilized by, 464, AWG-based, 271–273
640 liquid crystal, 269–270
physical topology of, 144–145, 150, MEMS-based wavelength-selective,
462–464, 540–544, 640–641 270–271
spectrum reuse in, 150–151 MI filter-based, 271–272
waveband-routed, 542–543 as optical and photonic technology, 266–274
wavelength-routed, 462–464 realization of, 266–267
Multistar physical topology, 150, 462–464, wavelength-dilated, 273–274
540–544 Multiwavelength transport network (MWTN)
Multitier networks project, 839
definition of, 581 MWNA. See Multiwavelength network
grooming of, 581–585 architecture
inefficiencies of, 584–585 MWS. See Multiwaveband switch;
OEO switches in, 582–583 Multiwavelength switch
two-tier, 582–584 MWTN project. See Multiwavelength
VCs in, 582 transport network project
Multivendor networks, 19 MZI. See Mach-Zehnder interferometer
Multiwaveband directional couplers, 57 MZ interferometer. See Mach-Zehnder
Multiwaveband switch (MWS), 56–57 interferometer
power transfer relations of, 57 MZ switch. See Mach-Zehnder switch
wavelength-dilated switch v., 273–274
Multiwavelength network architecture NAS. See Network access station
(MWNA) National Transparent Optical Network
connectivity in, 23, 91–163 Consortium, 840
current applications of, 756 NC. See Network connections
as framework, 34 n-cube graph, 436
high-bandwidth optical channels in, 12 NDF. See Nonzero dispersion fiber
history of, 18 Negative dispersion fiber, 214
MAC protocol positioning in, 120–121 Neighbor discovery
multiple layers in, 6–8, 12 adjacency discovered through, 731
nonlinear effects in, 186–187 approaches to, 731–732
origins of, 4–5 automated, 731
resources of, 28–87 implemented on transport links, 720n3
sublayers in, 29, 31, 75 in link management, 731–732
942 Index

Neighbor discovery (cont.) Network nodes. See also Dynamic nodes;

in OOO switch, 733 Logical switching nodes; Optical network
as optical control plane function, 721 nodes
provisioning system for, 732 architectures of, 783, 793–795, 814, 851
Nested label-switched path (LSP), 741–742 buffered, 770–774
NETRATS protocol. See Network Restoration chooser, 667–668
Algorithm for Telecommunication for contention resolution, 783
Systems protocol core, 818–819
NetSolver, 665 end, 442
NETSPAR protocol, 665–666 failure at, 648, 650, 656–657, 675, 685
Network access station (NAS), 7–8, 104 fiber to, 20
access gateways attached to, 14–15 gateway, 442
in ATM networks, 83 grooming, 583–584
as boundary, 30 for KEOPS testbed, 793–795
in broadcast star networks, 40–41, 324–327 LTN, 157–161
elementary, 448–449, 510, 521, 527–528, maximum throughput per, 588–589
636 in OLS, 812–814
end nodes connected to, 442 ONC for, 99–100
as LC port/transceiver interface, 67 OPS, 760–763
as LS interface, 78 p-cycles encircling, 688
modulation process of, 70–71 primary, 658
in MPS, 95, 143 root, 140, 145
nonblocking, 128–129, 131–132, 448–449, secondary, 658
510, 518–519, 527–528 sender, 667–668
OLPs in, 75, 77 static, 40–46, 432
ONN connected to, 30, 128, 131 in unbuffered networks, 760–763
OT side of, 70–71 vendor generic WADM, 851
packet switching by, 101, 116–117 in WDM ring, 458
in point-to-point connections, 69 Network Restoration Algorithm for
OR side of, 71–74 Telecommunication Systems
in static networks, 102–103, 324 (NETRATS) protocol, 668
traffic type characterized by, 101 Network state, 484
transceiver functions implemented by, 67, Next generation access, 849
77 Next Generation Internet (NGI), 812
typical, 67–68 NGI. See Next Generation Internet
WADM combined with, 58, 70, 131 NMS. See Network management system
wavelength assignment determined by, NNI. See Network-Network Interface
128 Node-encircling p-cycles, 688
Network connections (NCs), 33, 91–163 Nodes. See Network nodes
LLN, 94–95, 133–151, 558–567 Noise
management and control of, 96–102 ASE, 195–197, 200–201, 227, 241
static, 94–95, 102–122 in CNR, 227
in waveband-routed networks, 133–151 in digital signal detection, 221–226
in wavelength-routed networks, 94–95, in EDFAs, 195–197, 200–201
122–133 heterodyne ORs reducing, 74
Network management, 9, 17–18, 20, 39 OBI, 342
Network management system (NMS), 99–100, power, 48
735, 859 in RAs, 200–201
Network-Network Interface (NNI) RIN, 227
exterior v. interior, 720, 744 shot, 196, 220, 227, 235
signaling and, 742–745 in SNR, 167, 219–221
 Index 943

NOLM. See Nonlinear optical loop mirror Objects

regenerator of RSVP, 747
Nonblocking access links, 131–132 of RSVP-TE, 748–750
Nonblocking network access station (NAS), OBS. See Optical burst switching
128–129, 131–132, 448–449, 510, OC. See Optical connection
518–519, 527–528 OCDMA. See Optical code division multiple
Nonblocking stations access
in LLNs, 510, 518–519, 527–528 OCG. See Optical connection graph
in RCA, 128–129, 131–132, 448–449 OCH. See Optical connection hypergraph
Nonblocking switches. See Strict-sense OCh. See Optical channel layer
nonblocking switches; Wide-sense OCSS. See Optical carrier suppression and
nonblocking switches separation
Noncoherent systems OCT. See Optical coherence tomography
CDMA, 354, 361–366 O-E-O switch
intensity-modulated, 360–361 in multitier networks, 582–583
optical processing in, 360–363 O-O-O switches coexisting with, 585
Nonlinear effects, on transmission in OXCs, 59, 78, 275, 294–295, 582–583
effective interaction length, 185–186 Offered traffic, 387, 389, 391
large effective area fibers and, 186–187 OIC. See Optical integrated circuit
self- and cross-phase modulation, 185 OLP. See Overlay processors
stimulated Brillouin scattering, 184 OLS. See Optical label switching
stimulated Raman scattering, 183–184 OLT. See Optical line terminal
Nonlinear optical loop mirror regenerator OMS layer. See Optical multiplex section layer
(NOLM), 280 ONC. See Optical node controller
Nonlinear Schrodinger equation, 187 One-for-N protection, 650, 653–654
Nonplanar case, of cycle double cover (CDC), electrical layer, 654
685–686 optical, 671–672
Nonreciprocal devices, 257 One-for-one facility protection, 650
Nonreflecting broadcast star network, optical, 670–671
522–523 two routes in, 653
Non-return-to-zero (NRZ) keying, 229–231 One-plus-one automatic protection switching
Nonrevertive mode, of one-plus-one facility (APS), 650–653
protection, 655 One-plus-one dedicated backup path
Nonzero dispersion fiber (NDF), 182 protection, 692–693
Normalized line sum, 875 One-plus-one facility protection, 650–651
Normalized maximum hyperedge load, 627 electrical protection in, 670
Normalized quasidoubly stochastic (NQDS) nonrevertive mode of, 652–653, 655
matrix, 875–878 optical, 669–670
Notification request object, 750 revertive mode of, 653–655
No wavelength interchange, 465–468, 470, SONET, 652, 670
474 One-to-many logical connectivity, 102, 104,
NQDS matrix. See Normalized quasidoubly 162
stochastic matrix One-to-one logical connectivity, 104, 110, 162
NRZ. See Non-return-to-zero keying ONM. See Optical network manager
Numerical aperture, 170 ONNs. See Optical network nodes
On-off keying (OOK), 67, 221, 228–230
OA. See Optical amplifier ONTC. See Optical Networks Technology
OADM. See Optical add/drop multiplexer Consortium
OAM field. See Operations, administration ONU. See Optical network unit
and maintenance field OOC. See Orthogonal optical code
OBI. See Optical beat interference OOK. See On-off keying
944 Index

