on the co-existence of 10g-epons and wdm-pons: a scheduling and bandwidth allocation approach

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 29, NO. 10, MAY 15, 2011 1417

On the Co-Existence of 10G-EPONs andWDM-PONs: A Scheduling and Bandwidth

Allocation ApproachMohammad S. Kiaei, Chadi Assi, Senior Member, IEEE, Lehan Meng, and Martin Maier, Member, IEEE

Abstract—Dynamic bandwidth allocation and grant schedulingare important factors when designing passive optical networks. Inthis paper, we investigate the problem of optimal scheduling andbandwidth allocation in next generation 10G-EPON coexistingwith 1G WDM-PONs. We propose a network architecture forproviding this coexistence and derive an ILP model for offlinejoint scheduling and bandwidth assignment for 10G-TDM and1G-WDM ONUs. To address the scalability of the ILP model, weintroduce a Tabu Search based heuristic for efficiently obtainingnear optimal solutions. We further explore the tradeoff whichexists in terms of delay, scheduling length, and channel utilization,when separate or the same DBA modules are used for 1G- and10G-ONUs.

Index Terms—DBA, optimization, scheduling, WDM-PONs.

I. INTRODUCTION

D URING the past decade, various standards have been de-veloped for realizing passive optical networks (PONs).

Given that 90% of the traffic is in the form of Ethernet frames,it is desirable to have an Ethernet based PON structure in orderto reduce the adaptation required for exchanging data betweenthe LAN and the access network [1]. Accordingly, the IEEE hasdeveloped the 802.3ah Ethernet PON (EPON), also referred toas 1G-EPON standard for carrying traffic in the form of Ethernetframes. An EPON consists of one optical line terminal (OLT) lo-cated at the central office of the service provider which is con-nected to several optical network units (ONUs). Recently, the10G-EPON has emerged as a promising candidate for next-gen-eration high data rate access systems [2] and is expected to pro-vide ten-fold data rates of 10 Gb/s. This new PON has beenstandardized under the IEEE 802.3av task force with the aimof developing the physical layer specification and managementparameters. The 10G-EPON standard provides symmetric 10Gb/s downstream and upstream, as well as asymmetric 10 Gb/sdownstream and 1 Gb/s upstream data rates. In order to pro-vide backward compatibility with the existing and widely de-ployed 1G-EPON, the OLT in a 10G-EPON is equipped with

Manuscript received October 18, 2010; revised January 05, 2011; acceptedFebruary 02, 2011. Date of publication February 28, 2011; date of current ver-sion April 29, 2011.M. S. Kiaei and C. Assi are with the Department of Engineering and Com-

puter Science, Concordia University, Montreal, QC H3G 1M8, Canada.L. Meng is with the Department of Computer Science, University of Wa-

terloo, Waterloo, ON N2L 3G1, Canada.M. Maier is with the Optical Zeitgeist Laboratory, Institut National de la

Recherche Scientifique (INRS), Montreal, QC H5A 1K6, Canada.Color versions of one or more of the figures in this paper are available online

at http://ieeexplore.ieee.org.Digital Object Identifier 10.1109/JLT.2011.2121054

dual rate receivers for receiving data from 1G and 10G-ONUs.Furthermore, the downstream transmission channels are sepa-rated for sending downstream data and control traffic to 1G- and10G-ONUs.Another alternative to cope with the increasing bandwidth

demand is to use WDM PONs; however, these systems havenot been standardized yet. A WDM-PON can support multiplewavelengths in either or both upstream and downstream direc-tions. WDM-PONs can also be combined with TDM techniquesused by the EPON standard in order to further reduce costs andachieve larger bandwidth support. This combination leads to hy-brid WDM-TDM PONs, which improve scalability by allowingsplitting ratios of up to 1:1000, at the expense of optical ampli-fiers [3].In PON systems, resource management and allocation is

carried out by the process of dynamic bandwidth alloca-tion (DBA), which normally consists of two subproblems:grant sizing and grant scheduling. Grant sizing determinesthe length of the transmission window assigned to an ONUper each polling cycle, while grant scheduling indicates theorder of ONU grants during a given cycle [1]. Several grantsizing and scheduling methods have been introduced in theliterature for 1G WDM/TDM PONs. A straightforward sched-uling method for WDM PONs is the online Next AvailableSupported Channel (NASC), where the OLT schedules theupstream transmission of an ONU on the earliest availablewavelength channel supported by the ONU. Alternatively, inoffline scheduling, the OLT makes the schedule after receivingbandwidth requests from all ONUs [4]. In this work, we in-vestigate the problem of optimal scheduling and bandwidthallocation in next generation 10G-EPON coexisting with 1GWDM-PONs. We first propose a network architecture for sup-porting the coexistence. Then, we derive an ILP model for of-fline joint scheduling and bandwidth assignment for 10G-TDMand 1G-WDM ONUs. Our goal is to develop efficient band-width allocation and scheduling algorithms in this system withmulti-rate ONUs. Based on the choice of wavelength chan-nels, the OLT may use separate or the same DBA modules for1G- and 10G-PONs. To address this, we study two schedulingscenarios where the 10G TDM channel is either shared be-tween 1G- and 10G-ONUs, or it is dedicated to 10G-ONUs.We exploit the tradeoff which exists in terms of delay, sched-uling length, and channel utilization, when separate or thesame DBA modules are used for 1G- and 10G-ONUs. Weintroduce a Tabu Search based heuristic for obtaining nearoptimal solutions in remarkably shorter computation times.The rest of this paper is structured as follows. In Section II,

0733-8724/$26.00 © 2011 IEEE

1418 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 29, NO. 10, MAY 15, 2011

we review the relevant existing work. We address the majorspecifications of the new 10G-EPON standard in Section III.The network architecture is given in Section IV. We presentin Section V the problem statement and motivate the work.We discuss about the bandwidth allocation, delay analysis andpresent the ILP models for joint scheduling and bandwidthallocation in Section VI. The Tabu heuristic is explained inSection VII, followed by the numerical results in Section VIII.Finally, we conclude our paper in Section IX.

