JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 23, NO.

1, JANUARY 2005

277

A Highly Flexible and Efficient Passive Optical

Network Employing Dynamic Wavelength Allocation
Yu-Li Hsueh, Student Member, IEEE, Matthew S. Rogge, Student Member, IEEE,
Shu Yamamoto, Senior Member, IEEE, and Leonid G. Kazovsky, Fellow, IEEE

AbstractA novel and high-performance passive optical

network (PON), the SUCCESS-DWA PON, employs dynamic
wavelength allocation to provide bandwidth sharing across multiple physical PONs. In the downstream, tunable lasers, an arrayed
waveguide grating, and coarse/fine filtering combine to create
a flexible new optical access solution. In the upstream, several
distributed and centralized schemes are proposed and investigated. The network performance is compared to conventional
TDM-PONs under different traffic models, including the selfsimilar traffic model and the transaction-oriented model. Broadcast support and deployment issues are addressed. The networks
excellent scalability can bridge the gap between conventional
TDM-PONs and WDM-PONs. The powerful architecture is a
promising candidate for next generation optical access networks.
Index TermsDynamic wavelength allocation (DWA), optical
access network.

I. INTRODUCTION

HE exponential growth of Internet traffic volume directly

relates to the increase in high-speed connections to end
users. Still, the demand for ever higher data rates remains
strong. In the backbone networks, capacities greater than 1 Tb/s
over a single fiber have been achieved by means of wavelengthdivision-multiplexing (WDM) technology with hundreds of
channels. The emergence of low-cost and high-speed Ethernetbased local area networks (LANs) is accelerating the demands
for high-speed connections as well. To break the bottleneck between the LAN and the ultrahigh-capacity backbone networks,
cost-effective and high-performance access solutions are desirable. Passive optical networks (PONs) have been identified as
one of the promising access solutions, as shown by the many
research and standardization efforts focusing on this area.
Currently, the broadly accepted optical access solution is the
time-division multiplexed (TDM) PON. TDM-PONs combine
the high capacity afforded by optical fiber with the low installation and maintenance cost of a passive infrastructure. The
broadcast transmission in the downstream and the time-sharing
transmissions in the upstream limit the bandwidth of individual
users, but the resulting low transceiver count minimizes the
cost enough to justify the tradeoff.

Manuscript received December 16, 2003; revised September 21, 2004.

Y.-L. Hsueh, M. S. Rogge, and L. G. Kazovsky are with the Electrical Engineering Department, Stanford University, Stanford, CA 94305 USA.
S. Yamamoto is with KDDI R&D Laboratories, Inc., Saitama 356-8502,
Japan.
Digital Object Identifier 10.1109/JLT.2004.838811

One straightforward approach to increase the capacity is to

assign a set of wavelengths to each user for down/upstream
transmissions, which leads to the WDM-PON [1][3]. WDMPONs create point-to-point links between the central office (CO)
and each user, so no sharing is needed. However, high performance is not without high costwhenever the user shuts down
his connection, the corresponding transceiver in the CO is idle
and cannot support other users, in which case the network resource is left unused. WDM-PONs are high-performance but
luxurious access solutions today. However, as bandwidth demands increase and optical component costs decrease, WDMPONs will become more practical and significant.
Unfortunately, the migration path from TDM to WDM is not
trivial. A full WDM-PON has dedicated transmitters/receivers
for each end user as well as wavelength-routing devices in the
infrastructure. The cost required to add these new wavelengths
and modify the infrastructure to support WDM constitute a
major roadblock to adoption. Ideal solutions provide cost-effective and smooth service upgrades with minimal impact on
the existing TDM infrastructure.
In [4] we initiated the concept of a novel optical access network architecture, the SUCCESS-DWA PON, developed under
the Stanford University Access networking project at the Photonics and Networking Research Laboratory (PNRL). It employs dynamic wavelength allocation (DWA) to achieve a flexible, cost-effective, and high-performance PON architecture. In
this paper, we describe the architecture in detail, investigate the
performance, and comment on several implementation issues.
One significant feature of the SUCCESS-DWA PON is that its
excellent scalability can easily bridge the large gap between
TDM and WDM PONs. In addition, the architecture provides
excellent cost efficiency and network performance by sharing
bandwidth across multiple physical PONs. Existing arbitrary
field-deployed PON infrastructures remain in tact when brought
together into a SUCCESS-DWA PON, making the architecture
an ideal candidate for upgrading existing PONs. The inherent
flexibility and high performance allow the architecture to serve
equally well as high-end access networks serving large businesses and campuses.
The material and contributions of this paper are organized as
follows: Section II gives the description of the proposed network; analysis and performance comparisons are shown in Section III; the support for broadcast, the wavelength plan, and implementation issues are discussed in Section IV; the conclusion
is drawn in Section V.

0733-8724/$20.00 2005 IEEE

278

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 23, NO. 1, JANUARY 2005

Fig. 1. Basic architecture of SUCCESS-DWA PON in the downstream. The

grayed-out AWG and TLs demonstrate the system upgrade path.

time. If TL1 were to tune his laser through consecutive AWG

channels, the AWG would route channels {1,5,9, } toward
PON1, assuming channels {1,5,9, } are the straight through
channels of the AWG. Similarly, channels {2,6,10, } would
be routed to PON2, and likewise for PON3 and PON4. If a TL
wishes to communicate with a user on a particular PON, he
must determine what wavelength (a) falls in the passband of the
user and (b) exits the AWG destined for the correct PONfor
each TL-user combination, there exists exactly one wavelength
for which both (a) and (b) are satisfied.
For example, consider TL1 and User 18 (user channel 2,
PON2). If TL1 wishes to communicate with User 18, he compares the AWG channels acceptable by User 18, which are
{5,6,7,8}, and those that reach PON2 from TL1, which are
{2,6,10, }. The common wavelength in the two sets determines that TL1 should transmit on AWG channel 6. Generally,
communicating with the user channel
for a TL numbered
number , the corresponding AWG channel can be expressed
as

(1)
Fig. 2. Wavelength bands for AWG channels and thin-film WDM filter
channels.