O-O-O switch, 58–59, 78, 583 Optical add/drop multiplexer (OADM). See
neighbor discovery in, 733 also Wavelength add/drop multiplexer
O-E-O switches coexisting with, 585 B&S, 291–293
in opaque networks, 59, 294–296 as network building block, 289
in OXCs, 294–296 parallel, 289–290
OP. See Optical path serial, 289–290
Opaque networks, 8, 10, 12, 21, 60, 757. See wavelength-based, 289–291
also Optical packet-switched networks working and protection fibers of,
G.872 standard for, 33 291
hardware of, 58 Optical amplifier (OA), 30
logical layers in, 8, 10, 12 Optical and photonic device technology
O-O-O, 59, 294–296 all-fiber, 242
open WDM network architecture as, 837 controllability of, 242
optical switches in, 13n7, 21 couplers as, 242–255
OXC, 59, 294–296 as enabling technology, 241–274
practicality of, 583 multiwavelength switch technology as,
transparent networks v., 296 266–274
Opaque wavelength conversion, 278–281 optical filtering technology as, 257–266
Open Shortest Path First (OSPF) protocol, reciprocity and, 255–257
726 switches as, 242–255
adjacency in, 738 Optical beat interference (OBI), 340–345,
Djikstra’s algorithm used in, 738 362
link bundling in, 739 CIR measuring performance of, 342–345
link state information dissemination in, effects of, 341–342
739–741 experiment verifying, 344–345
nested LSPs in, 741 as noise, 342
traffic and, 737–738 packet switching and, 347–348
with traffic engineering extensions, 730, total power of, 343
738–742 Optical buffering. See Buffering
in transparent networks, 741–742 Optical burst switching (OBS), 798–808
Open Shortest Path First protocol with Traffic advantages of, 801
Engineering extensions (OSPF-TE), 730, architecture of, 800
738–742 benefits of, 757
Open Systems Interconnection (OSI) buffering in, 799–800
Reference Model contention resolution and, 806–808
DLC layer in, 84, 120 for Ethernet access, 821
TCP/IP compared with, 83–84 JET protocol in, 801–803
Open walk, 870–871 JIT protocol in, 803–806
Open WDM installations, 835–837 OPS v., 798–800, 821
OPERA. See Optical Packet Experimental packet loss probabilities in, 808
Routing Architecture routing in, 801
Operations, administration and maintenance wavelength conversion in, 801
(OAM) field, 763 wavelength-routed, 805–806
Operations support system (OSS), 859 WDM and, 800–801
OPS. See Optical packet switching Optical carrier suppression and separation
OPSnet (OCSS), 231, 818–820
core router architecture in, 817 Optical channel layer (OCh), 32–33
development of, 815–818 Optical code division multiple access
edge router architecture in, 816 (OCDMA), 352–353, 363–366
OPS networks. See Optical packet-switched Optical code label processing, 818
networks Optical coherence tomography (OCT), 209
 Index 945

Optical connection (OC), 30–31, 85–86, 432 solitons, 187–188

in channel assignment, 516–518, 528–531 spectrum of, 31n3, 34–39, 41, 46, 333–335,
management of, 100–101 464, 557, 640
many-to-one, 533 in Trans-Atlantic cables, 18
multipoint, 31–32 transmission impairments of, 174–187
notation for, 103 Optical filtering technology
originating in OTs, 69 AWG as, 265–266
point of origin of, 69 concatenation in, 240–241, 307
point-to-point, 29–32, 69 FP filters as, 259–262
routing table for, 517 gratings as, 263–265
time required for establishing, releasing, MI filters as, 259, 261–263, 271–272
100 as optical and photonic device technology,
Optical connection graph (OCG), 131 257–266
Optical connection hypergraph (OCH), ORs in, 258–259
517–518, 533–535 packet, 107, 117
Optical control plane static, 267
adjacency in, 719, 731, 738 wavelength-selective switches in, 258–259
advantages of, 716 Optical header processing, 789–790
architecture of, 719–720 Optical hop, 31–32, 438–439
control paradigms in, 722 Optical integrated circuit (OIC), 243
economics influencing, 753 Optical label switching (OLS). See also
functions of, 721–722, 727 Label-switched path
GMPLS in, 729–751 advantages of, 757–758, 821
interfaces of, 719–721 AOLS as, 809–811
MPLS in, 717, 719, 722–729 based on optical code label processing, 818
overview of, 714–722, 751–753 contention resolution in, 811
provisioning of connection between two LSR for, 85, 723–724, 727
routers in, 717 network implementations of, 811–820
reliability of, 752 network nodes in, 812–814
routing models in, 717–719 in OPS networks, 808–820
scalable addressing scheme for, 716n2 scalability of, 808
Optical crossbar switch, 282–284 subcarrier transmission system of, 811–812
Optical cross-connects (OXCs) testbed demonstrations of, 811–820
add/drop traffic and, 78–79 WDM network example of, 808–809
buffer arrays emulating, 776–778 Optical layer. See Packet switching, in optical
functions of, 78 layer
O-E-O arrangement of, 59, 78, 275, Optical-layer protection
294–295, 582–583 in mesh architectures, 677–703
O-O-O arrangement of, 294–296 path-based, 692–700
opaque, 59, 294–296 in point-to-point systems, 669–672
switch fabric of, 59, 243 shared (line-based), 679–692
transparent, 58, 293–295 in SHRs, 672–677
waveband-selective switch as, 57 Optical line terminal (OLT), 91n1, 410–411,
wavelength-selective switch as, 293–297 849
WIXC and, 65 in BPON, 412–414
Optical domain, 10 in EPON, 415
Optical fibers Optical multicast path. See Multicast optical
as enabling technology, 168–190 path
photonic crystal fibers, 188–190 Optical multiplex section (OMS) layer, 32–33
principles of guided wave propagation, Optical network manager (ONM), 99, 124,
168–174 520–521, 545
946 Index

Optical network nodes (ONNs), 7–8, 37, Optical network unit (ONU), 410–411,
39–66 422–424, 849
λ-channels switched by, 30 in BPON, 412–414
circuit-switched operation of, 14, 60 in EPON, 415–416
dynamic nodes of, 40, 46–48 Optical node controller (ONC), 99–101
example of, 715–716 Optical Packet Experimental Routing
functionality of, 39–40, 85–86, 133 Architecture (OPERA), 811–813
in hierarchical networks, 17 Optical packet-switched (OPS) networks. See
limitations of, 36 also Optical label switching; Optical
LS interfaced with, 77 packet switching; Packet switching, in
NAS connected to, 30, 128, 131 optical layer
static nodes of, 40–46 activities of, 758–759
in transparent optical networks, 12 architectures of, 758–787
wavelength assignment determined by, asynchronous, 782–783
128 buffering in, 61, 759, 770–791, 799–800,
wavelength selectivity in, 123 821
Optical networks. See also Multiwavelength enabling technology in, 757, 787–791,
optical networking; Optical network 821
nodes; Optical packet-switched networks; feasibility of, 757
Passive optical networks; Synchronous nodes in, 760–763
optical network; Testbeds, of optical OLS in, 808–820
networks overview of, 756–758, 820–822
bottlenecks in, 576–577 testbeds of, 787, 791–798, 818–821
client service associations in, 32–33 versions of, 757–758
commercial viability of, 2 wavelength-selective cross-connects in, 57
control and management of, 3, 9, 17–18, 20, Optical packet switching (OPS), 4, 11–12,
39, 96–102 60–62, 83
definition of, 1 buffering and, 61
functionality of, 22, 63 contention resolution in, 763–764
general characteristics of, 30 enabling technology for, 757, 787–791, 821
granularity in, 12, 16 with input buffering, 771–772
history of, 18–22 nodes in, 760–763
hybrid approach to, 13–16, 20, 60, 87 OBS v., 798–800, 821
LSNs in, 152–153 packet switching in optical layer v., 112
nonlinear operations performed in, 10–11, promise of, 820–821
21 in unbuffered networks, 759–764
objectives of, 4–9 in WDM, 61
ONMs in, 99, 124, 520–521, 545 Optical Pan-European Network, 839
ONUs in, 410–411, 422–424, 849 Optical path (OP). See also Optical path-based
provisioning of connection between two protection
routers in, 717 coalescing of, 518–519, 527, 546–547
road map for, 23–25 establishment of, 31
signaling in, for protection and restoration, expanded, 511
750–751 feasible, 487–488, 511, 546, 559
standardization of, 3, 20 multicast, 100, 105–106, 383, 522–525, 531
structural features of, 9 notation for, 103
switching requirements of, 52 in RCA, 443–444
typical connection of, 31–32 RWA of, 509–516, 526–528
versatility of, 3–4 signal, 7
Optical Networks Technology Consortium termination of, 60
(ONTC), 838 unfeasible, 511
 Index 947

Optical path-based protection, 692–700, opaque, 13n7, 21

705 technology of, 242–255
one-plus-one dedicated backup, 692–693, Optical threshold function (OTF), 789–790
695 Optical time to live (OTTL) function, 813
redundant trees for APS in, 697 Optical transmission section (OTS) layer,
shared backup, 693–697 32–33
spanning trees in, 697 Optical transmitter (OT)
SRGs in, 697–700 bandwidth constraints of, 31
Optical photon, 198–199 directly modulated, 70
Optical processing, with CDMA, 352–353 as enabling technology, 205–217
in coherent systems, 363–366 externally modulated, 70
in noncoherent systems, 360–363 fabrication of, 68
reasons for, 359 lasers in, 68, 70–71, 205–211
Optical receiver (OR) in NASs, 70–71
arrayed, 72–73 optical connection originating in, 69
bandwidth constraints of, 31 Optimal grooming, 599–600
BPF in, 73–74 Optimization problem, static RCA as, 464–474
conflict-free tuning schedule for, 111 algorithms for, 465–468
digital signal detection in, 221–226 complexity of, 443
direct detection, 71–73, 233n26 flow conservation equation in, 466–467
as enabling technology, 217–227 full wavelength interchange and, 468–469,
front-end amplifiers in, 219–221 470–471, 493–494
heterodyne, 73–74, 234–235 MIP formulations used in, 464–465
in IM/DD systems, 217–227 no wavelength interchange and, 465–468
in NASs, 71–74 path interference graph in, 467–468
in optical filtering technology, 258–259 RS version of, 471–474
optical power converted by, 68 sparse wavelength interchange and,
photodetectors in, 71–74, 217–218 469–470
savings of, 535 Optoelectronic wavelength conversion, 275
tunability of, 73–74, 110–111, 258 OR. See Optical receiver
Optical resource usage, 612, 627–628 Orientable cycle double cover (CDC), 460,
Optical single-sideband (OSSB) subcarrier, 685–686
809 Orthogonal optical code (OOC), 357–358
Optical spectral efficiency OSI Reference Model. See Open Systems
in ideal cases, 335 Interconnection Reference Model
illustration of, 334 OSPF protocol. See Open Shortest Path First
in multistar networks, 464, 640 protocol
transmission constraints and, 333–335 OSPF-TE. See Open Shortest Path First
Optical splitters, 500, 503, 505–507 protocol with Traffic Engineering
Optical switch architecture extensions
Benes switches, 285–286 OSS. See Operations support system
as enabling technology, 281–297, 788–789 OSSB subcarrier. See Optical single-sideband
for KEOPS, 793–795 subcarrier
router/selector, 284–285 OT. See Optical transmitter
space switches, 281–284 OTF. See Optical threshold function
switch cross-talk and, 286–288 OTS. See Optical transmission section layer
Optical switches. See also Optical switch OTTL function. See Optical time to live
architecture function
granularity and, 13–14, 78 Out-of-band optical service channel, 671
IOS, 857 Output buffering, 772, 780
maturation of, 2 Overlapping signal spectra, 34–35
948 Index