II. RELATED WORK

Substantial work exists for addressing the problem of dy-namic bandwidth allocation in EPON. One of the interesting al-gorithms for this purpose is the interleaved polling with adaptivecycle time (IPACT) [5] for single-channel EPON and extendedin [6] for multichannel networks. Another grant scheduling ap-proach is proposed in [7] where the under-loaded ONUs are im-mediately scheduled using the first available wavelength whilehighly loaded ONUs are deferred until the arrival of all RE-PORT messages. In [8], a control plane was presented for next-generation multichannel access networks which provides a flex-ible upstream wavelength allocation. This control plane oper-ates at two time-scales; one is the microscopic time-scale whichdeals with traditional packet access, the other one is the macro-scopic time-scale which assigns connections to upstream opticalchannels with the objective of optimizing network utilization.In [4], the grant scheduling problem is considered for the evo-lutionary upgrade of WDM PONs where an evolving numberof transmission channels are available. There, the offline sched-uling problem is formulated as an optimization problem and anILP model is presented based on the parallel machine sched-uling model given in [9]. An efficient grant scheduling methodfor a multichannel EPON has been introduced in [10] based onthe so-called Just-in-Time (JIT) scheduling framework. Morerecently, the authors of [11] presented several online schedulingapproaches for WDM PONs with the objective of reducing theidle gaps on each wavelength channel. The authors focused ononline scheduling of colorless ONUs, and the size of grant mes-sages is determined in advance based on a simple gated ap-proach. A hybrid TDM/WDM PON architecture has also beenintroduced in [12] by means of free spectral range (FSR) peri-odicity of the arrayed waveguide grating (AWG). The requiredenabling technologies for realizing state of the art future (staticand dynamic) WDM-PON andWDM/TDM-PON systems havebeen presented in [13]. In [14], grant scheduling problem forSTARGATE multichannel EPON (SG-EPON) [15] is formu-lated as an open shop scheduling problem. In most of the pre-vious work, the two subproblems of grant sizing and schedulinghave been considered separately yielding to non-joint sched-uling and bandwidth allocation methods, where the allocatedbandwidth is determined in advance using a grant sizing tech-nique such as “limited service” or “gated service” [4]. On thecontrary [16], the authors proposed joint scheduling and band-width allocation methods for WDM-PON. Similar to [4], themethods presented in [16] are based on the assumption thateach ONU cannot transmit its upstream data on more than onechannel simultaneously.

Fig. 1. Waveband allocation in 1G- and 10G-EPON. (a) 1G-EPON. (b) 10G-EPON.

III. IEEE 802.3 AV SPECIFICATIONS

The IEEE 802.3 av task force focused only on the physicallayer while maintaining complete backward compatibilitywith 1 Gb/s EPON equipment. Therefore, the MAC protocol of1G-EPON remains unchanged. Recall that in the IEEE 802.3 ahEPON (1G-EPON) standard, the so-called Multipoint ControlProtocol (MPCP) is implemented at the MAC layer. In MPCP,a GATE message is used by the OLT to convey information tothe ONU about the size of the allocated transmission windowand the schedule of its transmission and a REPORT messageis used by each ONU to transmit information to the OLTabout its queue occupancies. One major difference between1G- and 10G-EPON is that the latter supports both symmetric10 Gb/s downstream and upstream, and asymmetric 10 Gb/sdownstream and 1 Gb/s upstream data rates, while 1G-EPONprovides only a 1 Gb/s symmetric data rate. The line codingfor the 10G- and 1G-EPON is also different. The 64 B/66 Bline coding in 10G-EPON reduces the bit-to-baud overheadto 3%, compared to the 25% overhead in 1G-EPON which isincurred by 8 B/10 B line encoding. The burst signal format of10G-EPON is similar to that of 10G-EPON, except that the re-ceiver settling time of 10G-EPON is twice of that in 1G-EPON,i.e., 800 ns in 10G and 400 ns in 1G-EPON. The laser on/offtime and clock data recovery time of both standards are thesame (512 ns and 400 ns, respectively) [2]. The wavebandsutilized for upstream (US) and downstream (DS) transmissionsof 1G- and 10G-EPON standards are illustrated in Fig. 1.As can be seen in Fig. 1(a), 1G-EPON allocates a 100 nmwaveband centered at 1310 nm for upstream (US) transmissionand a 20 nm window centered at 1490 nm for downstream(DS) transmission. As shown in Fig. 1(b), the downstreamwavelength of 10G-EPON is allocated in a window between1575 and 1580 nm (with a typical value of 1577 nm), which isoutside of the analog RF video distribution band. Conversely,the upstream wavelength of 10G-EPON is allocated in a 20 nmwindow centered at 1270 nm which is completely covered by apart of the 1G-EPON upstream waveband.The backward compatibility requirement of the new and ex-

isting EPONs introduces several technical challenges and diffi-culties on the specification work such as a high power budgetexceeding 30 dB for symmetric 10 Gb/s transmission, conflictsin wavelength allocation, and dual-rate burst-mode operation

KIAEI et al.: ON THE CO-EXISTENCE OF 10G-EPONS AND WDM-PONS 1419

Fig. 2. Coexistence of 10G-TDM and 1G-WDM PONs.

