See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/221288659

Dynamic Routing Optimization in WDM Networks

Conference Paper · December 2010

DOI: 10.1109/GLOCOM.2010.5683270 · Source: DBLP

CITATIONS READS
12 121

5 authors, including:

Jorge Crichigno Nizam bin abd Ghani

University of South Carolina Universiti Sultan Zainal Abidin | UniSZA
123 PUBLICATIONS 1,532 CITATIONS 60 PUBLICATIONS 532 CITATIONS

SEE PROFILE SEE PROFILE

Joud Khoury Wei Shu

Raytheon Technologies University of New Mexico
42 PUBLICATIONS 366 CITATIONS 237 PUBLICATIONS 3,158 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Optical Netwoks View project

Studies on multi-channel networks using directional antennas View project

All content following this page was uploaded by Wei Shu on 17 February 2014.

The user has requested enhancement of the downloaded file.

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE Globecom 2010 proceedings.

Dynamic Routing Optimization in WDM Networks

J. Crichigno∗ , N. Ghani† , J. Khoury† , W. Shu† , M. Y. Wu‡
∗
Engineering Department, Northern New Mexico College, Espanola - NM, USA
†
Electrical & Computer Engineering Dept., University of New Mexico, Albuquerque - NM, USA
‡
Dept. of Computer Science & Engineering, Shanghai JiaoTong University, Shanghai, China
[email protected], {nghani, jkhoury, shu}@ece.unm.edu, [email protected]

Abstract—We present a multi-objective optimization approach In this paper, we present a multi-objective Integer Linear
for joint throughput optimization and traffic engineering, where Program (ILP) which jointly considers the throughput opti-
the routing request of traffic arrives one-by-one. We provide mization and the traffic engineering problems in wavelength
an Integer Linear Program (ILP) that simultaneously i) max-
imizes the aggregate throughput, ii) minimizes the resource convertible WDM network. In this kind of networks, converter
consumption, and iii) minimizes the maximum link utilization. devices are available at any node, and end-to-end traffic
We study the impact of optimizing the three different objectives requests must be routed such that no two connections use
simultaneously in dynamic environments, and show that better the same wavelength on a same link. We assume that a
solutions than those of mono-objective approaches can be ob- traffic request is composed of a source node, destination node,
tained. Because of the complexity of the ILP, we also propose
another ILP with reduced complexity, and study its performance and traffic demand. The proposed program simultaneously
and the optimality gap between it and optimal solutions. maximizes the aggregated throughput, minimizes the resource
Index Terms—WDM networks, integer linear programming, consumption and MLU. The ILP can generate optimal solu-
multi-objective optimization. tions for practical sized networks in a reasonable time. For
those scenarios where path disruptions are not permitted, we
I. I NTRODUCTION
also propose the application of another ILP, which works on
Dynamic throughput optimization and traffic engineering an on-demand basis. With this approach, the optimization
problem in Wavelength-Division Multiplexing (WDM) net- metrics quantify the solution for the new traffic request only
works refers to the problem of set up paths between nodes according to the current state of the network, avoiding full
in the network, called optical cross-connects, such that the re-optimization and path disruptions.
number of connection requests successfully routed (aggregate The paper is organized as follows. Section II discusses
throughput) are maximized while resource consumption and related work. Section III formulates the joint throughput op-
maximum link utilization (MLU) are minimized. timization and traffic engineering problem in WDM networks
Routing in WDM networks has been an active research area as an ILP. Section IV presents a suboptimal alternative ILP
for several years. Much of the work is focused on maximizing with reduced complexity, which avoids path disruptions and
the aggregated throughput in terms of accepted connections, full re-optimization. Section V shows performance studies, and
which is similar to minimizing the blocking probability [1]. Section VI concludes our work.
The latter is defined as the ratio of the rejected connections
to the total number of connection requests. Other approaches II. R ELATED W ORK
consider traffic engineering, which attempt to balance the load The throughput optimization problem in WDM networks
of the network and to improve the resource consumption. Load can be seen as an extension of the multi-commodity flow
balancing is obtained by minimizing the utilization of the most problem in a graph, where flows and link capacities take
heavily used link in the network, or MLU. This objective is integer values. Flows represent connection requests, while a
fundamental to minimize the link utilization throughout the link capacity represents the number of wavelengths available
network, so that no bottleneck link exists. Although minimiz- at the corresponding link. Banerjee et al. [1] presented an
ing the MLU avoid hot-spots, the total resource consumption ILP that minimizes the average number packet hop distance.
or sum of assigned wavelengths at links of paths may be higher The authors claimed that optimizing the above objective is
than needed. equivalent to maximizing throughput (i.e., the number of
The problem is further motivated by the fact that there accepted connections). However, in general these two metrics
may exist the possibility of having new routing requests in may conflict with each other. Krishnaswamy et al. [2] provided
the future (dynamically), which cannot be excluded. Thus, two ILPs for the routing problem in WDM networks, and
the routing scheme must be an on-line algorithm capable of relaxed LP formulations with proper rounding techniques to
handling requests in an optimal manner when the requests approximate to optimal solutions. Cavendish et al. [3] studied
are not all present at once. Since traffic requests may not the routing and wavelength assignment (RWA) problem in
be known in advance, the optimal solution may need to re- WDM mesh networks with full wavelength conversion capa-
establish paths for previous requests and re-optimization may bilities, and attempted to minimize multiple objectives such
be needed whenever a new request arrives or departs. as the number of wavelength converters, the number of hops