II. NETWORK ARCHITECTURE

A. Downstream
Fig. 1 shows the network architecture. Tunable lasers (TLs),
the arrayed waveguide grating (AWG), and thin-film WDM filters (Ch1Ch16 in Fig. 1) constitute the key parts of this network. TLs and the AWG reside in the CO, while the WDM filters
are within the Optical Network Units (ONUs). The AWG which
couples the TLs together is cyclic or near-cyclic. Notice that
the field-deployed infrastructure is left untouched. This illustrative network includes four TLs and sixty-four end users located
across four physical PONs. The architecture can be extended in
numerous ways, e.g., to encompass more physical PONs or to
provide more TLs per PON, which is addressed later. The dotted
PON boxes correspond to arbitrary passive fiber distribution networks of existing PONs. Traditional PONs may be physically
deployed with different distribution techniques, including multistage splitting (e.g. PON1 in Fig. 1), bus tapping (e.g., PON2),
single wide-band splitters (e.g., PON3), or a hybrid combination (e.g., PON4). Regardless of the physical configuration, the
design requires only that all wavelengths from the optical line
terminal (OLT) reach all ONUs. The upstream traffic is separated from the downstream traffic with a wide-band WDM filter
between the AWG and the PON.
In the basic architecture, each ONU within a single PON contains a unique fixed-wavelength filter and a burst-mode receiver.
The key lies in the fact that the passband of the ONU filter encompasses the free spectral range of the AWG. For example,
200 GHz ONU filters would work with a 4 4 cyclic 50 GHz
AWG. The relative filter shapes are illustrated in Fig. 2.
The seemingly complicated architecture provides a very
powerful capabilityany tunable laser can individually address any ONU across separate physical PONs at any given

in which
and
denote the total number of user
channels, and the number of input AWG ports (the maximum number of TLs which can be coupled via the AWG),
respectively.
Since each ONU contains only one photodetector (PD), two
TLs must not access the same ONU simultaneously and a suitable media access control protocol is necessary. Notice that the
AWG does allow all TLs to simultaneously transmit on the same
wavelength. In this case, each TL will be routed to a different
PON. For each downstream frame, the TLs tune to appropriate
wavelengths and transmit the data to corresponding end users.
The transmission durations for end users are globally managed
to achieve optimal performance. This is very useful especially
when accommodating very bursty Internet traffic. All TLs share
the load, shifting bandwidth back and forth across the separate
physical PONs as necessarywe call this technique dynamic
wavelength allocation.
To illustrate the flexibility of this architecture, compare an
initial deployment of four TDM PONs to a SUCCESS-DWA
PON that spans four physical PONs. The first several subscribers would likely occupy physical locations across more
than one PON, and in the worst case they may be spread across
all four PONs. In the four TDM PON case, then, all four OLTs
(lasers) must be activated, despite the fact that some OLTs may
only be serving a few subscribers. With the SUCCESS-DWA
PON, on the other hand, only one TL and AWG are initially
added to the central office, and the subscribers across multiple
PONs are all serviced by the single TL. As demand grows, additional TLs can be added to the AWG. When the subscription
rate is high enough, the two scenarios seem to convergeboth
have four transmitters serving all subscribers. However, the
SUCCESS-DWA PON enjoys the benefit of statistical multiplexing over a larger customer base, so its performance will
exceed that of the four TDM PONs.

HSUEH et al.: HIGHLY FLEXIBLE AND EFFICIENT PON EMPLOYING DYNAMIC WAVELENGTH ALLOCATION

279

Fig. 4. Tunable demultiplexer: (a) cascade implementation composed of three

2 wavelength demultiplexers with complementary outputs;
single-stage 1
(b) arbitrary wavelength demultiplexing; and (c) consecutive wavelength
demultiplexing.

Fig. 3. Upstream scenarios: (a) conventional TDM-PON; (b) SUCCESSDWA with distributed upstream schemes; and (c) SUCCESS-DWA with
centralized upstream schemes.