at the OLT receiver [17]. Two main techniques are employedfor achieving the coexistence of 10G-EPON with 1G-EPON(and analog RF video distribution) systems: WDM overlay inthe downstream direction and a dual-rate burst-mode receiverin the upstream direction to support a dual-speed TDM. In thedownstream direction, since the wavelength bands are distinct,a WDM-overlay is a straightforward way to provide the coexis-tence with the legacy 1G-EPON. On the contrary, as depicted inFig. 1, the upstream wavelength band of 10G-EPON is in facta subset of the 1G-EPON waveband; hence a dual-rate burst-mode operation is the only remaining option to retain the coex-istence requirement by using dual-speed TDM [2]. For this pur-pose, the OLT is equipped with dual-rate mode receiver. More-over, the OLT provides three kinds of MAC instances for oper-ating on symmetric and asymmetric data rates; namely, the OLTsupports 1/1 Gb/s, 10/1 Gb/s, and 10/10 Gb/s MAC instances.

IV. NETWORK ARCHITECTURE

As illustrated in Fig. 2, our PON comprises one OLT con-necting in a tree topology to multiple 1G- and 10G-ONUs. Inorder to enjoy the benefits of multi-channel upgraded PON orhybrid WDM-TDM PONs, the upstream transmission wave-band should be split into multiple wavelengths. We note thatthe upstream waveband for 10G-EPON is too narrow to be splitinto multiple wavelengths, whereas the 100 nm waveband of1G-EPON can be more easily split into multiple channels forONU upstream transmissions. Therefore, we consider a dual-rate EPON architecture with 10G-TDM ONUs coexisting withfuture 1G-WDM ONUs. One of the most cost-effective tech-nologies for realizingWDMONUs is to utilize so-called “color-less ONUs” which are wavelength-independent, and make useof a reflective semiconductor optical amplifier (RSOA) at theONU for remote modulation of the upstream data. In this ap-proach, the OLT is equipped with laser diodes to send opticalcontinuous wave (CW) signals to the attached reflective ONUs,where the CW signal is modulated and sent back to the OLT;hence, no light source is required at the ONU. In our architec-ture, the OLT is equipped with an array of fixed-tuned receiversand fixed-tuned transmitters for receiving from and sending outdata to the ONUs. Two types of receivers are deployed at theOLT. One is denoted by which is used at one of theUS channels for receiving data from 1G-ONUs. Theother is the dual-rate receiver tuned to the center of10G-EPON US waveband ( nm) for receivingdata from 10G-ONUs and from those 1G-ONUs sharing theUS channel with 10G-ONUs. Each of the transmitters at the

OLT are either fixed tuned to one of wavelengthsfor sending CW signals to the reflective 1G-ONUs, or they aretuned to one of the wavelengths for sendingDS data and control traffic to 1G-ONUs. Also, there is a 10 Gb/stransmitter at the OLT fixed tuned to for transmitting DSdata to 10G-ONUs.The OLT provides three kinds of MAC instances; namely,

1/1 Gb/s, 10/1 Gb/s, and 10/10 Gb/s. The 10G-ONUs are TDMONUs working on and channels for their upstreamand downstream transmissions, respectively. As shown inFig. 2, a given 10G-ONU generates either a 1 or 10 Gb/s signal,depending on which one of the two specified transmit pathsis implemented at the ONU [17]. Conversely, the 1G-WDMONUs are equipped with an RSOA, which can be tuned to allthe existing upstream channels including for transmittingupstream data and control traffic to the OLT. This way, the1G-ONUs are capable of transmitting on all available channelsincluding the 10G channel; we do not however allow simulta-neous transmissions on multiple channels. Also, each 1G-ONUemploys an array of fixed-tuned receivers, each tuned to one ofthe wavelengths for receiving downstreamdata and control traffic from the OLT.

V. MOTIVATION AND PROBLEM STATEMENT

It is important to note that even though the upstream wave-band for 1G-EPON standard spans 1260 nm to 1360 nm, somenetwork operators may restrict the waveband of 1G-EPON usersnot to extend below 1300 nm in order to avoid inventory prob-lems [18]. Thus, the upstream coexistence can be achieved usingWDM. Whether the upstream waveband is restricted or not,10G-ONUs may or may not share their upstream channel with1G-WDM ONUs. When the 1G-ONUs and 10G-ONUs operateon the same channel and the dual-rate burst-mode is used, allONUs should be controlled by a single scheduler and DBAmodule at the OLT. Conversely, if the allocated wavebands for1G and 10G-ONUs are different, the OLT can deploy separateDBA and scheduling modules for 1G and 10G-ONUs. In thefollowing example, we illustrate these two scenarios and theireffects on the channel utilization and the length of the sched-uling period.We note that for a lightly loaded network, the online NASC

provides better scheduling solutions in terms of delay andpacket loss compared to offline scheduling [4]. Also, it wouldnot be reasonable to share the 10G channel with 1G ONUswhen the 10G ONUs are highly loaded. Therefore, we considera traffic scenario where the 10G-ONUs are lightly loaded and1G-ONUs have different load levels. We consider a networkwith 10 1G-WDM ONUs indexed by and 210G-TDM ONUs indexed by and . Thereare four 1G WDM wavelengths and one 10Gwavelength for 10G ONUs. The round-trip time (RTT)between each ONU and OLT is assumed to be 100 sec, whichcorresponds to a 10 km distance. The allocated time slots ofeach ONU are illustrated in Fig. 3. In Fig. 3(a), the 10G channelis dedicated to 10G-ONUs, which are arbitrated accordingto the online NASC method. In this case, the OLT polls the10G-ONUs every 100 sec and grants the requested bandwidth.The 1G-WDM ONUs are scheduled using the non-joint offline