Overlay model, for routing, 718, 751–752 encapsulation in, 114, 117
Overlay processors (OLPs) high logical connectivity and, 111,
bridges as, 461–462 121–122, 399
functionality of, 85–86 lossless scheduling in, 405–407
in NASs, 75, 77 MAC protocol for, 111–121, 399–408
regeneration of, 76 by NASs, 101, 116–117
RPs replaced by, 75 OPS v., 112
as shim between layers, 87 packet filters in, 107, 117
TPs and, 69, 75 perfect scheduling in, 407–408
transceiver interface with, 75 scheduling with loss in, 403–405
wavelength continuity through, 76 using SCMA, 347–348
wavelength-interchanging, 76–77 in static networks, 111–121, 399–409
Overmodulation, 345 TDM/T-WDMA case of, 117–119
OXCs. See Optical cross-connects trade-offs of, 400
uncontrolled scheduling in, 401–403
Packet loss Packet traffic
from contention, 763–764 connection management of, 101–102
in deflection routing, 766–770 fixed-frame scheduling for, 383–389
by hybrid system, 786 MPLS and, 722–727
in LDR, 768–770 NAS characterizing, 101
in OBS, 808 queues for, 385–386, 404–405
in SP routing, 767–768 retransmitted, 401, 405
in wavelength conversion, 767–770, stream traffic v., 101–102
780–781 unpredictability of, 384
Packets Pan-European Photonic Transport Overlay
delineation procedure for, 787 Network, 840
dropping of, 778 Partitioning. See Spectrum partitioning
filters for, 107, 117 Passband, 259–261
synchronization of, 788 Passive optical networks (PONs), 91n1
Packet scheduling. See also Fixed-frame APON, 412–414
scheduling BPON, 412–414
alternatives for, 776–778 channel reservations in, 411–412
output queuing v., 778 decentralized control in, 415–416
switches’ role in, 775 DWDM/TDM, 422–424
Packet switching end systems in, 410
circuit-switched operation v., 756 EPON, 414–416, 849
demand-assigned connections compared fixed frame, 412–414
with, 399 GPON, 414, 849
LCs and, 111, 121–122, 399 light sources in, 422
OBI and, 347–348 MAC protocol for, 412–416
static multipoint networks and, 111–121, recent trends in, 422–424
399–409 as static multipoint network, 409–425
Packet switching, in optical layer, 92–93, 102. two-fiber v. single fiber links in, 411
See also Optical packet switching WDM, 416–420, 422–424
antecedants of, 113 wireless access and, 420–422
in broadcast star networks, 112–113, 117 Path admissibility, in RCA, 450–452,
burstiness and, 111, 387, 389 485
using CDMA, 366–367 Path AIS. See Path alarm indication signal
channel reservations in, 406, 411–412 Path alarm indication signal (AIS), 656
dynamic capacity allocation and, 112, Path-based survivability schemes, 678,
399–400 695–697, 705
 Index 949

Path coalescence, 518–519, 527 enabling technology and, 235–241

inseparability causing, 559 PMD as, 241
interference caused by, 546–547 prediction of, 297–298, 302–308
PathErr message, 746 signal power divergence as, 239–240
Path-independent loss crossbar switch, 282 Permutation matrices, 877, 879
Path interference graph, 449–452 Permutation routing, 463
joint routing and wavelength minimization Permutation switches
solved in, 451 binary switching elements of, 48–52
minimal vertex coloring in, 450–451 blocking characteristics of, 50–52
in RCA optimization problem, 467–468 complexity of, 48–52
Path length crossbar, 48
constraints on, 467 LDC and, 54
effective, 489 rearrangeably nonblocking, 50–52, 66
Path messages, 745–748, 750 static routers as, 45
Path protection strict-sense nonblocking, 51, 59
dedicated, 695 wide-sense nonblocking, 51–52
in GMPLS, 736–737 Petersen graph, 434, 437, 440–442, 468,
ILPs in, 694–697 508–509, 766
line v., 651–652 Petersen network
optical, 692–700, 705 connection interference graphs for, 530–531
routing and, 736–737 deflection routing in, 765–770
shared, 693–697 demand-assigned LCs in, 539–540
Path restoration, 668 full connectivity in, 521, 538
Path selector, 658, 659, 661–662 LASs on, 519–521
Path state, 746, 748 many-to-many OCs and, 533–535
PathTear message, 746–747, 750 many-to-one OCs and, 533
PBG fiber. See Photonic band gap fiber MPS in, 522
PCFs. See Photonic crystal fibers multicast OPs in, 531
p-Cycles performance comparisons of, 535–538
failure recovery using, 688 point-to-point connections in, 525
ILP for, 689–690 point-to-point OPs in, 526–528, 531
link, 688 RWA on, 518–519, 521–528
node-encircling, 688 transmission channel assignment in,
shared optical protection by, 688–690 532–533
trade-offs of, 688–689 trees in, 521–528
PD. See Photodetector PHASAR. See Phased array
PDG. See Polarization-dependent gain Phase conjugation, 277–278
PDH. See Plesiochronous digital hierarchy Phased array (PHASAR), 265
PDL. See Polarization-dependent loss Phase mask, 363–365
Peer model, for routing, 718, 751–752 Phase shift keying (PSK), 67, 229–230
Penalties Photocurrent, 234
chirp-induced, 240 Photodetector (PD), 71–74, 217–218
distortion-induced, 240–241 Photodiode, 217–219
power, 238–239 APD, 218–220, 422
Q, 307–308 PIN, 218
Perfect scheduling, 407–408 Photonically integrated laser arrays, 211
Performance impairments, in network Photonic band gap (PBG) fiber, 189
environment Photonic cross-connect (PXC), 58
chirp-induced penalty as, 240 Photonic crystal fibers (PCFs), 21, 188–190
cross-talk as, 235–239 air-clad core, 189
distortion-induced penalty as, 240–241 effective-index, 188–189
950 Index

Photonic crystal fibers (PCFs) (cont.) symmetry of, 435–436

photonic band gap, 189 of transparent networks, 12, 30, 41
semiconductor technology combined with, of WANs, 438–439
190 of wavelength-routed networks, 434–442
Photonic device technology. See Optical and PIN photodiode, 218
photonic device technology pJET protocol. See Priority JET protocol
Physical Layer Operations Administration and Plain old telephone service (POTS), 6
Maintenance (PLOAM) cells, 412 Planarity-testing algorithm, 684
Physical layer, 6–8 Planar lightwave circuit (PLC), 243, 249
adjacency in, 731 Planar waveguide, 264–265
control techniques for, 21 PLC. See Planar lightwave circuit
design of, in hypernets, 634–637 Plesiochronous digital hierarchy (PDH), 893
design of, in multihop networks, 594–596, PLOAM cells. See Physical Layer Operation
642 Administration and Maintenance cells
logical layer interface with, 30 PMD. See Polarization mode dispersion
of multilayered networks, 7n4, 10–11, 11n6, PM layer. See Physical media layer
13, 21–22 Point-and-click provisioning, 714, 719
nonlinear functionality in, 75 Point process, 880
optical and fiber layers of, 13 Point-to-point connections
reconfigurable, 16, 79–80, 82 algorithms for, 545–549, 551–557
routing with constraints on, 548–549 in ATM networks, 82
signal processing in, 38 dynamic routing rules for, 544–558
in static networks, 102 in LRNs, 578–579, 585–591
topology of, 22, 30, 41, 79, 85–86, 93 multicast connections compared with, 106
transparency of, 8, 10–11, 11n6 NASs in, 69
Physical-layer simulation OCs as, 29–32, 69
of DWDM network, 308–311 in Petersen networks, 525
three-step approach to, 298–301 in SDH, 80
of WADM chain, 301–302 in SONET, 80
of WDM network, 301–311 in static networks, 103–111
Physical media (PM) layer, 32 terminology of, 29–32
Physical topology (PT), 12, 22, 30, 85–86. See throughput from, 531–532, 586–589
also Ring topology; Tree topology in wavelength-routed networks, 122–123
control over, 144–145 wavelength routing and, 145
density of, 434, 445 Point-to-point network
folded bus, 41, 155 hypernets reduced to, 612
growth of, 440, 478–479 LRNs as, 153–156, 578–579, 585–591, 641
of hypernets, 158 protection in, 650–654, 669–777
logical topology embedded in, 91, 594–596, restoration in, 650–654
634–637, 642 static, 103–111
logical topology independent of, 79, 94 WDM, cost issues for, 831–832
logical topology matched with, in static Point-to-point optical path (OP), 526–528, 531
RCA, 444–448 Point-to-point topology, 145, 153–156
of LRNs, 436–437, 580–581, 585 grooming in, 597–607
multistar, 144–145, 150, 462–464, multihop networks and, 585–591
540–544, 640–641 Poisson process, 385, 392, 881, 884
RCA and, 444 Polarization
ring, 150, 452–458 flipping of, 267–268
spectrum reuse and, 41, 642 PDG and, 203, 205
star, 41, 91–93 PDL and, 180, 239–240
survivable, 649 PMD and, 179–180, 241
 Index 951