When demand dictates, the SUCCESS-DWA PON can be

scaled far beyond the conventional TDM PON. The grayed-out
AWG in Fig. 1 illustrates the concept. Additional AWGs can
be added to shift from four TLs serving four PONs to four TLs
serving each single PON. If 8 8 AWGs are utilized, sixteen
end users can be served by eight TLs, which results in a very
high-performance network close to a WDM-PON. The excellent scalability provides a smooth and graceful upgrade from a
TDM-PON all the way toward a WDM PON. The step-by-step
system upgrade easily tracks user demands, and the initial overhead can be even lower than that of conventional TDM-PONs. In
addition, the TLs provide protection for each other, maintaining
service to all physical PONs in the event of a failure.
B. Upstream
Depending on the performance requirements, there is a wide
range of possible scenarios for upstream transmission. In this
subsection, several different upstream schemes are investigated
and reported.
One inexpensive upstream scenario requires only a fixedwavelength transmitter in each ONUa cost-effective 1310 nm
FabryProt (FP) laser can serve as the light source. Between
the AWG and the physical PON, the upstream wavelength is
extracted and a single PD terminates all the traffic for that PON.
Within each physical PON, then, bandwidth is shared amongst
all users on that PONthis is precisely the upstream scheme
in a traditional TDM-PON, as illustrated in Fig. 3(a). The performance is limited, of course, but the cost is minimized. In [5],
however, it was reported that the measured upstream traffic rates
were highly related to the downstream traffic rates. Users enjoying high-speed downloads also expect a commensurate data
rate in the upstream. This suggests that a high-performance upstream scheme is appropriate.
Ideally, the concept of DWA could also be employed in the
upstream, allowing similar features in terms of cost, scalability,
and performance. To realize DWA in the upstream, there must
be tunability in the network. Fig. 3(b) shows the scenario with

tunable devices in the ONU. In this example, a demultiplexer

(DeMux) in the OLT routes the upstream wavelengths to the
four corresponding PDs. The light source at each ONU can be
i) a fixed-wavelength FP; ii) two or more FPs at different wavelengths; or iii) a TL. For i), the ONU is virtually grouped with
other ONUs at the same wavelength, and shares the same PD in
the time domain. For ii), the ONU can choose one of the FPs
for transmission. For iii), the ONU can use any of the upstream
wavelengths.
The tunablility can also be in the OLT. In Fig. 3(c), the DeMux
at the OLT is tunable, which may potentially be realized by
MEMS [6], tunable thin-film filters [7], tunable fiber couplers
[8], or other technologies. Each end user has a specific upstream
wavelength , and cost-effective VCSELs would be ideal light
sources.
Conceptually, a
optical DeMux could be realized by casDeMux as shown in Fig. 4(a). The function of the
cading
DeMux is partitioning the incoming wavelengths into two
output groups, such that the two output ports would have complementary output spectra. Tunable DeMux with different wavelength grouping abilities could affect the network performance.
Fig. 4(b) illustrates the spectrum of a tunable DeMux with arbitrary wavelength grouping ability, while Fig. 4(c) illustrates that
with consecutive wavelength grouping ability, which shows a relatively easier but limited wavelength demultiplexing function.
Based on the illustrated setups in Fig. 3, we investigate five
upstream schemes and evaluate each in terms of cost, performance, and scalability.
Scheme A: Fixed grouping of end users with FPs at several
different wavelengths.
Scheme B: Some users are equipped with dual FPs at different wavelengths, and the rest are equipped with a single
FP.
Scheme C: Some users are equipped with TLs, and the
others are equipped with a single FP.
Scheme D: A tunable DeMux in the OLT that performs
consecutive wavelength demultiplexing.
Scheme E: A tunable DeMux in the OLT that performs
arbitrary wavelength demultiplexing.
For Scheme A, in the case of 16 users and four wavelengths,
the users are divided into four groups. Users within a group
transmit on the same upstream wavelength and time-share the
same PD in the OLT. This simple grouping among users by

280

Fig. 5. The basic SUCCESS-DWA upstream architecture for users across

multiple PONs. The grayed-out PDs and AWG demonstrate the system upgrade
path.

wavelengths can be viewed as a straightforward upgrade from

the conventional TDM-PON. It is static in wavelength allocation
and provides a baseline for comparison with the DWA schemes.
Schemes B and C are distributed DWA schemes, in which the
tunability is spread across the ONUs. The distributed schemes
allow more flexibility since the deployment of tunable devices
can be judged by user demands. On the other hand, Schemes D
and E are centralized DWA schemes, in which the tunability resides in the OLT.
Qualitatively, full tunability results in the best performance.
Scheme E exhibits full-tunability, but at the cost of an expensive, centralized tunable DeMux. Scheme C can provide equal
performance if all users are equipped with TLs. While the centralized schemes subject all ONUs to the high cost of the tunable
device, the distributed schemes require only those ONUs which
demand high performance to be upgraded. The added design
flexibility of distributed schemes makes them more preferable.
Ideally, the upstream scenario should have a strong and
smooth migration path similar to the downstream scheme.
Initially, each user is equipped with a fixed-wavelength transmitter that corresponds to the upstream group to which he has
been assigned. If the groups are created carefully, one can still
achieve scalability similar to the downstream SUCCESS-DWA
scheme. A complete upstream SUCCESS-DWA PON architecture is shown in Fig. 5. Four physical PONs are connected to
the OLT, and a
AWG functions as the DeMux in Fig. 3
which is responsible for routing the incoming wavelengths to
the corresponding PDs. It is worth noting that this AWG does
not require cyclic features, therefore a thin-film based AWG
could be adopted, and lower insertion losses and costs can
be expected. Additionally, upstream and downstream AWGs
pass completely different wavelengths and require different
channel spacings. Therefore, separate AWGs are necessary for
up/downstream. More details about the wavelength plan are
described in Section IV.
Similar to the downstream, only one PD and its corresponding
receiver module are activated in the initial deployment. To cover
all end users located on different physical PONs, the first several