Fig. 3. An illustrative example for comparing different scheduling frameworks. (a) Separate scheduling modules for 1G- and 10G-EPON. (b) Same schedulingmodule for 1G- and 10G-EPON. (c) Joint scheduling and bandwidth allocation for 1G- and 10G-EPON.

scheduling method presented in [16]. The initial gap repre-sents the inter-scheduling cycle gap (ISCG) which is mainlydetermined by the RTT of the first ONU scheduled on eachchannel [19]. We observe that almost 60% of the 10G channelis wasted, whereas this channel could have been utilized moreefficiently if it had been shared by 1G-ONUs. In Fig. 3(b), wesee that a more efficient schedule with shorter polling cycle (ormakespan) and higher channel utilization can be obtained whenthe 10G channel is shared with the 1G-ONUs. As the pollingcycle increases for 10G-ONUs, they will have larger bandwidthrequests compared to the online scheduling in Fig. 3(a). Conse-quently, the average packet delay will increase for 10G-ONUs.To mitigate this problem, we can further reduce the makespanusing joint scheduling and bandwidth allocation for 1G-ONUs[Fig. 3(c)]. The transmission window size of highly loaded1G-ONUs is reduced based on the minimum guaranteed band-width of each channel. We observe that in the joint method, thescheduling length and therefore the packet delay for 10G-ONUsare decreased at the expense of a larger delay for the 1G-ONUtransmissions. In summary, Fig. 3 illustrates a clear tradeoffbetween transmission delay, channel utilization and schedulingperiod when using different scheduling methods.

VI. JOINT SCHEDULING AND BANDWIDTH ALLOCATION

We assume that each 1G-ONU can not transmit on more thanone channel per cycle. This way, during each polling cycle, theOLT has to send only one CW signal to the WDM-ONU inorder to remotely modulate the upstream data; therefore, theplanning cost decreases compared to a scenario where the ONUtransmissions per cycle are allowed to be bifurcated into dif-ferent channels and the OLT has to send multiple CW signals toWDM-ONUs in each polling cycle.

A. Bandwidth Allocation

In order to determine the allocated bandwidth for each ONU,we should first determine the minimum guaranteed bandwidthon each wavelength. Namely, we should determine the min-imum guaranteed bandwidth for 10G ONUs on 10G channel

denoted by and that for 1G-ONUs on 1G and 10Gchannels denoted respectively by and .Formally, the value of the minimum guaranteed bandwidth isdetermined by the polling cycle length , which is in turndetermined by the scheduling algorithm. Therefore, to obtaina reasonable estimate for the minimum guaranteed bandwidth,we have to assume a typical value for the cycle time, e.g.,

msec. Considering the transmission windows on the10G channel, we find

(1)

where is the total number of 10G-TDM ONUs,is the effective data rate of 1G (10G) chan-

nels, and is the number of 1G-ONUs which are decided tobe scheduled on the 10G channel. We assume that a 10G-ONUcan transmit up to 10 times more bytes during a period thana 1G-ONU

(2)

Then, we obtain

(3)

The minimum guaranteed bandwidth for the rest of 1G-ONUson the 1G wavelength channels can be determined as follows:

(4)

where is the total number of 1G channels, and is thetotal number of 1G-WDM ONUs. Next, we determine the allo-cated bandwidth for each ONU. Let be the requested band-width and be the allocated bandwidth for . If the re-quested bandwidth is less than the minimum guaranteed band-width, then the whole request will be granted, i.e.,

(5)


Fig. 4. Dynamics of bandwidth allocation.

where is the minimum guaranteed bandwidth ofon channel that can be determined from one of the

expressions (2), (3), or (4) (it is obvious thatfor 10G-ONUs on 1G channels). Otherwise, if the requestedbandwidth is greater than , the grant size of ONUswill be reduced to meet the following inequality:

(6)

B. Delay Analysis

To obtain a reasonable estimate on the expected transmissiondelay per each ONU, one needs to know the behavior of thebandwidth requests.We assume that each ONU request is gener-ated according to its instantaneous data rate on the upstreamchannel; that is, during each polling cycle, the ONU generatesa bandwidth request of . The delay analysis is illus-trated in Fig. 4. The upper axis illustrates the buffer occupancyof ONU , while the lower axis shows the data transmissionon the channel assigned to ONU . In the first cycle, the ONUtransmits the packets stored in its buffer and sends its request forthe next cycle inside a REPORT message. In the second cycle,the ONU transmits what has been requested (and allocated) inthe first cycle; meanwhile, the ONU buffer is being filled withnewly generated packets, which will be scheduled for transmis-sion in the third scheduling cycle. Therefore, the data generatedat time will be transmitted at , where is thestart time and is the length of the window allocated for ONUon the supported channel. If the ONU data request is small

enough to be transmitted in one cycle, then this would translateinto the maximum delay, i.e.,

(7)

Theorem: If the allocated bandwidth is less than the ONUrequest, the maximum packet delay can be obtained from

(8)

in which is the number of polling cycles elapsed until thebuffer of ONU is full.

Proof: If the granted bandwidth is less than the request,some packets will stay in the ONU’s buffer and will be trans-mitted in subsequent cycles. As shown in Fig. 4, the data whichis generated at time has to be transmitted at

. This leads to a delay of and the

ONU buffer keeps an amount of bits for the subsequenttransmission. Similarly, in the third cycle, bits willremain in the ONU buffer to be transmitted on the fourth cycle.In general, there will be a maximum delay of

in the th scheduling cycle, whereis accumulated in the ONU buffer. This accumulation continuesuntil cycles after which the buffer of ONU is full, and thesubsequent generated packets will be lost at the ONU. In otherwords

(9)

where is the buffer size of ONU . Therefore, the maximumdelay can be derived from (8).

