Optical Switching and Networking 6 (2009) 151–162

Contents lists available at ScienceDirect

Optical Switching and Networking

journal homepage: www.elsevier.com/locate/osn

A survey of dynamic bandwidth allocation algorithms for EthernetPassive Optical NetworksJun Zheng a,∗, Hussein T. Mouftah ba School of Information Science and Engineering, Southeast University, Nanjing, Jiangsu 210096, Chinab School of Information Technology and Engineering, University of Ottawa, Ottawa, Ontario K1N 6N5, Canada

a r t i c l e i n f o

Article history:Received 1 October 2008Received in revised form 20 February 2009Accepted 17 March 2009Available online 31 March 2009

Keywords:Dynamic bandwidth allocationEthernetPassive Optical Network

a b s t r a c t

Ethernet Passive Optical Network (EPON) has been widely considered as a promisingtechnology for implementing the FTTx solutions to the ‘‘last mile’’ bandwidth bottleneckproblem. Bandwidth allocation is one of the critical issues in the design of EPON systems. Inan EPON system,multiple optical network units (ONUs) share a common upstream channelfor data transmission. To efficiently utilize the limited bandwidth of the upstream channel,an EPON systemmust dynamically allocate the upstreambandwidth amongmultiple ONUsbased on the instantaneous bandwidth demands and quality of service requirements ofend users. This paper introduces the fundamental concepts on EPONs, discusses the majorissues related to bandwidth allocation in EPON systems, and presents a survey of the state-of-the-art dynamic bandwidth allocation (DBA) algorithms for EPONs.

© 2009 Elsevier B.V. All rights reserved.

1. Introduction

With the deployment of fiber optic technology over thepast decade, the telecommunication infrastructure has ex-perienced a tremendous growth in the bandwidth capacityof its backbone networks. This began with the wide areanetworks (WANs) that provide connectivity between citiesthrough the metropolitan area networks (MANs) that con-nect service providers’ central offices. However, subscriberaccess networks, which cover the ‘‘last mile’’ areas, andserve numerous residential and small business or organi-zation users, have not been scaled up commensurately. Thelocal subscriber lines for telephone and cable television arestill using twisted pairs and coaxial cables. Many residen-tial connections to the Internet are still through dial-upmodems operating at a low speed on twisted pairs. Withthe ever-increasing users’ demands for various broadbandapplications, such as Internet telephony, high-definition

∗ Corresponding address: School of Information Science and Engineer-ing, Southeast University, No. 2 Si Pai Lou, Nanjing, JiangSu 210096, China.Tel.: +1 6135625800x6243; fax: +1 6135625664.E-mail address: jzheng@ieee.org (J. Zheng).

1573-4277/$ – see front matter© 2009 Elsevier B.V. All rights reserved.doi:10.1016/j.osn.2009.03.003

television (HDTV), interactive games, and video on de-mand, the ‘‘last mile’’ segment has become a bandwidthbottleneck in today’s telecommunications infrastructure,which has largely limited the development of broadbandservices to subscriber users [1]. Although recent deploy-ment of innovative xDSL and CaTV technologies has sig-nificantly upgraded this segment, they are still insufficientfor meeting the ever-increasing bandwidth demand ofsubscriber users. To alleviate this bottleneck, fiber to thehome/curb/building (FTTH/FTTC/FTTB) technologies havebeen long envisioned as a preferred solution, and passiveoptical networks (PONs) have been widely considered as apromising technology for implementing various FTTx solu-tions.As one of the promising solutions, Ethernet Passive

Optical Network (EPON) has received great attentionfrom both industry and academia in recent years. EPONcombines low-cost Ethernet equipment and low-costpassive optical components and thus has a number ofadvantages over traditional access networks, such as largerbandwidth capacity, longer operating distance, lowerequipment and maintenance cost, and easier update tohigher bit rates [2]. In an EPON system, all Optical Network