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 23, NO. 1, JANUARY 2005

subscribers from different PONs are assigned different upstream

wavelengths, i.e., the first several subscribers from PON1 are assigned , the first several subscribers from PON2 are assigned
, and so on, as shown in Fig. 5. When the number of users
increases, a second PD and its corresponding receiver module
can be installed in the OLT. For new subscribers to be served
by the second PD, they are assigned on PON1, on PON2,
and so on, as in Fig. 5. In general, with a
AWG, the
first assigned wavelength for subscribers from the th PON is
.
This architecture can scale far beyond its TDM-PON counterpart, which is illustrated by the grayed-out AWG in Fig. 5. When
demand dictates, similar AWGs can be added to shift from four
PDs serving four PONs to four PDs serving each PON. For
the bandwidth-demanding users, tunable devices can replace the
fixed-wavelength transmitters, as in Schemes B and C. Unlike
the SUCCESS-DWA scheme in the downstream which makes
use of fast TLs with 10 ns tuning times to achieve DWA on
the order of , the upstream DWA schemes perform relatively
slower wavelength reallocations on the order of ms, due to the
more involved communications between the OLT and ONUs for
the upstream transmission scheduling.
III. ANALYSIS AND RESULTS
In this section, we compare the performance of the
SUCCESS-DWA PON with the aggregate performance of a
group of conventional TDM PONs. Together, the conventional
TDM-PONs serve the same number of users with the same
number of transmitters as the SUCCESS-DWA PON. Under
these constraints, the performance gain is quantified for many
different traffic scenarios. Using the same constraints to compare a conventional WDM PON to the SUCCESS-DWA PON
is not very informative, as a fully upgraded SUCCESS-DWA
PON with the same number of TX/RX as end users perfectly
emulates a conventional WDM-PON.
It is well known that the arrival of Internet packets can exhibit extreme rate variations in multiple time scales. The main
reason for rate fluctuations at time scales greater than a few hundred milliseconds is due to extreme variability in the flow sizes,
while the variability which occurs at smaller time scales is due
to the burstiness induced by TCP [9], [10]. In the first set of
simulations, we employ the -stable self-similar traffic model
developed in [11] to capture the burstiness and long-range correlations of realistic Ethernet packets. The characteristic exponent is chosen to be 1.63 and the self-similarity parameter
, which were extracted from an actual file transmission over Ethernet in [11]. The traffic dispersion and the mean
traffic are chosen at different values for different traffic loads
.
while keeping a constant ratio of
The first simulations compare downstream performance of
a four-TL SUCCESS-DWA PON to four TDM-PONs. In both
cases, sixty-four users are evenly distributed across the four
physical PONs. Fig. 6(a) and (b) plot the latency and queue
depth characteristics, respectively, for the SUCCESS-DWA and
TDM-PON architectures. Clearly, the SUCCESS-DWA PON
outperforms the TDM-PONs by a sizeable margin. Note that
both TDM and SUCCESS-DWA PONs are subject to the exact
same traffic patterns in any given simulation run. Due to the

HSUEH et al.: HIGHLY FLEXIBLE AND EFFICIENT PON EMPLOYING DYNAMIC WAVELENGTH ALLOCATION

Fig. 6. Downstream characteristics (a) average packet latency (b) average

queue depth.

fixed
ratio, the variance of the queue lengths increases
with the traffic load. This results in increased statistical multiplexing gain for the SUCCESS-DWA PON as the traffic load
increases. If more TLs were coupled together, as in an
SUCCESS-DWA PON, the performance would be even better,
since more transmitters can shift their wavelengths back and
forth across multiple PONs when necessary.
For the upstream, Fig. 7(a) and (b) show the latency and queue
depth characteristics versus the traffic load for the five schemes.
For distributed Schemes B and C, half of the ONUs are equipped
with tunable transmitters. Sixteen ONUs served by four PDs
are considered in this analysis. Clearly, all the SUCCESS-DWA
schemes B-E outperform the fixed-grouping scheme A. There
is no significant difference among the upstream DWA schemes,
even though schemes B and C are only partially populated with
tunable devices.
Much of the work on traffic modeling has been derived
from measurements taken either on the aggregated links of
the Internet, or at the servers or campus networks attached to
the Internetthe -stable self-similar traffic model used up to
now is one example. With the -stable traffic model, every end
user is assumed continuously downloading/uploading data with

281

Fig. 7. Upstream characteristics: (a) average packet latency and (b) average
queue depth.

self-similar packet arrivals. Optical access networks, however,

sit at the far edge of the networkclose to the users, not the
servers. Additionally, they have not been widely deployed and
characterized.
The Markov process has been widely employed to model Internet traffic [12][15] and variable bit rate video traffic. A large
number of ON/OFF sources with application specific transaction
lengths or data volume distributions can be superposed to constitute a batch process [12][15] to emulate Internet traffic. In
an attempt to approximate a sequence of transactions as initiated
by an actual end user, we adopted a transaction-oriented Markov
traffic model with specific transaction lengths corresponding to
the popular user applications. In this approach, aggregation of
the traffic from or to multiple end users results in a superposition of Markovian processes at the OLT, in which the traffic
buffering and dispatching behaviors are of interest. Specifically,
we consider three different families of transactions roughly corresponding to (a) web content [100 kB300 kB], (b) larger files
[34 MB] (i.e., MP3), and (c) significant data transfers [100
MB200 MB] (i.e. multimedia content). The bursty nature of
user transactions is modeled with discrete-time Markov chains.

282

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 23, NO. 1, JANUARY 2005

Fig. 9. Latency comparison between SUCCESS-DWA PON and TDM PONs

for different traffic scenarios. For TS1,  =  = 1, q = 0:5, q = 0:9, and
q = 0:6. For TS2,  =  = 0:1, q = 0:8, q = 1, and q = 0:95.