C. ILP Model

In [4], the offline scheduling problem in a WDM-PON is for-mulated as a non-joint optimization problem, where the grantsizing is done in advance using a bandwidth allocation methodsuch as “limited service” or “gated service” [1]. The authorspresented an ILP model based on the scheduling theory, whereeach ONU is considered as a job, its grant size defines the pro-cessing time, and the channels used for transmission on the PONrepresent machines. Therefore, the problem reduces to a Par-allel Machine (PM) scheduling problem, where a set of jobs,with specific processing times, are executed on a set of ma-chines. Using the same concept, we derive an ILP model forjoint scheduling and bandwidth allocation in 10G-EPON coex-isting with 1G WDM-PON. Our only channel restriction is thatthe 10G-ONUs can only be granted on the 10G channel. In ourmodel the size of transmission window for ONU on channelis a variable which is determined inside the model along withthe schedule.SETS:

set of 1G-ONUsset of 10G-ONUs

set of all existing ONUsset of WDM 1G channelsset of all transmission channels

INPUT PARAMETERS:requested bandwidth of ONU

minimum guaranteed bandwidth for ONUon channel

maximum affordable delay per ONU


buffer size of ONUtransmission line rate for ONU (1Gb/s for

1G-ONUs and 10Gb/s for 10G-ONUs)

if channel is supported by ONUotherwise

OUTPUT VARIABLES:maximum completion time or the makespanlength of transmission window for ONU on

channel

if position on channel is selectedfor ONUotherwiseifotherwise.

if channel is assigned to ONUotherwise

Our objective is to minimize the maximum completion time

(10)

Themodel should guarantee that themakespan is not less thanthe completion time on any wavelength

(11)

In order to avoid quadratic terms, we define variables. Using these variables, we write

(12)

to assure that the makespan is greater than the completion time.Variables are determined in the following set of lineariza-tion constraints where is a large positive number

(13)

(14)

(15)

The following constraints determine the channel and time slotassignment of upstream bandwidth request of each ONU basedon the parallel machine model presented in [4]:

(16)

(17)

Constraint (16) ensures that each ONU is assigned to only onescheduling position and constraint (17) guarantees that eachscheduling position is assigned to no more than one ONU. Thefollowing equation indicates that, on each scheduling round,only one channel is assigned to each ONU:

(18)

By involving , (18) implies that the 10G-ONUs can not beallocated on 1G channels. The following equations are requiredfor bandwidth allocation as discussed in Section VI-A:

(19)

(20)

For all and , we have

(21)

Constraint (23) is for limiting the maximum packet delay pereach ONU. Note that (8) and (9) include nonlinear terms whichcan not be involved in our ILP model. Hence, to keep the modellinear, we approximate as

(22)

This approximation is based on the fact that in the joint model,the allocated bandwidth is desired to approach the minimumguaranteed bandwidth; after replacing with , we rewrite(9) in the form of an ILP constraint as follows:

(23)

where is the total scheduling length considering the initialgap on each cycle, i.e., . This constraintimplies that the delay for ONU is less than a predeterminedvalue , when ONU is scheduled on .We note that this model can be used for the case that the 10G

channel is dedicated to 10G-ONUs, as well as the case that itis shared with 1G-ONUs. The required changes are reflected inparameter that determines whether a 1G-ONU can transmiton the 10G channel or not. Accordingly, the value ofchanges depending on whether the 10G channel is shared ordedicated.

VII. TABU SEARCH HEURISTIC METHOD

Clearly, the ILP model developed in Section VI-C can onlybe solved for small network instances, as will be shown later.In addition, we have to make an approximation for the delayexpression in (8) in order to avoid nonlinearity in the model.Thus, it is vital to develop a heuristic in order to involve thenonlinear terms and obtain promising solutions with a toler-able computational cost. To this end, we develop a Tabu searchmethod for solving the joint scheduling and bandwidth alloca-tion problem. Our Tabu search heuristic starts from an initialfeasible solution and iterates using two types of moves for ob-taining the neighbor solution. One is reordering and moving theONU grants from one wavelength to another. In this move, theneighborhood of the current solution is obtained by moving atransmission window of an ONU from its assigned wavelengthto another (if there is any). Note that this move can not be ap-plied for the grants of 10G-ONUs, since they can only transmit


TABLE ILOAD DISTRIBUTION FOR 1G- AND 10G-ONUS IN DIFFERENT EXPERIMENTS

on the 10G channel. The other move is reducing the transmis-sion windows of ONUs whose bandwidth request is larger thanthe minimum guaranteed bandwidth. For these ONUs, we chosea grant size in the interval given by inequality (6) in a waythat the imposed delay in expression (8) is minimized. Usingthese two moves, our Tabu search algorithm performs a localsearch to explore new feasible solutions. In each iteration, bothmoves are assessed, and the one that yields the best result ischosen as the final move to be performed. Our Tabu methodalso makes use of a short term memory (Tabu list) that stores in-formation associated with recently explored solutions to avoidcycling. For example, the Tabu list contains the positions of theswapped grants, and any move that schedules an ONU grantback to its old position is considered Tabu (i.e., forbidden). TheTabu search algorithm needs some stopping criteria. One stop-ping criterion is to iterate for a certain number of iterations de-pending on the number of ONUs. Furthermore, we note that theutilization of different channels varies as the grants of ONUsare resized and reordered among different channels. Therefore,we consider another stopping criterion such that the algorithmruns until the last scheduled ONUs on all channels have thesame finishing time. This is equivalent to maximizing the av-erage bandwidth utilization (measured as the ratio of the sumof ONU transmission times to the total scheduling length of allchannels).