principal states of, 241 Pulse broadening, 178–180

states of, 174 PXC. See Photonic cross-connect
transverse, 267
Polarization-dependent gain (PDG), 203, 205 QDS matrix. See Quasidoubly stochastic
Polarization-dependent loss (PDL), 180, matrix
239–240 QD SOAs. See Quantum dot SOAs
Polarization-maintaining fibers, 174 QoS requirements. See Quality-of-service
Polarization mode dispersion (PMD), requirements
179–180, 241 Q penalty, 307–308
PON. See Passive optical networks Quality-of-service (QoS) requirements, 716
Population inversion, in EDFA, 193 enforcement of, 759
POTS. See Plain old telephone service MPLS applications for, 728–729
Power amplifiers, 190 Quantum dot (QD) SOAs, 203
Power penalty, 238–239 Quantum efficiency, 218
Power transfer relations Quantum limit, 220
of directional couplers, 42–44 Quasidoubly stochastic (QDS) matrix,
of generalized switches, 53 875–879
of LDCs, 55 Queues
of multiwaveband switches, 57 DQDB and, 458
static routers and, 45 in fixed-frame scheduling, 384–388
of switch fabrics, 50 Little’s formula for, 885
of waveband-selective switches, 57 Markov chains and, 884–885
Preamplifiers, 190 M|D|1, 884–885
Prechirp techniques, 183 M|G|1, 884–885
Precomputed backup route computation, 678 M|M|1, 884
Principal states of polarization (PSP), 241 model of, for hybrid electric and optical
Priority JET (pJET) protocol, 803 buffering, 786
Protection. See also Automatic protection packet scheduling v., 778
switching; Path protection; Shared for packet traffic, 385–386, 404–405
optical protection for perfect scheduling, 407–408
current techniques of, in logical layer, terminology of, 883–885
650–669
cycles, 681–683 RA. See Raman amplifier
electrical, 670 RACE-MWTN consortium, 839
equipment, 650 Radiative electron-hole recombination, 202
line-based, 651–652, 657, 674, 679–692 Radio-over-fiber (ROF), 420–422
link failure, 681–683, 685–687, 690–692, RAM. See Random access memory
750–751 Raman amplifier (RA)
objectives of, 648–649 distributed, 199–200
in point-to-point networks, 650–654, drawbacks of, 200–201
669–777 EDFAs compared with, 198–199, 201
restoration v., 648–649, 704 gain coefficient of, 198–199
segment, 679, 700–702, 705 hybrid distributed-discrete amplification
signaling for, in optical networks, 750–751 with, 199–200
in SONET, 650–669 noise in, 200–201
Protection information object, 749 renewed interest in, 191
Protocol stack, of MPLS, 727–728 Raman gain coefficient (RGC), 198–199
PSK. See Phase shift keying Random access. See Uncontrolled scheduling
PSP. See Principal states of polarization Random access memory (RAM), 758–759,
PT. See Physical topology 821
Public Switched Telephone Network, 647 Randomly connected network (RCN), 477
952 Index

Random processes, 880–883 Reed-Solomon (RS) codes, 232

Rare-earth-doped fiber amplifiers, 191 Reflecting broadcast star network, 523
Rayleigh scattering, 175–176, 200–201 Reflective semiconductor optical amplifier
Rays (R-SOA), 423
geometric optics and, 168–171 Refractive index profiles, 168–169
in graded-index fiber, 169–171 Regeneration. See Signal regeneration
largest possible angle of incidence for, 170 Regenerative repeater, 76
propagation of, 170–171 Relative intensity noise (RIN), 227
refractive index profiles and, 168–169 Representative multiplexing and multiple
Snell’s law and, 169–170 access schemes
in step-index fiber, 168–170 CDMA, 352–367
RCA. See Routing and channel assignment SCM/SCMA, 336–352
RCN. See Randomly connected network in shared-medium networks, 327–328
Real-time backup route computation, 678 TDM/TDMA, 106–109, 116–117, 119, 134,
Rearrangeably nonblocking permutation 147, 161, 328–335
switches, 50–52, 66 Reservation. See Channel reservation
Receiver. See Optical receiver Resilient packet ring (RPR), 849, 852
Receiver sensitivity, 226 Resource discovery, 721
Reception processor (RP), 330–331 Resource management, as optical control
demultiplexing in, 70 plane function, 721
functionality of, 85–86 Resource Reservation Protocol (RSVP). See
OLP replacing, 75 also Resource Reservation Protocol with
in overlay processor, 69 Traffic Engineering Extensions
transmission signal converted in, 31, 68 basic entities in, 745–746
Reciprocity, 255–257 main objects of, 747
Reconfigurable optical add/drop multiplexer message types in, 746–748, 750
(ROADM). See Wavelength add/drop signaling and, 745–749
multiplexer Resource Reservation Protocol with Traffic
Reconfiguration Engineering Extensions (RSVP-TE),
hitless, 596 726, 730
of layers, 16, 79–80, 82, 94 with enhancements for GMPLS support,
in response to traffic, 596 749–750
in wavelength-routed networks, 432 objects of, 748–750
Record Route object, 748 signaling and, 747–750
Recovery Restoration
from failure, 580, 647, 649, 651–652, centralized, 665–666
688 current techniques of, in logical layer,
in LRNs, 580 650–669
multicarrier, 860–861 DCS-based, 665–669
speed of, 649 distributed, 666–669
using p-cycle approach, 688 dynamic, 705
Recursive design, 440–442 by FASTAR system, 665–666
Recursive grid, 440, 442 objectives of, 648–649
Redundancy path, 668
in CDCs, 680 in point-to-point networks, 650–654
CRC for, 114 protection v., 648–649, 704
in generalized loopbacks, 691 RCA and, 127
in ring covers, 664, 669 sender-chooser method of, 667–668
survivability influenced by, 649–650 signaling for, in optical networks, 750–751
trees and, 697 in SONET, 650–670
Redundant trees, 697 in survivability, 647–705
 Index 953

ResvConf message, 746 edge, 723–725, 816

ResvErr message, 746 IP, 4, 83–84
Resv message, 745–746, 748, 750 Latin, 45, 265
ResvTear message, 746–747 LER as, 723–725
Return-to-zero (RZ) keying, 229–231 LSR as, 85, 723–724, 727
Reuse. See Spectrum reuse in OPSnet, 816–817
Reverse bias voltage, 216–217 static, 123–124, 442–484
Revertive mode, of one-plus-one facility terminology of, 734
protection, 653–655 Router/selector switch, 284–285
RGC. See Raman gain coefficient Routing. See also Deflection routing; Routers;
RIN. See Relative intensity noise Routing and channel assignment;
Ring cover Routing and waveband assignment
algorithms for, 664–665 address shift, 591, 623–625
in mesh architectures, 664–665, 669 AWG properties of, 265
redundancy in, 664, 669 channel assignment compared with, 124
trade-offs of, 663 characterization of, 485–486
Ring topology. See also Bidirectional ring; constraints on, 135–140, 509, 548–549
Cycle double cover; Interconnected rings; deflection, 807–808
Ring cover; Self-healing ring; distributed, 737
Shared-protection ring; Unidirectional evolution of, 19
path-protected ring; Unidirectional fortuitous destinations in, 135–136, 513,
path-switched ring; Wavelength 515
division-multiplexed ring garbling in, 137
advantages of, 458 within given waveband, 546–550, 558–560
arbitrary connectivity on, 457–458 in GMPLS, 730, 734–742
collapsed, 703–704 header, 624
decomposition of mesh networks and, hybrid model for, 718–719, 751–752
458–462 joint, 451
in MANs, 844–845, 850–851 limited deflection, 768–770
as Moore graph, 452 link properties and, 737
RCA and, 452–462, 599–600 in LLNs, 135–140, 507–568, 890–892
RPR, 849, 852 of multicast connections, 497–507, 558–567
SONET, 452–453, 598–602, 654–663, 672 in multihop network design, 591–593
SONET/SDH, 842–843 multiple waveband, 555–557
with sparse wavelength interchange, 494, in OBS, 801
496 OC table for, 517
switching and, 674 operational considerations in, 736–737
wavelength assignment and, 456–457 as optical control plane function, 721
wavelength requirements for full overlay model for, 718, 751–752
connectivity and, 456 path protection and physically diverse paths
RIP. See Routing Information Protocol in, 736–737
Ripple effect, in interconnected rings, 663 peer model for, 718, 751–752
RITE-Net, 418 permutation, 463
ROF. See Radio-over-fiber protocols for, 734–736
Rooted routing, 534 resource discovery and path computation
LDCs in, 564–565 function of, 721
on tree, 564–566 rooted, 534, 564–566
Root node, 140, 145 rules of, 566, 591
Routers. See also Static routers SP, 767–768
core, 817–818 tree, 561–566, 634–636
cross-talk of, 286–288 waveband, 553–555
954 Index