Fig. 8. Transaction-oriented traffic model based on Markov chains.

In Fig. 8, S1, S2, S3, and S0 correspond to transactions (a),

(b), (c), and the idle state, respectively. The parameters are related by the equations
(2)
(3)
(4)
where is the average traffic load of the user; and denote the
ratios of transaction types (b) to (a) and (c) to (a), respectively.
For each time slot in the idle state, a user can initiate any one
of the three transaction types or remain in the idle state. Once a
transaction has been initiated, the users demand remains constant
until the transaction is complete. By varying the probabilities of
each type of transaction and its degree of burstiness, we are able
to examine several different traffic scenarios. Different offered
loads and subscriber scenarios are considered as well. Since the
optimal total buffer size depends heavily on both the traffic and
the practical limitations of the design, we chose to use infinite
buffer sizes and focus on latency as a measure of performance.
Two different sets of traffic parameters were chosen: traffic
scenario 1 (TS1) provides an approximately equal mix of transaction types; TS2 emphasizes the amount of small (web) transactions. In TS1, the transaction duration for web content varies
considerably while larger files and significant transfers have relatively constant durations. In TS2, the transaction duration variation for web content is substantially reduced. Fig. 9 plots the
latency characteristics for TDM and SUCCESS-DWA PON architectures for both TS1 and TS2. The SUCCESS-DWA PON
significantly outperforms the TDM PONs. Under this traffic
model, the distribution of active users tends to be more uneven when the traffic load is low, resulting in more substantial gains at lower traffic loads. Several other traffic scenarios

Fig. 10. Latency versus user distribution shows consistently good performance
for the SUCCESS-DWA PON and strong dependence on user distribution for
TDM-PON. Total number of users on PON1 and PON2 is 16.

were simulatedall showed appreciable performance advantages for the SUCCESS-DWA PON. Note that both TDM and
SUCCESS-DWA PONs are subject to the exact same traffic patterns in any given simulation run.
Another key benefit of the SUCCESS-DWA PON is its
indifference to user distribution. Consider four TDM-PONs
where the subscribers are unevenly distributedmost of the
active subscribers are attached to PON1, for example. Since the
PONs are disjoint, the available resources from PONs 2, 3, and
4 cannot be used to improve the performance of PON1. The
SUCCESS-DWA PON treats all users fairly, regardless of their
location across the physical PONs. To illustrate this feature,
simulations with varying user distributions were performed. In
these simulations, two physical PONs support a total of sixteen
users, i.e., an average subscription rate of 50%. In Fig. 10, the
-axis represents different possible distributions of users across
the two PONs. As Fig. 10 indicates, not only do the TDM PONs
perform poorly in general, but the TDM-PON with the most

HSUEH et al.: HIGHLY FLEXIBLE AND EFFICIENT PON EMPLOYING DYNAMIC WAVELENGTH ALLOCATION

283

Fig. 12. User filter passband spectra with regular user channels and the broadcast channel.

Fig. 11. (a) Average packet latency versus traffic load for all ONUs in Schemes
AE; half of the ONUs are equipped with tunable transmitters in Schemes B and
C. (b) Average packet latency for fixed and tunable ONUs in Schemes B and C,
when the number of tunable ONUs increase from 0 to 16. Traffic load is 0.5.

users suffers considerably. As expected, the SUCCESS-DWA

PON performs equally well in all cases.
This model is also applied to the upstream simulations.
Fig. 11(a) shows the latency characteristics for each of the five
upstream schemes described in Section II. Scheme A exhibits the
largest latencies due to the least flexibility, while the centralized
tuning Schemes D and E show the smallest latencies because
of full wavelength tunability. The distributed tuning Schemes B
and C significantly outperform Scheme A when only half of the
users are equipped with tunable devices. Fig. 11(b) illustrates
the dependence of tunable and fixed ONU performances on
the number of tunable devices in the network. Interestingly,
as more users upgrade to the tunable devices, even the fixed
ONUs benefit from the added tunability in the network.
IV. PRACTICAL ISSUES
A. Broadcast Support
The large-scale deployment of full-service optical access networks will come to reality if its cost can be low enough, it can

enhance revenues, and reduce operating costs [16]. Delivery of

digital video or multimedia services to the end users could be
one of the driving forces for PONs. As high-speed access technologies become available to most end users, it is also expected
that high-speed multicast applications will evolve quickly in
most network environments. Therefore, in the physical layer architecture, it will be beneficial to incorporate broadcast or multicast abilities.
TDM-PONs broadcast data downstream by design, and thus
it is straightforward to realize broadcast when necessary. On the
contrary, a full WDM-PON does not inherently support multicast or broadcast in the physical layer. For the SUCCESS-DWA
PON, broadcast can be realized by having two passbands on
the user WDM filters at the ONUs, as shown in Fig. 12. One
passband is the unique user channel for receiving regular downstream traffic, and the other passband, common to all users, is a
broadcast channel (BC).
For the illustrative architecture in Fig. 1, whenever the broadcast is activated on a certain PON, e.g., TL1 is transmitting on
AWG channel 65, which is routed to PON1 and falls into the
BC, all users on PON1 receive this signal while users on the
other PONs are not affected. Therefore, we can realize broadcast in a PON-by-PON sense. Note that whenever broadcast is
activated on a certain PON, all users on that PON are subject to
the broadcast signal and cannot receive the regular downstream
traffic. This is because an extremely cost-sensitive ONU would
utilize only one photodetectorthe normal and broadcast transmissions would have to share in the time domain. If instead,
ONUs were fitted with two receivers, one dedicated to broadcast traffic, this constraint could be lifted.
B. Wavelength Plan
The TLs in the OLT can be accurately controlled in wavelength to allow DWDM in the downstream transmission. For the
SUCCESS-DWA PON, consider a
AWG with 100 GHz
channels and 16 user WDM channels. Together with the BC,
the total wavelength range can be estimated as
. Commercially available [17] fast TLs for
telecommunications have output wavelengths falling within the
C-band (15251562 nm) and L-band (15701615 nm), with a
60 nm. There are many possible ways to
tuning range of
allow more user channels or more TLs, as shown in Fig. 13.
With a finer spacing of the AWG channels, more user channels
or more TLs can be added into the system. A tradeoff lies between the number of channels and the maximum modulation