VIII. NUMERICAL RESULTS

We implemented the ILP models for the joint and non-jointscheduling and bandwidth allocation problem in C++, usingthe “CPLEX Concert Technology” and solved them using thesolver CPLEX 11.0.1 [20]. We consider different groups ofexperiments, by varying the number of 1G- and 10G-ONUs

as well as the number of available 1G chan-nels . In each group, we conduct various experimentsassuming that the bandwidth requirement of each ONU israndomly uniformly distributed over the intervals listed inTable I. For obtaining the minimum guaranteed bandwidthin expressions (3) and (4), we assume ; thatis, 20% of 1G-ONUs are assumed to be scheduled on the10G channel when it is shared between 1G- and 10G-ONUs.Obviously, when the 10G channel is dedicated to10G-ONUs. The rest of the network parameters are as fol-lows: transmission rates of 1G- and 10G-ONUs are 1 Gbpsand 10 Gbps. The buffer sizes are 1 MBytes and 10 MBytesrespectively. The guard bandwidth between adjacent slots is1.5 s and the Inter-Scheduling Cycle Gap (ISCG) of eachchannel is 110 s.

TABLE IIMAKESPAN ( SEC)

TABLE IIIAVERAGE CHANNEL UTILIZATION (PERCENTAGE)

A. Evaluation of Different Methods

In the first set of our experiments, we consider two networkscenarios; i.e., with 2 10G-ONUs, 8 1G-ONUs, and two 1GWDM channels and with 4 10G-ONUs, 16 1G-ONUs, andthree 1G WDM channels. We compare our joint ILP model andTabu heuristic with the non-joint ILP model presented in [16]which is a modified version of the model presented in [4]. Forthe non-joint model, we assume that the bandwidth allocation iscarried out in advance using the “gated” grant sizing technique.Each scheduling method is employed for two cases; i.e.,is either shared with 1G-ONUs or it is dedicated to 10G-ONUs.Tables II presents the obtained makespan of different methodsfor two network scenarios, when a separate or shared DBA isused; that is, is dedicated to 10G-ONUs or it is sharedwith 1G-ONUs. Similarly, the average channel utilization of dif-ferent methods are listed in Table III for different schedulingscenarios. The results for the joint ILP model could not be ob-tained for larger network instances and higher traffic loads, dueto high computational time (more than a few hours). First, weobserve that the Tabu outperforms the joint ILPmodel inmost ofthe cases. This is because the Tabu method can effectively deal


TABLE IVPERFORMANCE RESULTS

with the nonlinear term in (8), rather than making approxima-tion as in the ILP method, which results to underestimating themaximum delay. Second, we can clearly recognize the consis-tent improvement achievedwhen the 10G channel is shared with1G-ONUs. As expected, a shared DBA module yields higherchannel utilization and shorter scheduling periods, as ismore efficiently utilized by some of the 1G-ONUs.

B. Upgrading ONUS From 1 Gbps to 10 Gbps

We investigate the performance benefits obtained by the userswhen they undergo rate upgrade from 1G to 10G. We evaluatethe joint Tabu when the 1G- and 10G-ONUs are using the sameDBA module. We consider a network consisting of 6 1G WDMchannels and a total of 40 ONUs. We conduct 5 experimentswhere the 1G-ONUs are gradually being upgraded to 10G. Ascan be seen in Table IV, increasing the number of 10G-ONUsyields a smaller makespan and shorter maximum expected de-lays (measured as the ratio of the sum of maximum delay perONU obtained from (7) and (8) to the total number of ONUs).The reason is that when an ONU is upgraded to 10G, its re-quested bandwidth becomes much less than the minimum guar-anteed bandwidth of 10G-ONUs which is given by (2). There-fore, the upgraded ONUwould be allocated its whole request onthe 10G channel with a 10 times smaller transmission window.However, the average channel utilization decreases, since therequest of new upgraded 10G-ONUs can only be transmittedon 10G channel and therefore the utilization of 1G channels de-creases. This means that the network becomes under-utilizedand its potential bandwidth becomes wasted if the offered loadis not increased to catch upwith the line upgrade to 10G. In otherwords, to exploit the new capabilities of the upgraded system,more subscribers should be granted, resulting in higher loads.To better illustrate this issue, we conduct a set of experiments

where the load is gradually increased while the 1G-ONUs areupgraded to 10G. The obtained results are presented in Fig. 5.Experiment in Table I is evaluated for each of the sce-narios in Table IV. We note that ONU requests are generatedon a constant bit rate and the total load can be computed as thesum over the instantaneous bitrates of all ONUs. The overallload (measured in Gb/s) is given in Fig. 5(a) and the corre-sponding makespan, maximum delay, and channel utilizationare presented in Fig. 5(b)–(d), respectively. We observe that byincreasing the load from 13.5 Gb/s to 17 Gb/s, there is a slightincrease in the makespan for the upgraded system, while themaximum delay and average channel utilization stay in an ac-ceptable interval.