Routing and channel assignment (RCA). See RS codes. See Reed-Solomon codes
also Dynamic routing and channel RS formulation, for optimization problem,
assignment; Static routing and channel 471–474
assignment R-SOA. See Reflective semiconductor optical
algorithms for, 51–52, 127, 139, 455–468, amplifier
474–591, 623–625, 637 RSVP. See Resource Reservation Protocol
in bidirectional ring, 129–132 RSVP-TE. See Resource Reservation
channel assignment constraints on, Protocol with Traffic Engineering
125–128, 131, 138–139 Extensions
as control problem, 127 RWA. See Routing and waveband assignment
decoupling of, 452 RZ. See Return-to-zero keying
dedicated connections in, 124–125, 127
examples of, 128–133 SA. See Saturable absorber; Source address
fiber topology/spectrum trade-off in, 133 SaD switch. See Split-and-deliver switch
implementation of, 124 SAN. See Storage area network
in LLNs, 890–892 Saturable absorber (SA), 279
in mesh network, 131–133, 494–495 Saturation degree, 872
Monte Carlo approach to, 479–484 Saturation, gain, 194–195
nonblocking stations in, 128–129, 131–132, SBS. See Stimulated Brillouin scattering
448–449 Scalability, 122
OPs created in, 443–444 of MEMS, 250–251
path admissibility in, 450–452, 485 of OLS, 808
path interference graph in, 467–468 for optical control plane, 716n2
physical topology and, 444 spectrum reuse and, 31n3, 122, 162
rearrangeability in, 124–126 Scattered wavelength path, 798
restoration and, 127 SCH. See Set covering heuristic
ring topology and, 452–458, 599–600 Scheduling. See also Channel allocation
routing constraints on, 135–140 schedule; Fixed-frame scheduling; Fixed
shortest path routing in, 131 scheduling algorithm; Packet scheduling
switched connections in, 124–125, 127 efficiency of, 332–333
traffic requirements and, 443 heterogeneous traffic, 379
in wavelength-routed networks, 124–133, with loss, 403–405
433, 442–544 lossless, 405–407
in WDM rings, 599–600 perfect, 407–408
Routing and waveband assignment (RWA), transmission, 785
508 uncontrolled, 401–403
algorithms for, 544–568 void-filling, 802–803
blocking probability measuring Schrodinger equation, nonlinear, 187
performance of, 545 SCM. See Subcarrier multiplexing
channel allocation in, 550–551 SCMA. See Subcarrier multiple access
channel assignment in, 516–518 SCM/SCMA. See Subcarrier
in hypernets, 635–637 multiplexing/subcarrier multiple access
with multicast OPs, 522–525 SCM/WDMA/SCMA. See Subcarrier
ONMs in, 520–521, 545 multiplexing/wavelength division
of OPs, 509–516, 526–528 multiple access/subcarrier multiple
on Petersen networks, 518–519, 521–528 access
in waveband-routed networks, 442–484 SCS. See Single-crystal silicon
waveband selection in, 545–546 SDDP. See Segment, Deflect, and Drop Policy
Routing Information Protocol (RIP), 735 SDH. See Synchronous digital hierarchy
RP. See Reception processor SDP. See Segment and Drop Policy
RPR. See Resilient packet ring Segment and Drop Policy (SDP), 807
 Index 955

Segment, Deflect, and Drop Policy (SDDP), Shared medium networks, 103, 117, 324–327
807 Shared optical protection
Segment protection, 679, 700–702, 705 advantages of, 679–680
SelfHealing Network, 666–667 by CDCs, 680–687
Self-healing ring (SHR). See also by generalized loopbacks, 690–692
Shared-protection ring in mesh architectures, 679–692
BLSR, 655, 656–657, 659–663 path v., 651–652
interconnection techniques of, 657–663 by p-cycles, 688–690
optical-layer protection in, 672–677 switching, 657, 674
SONET, 654–663, 672 Shared path-based protection, 693–697
SPRING, 673–677 Shared-protection ring (SPRING)
survivability and, 649 four-fiber, 453–458, 673–675
UPPR, 672–673 trade-offs of, 677
UPSR, 655–656, 659–663 two-fiber, 454–455
in WDM, 672–677 WDM, 453–458, 673–675
Self-phase modulation, 185 Shared risk groups (SRGs)
Self-saturation, of amplifiers, 195 classification of, 698–699
Semiconductor laser amplifier in a loop ILPs for, 700
optical mirror (SLALOM), 280, 789 in optical path-based protection, 697–700
Semiconductor laser diode, 18, 205–207 Shared risk link group (SRLG), 736–737,
Semiconductor optical amplifier (SOA), 740
201–203, 279 Shared wavelength converters, 781–782
development of, 190 Shortest path algorithm. See k shortest path
drawbacks of, 202–203 algorithm
electron-hole recombination in, 201–202 Shortest path tree (SPT) heuristic, 501–502
PDG impacting, 205 advantages of, 561
quantum dot, 203 color clashes caused by, 562–564
reflective, 423 Shortest path with deletions (SPD) algorithm,
structure of, 201 487–488
using MZIs, 788, 814–815 Short leap shared protection (SLSP), 700–701
wavelength conversion based on, 790 Shot noise, 196, 220, 227, 235
Semiconductor technology. See also SHR. See Self-healing ring (SHR)
Semiconductor optical amplifier ShuffleNet, 154, 586–589
controllable directional couplers and, bidirectional ring, 155–156
245–246 deflection probability in, 770
laser diode as, 18, 205–207 directed Kautz hypernets compared with,
PCFs combined with, 190 630–631
SLALOM as, 280, 789 disadvantages of, 588, 630
Sender-chooser restoration method, 667–668 edge grouping and, 614–615
Sender node, 667–668 embedding of, 155, 594–595
Service level agreements (SLAs) end systems communication in, 588
with ISPs, 5–6 folded bus, 155
requirements of, 648 internodal distances in, 439–440
traffic engineering and, 6 Kautz digraphs compared with, 630–631
Service rate, 385 LCG of, 593
Service selector, 658, 659, 662 LSNs in, 157
Session Attribute object, 748 maximum throughput per node for, 588–589
Set covering heuristic (SCH), 502 as multihop network, 587–589
SFP module. See Small form factor pluggable undirected Kautz hypergraph compared
module with, 628–630
Shared backup path protection, 693–697 wavelength assignments for, 156
956 Index

Signaling Solitons, 187–188

in GMPLS, 730, 742–751 SONET. See Synchronous optical network
in JIT protocol, 804–806 SONET/SDH standard, 841–843, 893
for label distribution, 745 Source address (SA), 114
LSRs exchanging information through, 727 Source-destination light-tree, 498
NNI and, 742–745 Source grooming, 597
as optical control plane function, 721, 727 Space dilation, 51, 288
for protection and restoration in optical Space division multiplexing, 34
networks, 750–751 Space switches, 281–285
RSVP and, 745–749 Spanning subgraph, 870
RSVP-TE and, 747–750 Spanning trees, 560–561, 697, 871
UNI and, 742–745 Span switching, 657
Signal power divergence, 239–240 Sparse multicasting, 505–507
Signal regeneration Sparse wavelength interchange, 469–470
all-optical 2R or 3R, 788 in mesh networks, 460–461
BER in, 274–275 rings with, 494, 496
devices for, 279–281, 789 Spatial light modulator (SLM), 254–255,
as enabling technology, 274–281, 788 269
islands of transparency and, 856 SPD algorithm. See Shortest path with
levels of, 274–275 deletions algorithm
opaque wavelength conversion and, Spectral inversion, midsystem, 182
278–281 Spectra, subcarrier, 535
Signal-to-noise ratio (SNR), 167, 219–221 Spectrum partitioning, 34–39
Silicon-on-insulator (SOI) wafer, 252 in LLNs, 557–558
Single-access ring interconnection one-tiered approach to, 37
configurations, 657–663 optical spectrum usage increased by, 557
Single-crystal silicon (SCS), 252–253 two-tiered approach to, 37–38
Single fiber pair access, 129–130, 456–457 waveband, 35, 38–39
Single-frequency lasers, 181, 207–211 wavelength, 34–35, 38
Single longitudinal mode operation, 207 Spectrum reuse
SLALOM. See Semiconductor laser amplifier in bidirectional ring, 93, 150–151
in a loop optical mirror DCA preventing, 103
SLM. See Spatial light modulator as key to scalability, 31n3, 122, 162
SLOB. See Switch with large optical buffers in mesh networks, 133, 150–151
Slot-by-slot channel reservations, 406 minimization of, 140
Slotted networks multistar, 150–151
ALOHA, 402–403 physical topology and, 41, 642
dump-and-insert buffering in, 775 scalability and, 122, 162
packet synchronization in, 788 static routers and, 45, 123–124
slot synchronization in, 793 throughput increased by, 538
unbuffered, 761–763 wavelength selectivity influencing, 123–124
Slow light, 790–791 Speed-up factor, 627
SLSP. See Short leap shared protection Split-and-deliver (SaD) switch, 498, 501–502
Small form factor pluggable (SFP) module, Splitter placement in wavelength-routed
850 network (SP-WRN), 506
Smectic crystal, 254 Splitters. See Optical splitters
Snell’s law, 169–170 Splitting losses, 504–505
SNR. See Signal-to-noise ratio Spread-spectrum technique, 352
SOA. See Semiconductor optical amplifier SPRING. See Shared-protection ring
Soft permanent connection, 744 SP routing, 767–769
SOI wafer. See Silicon-on-insulator wafer SPT. See Shortest path tree heuristic
 Index 957