284

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 23, NO. 1, JANUARY 2005

eight corresponding wavelengths are of interest, one trivial

approach employs 10 nm-spaced channels with each occupying
half of the ITU CWDM channel as in Fig. 14(b). Nonstandard light sources and WDM filters may be required. For the
centralized SUCCESS-DWA upstream schemes, since each
user needs a unique upstream wavelength, cost-effective and
narrow-linewidth VCSELs can serve as the upstream light
sources. However, wavelength control may be necessary to
achieve the required accuracy illustrated in Fig. 14(c). This can
impose extra costs, favoring the distributed SUCCESS-DWA
upstream schemes over the centralized ones.
C. Implementation Issues

Wavelength plan in the downstream (a) 4 2 16 2 100 GHz, (b) 4 2

2 50 GHz, (c) 8 2 16 2 50 GHz, and (d) 8 2 32 2 25 GHz.

Fig. 13.

32

Fig. 14. Wavelength plan in the upstream: (a) distributed upstream schemes
with four wavelengths falling on ITU CWDM channels, (b) distributed upstream
schemes with eight wavelengths, and (c) centralized upstream scheme with 16
wavelengths.

bandwidth. In addition, the quality and corresponding cost of

the required filters may make some scenarios suboptimal.
In the SUCCESS-DWA PON, only one set of user WDM
channels is defined, which is reused in every physical PON.
It is worth noting that the number of user channels is not a
hard limit for the number of end users. A system with sixteen
user channels can accommodate more than 16 end users on
each physical PON if necessary. In this case, some users may
be assigned user channels that are already occupied by other
users on the same PON and must therefore time-share the
bandwidth on that channel. The scheduling algorithm in the
OLT is responsible for managing the traffic flow for fair access
of end users.
The laser wavelength control in the ONUs is more challenging than in the OLT, and CWDM is more suitable for
upstream transmission. For the distributed SUCCESS-DWA
upstream schemes with four PDs in Fig. 3(b), four wavelengths corresponding to the ITU CWDM channels in the
O-band (12601360 nm) can be utilized, which have a channel
spacing of 20 nm as shown in Fig. 14(a). If eight PDs and

Implementation of the SUCCESS-DWA PON protocols requires additional considerations including: laser tuning time,
burst-mode preamble sequence, and additional guard time. The
penalty resulting from this overhead heavily depends on the design of the TL, the burst-mode subsystems, and the specific implementation of the OLT queuing manager. For example, concatenating the packets into larger groups before transmitting
causes only marginal penalties in latency but easily reduces the
overhead to 1% or less.
The scheduling algorithm of the SUCCESS-DWA PON
needs extra consideration to handle multiple transmitters.
In the downstream, at most one TL can address a particular
user at any instant in time. The scarce/shared resource in this
architecture is the TL, since the number of TLs is typically less
than the number of end users. The scheduler must maintain
fairness and avoid collisions while maximizing the use of the
TLs. One possible approach employs virtual output queues to
avoid head-of-line blocking, and longest-queue-first scheme
or other more sophisticated schemes to determine the order
in which to service the queues. Since many TLs service the
same set of queues, a suitable arbiter must be chosen to ensure
that no two TLs attempt to service the same queue. Virtual
output queues inherently avoid infrastructure collisions in this
architecture. However, scalability can become a significant
issue when the number of users is large. Another approach
employs queuing at the transmitters. The incoming packets are
assigned to the transmitter with the shortest queue to minimize
delay, for example. To ensure that no two packets headed to
the same end user overlap across the different TL queues, the
scheduler can keep track of the destinations of the last packets
in the queues and assign the incoming packet to the TL queue
if they have the same destinations. Quality of service can also
be achieved by prioritizing the packets when assigning to
the queues. In general, a multiple transmitter system allows
more flexibility to tailor the network performance than the
conventional TDM-PON.
In the upstream, the scheduling algorithm should also prevent packet collisions in each of the upstream wavelengths. As
in conventional TDM PONs, the ranging process is necessary to
measure the propagation delays from the ONUs to the OLT, and
the OLT allocates nonoverlapping time slots to the ONUs on
a per-wavelength basis. Dynamic bandwidth allocation (DBA)
schemes can be employed to maximize the throughput. If multiple lasers in a single ONU (see upstream scheme B) are permitted to transmit at the same time, the ONU can run at full