C. Packet Level Simulation

We carry out packet-level simulations to study the perfor-mance of the proposed scheduling methods in a tangible packet

Fig. 5. (a) Network traffic load (Gb/s). (b) Makespan ( s). (c) Expected max-imum delay (ms). (d) Average channel utilization.

level system. We simulated the operation of the schedulingmethods using OMNet++, a discrete event simulator.1 Weassume that the ONU loads are generated at a constant bitrate(CBR) based on their instantaneous data rates, and the packetsize is uniformly distributed between 64 and 1518 bytes. Thenumerical results are collected for experiments , and

in network instances and when is eithershared with 1G-ONUs or dedicated to 10G-ONUs. For eachexperiment, we take the scheduling solutions from non-jointILP [16], joint Tabu, and joint ILP as inputs to our simulator.The performance metrics are the average queuing delay andaverage packet loss, which are presented in Tables V andVI. The obtained results are in accordance with our previousdiscussions. As expected, sharing with 1G-ONUs can sig-nificantly improve the overall performance in terms of packetloss and average delay. Indeed, by sharing the 10G-channel,more bandwidth can be allocated to 1G-ONUs in shorterscheduling period. However, we should note that the min-imum guaranteed bandwidth of 10G-ONUs decreases as the10G-channel is shared with 1G-ONUs. Therefore, 10G-ONUsmay have larger drop rate and delay for higher traffic loads.Another conclusion is that our Tabu is a promising method forthe joint scheduling and grant sizing problem of 1G-WDM and10G-TDM ONUs; it provides close to optimal solutions withsignificantly smaller computational costs compared to the ILPmodels. In Fig. 6, we compare our joint Tabu method with theonline NASC and the offline scheduling method presented in[4], that is non-joint scheduling based on minimizing the totalcompletion time (we call this method NJS-TCT). The uppertwo figures show the average packet loss and the lower twofigures show the average queueing delay for two groups ofnetwork topologies. It can be recognized that our Tabu methodconsistently outperforms the offline method of [4] in terms ofpacket loss and average queueing delay. We also observe thatfor medium load experiment , our Tabu method exhibitsa better performance than the online NASC. It should also be

1[Online]. Available: http://www.omnetpp.org/


TABLE VAVERAGE QUEUING DELAY ( SEC)

TABLE VIAVERAGE PACKET LOSS RATE (PERCENTAGE)

Fig. 6. Comparison of our Tabu method with online NASC and NJS-TCT.(a) Group 1. (b) Group 2. (c) Group 1. (d) Group 2.

noted that the offline methods (either Tabu or ILP) have alwaysa higher channel utilization than online NASC.

IX. CONCLUSION

We studied the problem of optimal scheduling and band-width allocation in next generation 10G-EPON coexistingwith 1G WDM-PONs. We derived an ILP model for offlinejoint scheduling and bandwidth assignment for providing thiscoexistence. For large network instances, the size of the ILP

model becomes prohibitively large. Therefore, we introduceda Tabu Search heuristic to achieve near optimal solutions innotably shorter computation times. We explored the tradeoffwhich exists in terms of delay, scheduling length, and channelutilization, when separate or the same DBA and schedulingmodules are used for 1G- and 10G-ONUs. We also showd theinfluence of gradually upgrading the 1G-ONUs to 10G-ONUsin a network with a fixed number of ONUs. We conclude thatupgrading WDM 1G-ONUs to TDM 10G-ONUs can improvethe quality of service experienced by end users, yet it woulddecrease the channel utilization on the existing 1G channels.

REFERENCES[1] M. P. McGarry, M. Reisslein, and M. Maier, “Ethernet passive optical

network architectures and dynamic bandwidth allocation algorithms,”IEEE Commun. Surveys Tutorials, vol. 10, pp. 46–60, 2008.

[2] K. Tanaka, A. Agata, and Y. Horiuchi, “IEEE 802.3av 10G-EPONstandardization and its research and development status,” J. Lightw.Technol., vol. 28, no. 4, pp. 651–661, Feb. 2010.

[3] K. Grobe and J. P. Elbers, “PON in adolescence: From TDMA toWDMPON,” IEEE Commun. Mag., vol. 46, no. 1, pp. 26–34, Jan.2008.

[4] M. McGarry, M. Reisslein, M. Maier, and A. Keha, “Bandwidth man-agement for WDM EPONs,” OSA J. Opt. Netw., vol. 5, no. 9, pp.637–654, Sep. 2006.

[5] G. Kramer, B. Mukherjee, and G. Pesavento, “IPACT: A dynamic pro-tocol for an Ethernet PON (EPON),” IEEE Commun. Mag., vol. 40, no.2, pp. 74–80, Feb. 2002.

[6] K. Kwong, D. Harle, and A. Andonovic, “Dynamic bandwidth alloca-tion algorithm for differentiated services over WDM EPONs,” in Proc.9th ICCS, Singapore, Sep. 2004, pp. 116–120.

[7] A. Dhaini, C. Assi, M. Maier, and A. Shami, “Dynamic wavelengthand bandwidth allocation in hybrid TDM/WDM EPON networks,” J.Lightw. Technol., vol. 25, no. 1, pp. 277–286, Jan. 2007.

[8] M. Gagnaire and M. Koubaa, “A new control plane for next-gener-ation WDM-PON access systems,” in Proc., IEEE/ICST AccessNets,Ottawa, ON, Canada, Aug. 2007, pp. 116–120.

[9] M. Pinedo, Scheduling: Theory, Algorithms, and Systems, 3rd ed.Englewood Cliffs, NJ: Prentice-Hall, 2002.

[10] M. P. McGarry, M. Reisslein, C. J. Colbourn, M. Maier, F. Au-rzada, and M. Scheutzow, “Just-in-time scheduling for multichannelEPONs,” J. Lightw. Technol., vol. 26, no. 10, pp. 1204–1216, May2008.