978-1-4244-5637-6/10/$26.00 ©2010 IEEE

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE Globecom 2010 proceedings.

and the number of wavelengths used along the path; yet the   
Max F = w1 · fn − w2 · xn
ij − w3 · α
authors just optimized one metric at a time. Belgacem et al. [4] n|ts e
n ≤t<tn n|ts e
n ≤t<tn (i,j)∈E
proposed a heuristic to solve large instances of the through- ⎧
  ⎨ fn ; if i = sn
put optimization problem in WDM networks. The heuristic xn
ij − xn
ji = −fn ; if i = dn n|tsn ≤ t < ten (1)
⎩ 0; otherwise
partitions the problem into several smaller ILP subproblems j:(i,j)∈E j:(j,i)∈E

that are optimally solved. Jaekel et al. [5] presented two ILP xn
ij ≤ α · cij ; (i, j) ∈ E (2)
formulations which minimize the MLU and the resources used n∈N

to route the connection request. Christodoulopoulos et al. [6] fn ≤ rn ; n|tsn ≤ t < ten (3)
presented an ILP to optimize the number of wavelengths to xn
ij ∈ {0, 1, 2, ...}; n|tsn ≤ t < ten , (i, j) ∈ E (4)
satisfy certain connection requests, and a heuristic to solve it fn ∈ {0, 1, 2, ..., rn }; n|tsn ≤ t < ten , (i, j) ∈ E (5)
in a polynomial time. This problem is a dual of the throughput 0 ≤ α≤1 (6)
optimization; the goal is the minimization of the number of
wavelengths instead of the maximization of the throughput.
Fig. 1. Integer Linear Program 1 (ILP1).
Note that the latter assumes that the topology is fixed, and the
number of wavelengths per link is already known. A traffic
engineering scheme which minimizes congestion is proposed
the number of connections allocated to any flow n. Eqs. (4)
in [7]. To achieve this objective, traffic engineering schemes
and (5) restrict the variables to be positive integers, and Eq.
attempt to minimize the MLU. In our previous work [8], we
(6) states that the MLU is between 0 and 1.
presented an ILP for the joint throughput optimization and
traffic engineering problem for static scenarios, where all the ILP1 can be considered as a multi-objective problem, where
traffic requests are routed at once. Performance studies showed the objective function F at a time t may be expressed as an
that, by considering the multiple objectives simultaneously, objective vector F = (F1 , F2 , F3 ), where:

better solutions can be obtained for static cases. This paper F1 = fn (throughput) , (7)
extends the work in [8] to dynamic scenarios, where traffic n|tsn ≤t<ten
requests arrive one-by-one, and presents two ILPs schemes  
for these scenarios. F2 = − xnij (resource consumption) ,(8)
n|tsn ≤t<ten (i,j)∈E
III. J OINT T HROUGHPUT O PTIMIZATION AND T RAFFIC F3 = −α (MLU) . (9)
E NGINEERING P ROBLEM
The relative importance of each objective can be varied
A. Problem Formulation according to the weight vector w  = (w1 , w2 , w3 ). We will
We represent a wavelength convertible WDM network as a consider the routing of all connection requests as the main
graph G = (V, E), where V is the set of nodes and E the set of goal. Thus, we will set w1 >> w2 and w1 >> w3 .
links. Let N be the set of end-to-end flows or traffic demands. In next sections, we will use Pareto optimality concepts
Each demand is characterized by a 5-tuple (sn , dn , rn , tsn , ten ), [9]. Let w(a),
 w(b)
 be two weight vectors, and F (a), F (b) be
which denotes the source node, the destination node, the the objective vectors corresponding to the solutions of ILP1
number of connection requests, the start time, and the end time with w(a)
 and w(b)
 respectively. We will say that the solution
of the demand. Flows are not known in advance, but arrives obtained by setting ILP1 with w(a)  is better than that with
dynamically on time. Let xnij be a variable representing the w(b)
 if and only if (iff) Fi (a) > Fi (b) for some i ∈ {1, 2, 3},
number of connection requests of the nth flow routed on link and Fi (a) ≥ Fi (b) for the remainder objectives. We denote
(i, j), and cij be the capacity of the link (i, j) in number this relation as F (a) > F (b).
of wavelengths. Define fn ≤ rn , n ∈ N , as the number of ILP1 is a generalized model for the joint throughput op-
connection requests allocated to flow n. Note that in typical timization and traffic engineering problem. The flexibility of
throughput optimization problems, rn is not given beforehand, the model permits us to solve not only the joint problem but
and the objective is to maximize fn . At a time t, the problem also either of them by appropriately setting the weight vector
is defined as shown in Fig. 1.  = (w1 , w2 , w3 ).
w
The objective function F consists of three terms with their
corresponding weights w1 , w2 and w3 . The first term is B. Complexity of ILP1
the aggregated throughput. The second term is the resource Variables xnij , Eq. (4), and fn , Eq. (5), are integer vari-
consumption (i.e., wavelengths) utilized by the scheme, and ables. The objective function and constraints are linear in
the third term, α, is the MLU. Minimizing the last two the variables xnij , fn and α, (i, j) ∈ E and n ∈ N .
objectives is equivalent to maximizing the negative of them. Therefore, the model is an integer linear program and is NP-
Eq. (1) is the flow conservation constraint, and Eq. (2) is the hard. The number of constraints from Eqs. (1), (2), and (3)
link capacity constraint, where capacities are multiplied by are O(|V | · |N |) (because at most |N | traffic requests can
the MLU. Eq. (3) restricts the number of connection requests be simultaneously active), |E|, and O(|N |). The number of
allocated to flow n to be no more than its requested number variables from Eqs. (4) and (5) are O(|E| · |N |) and O(|N |).
of connections rn . This constraint permits us to upper bound Thus, the number of constraints and variables are dominated

978-1-4244-5637-6/10/$26.00 ©2010 IEEE

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE Globecom 2010 proceedings.