SP-WRN. See Splitter placement in ONN functionality in, 133

wavelength-routed network packet switching in, 111–121, 399–409
Squelching, 656–657 physical layer in, 102
SRGs. See Shared risk groups point-to-point, 103–111
SRLG. See Shared risk link group star, 102–103
SRS. See Stimulated Raman scattering wavelength assignment rule in, 44–46
SSFBG. See Superstructure fiber Bragg Static nodes, 40–45
grating directional couplers building, 42–44
Star couplers, 40–41, 140 in N-node static network, 432
Banyan structure of, 43, 50 static routers and, 44–46
connectivity layers of, 104–105 Static RCA. See Static routing and channel
dynamic directional, 43–44 assignment
LDC as, 54 Static routers
nonreflecting v. reflecting, 141 in directed star networks, 45, 103n7
static router replacing, 44 as permutation switches, 45
Stark effect, 194 power transfer relations and, 45
Star networks. See also Broadcast star spectrum reuse and, 45, 123–124
network; Multistar network star coupler replaced by, 44
broadcast, 44, 103–106, 109–110, 112–113, static nodes and, 44–46
117, 122, 140–142, 145–147, 158, 160 wavelength-routed networks and, 442–484
directed, 45, 103n7 Static routing and channel assignment (RCA)
physical topology of, 41, 91–93 algorithms for, 465–468, 474–486
static, 102–103 dynamic RCA v., 557–558
undirected, 103n7 as graph coloring problem, 449–452
wavelength-routed, 123–124, 149 grooming in, 600–602
State processes, 880 heuristics for, 474–486
State transition, 880–882 LLN rules for, 507–544
Static directional couplers, 42–44 logical and physical topologies matched in,
Static filter, 267 444–448
Static multipoint networks, 103–111 in multistar WRNs, 462–464
broadcast star, 103–106, 324–327 nonblocking stations in, 448–449
capacity allocation for dedicated as optimization problem, 443, 464–474,
connections and, 371–389 493–494
constraints on, 367–371, 424 rings and, 452–462
demand-assigned connections and, 389–399 in waveband-routed networks, 507–544
NAS role in, 324 in wavelength-routed networks, 442–484,
packet switching in optical layer and, 507–544
111–121, 399–409 Static shared medium network, 103
PON, 409–425 Static single-waveband networks. See Static
randomness in, 424 networks
representative multiplexing and multiple Steiner forest, 635
access schemes in, 327–367 Steiner tree, 501–502, 635
Static networks, 162–163. See also Static Step-index fiber
multipoint networks intermodal dispersion in, 177–178
connection management in, 102 rays in, 168–170
embedding problem in, 594 Stimulated Brillouin scattering (SBS), 184
LAMBDANET as, 20 Stimulated Raman scattering (SRS), 183–184,
limitations of, 95, 122 186–187, 198
NAS in, 102–103, 324 STOLAS. See Switching Technologies for
NCs in, 94–95, 102–122 Optimally Labeled Signals
nodes in, 40–46, 432 Storage area network (SAN), 858
958 Index

Stream traffic Survivability. See also Protection

connection management of, 101–102 FITNESS protocol for, 667
fixed-frame scheduling for, 371–383 line-based, 651–652, 657, 674, 678–692,
packet traffic v., 101–102 705
in SONET, 101 of logical topology, 649
Strict-sense nonblocking switches of multicast connections, 702–703
conversion and, 59 path-based, 678, 695–697, 705
cross-points in, 51 of physical topology, 649
Stripe geometry, 207 redundancy influencing, 649–650
STS. See Synchronous transport signal rerouting around fault conditions ensuring,
Subcarrier multiplexing/subcarrier multiple 649
access (SCM/SCMA), 348–349 restoration in, 647–705
Subcarrier multiple access (SCMA), ring covers for, 664–665
345–352 SHRs and, 649
modulation in, 71, 337–339 taxonomy of schemes for, 677–678
packet switching using, 347–348 topology influencing, 649
principles of, 336–339 Switch controller
SCM/SCMA, 348–349 transmission scheduling by, 785
SCM/WDMA/SCMA, 349–352 in unbuffered networks, 763
trade-offs of, 339, 352 Switch delay line architecture. See
transmitting and receiving stations in, Feed-forward line delay architecture
337–339 Switched connections
Subcarrier multiplexing (SCM), 348–352, dedicated connections v., 124
811–812 development of, 166
Subcarrier multiplexing/wavelength division in RCA, 124–125, 127
multiple access/subcarrier multiple rearrangement of, 125
access (SCM/WDMA/SCMA), 349–352 terminology of, 744
Subcarriers Switches. See also Guided-wave switches;
advantages of, 336 Multiwavelength switch; Permutation
channels of, 346–347 switches
in dispersion, 181 AWG-based, 271–273
frequency of, 336 Benes, 49–51, 66, 285–288
microwave, 227 comparisons of, 243
for multiplexing v. multiple access, 349 crossbar, 48, 52, 55, 281–284
in OLS, 811–812 cross-talk in, 52, 62, 247, 286–288
OSSB, 809 early examples of, 242–243
packet switching using, 347–348 generalized, 52–53
receiver array, 535 hybrid, 296–297
SCMA, 71, 336–339, 345–352 IOS, 857
SCM/SCMA, 348–349 liquid crystal architectures of, 254–255,
SCM/WDMA/SCMA, 349–352 269–270
spectra of, 339–345, 535 MC-LGS, 637–639
TDM/T-SCMA, 345–347 MEMS, 20, 242, 250–253, 270–271
tunability of, 351–352 MI filter-based, 271–272
Subgraphs, 869–870 O-E-O, 59, 78, 275, 294–295, 582–583, 585
Subrate tributaries, 583 O-O-O, 58–59, 78, 294–296, 583, 585, 733
Suggested label object, 749 optical, 2, 13–14, 13n7, 21, 78, 242–255,
Supercontinuum generation, 208–209 857
Superhub, 845–846, 850–851 as optical and photonic device technology,
Superstructure fiber Bragg grating (SSFBG), 242–255
365–366 in packet scheduling, 775
 Index 959

router/selector, 284–285 SHRs in, 654–663, 672

SaD, 498, 501–502 in SONET/SDH standard, 841–843, 893
SLOB, 795–796 stream traffic in, 101
space, 281–285 STS as building block of, 893–896
with tunable wavelength converters, terminology of, 651
779–781 over WDM rings, 598–602
WASPNET optical packet, 796–797 WDM terminals in, 80–81
wavelength-dilated, 273–274 Synchronous transport signal (STS), 79
wavelength-selective, 288–297 concatenated, 894–896
Switch fabrics in LRNs, 580
Benes, 49–51, 66, 285–286 as SONET building block, 893–896
cost reduced by, 55
cross-talk reduced by, 52, 62 TCP. See Transmission Control Protocol
large-scale free-space integrated, 250–253 TCP/IP suite
multistage, 48–50 network interface layer of, 83
of OXCs, 59, 243 OSI compared with, 83–84
power transfer relations of, 50 UDP and TCP in, 727
transceivers and, 59 TDM. See Time division multiplexing
Y-branch, 248, 284 TDMA. See Time division multiple access
Switching Technologies for Optimally TDM/TDMA. See Time division
Labeled Signals (STOLAS), 813–815 multiplexing/time division multiple
Switch with large optical buffers (SLOB), access
795–796 TDM/T-SCMA. See Time division
Synchronization multiplexing/time-subcarrier multiple
circuit, 762 access
packet, 788 TDM/T-WDMA. See Time division
slot, 793 multiplexing/time wavelength-division
timing, 108–109 multiple access
Synchronous digital hierarchy (SDH), 4 TE. See Traffic engineering
history of, 19 Telecommunications Management Network
point-to-point connections in, 80 (TMN), 839
in SONET/SDH standard, 841–843, 893 TeleManagement Forum (TMF), 859
Synchronous optical network (SONET), 4, Telephone networks
679–680 circuit-switched mode of, 4
ADM in, 79–81, 598–599 history of, 18–19
in ATM-over-SONET structure, 82–83 POTS provided by, 6
concatenated frame of, 894–896 Public Switched Telephone Network as, 647
cost issues of, 833–835 Tell-and-go protocol, 403–405
DCS in, 79–81, 582–583 Ten-gigabit Ethernet, 847–848
granularity in, 81 TE polarization. See Transverse electric
grooming in, 896n1 polarization
history of, 19 Terminal connection point, 33
linear APS in, 650, 653 Testbeds, of optical networks, 21, 39, 787
LP v., 651 AON Consortium, 839
as LRN, 79–81 contention resolution and, 791–798
in multilayered networks, 6–8 CORD, 791–793
paths in, 651, 659, 893 early OLS, 811–813
point-to-point connections in, 80 European, 839–840
protection in, 650–669 IP/WDM optical label-switching, 812
restoration in, 650–670 KEOPS, 793–796
ring, 452–453, 598–602, 654–663, 672 MONET, 840–841
960 Index