HSUEH et al.: HIGHLY FLEXIBLE AND EFFICIENT PON EMPLOYING DYNAMIC WAVELENGTH ALLOCATION

speed when all its lasers are turned on. This allows the OLT
more flexibility to allocate bandwidth among ONUs thereby accommodate very bursty Internet traffic.
TLs have been considered as light sources in optical access
networks for several years [18]. As technology progresses,
fast TLs with tuning times within 2030 ns [19] have been
experimentally demonstrated and can serve as the downstream
light sources in SUCCESS-DWA PONs. On the other hand,
active research efforts are focusing on tunable VCSELs [20],
[21] whose potential low cost and high reliability would make
them promising candidates for deployment in ONUs.
High-speed burst-mode receivers are one of the crucial components in optical packet-switched networks. Many research efforts continue in this field and promising achievements have
been demonstrated [22][24]. Motivated by packet-switched applications, one can expect monolithic burst-mode receivers to be
commercially available with costs low enough for large-scale
deployment. Recall that the existing TDM-PONs require burstmode receivers in the upstream.
In its basic form, the SUCCESS-DWA PON requires approximately 68 dB more additional power (AWG + up/downfilter
+ user filter) than a traditional TDM PON in the downstream.
This additional power can be supported by lasers with higher
output powers, optical amplifiers, or forward-error-correction
(FEC) codes. In the upstream, the extra power loss is due to the
insertion loss of the DeMux, for which the same techniques can
be applied.
V. CONCLUSION
We propose a novel PON that employs dynamic wavelength allocation to efficiently provide services across several
physical PONs. The design is extremely flexible in terms of
capacity and immune to uneven user distributions across PONs.
Compared to WDM-PONs, the SUCCESS-DWA PON enjoys
flexibility, resource reallocation, and potentially lower costs.
The SUCCESS-DWA PON also provides a gradual and costeffective means to scale capacity as demand increases. The
SUCCESS-DWA PON is designed to move bandwidth freely
among multiple physical PONs, potentially greatly improving
performance. The network can be scaled from one TL per
PONs all the way toward
TLs per one PON, where
is
the number of AWG input/output ports. In addition, multiple
physical PONs enjoy shared protection against equipment
failures at the OLT. Judging from user demands, the fixed transmitters at the ONUs can be upgraded to tunable transmitters
for better performance. It is demonstrated that even the fixed
ONUs benefit from the upgrade. With excellent scalability, the
SUCCESS-DWA PON can be configured to span the range of
capacities between conventional TDM-PONs and full WDM
PONs. Broadcast can be supported with the broadcast channel
in the WDM filters. DWDM in the downstream and CWDM in
the upstream constitute a reasonable compromise between performance and cost. The field-deployed physical infrastructure
is kept untouched and need not be a specific topology. The powerful architecture is a promising candidate for next generation
optical access networks. A testbed is currently on-going for the
experimental demonstration of the SUCCESS-DWA PON.