by the terms O(|V | · |N |) and O(|E| · |N |). If these terms are

not large (i.e., the number of nodes and links in the network,
|V | and |E|, and the number of simultaneously active traffic
demands are not too large), then, ILP1 can be solved in a
reasonable amount of time.
IV. S UBOPTIMAL A PPROACH
Note that the optimal solution of ILP1 at any time t may Fig. 3. NSF network topology.
require the re-routing and re-establishment of paths whenever
a new request arrives or a previous request ends. Finding
the optimal solution of ILP1 may not be a viable solution the inter-arrival time is randomly distributed between 0 and 50
when path disruptions are not permitted, or when the number seconds. We solved ILP1 and ILP2 with four different weight
of active traffic requests, n|tsn ≤ t < ten , is large, since vectors: w(a)
 = (1000, 0, 0), w(b)
 = (950, 50, 0), w(c)
 =
the computational time increases rapidly. As an alternative (950, 0, 50), and w(d)
 = (900, 50, 50). The first objective of
approach, we also evaluate a suboptimal approach that routes the problem is the throughput maximization. Thus, by setting
traffic requests one-by-one, as they arrive. Let k ∈ N be a w1 >> w2 and w1 >> w3 , we give more importance to F1
traffic request currently arriving. Then, the request k is routed than F2 and F3 .
according to ILP2:
A. Impact of w
 in ILP1
 Consider Fig. 4(a). The curves for ILP1 were obtained with
Max F = w1 · fk − w2 · xkij − w3 · α
(i,j)∈E
w(a)
 (F1 (b), F1 (c), and F1 (d) are equal to F1 (a), because
⎧ w1 >> w2 and w1 >> w3 , for all weight vectors). For ILP2,
  ⎨ fk ; if i = sk
xn
ij − xn
ji = −fk ; if i = dk (10) the curves were obtained with w(d),
 which produced highest
⎩ 0; otherwise
j:(i,j)∈E j:(j,i)∈E throughput (to be explained in Section V-B). The parameter r
xkij ≤ α · crij ; (i, j) ∈ E (11) is the per-flow connection requests; i.e., for all flow n ∈ N ,
fk ≤ rk ; (12) rn = r. Note that the gap between ILP1 and ILP2 increases
xkij ∈ {0, 1, 2, ...}; (i, j) ∈ E (13) as r increases. For r = 1, the curves are overlapped. The
fk ∈ {0, 1, 2, ..., rn }; (i, j) ∈ E (14) impact of the w  can be noted in Figs. 4(b)-(d), which shows
0 ≤ α≤1 (15) the resource consumption in wavelengths (F2 ), the MLU (F3 ),
and the average link utilization (ALU), when r = 1. The ALU
is defined as
Fig. 2. Integer Linear Program 2 (ILP2). 1   xnij
,
|E| cij
(i,j)∈E n∈N
ILP2 has the same structure as ILP1; however, it only
establishes a set of paths for the traffic request k considering and measures the link utilization in the network, on average.
the current state of the network. The link capacity constraint Fig. 4(b) shows the resource consumption. If F2 is ignored
given by Eq. (11) includes the term crij , which is the residual as an objective function (w(a)
 = (1000, 0, 0) and w(c)
 =
capacity of link (i, j) ∈ E at the time the traffic request (950, 0, 50)), the resource consumption increases with respect
k arrives. The optimal solution of ILP2 may represent a to those results obtained by slightly weighting F2 with a
suboptimal solution for the multi-commodity problem (ILP1); positive weight of 50. Similarly, Figs. 4(c) and 4(d) show that
however, it has a lower complexity than ILP1. the MLU and ALU are considerably reduced when a positive
weight w3 > 0 is used. Note that even w(c)  = (950, 50, 0)
A. Complexity of ILP2 reduces the MLU and ALU; i.e., the minimization of the
Similar to ILP1, variables xnij , Eq. (13), and fn , Eq. (14), resource consumption indirectly reduces MLU and ALU.
are integer variables. The objective function and constraints are
linear in the variables xnij , fn and α, (i, j) ∈ E and n ∈ N , B. Impact of w
 in ILP2
which makes ILP2 also an integer linear program and NP-hard Consider now Figs. 5(a)-(d), which show the results ob-
problem. However, since only one traffic request is routed at a tained with ILP2. Here, we set r = 5 to better illustrate the
time, the number of constraints and variables are O(|V |) and effect of difference weight vectors. In contrasts with ILP1
O(|E|) instead of O(|V | · |N |) and O(|E| · |N |). Thus, the where the four different weight vectors produced the same
running time of solving an instance of ILP2 is inferior than value of F1 , the throughput obtained with ILP2 varies with the
that of solving an instance of ILP1. weight vector. According to Fig. 5(a), the higher throughput is
obtained with w(b)
 = (950, 50, 0) and w(c)
 = (950, 50, 50).
V. P ERFORMANCE S TUDIES These weight vectors produced even higher throughput than
Consider the network network topology shown in Fig. 3. We that obtained with w(a)
 = (1000, 0, 0), which optimizes F1
generated 200 hundreds traffic demands. The duration of each only. This example demonstrates that to optimize through-
traffic demand is exponentially distributed (10 seconds), and put, other objectives such as resource consumption must be

978-1-4244-5637-6/10/$26.00 ©2010 IEEE

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE Globecom 2010 proceedings.

(a) (b)

(c) (d)

Fig. 4. (a) Throughput results. The curves for ILP1 were obtained with w(a),
 while those for ILP2 were obtained with w(d).
 r is the per-flow traffic request.
(b) Resource consumption, in number of wavelengths, for ILP1. The per-flow traffic requests is r = 1. (c) MLU obtained with ILP1, for r = 1. (d) ALU
obtained with ILP1, for r = 1.