Testbeds, of optical networks (cont.) techniques for, 106–109, 122n16

for multihop packet transmission and time scales of, 106–107
multirate payload, 819–820 transmission channel assignment and,
in multiwavelength optical networking, 828, 532
838–841, 861 transmission constraints in, 333–335
National Transparent Optical Network Time division multiplexing/time division
Consortium, 840 multiple access (TDM/TDMA), 106–109,
NGI, 812 116–117, 119, 134, 147, 161, 328–335
OLS, 811–820 Time division multiplexing/time-subcarrier
ONTC, 838 multiple access (TDM/T-SCMA),
OPERA, 811–813 345–347
OPS, 787, 791–798, 818–821 Time division multiplexing/time-wavelength
OPSnet, 815–818 division multiple access
STOLAS, 813–815 (TDM/T-WDMA), 109–111, 116–119,
telecommunications infrastructure driven 135, 147, 161, 328–335, 395–399
by, 840–841 Time division techniques. See Time division
WASPNET, 796–798 multiple access; Time division
T-F-CDMA. See Two-dimensional multiplexing
time-frequency code division multiple Time domain contention resolution, 770–778
access Time processes, 880
Threefold multicast connections, 531 Time slot continuity, 61
Three-wave mixing, 64 Time slot interchange (TSI), 61, 656
Threshold current, 207 Time-subcarrier multiple access (T-SCMA),
Time division multiple access (TDMA). See 345–347
also Time division multiplexing/time Time-wavelength division multiple access
division multiple access (T-WDMA). See also Time division
frame schedule for, 108 multiplexing/time-wavelength division
guard times for, 108 multiple access
multiple access schemes and, 328–335 blocking calculations and, 395–399
ranging and, 108–109 for broadcast star network, 109–110
TDM/TDMA example of, 106–109, CAS for, 328–332
116–117, 119, 134, 147, 161, 328–335 downside of, 336
techniques for, 106–109, 122n16 frame schedule for, 110
time scales of, 106–107 scheduling efficiency in, 332–333
Time division multiplexing (TDM), 36, TDM/T-WDMA example of, 109–111,
836–837 116–119, 135, 147, 161, 328–335,
CAS for, 328–332 395–399
circuit-switched cases of, 106–107 transmission constraints in, 333–335
downside of, 336 Time/wavelength domain contention
DWDM/TDM PON and, 422–424 resolution, 778–782
frame schedule for, 108, 110 TM. See Transmission matrix
guard times for, 108 TMF. See TeleManagement Forum
multipoint cases of, 106–107 TMN. See Telecommunications Management
propagation delay in, 535 Network
ranging and, 108–109 TM polarization. See Transverse magnetic
scheduling efficiency in, 332–333 polarization
TDM/TDMA example of, 106–109, Topology. See Logical topology; Physical
116–117, 119, 134, 147, 161, 328–335 topology
TDM/T-SCMA, 345–347 Total internal reflection, 169–170
TDM/T-WDMA, 109–111, 116–119, 135, TOWC module. See Tunable optical
147, 161, 328–335, 395–399 wavelength converter module
 Index 961

TP. See Transmission processor Translucent networks, 77, 296

Traffic. See also Fixed-frame scheduling; Transmission bands, 35
Packet switching, in optical layer; Packet Transmission channel, 107. See also
traffic; Traffic constraints; Traffic End-to-end transmission channel
engineering channel assignment and, 532–535
add/drop, 78–79 as dedicated pipe, 101
balance factor of, 369–370, 388, 611–612 LCs and, 31
carried, 387, 389, 391, 471–474, 539–540 nonlinear functionality inserted at, 75
demand-assigned, 538–540 in static routing rules, 507
in fiber-optic transmission systems, 647 Transmission constraints, 333–335
fixed-frame scheduling for, 371–389, 593 Transmission Control Protocol (TCP), 727
heuristic algorithms for, 600–602, 604–605 Transmission impairments, of optical fibers
intensity of, 385, 387, 389, 393–395 attenuation, 175–177
loss of, in UPPRs, 673 dispersion, 177–183
Markovian models for, 391 fiber geometry and fiber cables, 174–175
matrix scaling of, 379–380, 593 nonlinear effects, 183–187
in multihop network design, 591–593 in WANs, 75
networks reconfigurable in response to, Transmission matrix (TM), 875
596 Transmission processor (TP)
offered, 387, 389, 391 functionality of, 85–86
OSPF for, 737–738 logical signal converted in, 31, 67–68
RCA and, 443 OLPs and, 69, 75
stream, 101–102, 371–383 in overlay processor, 69
in switches with tunable wavelength Transmission scheduling, 785
converters, 780 Transmission signal, 31, 68
unbalanced, 370–371 Transmitter edge, 606
Traffic constraints Transparent networks, 87
balance factor of, 369 advantages and disadvantages of, 10–11,
in fluid-flow traffic, 367–371 59
synchronous v. asynchronous, 367–370 definition of, 10–11, 11n6
tunability and, 368 fault isolation in, 733
Traffic engineering (TE) hardware of, 58
ISIS-TE for, 730 with logically routed overlay, 94
link, 739 MAN, 844–845
MPLS applications for, 729 multicast connections in, 500
OSPF-TE for, 730, 738–742 National Transparent Optical Network
RSVP-TE for, 726, 730, 747–750 Consortium and, 840
SLAs and, 6 ONNs in, 12
Transceivers opaque networks v., 296
in broadcast star networks, 112 origins of, 20
in CDMA, 358–359 OSPF in, 741–742
functionality of, 85–86 OXC, 58, 293–295
LC port interface of, 67 physical topology of, 12, 30, 41
NAS functions implemented by, 67, waveband basis of, 60
77 wavelength-routed, 122
OLP interface with, 75 WDM, 829, 852–853
switch fabrics and, 59 Transparent wavelength conversion,
tunability of, 109–111 275–278
types of, 109 Transponder, 63, 76
Transient chirp, 213 Transport networks, 91–93, 715
Transimpedance amplifier, 219 Transport plane, 714
962 Index

Transversal filter equalizer, 233 in wavelength conversion, 276, 779–782,

Transverse electric (TE) polarization, 267 793–794
Transverse magnetic (TM) polarization, 267 of wavelength converters, 793–794
Tree decomposition Tunable optical wavelength converter (TOWC)
illustration of, 564–565 module, 779–782, 793–794
LDCs and, 566 Tunable transmitters with fixed receivers
using MBFS-d, 562–563 (TT-FR), 109, 118–119, 331, 346,
objectives of, 561 394–395
procedure for, 560–561 Tunable transmitters with tunable receivers
Tree topology, 41, 85–86, 95, 100, 103. See (TT-TR), 109, 118–119, 331, 346,
also Light-tree; Tree decomposition 394–395
for APS, 697 Tunable wavelength converters, 779–782,
broadcast star emulated by, 145–146 793–794
channel allocation on, 566–567 Tunnel Identification, 748
choosing of, for connection, 566 T-WDMA. See Time division
coloring of, 522 multiplexing/time wavelength-division
edge-disjoint, 560–561 multiple access
embedded in mesh network, 140–142, Twisted Benes switch fabric, 66
147–149 Twisted nematic crystal, 254
fiber-disjoint, 634–636 Two-dimensional time-frequency code
in hypernets, 158, 634–636 division multiple access (T-F-CDMA),
in LLNs, 513–515 353
MBFS-d generating, 567 Two-fiber bidirectional line-switched ring
multicast, 100, 559–560, 631–632 (BLSR), 656–657, 676
in optical path-based protection, 697 Two-fiber shared-protection ring (SPRING),
in Petersen networks, 521–528 454–455
realization of, 145–149 Two-fiber unidirectional path-switched ring
routing and, 561–566, 634–636 (UPSR), 660
spanning, 560–561, 697, 871 Two-fiber UPSR. See Two-fiber unidirectional
Steiner, 501–502, 635 path-switched ring
waveband disjoint, 522 Two-tier networks
TSI. See Time slot interchange advantages of, 583
TT-FR. See Tunable transmitters with fixed grooming in, 582–583
receivers protection model in, 584
TT-TR. See Tunable transmitters with tunable
receivers UDP. See User Datagram Protocol
Tunability Ultralong-haul networks
of FBGs, 264 commercial network deployments and,
in fixed-frame scheduling, 373–376 856–858
of FP filters, 261 mesh architectures and, 856
of lasers, 210–212 WADM architectures and, 854–856, 862
of MI filters, 262 Unamplified systems, 185–186
of ORs, 73–74, 110–111, 258 Unbalanced traffic, 370–371
of subcarriers, 351–352 Unbuffered networks
traffic constraints and, 368 general packet format for, 762–763
of transceivers, 109–111 nodes in, 760–763
of transmitters, 109, 118–119, 331, 346, OPS in, 759–764
394–395 slotted, 761–763
of TT-FRs, 118–119, 331, 346, 394–395 switch controller rule in, 763
of TT-TRs, 118–119, 331, 346, 394–395 unslotted, 761, 763, 784–785
VCSEL, 212 Unconstrained algorithms, 486
 Index 963