285

REFERENCES
[1] S.-J. Park, S. Kim, K.-H. Song, and J.-R. Lee, DWDM-based FTTC
access network, J. Lightw. Technol., vol. 19, pp. 18511855, Dec. 2001.
[2] G. Maier, M. Martinelli, A. Pattavina, and E. Salvadori, Design and cost
performance of the multistage WDM-PON access networks, J. Lightw.
Technol., vol. 18, pp. 125143, Feb. 2000.
[3] R. D. Feldman, E. E. Harstead, S. Jiang, T. H. Wood, and M. Zirngibl,
An evaluation of architectures incorporating wavelength division multiplexing for broad-band fiber access, J. Lightw. Technol., vol. 16, pp.
15461559, Sep. 1998.
[4] M. S. Rogge, Y.-L. Hsueh, and L. G. Kazovsky, A novel passive optical
network with dynamic wavelength allocation, in Proc. OFC Tech. Dig.,
2004, Paper FG1.
[5] N. Vicari and S. Khler, Measuring Internet user traffic behavior dependent on access speed, in Proc. ITC Specialist Seminar on IP Traffic
Measurement, Modeling and Management, 2000, pp. 5-15-7.
[6] K. Yu and O. Solgaard, MEMS optical wavelength deinterleaver with
continuously variable channel spacing and center wavelength, IEEE
Photon. Technol. Lett., vol. 15, pp. 425427, Mar. 2003.
[7] L. Domash, E. Ma, N. Nemchuk, A. Payne, and M. Wu, Tunable thin
film filters, in OFC Tech. Dig., 2003, pp. 522524.
[8] W. Shin and K. Oh, The broadband wavelength selective cross-connect
switch based on a fused optical fiber coupler, in Proc. CLEO Tech. Dig.,
2002, pp. 79.
[9] S. Floyd and V. Paxson, Difficulties in simulating the Internet,
IEEE/ACM Trans. Netw., vol. 9, pp. 392403, Aug. 2001.
[10] J. W. Roberts, Traffic theory and the Internet, IEEE Commun. Mag.,
vol. 39, pp. 9499, Jan. 2001.
[11] A. Karasaridis and D. Hatzinakos, Network heavy traffic modeling
using -stable self-similar processes, IEEE Trans. Commun., vol. 49,
pp. 12031214, Jul. 2001.
[12] L. Muscariello, M. Mellia, M. Meo, R. L. Cigno, and M. A. Marsan, A
Simple Markovian Approach to Model Internet Traffic at Edge Routers.
[13] S. H. Kang and D. K. Sung, Real time cell loss estimation for ATM
multiplexers with heterogeneous ON/OFF sources, in Proc. ICC97,
Jun. 1997, pp. 16791683.
[14] W. Willinger, M. S. Taqqu, R. Sherman, and D. V. Wilson, Self-similar
through high-variability: statistical analysis of Ethernet LAN traffic at
the source level, IEEE/ACM Trans. Networking, vol. 5, no. 1, pp. 7186,
Feb. 1997.
[15] A. Klemm, C. Lindemann, and M. Lohmann, Traffic modeling of IP
networks using the batch Markovian arrival process, in TOOLS2002,
London, U.K., Apr. 2002.
[16] N. Chand, P. D. Magill, S. V. Swaminathan, and T. H. Daugherty, Delivery of digital video and other multimedia services (>1 Gb/s bandwidth) in passband above the 155 Mb/s baseband services on a FTTx
full service access network, J. Lightw. Technol., vol. 17, pp. 24492460,
Dec. 1999.
[17] B. Mason, J. Barton, G. A. Fish, L. A. Coldren, and S. P. DenBaars,
Design of sampled grating DBR lasers with integrated semiconductor
optical amplifiers, IEEE Photon. Technol. Lett., vol. 12, pp. 762764,
Jul. 2000.
[18] N. J. Frigo, P. P. Iannono, P. D. Magill, T. E. Darcie, M. M. Downs,
B. N. Desai, U. Koren, T. L. Koch, C. Dragone, H. M. Presby, and G.
E. Bodeep, A wavelength-division multiplexed passive optical network
with cost-shared components, IEEE Photon. Technol. Lett., vol. 6, pp.
13651367, Nov. 1994.
[19] K. Shrikhande, I. M. White, M. S. Rogge, F.-T. An, E. S. Hu, S. S.-H.
Yam, and L. G. Kazovsky, Performance demonstration of a fast-tunable
transmitter and burst-mode packet receiver for HORNET, in Proc. OFC
Tech. Dig., vol. 4, 2001, pp. ThG2-1ThG2-3.
[20] L. Chrostowski, C.-H. Chang, R. Stone, and C. J. Chang-Hasnain, Uncooled injection-locked 1.55  tunable VCSEL as WDM transmitter,
in Proc. OFC Tech. Dig., 2003, pp. 753754.
[21] D. Vakhshoori, J.-H. Zhou, M. Jiang, M. Azimi, K. McCallion, C.-C.
Lu, K. J. Knopp, J. Cai, P. D. Wang, P. Tayebati, H. Zhu, and P. Chen,
C-band tunable 6 mW vertical-cavity surface-emitting lasers, in Proc.
OFC Tech. Dig., 2000, pp. PD13-1PD13-3.
[22] Y. Yamada, S. Mino, and K. Habara, Ultra-fast clock recovery for burstmode optical packet communication, in Proc. OFC Tech. Dig., 1999,
pp. 114116.
[23] K. Murata and T. Otsuji, A novel clock recovery circuit for fully monolithic integration, IEEE Trans. Microwave Theory Tech., vol. 47, pp.
25282533, Dec. 1999.

286

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 23, NO. 1, JANUARY 2005

[24] S. Brigati, P. Colombara, L. DAscoli, U. Gatti, T. Kerekes, and P.

Malcovati, A SiGe BiCMOS burst-mode 155-Mb/s receiver for PON,
IEEE J. Solid-State Circuits, vol. 37, pp. 887894, Jul. 2002.

Yu-Li Hsueh (S03) received the B.S. degree in electrical engineering and the M.S. degree in electrooptical engineering from National Taiwan University,
Taipei, R.O.C., in 1996 and 1998, respectively. He is
currently pursuing the Ph.D. degree in electrical engineering at Stanford University, Stanford, CA.
His research interests include optical access networks, photonic crystal fibers, and optical waveguide
modeling.

Shu Yamamoto (SM95) received the B.S., M.E.,

and Ph.D. degrees in electronics engineering from
the University of Tokyo, Tokyo, Japan, in 1977,
1979, and 1989, respectively, and the M.S.E.E.
degree from California Institute of Technology,
Pasadena, CA, in 1983.
He joined Kokusai Denshin Denwa (KDD) Company, Limited (currently KDDI Corporation), Tokyo,
Japan, in 1979. He has worked on optical fiber
submarine cable systems and optical networking
at KDDI Research and Development Laboratories.
From 2000 to 2004, he was with KDDI Laboratories USA, Inc., Palo Alto, CA.
In 2004, he returned to the KDDI R&D Laboratories, Inc., Saitama, Japan.

Matthew S. Rogge (S03) received the B.S. degree

in electrical and computer engineering from the University of Missouri at Columbia in December 1997.
Since then, he has been at Stanford University, Stanford, CA, working toward the M.S. and Ph.D. degrees
in electrical engineering.
He has worked on optical networking projects in
the metropolitan and access spaces. He contributed
to the HORNET project, a metropolitan area network
that employs tunable transmitters to achieve efficient,
scalable packet transport and is currently working on
optical access networks.

Leonid G. Kazovsky (M80SM83F91) received

the M.S. and Ph.D. degrees in electrical engineering
from the Leningrad Electrotechnical Institute of
Communications in 1969 and 1972, respectively.
From 1974 to 1984, he taught and was engaged
in research at Israeli and United States universities.
From 1984 to 1990, he was with Bellcore, Red
Bank, NJ, researching coherent and WDM optical
fiber communications systems. In 1990, he joined
Stanford University as Professor of Electrical Engineering and founded the Photonics and Networking
Research Laboratory.