simultaneously considered. Fig. 5(b) shows the resource con- of throughput between ILP2 and the optimal increases up to
sumption. Note again the good performance of ILP2, obtained 60 % for r = 5. Note, however, that until t = 50 s, δ is mostly
with w(b)
 and w(c):
 with these weight vectors, not only the below 40 %. For t ≥ 50, there are no more connection request
throughput is maximized but also the resource consumption arrivals; remember that ILP1 considers, at any time t, all the
is minimized. Fig. 5(c) illustrates the MLU. Since r = 5, active flows n ∈ N |tsn ≤ t < ten , while ILP2 does not reroute
the scenario represents a high-loaded case, and the MLU traffic requests previously arrived (no flow disruption). This
is 1 most of the time, independently of w.  As shown in is one of the reason of the surcharge gap. Another reason
Fig. 5(d), the ALU is also minimized with w(b)  and w(c).
 of the bigger gap when t > 50, for r ≥ 4, is the fact that
Thus, by considering the 3 objectives simultaneously, better ILP1 reroutes traffic requests that were not able to be routed
solutions are obtained. Note that these solutions may not be when the network was congested. E.g, let assume that a traffic
even tradeoff solutions, but Pareto solutions. request n arrived at tsn = 40, with an ending time ten = 70
s. At t = 40 s, assume that the network is congested, and n
C. Throughput Surcharge cannot be routed. Assume also that at t = 60 s, the congestion
Since the weight vectors used in this paper attempts to pri- is reduced and resources can be allocated to request n. ILP1
marily maximize throughput, we now compare the throughput then routes traffic request n, while ILP2 does not.
surcharge between the best solutions, in term of throughput,
of ILP2 (obtained with w(d))
 with those of ILP1 (for ILP1, VI. C ONCLUSION
any weight vector produced the same throughput). We define We have presented a multi-objective ILP for the joint
the throughput surcharge δ as the gap between the optimal throughput optimization and traffic engineering problem in
throughput obtained with ILP1, F1 (ILP1), and the throughput dynamic WDM networks (ILP1), where traffic requests arrive
obtained with ILP2, F1 (ILP2): one-by-one. The scheme considers the possibility of having
F1 (ILP 1) − F1 (ILP 2) new routing requests dynamically, and can be considered as
δ= . (16) an on-line scheme capable of handling requests in an optimal
F1 (ILP 1)
manner when the requests are not all presented at once. The
Fig. 6 shows the results for different values of r. When r = proposed program simultaneously maximizes the aggregated
1 and r = 2, ILP2 produces the same throughput as ILP1, and throughput, minimizes the resource consumption and MLU.
the throughput surcharge is 0. As r increases, the difference The ILP can generate optimal solutions for practical sized

978-1-4244-5637-6/10/$26.00 ©2010 IEEE

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE Globecom 2010 proceedings.

(a) (b)

(c) (d)

Fig. 5. (a) Throughput results obtained with ILP2, for r = 5. (b) Resource consumption, in number of wavelengths, obtained with ILP2, for r = 5. (c)
MLU obtained with ILP2, for r = 5. (d) ALU obtained with ILP2, for r = 5.

networks in a reasonable time. We have also presented a Acknowledgement. This work was funding in part by the US
suboptimal approach based on integer linear programming National Science Foundation under Award CNS-0806637. The
(ILP2), which does not require the re-establishment of paths authors are very grateful for this support.
for previous traffic requests already routed. The complexity of
ILP2 is much lower than the ILP1, and is a good alternative R EFERENCES
scheme according to simulation results. Performance studies [1] D. Banerjee and B. Mukherjee. Wavelength-routed optical networks:
linear formulation, resource budgeting tradeoffs, and a reconfiguration
show that, by considering the multiple objectives simultane- study. IEEE/ACM Trans. Netw., 8(5):598–607, 2000.
ously, better solutions can be obtained. For ILP2, we have [2] R. Krishnaswamy and K. Sivarajan. Algorithms for routing and wave-
noticed that, to maximize connection requests, the objective to length assignment based on solutions of lp-relaxations. IEEE Communi-
cation Letters, 5(10), 2001.
be optimized should not only be throughput but also resource [3] D. Cavendish, A. Kolarov, and B. Sengupta. Routing and wavelength
consumption and MLU. Future work includes the development assignment in wdm mesh networks. IEEE GLOBECOM 2004, Dec. 2004.
of a systematic procedure to determine appropriate values of [4] L. D. Belgacem and N. Puech. Solving large size instances of the rwa
problem using graph partitioning. In IEEE International Conference on
the weight vectors. Optical Networking Design and Modeling (ONDM), 2008.
[5] A. Jaekel and Y. Chen. Efficient distributed algorithm for rwa using path
protection. Journal of Networks, 1(3), July. 2006.
[6] K. Christodoulopoulos, K. Manousakis, and E. Varvarigos. Comparison
of routing and wavelength assignment algorithms in wdm networks. IEEE
GLOBECOM, Dec. 2008.
[7] R.M. Krishnaswamy and K.N. Sivarajan. Design of logical topologies:
A linear formulation for wavelength routed optical networks with no
wavelength changers. IEEE INFOCOM, Mar. 1998.
[8] J. Crichigno, W. Shu, and M. Y. Wu. Throughput optimization and traffic
engineering in wdm networks considering multiple metrics. In IEEE
International Conference on Communications, 2010.
[9] K. Deb. Multi-Objective Optimisation Using Evolutionary Algorithms.
John Wiley and Sons Ltd, 2001.

Fig. 6. Throughput surcharge obtained according to Eq. (16).

978-1-4244-5637-6/10/$26.00 ©2010 IEEE

View publication stats