Uncontrolled networks. See Static networks Virtual connection (VC), 11

Uncontrolled scheduling, 401–403 multicast, in hypernets, 631–632
Undirected hypergraphs, 613, 623, 625, in multitier networks, 582
628–630, 639–640, 873–874 Virtual path identifier (VPI), 82
Undirected hypernets, 608–610, 635–636 Virtual private network (VPN)
Undirected Kautz hypergraph MPLS allowing, 722, 728
address shift routing in, 625 optical, 858–859
constructed from Kautz digraph, 639–640 requirements of, 5–6, 8
converted to directed, 623 Virtual private optical network (VPON),
ShuffleNet compared with, 628–630 858–859
Undirected star network, 103n7 VOA. See Variable optical attenuator
Unfairness factor, 490 Voice-over IP (VoIP), 5–6, 8
Unfeasible paths, 511 Void-filling scheduling algorithms, 802–803
UNI. See User-Network Interface VoIP. See Voice-over IP
Unidirectional path-protected ring (UPPR), VPI. See Virtual path identifier
672–673 VPN. See Virtual private network
Unidirectional path-switched ring (UPSR) VPON. See Virtual private optical network
BLSR-UPSR interconnected architectures
of, 659–663 WADM. See Wavelength add/drop multiplexer
BLSR v., 655–657 (WADM)
fiber cuts and, 655–656 Walks, of graphs, 870–871, 874
nonrevertive v. revertive switching in, 655 WAN. See Wide area networks
two-fiber, 660 WASPNET testbed. See WAvelength Switch
UPSR-UPSR interconnected architectures optical Packet NETwork testbed
of, 659, 663 Waveband multiplexing, 37–38
Unsaturated gain, 194–195 Waveband reuse factor, 635
Unslotted networks Waveband-routed networks (WRNs), 28,
performance advantages of, 763 37–39
unbuffered, 761, 763, 784–785 constraints on, 38
UPPR. See Unidirectional path-protected ring multistar, 542–543
UPSR. See Unidirectional path-switched ring NCs in, 133–151
Upstream label object, 749 RWA in, 442–484
User Datagram Protocol (UDP), 727 static RCA in, 507–544
User-Network Interface (UNI), 720 wavelength requirements of, 432–434
parameters requested for, 743 Waveband routing, 553–555
signaling and, 742–745 Waveband-selective directional couplers, 57
in wavelength-on-demand, 858 Waveband-selective switch (WSS), 37, 56–57
Waveband-space switch (WSS)
Variable bit rate payload, 818 dynamic nodes and, 56–60
Variable optical attenuator (VOA), 241 hardware realization of, 58
VC. See Virtual connection three-stage realization of, 56–57
VCI. See Virtual channel identifier time scale of, 60–61
VCSEL. See Vertical cavity surface emitting Waveband spectrum partitioning, 35,
laser 38–39
Vertex coloring, 450–451 Waveband switching layer, 60, 296–297
Vertical cavity surface emitting laser Waveguide dispersion, 173
(VCSEL), 205 Waveguide grating, 259, 263–265
structure of, 211 Waveguide imperfections, 176
tunability of, 212 Wavelength add/drop multiplexer (WADM),
Virtual buffering, 765 58
Virtual channel identifier (VCI), 82 AWG-based, 271–272
964 Index

Wavelength add/drop multiplexer (cont.) Wavelength division multiple access

long-haul networks and, 854–856, 862 (WDMA), 390–395. See also Subcarrier
Mach-Zehnder, 264–265 multiplexing/wavelength division
in MANs, 842 multiple access/subcarrier multiple
MEMS-based, 270–271 access; Time division
MZ, 264–265 multiplexing/time-wavelength division
NAS combined with, 58, 70, 131 multiple access
physical-layer simulation of, 301–302 cost issues for, 833–836
vendor generic node architecture for, 851 economic case for, 836
Wavelength assignment electronic cross-connect in, 835
in double fiber pair access, 456 Wavelength division-multiplexed (WDM)
NAS/ONN connection determining, 128 mesh network
rule for, in static networks, 44–46 algorithms for, 603–606
for ShuffleNet, 156 grooming in, 602–607
in single fiber pair access case, 456–457 ILP for, 603–604
Wavelength bands, 176 lightpath edges in, 606
Wavelength-brokering operational model, Wavelength division-multiplexed (WDM) ring
859–860 algorithms for, 600–602
Wavelength bypass edge, 606 asymptotic relative costs for, 602
Wavelength continuity, 37–38 cost issues for, 832–834
channel assignment constrained by, 127, cross-connect capabilities of, 602
139, 450 economic case for, 834
λ-channels and, 63 fiber exhaust in, 832–833
through OLPs, 76 grooming in, 598–602
Wavelength conversion (WC), 801. See also interconnected, 302–308
Wavelength converters RCA in, 600
all-optical, 275–278 self-healing, 672–677
contention resolution by, 778–783 SONET over, 598–602
by DFC, 65, 276–277, 281 as SPRING, 453–458, 673–675
as enabling technology, 274–281, 789 as UPPR, 672–673
in OBS, 801 Wavelength division multiplexing (WDM), 34,
opaque, 278–281 37, 93. See also Dense wavelength
optical header processing combined with, division multiplexing; Wavelength
790 division-multiplexed mesh network;
optoelectronic, 275 Wavelength division-multiplexed ring
packet loss probabilities in, 767–770, advantages of, 828–829
780–781 amplets and, 204–205
shared, 781–782 coarse, 17, 177, 205
transparent, 275–278 cost issues for, 831–832
tunability in, 276, 779–782, 793–794 deflection routing and, 767
waveband switching layer implemented by, equalization in, 233–234
274 fiber link capacity increased by, 2, 10
Wavelength converters, 63–67 history of, 18–20
DFC as, 65, 276–277, 281 in MANs, 35–36, 843–847, 851–854
linear, 64 modulation formats in, 229
shared, 781–782 noise power and, 48
transponders as, 63 OBS and, 800–801
tunable, 779–782, 793–794 OLS network example of, 808–809
Wavelength cross-connect (WXC), 846 open v. closed installations of, 835–837
Wavelength demultiplexer (WDMUX), 44–45 OPS in, 61
Wavelength-dilated switch, 273–274 physical-layer simulation of, 301–311
 Index 965

in PONs, 416–420, 422–424 as dynamic single-wavelength-

in SONET, 80–81 per-waveband networks, 133
SPRINGs in, 673–675 embedding problem in, 594
supercontinuum light in, 209 full connectivity in, 479
transparency and, 829, 852–853 geographical reach limitation in, 577–578
in transparent networks, 852–853 growth of, 440, 478–479
trend toward, 80 LLNs compared with, 150–151, 527
UPPR in, 672–673 multicast-capable, 498, 501–505
WMUX for, 44–45, 56 multifiber links v. multiple wavelengths in,
Wavelength grid, 35 126
Wavelength interchange. See also multipoint connection set up in, 122–123,
Wavelength-interchanging cross-connect 122n16
blocking improved by, 497 multistar, 462–464
bridges in, 460–462 NCs in, 94–95, 122–133
fairness improved by, 491, 494, 498–499 ONN functionality in, 133
full, 468–469, 470–471, 493–495 optical connection set up in, 443
limited, 469–470 physical topology of, 434–442
in Monte Carlo approach to RCA, 484 point-to-point connections in, 122–123
by OLPs, 76–77 purely optical approaches in, 577, 578
as optimization problem, 465–471, RCA in, 124–133, 433, 442–544
493–494 reconfiguration in, 432
in SONET over WDM rings, 599 splitter placement in, 506
sparse, 460–461, 469–470, 494, 496 star, 123–124, 149
Wavelength-interchanging cross-connect static RCA in, 507–544
(WIXC), 63, 766 topologies of, 434–442
arbitrary connectivity and, 457–458 transparent, 122
bottlenecks removed by, 274 wavelength requirements of, 432–434,
implementation of, 65–66 471
OXCs and, 65 Wavelength-routed optical burst switching
Wavelength-link edge, 606 (WR-OBS), 805–806
Wavelength list, 486–487, 499 Wavelength-selective coupler (WSC), 191
Wavelength mixing, 789 Wavelength-selective cross-connect (WSXC)
Wavelength multiplexer (WMUX), 44–45, formation of, 855
56 in OPS networks, 57
Wavelength-on-demand, 858 Wavelength-selective switch (WSS), 288–297
Wavelength requirements conversion of, 273
extra fibers reducing, 476–477 design of, 259
for full connectivity, 456 MEMS-based multiwavelength, 270–271
of LLNs, 433 OADM, 289–293
of logical topology, 444 in optical filtering technology, 258–259
of LRNs, 433 OXC, 293–297
of waveband-routed networks, 432–434 Wavelength selectivity, 123
of wavelength-routed networks, 432–434, Wavelength spectrum partitioning, 34–35, 38
471 WAvelength Switch optical Packet NETwork
Wavelength-routed networks (WRNs), 23, 28, (WASPNET) testbed, 796–798
36, 38–39, 66–67, 94–95, 135–136, Wavelength terminal multiplexer (WTM), 80,
162–163 830
backward compatibility in, 461 Wavenumber, 172
λ-channels recognized in, 122 WC. See Wavelength conversion
connection rearrangement in, 124–126 WDM. See Wavelength division multiplexing
dynamic RCA in, 484–507
966 Index

WDM mesh network. See Wavelength WIXC. See Wavelength-interchanging

division-multiplexed mesh network cross-connect
WDM ring. See Wavelength WMUX. See Wavelength multiplexer
division-multiplexed ring WRN. See Waveband-routed networks;
WDMUX. See Wavelength demultiplexer Wavelength-routed networks
Wide area networks (WANs) WR-OBS. See Wavelength-routed optical
amplifiers in, 190–191 burst switching
current architecture considerations in, WSC. See Wavelength-selective coupler
854–856 WSS. See Waveband-selective switch;
hybrid approaches to, 14–16 Waveband-space switch;
MWNA in, 29–30 Wavelength-selective switch
OPS and, 61 WSXC. See Waveband switching layer;
optical approaches to, 14–16 Wavelength-selective cross-connect
physical topologies of, 438–439 WTM. See Wavelength terminal multiplexer
transmission impairments in, 75 WXC. See Wavelength cross-connect
Wide-sense nonblocking switches,
51–52 Y-branch switch, 248, 284