[11] I. T. K. Kanonakis, “Improving the efficiency of online up-streamscheduling and wavelength assignment in hybrid WDM/TDMAEPON networks,” IEEE J. Sel. Areas Commun., vol. 28, no. 6, pp.838–848, Aug. 2010.

[12] J. P. C. Bock and S.Walker, “HybridWDM/TDMPON using the AWGFSR and featuring centralized light generation and dynamic bandwidthallocation,” J. Lightw. Technol., vol. 23, no. 12, pp. 3981–3988, Dec.2005.

[13] J.-I. Kani, “Enabling technologies for future scalable and flexibleWDM-PON and WDM/TDM-PON systems,” IEEE J. Sel. TopicsQuantum Electron., vol. 16, no. 5, pp. 1290–1297, Sep./Oct. 2010.

[14] L. Meng, J. El-Najjar, and C. Assi, “A joint transmission grantscheduling and wavelength assignment in multichannel SG-EPON,”J. Lightw. Technol., vol. 27, no. 22, pp. 4781–1492, Nov. 2010.

[15] M. Maier, M. Herzog, and M. Reisslein, “STARGATE: The next evo-lutionary step toward unleashing the potential ofWDMEPONs,” IEEECommun. Mag., vol. 45, no. 5, pp. 50–56, May 2007.

[16] M. S. Kiaei, L. Meng, C. Assi, and M. Maier, “Efficient schedulingand grant sizing methods for WDM PONs,” J. Lightw. Technol., vol.28, no. 13, pp. 1922–1931, Jul. 2010.

[17] M. Hajduczenia, H. J. A. D. Silva, and P. Monteiro, “Development of10 Gb/s EPON in IEEE 802.3av,” IEEE Commun. Mag., vol. 46, no. 7,pp. 40–47, Jul. 2008.

[18] G. Kramer, “10G-EPON: Drivers, challenges, and solutions,” in Proc.35th Eur. Conf. Opt. Commun. (ECOC’09), Vienna, Austria, Sep.2009, pp. 1–3.

[19] M. P. McGarry, M. Maier, and M. Reisslein, “WDM ethernet passiveoptical networks,” IEEE Commun. Mag., vol. 44, no. 2, pp. S18–S25,Feb. 2006.

[20] “ILOG CPLEX 11.0 User’s Manual,” IBM ILOG, Sep. 2007.


Mohammad S. Kiaei was born in Tehran, Iran, in 1977. He received the B.Sc.and M.Sc. degrees in electrical engineering from Sharif University of Tech-nology, Tehran, Iran, in 2000 and 2003, respectively. He is currently workingtoward the Ph.D. degree at Concordia University, Montreal, QC, Canada.His research interests include network survivability, optimization of optical

networks, and multimedia security.

Chadi Assi (SM’08) received the B.Eng. degree from the Lebanese University,Beirut, Lebanon, in 1997, and the Ph.D. degree from the Graduate Center, CityUniversity of New York, New York, in 2003.He is an Associate Professor with the Concordia Institute for Information

Systems Engineering, Concordia University, Montreal, QC, Canada. Prior tojoining Concordia University, he was a Visiting Researcher for one year withNokia Research Center, Boston, MA, working on optical access networks. Heis an associate editor forWireless Communications and Mobile Computing. Hiscurrent research interests are in the areas of optical networks, multi-hop wirelessand ad hoc networks, and security.Dr. Assi was the recipient of the prestigious Mina Rees Dissertation Award

from the City University of New York in August 2002 for his research onWDM networks. He is on the Editorial Board of the IEEE COMMUNICATIONSSURVEYS AND TUTORIALS and serves as an associate editor for the IEEECOMMUNICATIONS LETTERS.

Lehan Meng was born in China. He received the B.Eng. degree in computerengineering from the Xidian University, Xi’an, China, in 2004, and the M.Sc.degree from Concordia University, Montreal, QC, Canada, in 2009. He is cur-rently working toward the Ph.D. degree in computer science at the Universityof Waterloo, Waterloo, ON, Canada.His research includes metro-access network architecture design, network re-

sources management, network optimization, and scheduling algorithms.

Martin Maier (M’03) received the Dipl.-Ing. and Dr.-Ing. degrees (both withdistinctions) in electrical engineering from the Technical University of Berlin,Berlin, Germany, in 1998 and 2003, respectively.He is an Associate Professor with the Institut National de la Recherche Sci-

entifique (INRS), Montreal, QC, Canada. He was a Visiting Researcher with theUniversity of Southern California (USC), Los Angeles, and Arizona State Uni-versity (ASU), Tempe. In the summer of 2003, he was a Postdoctoral Fellowwith the Masschusetts Institute of Technology (MIT), Cambridge. He is the au-thor of the book,Metropolitan AreaWDMNetworks—An AWGBased Approach(Springer, 2003). His research interests include network and node architectures,routing and switching paradigms, protection, restoration, multicasting, and thedesign, performance evaluation, and optimization of MAC protocols for opticalWDMnetworks, Automatically Switched Optical Networks (ASONs), with par-ticular focus on metro and access networks. Recently, his research interests haveconcentrated on evolutionary WDM upgrades of optical metro ring networksand access networks.Dr. Maier was a recipient of the two-year Deutsche Telekom doctoral

scholarship from June 1999 through May 2001. He was also a corecipientof the Best Paper Award presented at SPIE Photonics East 2000-TerabitOptical Networking Conference. He serves on the Editorial Board of the IEEECOMMUNICATIONS SURVEYS AND TUTORIALS.

on the co-existence of 10g-epons and wdm-pons: a scheduling and bandwidth allocation approach

Documents