152 J. Zheng, H.T. Mouftah / Optical Switching and Networking 6 (2009) 151–162

Units (ONUs) share a common upstream transmissionmedium with limited bandwidth. To efficiently utilizethe limited upstream bandwidth, an EPON system mustemploy a medium access control (MAC) mechanism toarbitrate the access to the shared medium in order toavoid data collisions in the upstream direction. For thispurpose, bandwidth allocation becomes one of the criticalissues in the design of an EPON system and a varietyof bandwidth allocation algorithms have been proposedin the literature. The purpose of this paper is to give anintroduction of the major issues in bandwidth allocationfor EPON systems and present a survey of the state-of-the-art dynamic bandwidth allocation (DBA) algorithmsproposed for EPONs.The remainder of the paper is organized as follows. In

Section 2, we introduce EPON architectures and discussrelated issues. In Section 3, we discuss the major issuesrelated to bandwidth allocation in EPON systems. InSection 4, we present a survey of the state-of-the-artdynamic bandwidth allocation (DBA) algorithms for EPONsystems. In Section 5, we conclude this paper.

2. EPON architectures

In this section, we introduce EPON architectures, anddiscuss the channel separation and multiple access issuesin EPON systems.

2.1. Network architectures

An EPON system is a point-to-multipoint fiber opticalnetwork with no active elements in the transmission pathfrom its source, i.e., an optical line terminal (OLT), to adestination, i.e., an optical network unit (ONU). It canuse different multipoint topologies, such as bus, ring, andtree [2], and different network architectures [2–5]. Themost typical EPON architecture is based on a tree topologyand consists of an OLT, a 1:N passive star coupler (orsplitter/combiner), and multiple ONUs, as shown in Fig. 1.The OLT resides in a central office (CO) that connects theaccess network to ametropolitan area network (MAN) or awide area network (WAN), and is connected to the passivestar coupler through a singe optical fiber. Each ONU islocated either at curbs or at subscriber premises, and isconnected to the passive coupler through a dedicated shortoptical fiber. The distance between the OLT and each ONUtypically ranges from 10 to 20 km. In an EPON system,all transmissions are performed between the OLT and theONUs. In the downstream direction, an EPON is a point-to-multipoint network, in which the OLT broadcasts datato each ONU through the 1:N splitter, where N is typicallybetween 4 and 64. Each ONU extracts the data destinedfor it based on its media access control (MAC) address. Inthe upstream direction, an EPON is a multipoint-to-pointnetwork, in which multiple ONUs transmit data to the OLTthrough the 1 : N passive combiner. The line data ratefrom an ONU to the OLT and the user access rate from auser to an ONU do not necessarily have to be equal and theline data rate is usually much higher than the user accessrate. Since all ONUs share the same upstream transmissionmedium with limited bandwidth, an EPON system must

OLT

ONU 1

ONU 21:N splitter/ combiner

Users

MAN/ WAN

CO

...

ONU N

Fig. 1. EPON architecture.

OLT

ONU 1

COSingle user

1:N splitter/ combiner

ONU 2

ONU 3Users

Users Sub-OLT 1

Sub-OLT 2

Sub-OLT 3

Fig. 2. Two-stage EPON architecture.

employ a MAC mechanism to arbitrate the access to theshared medium in order to avoid data collisions in theupstreamdirection and thus efficiently share the upstreamtransmission bandwidth among all ONUs.In [5], Shami et al. proposed a cascaded two-stage ar-

chitecture for an EPON, which introduces an intermedi-ate level of ONU nodes (called sub-OLT) to the network, asshown in Fig. 2. This architecture allows more end usersto share the upstream OLT bandwidth without incurringextra overhead for switch-over between end users. It alsoenables longer access reach/distances than the usual 25 kmbecause the intermediate sub-OLT nodes add another levelof electrical generation. Moreover, the introduction of anintermediate stage can help reduce theOLT hardware com-plexity significantly.

2.2. Channel separation

In an EPON system, the upstream and downstreamtransmission channels should be appropriately separatedin order to increase the transmission efficiency. A simplesolution is to use space division multiplexing, where twoseparate optical fibers and passive couplers are used, onefor upstream transmission and the other for downstreamtransmission. A more cost-effective solution is to use asingle coupler and a single fiber for both directions withone wavelength for upstream transmission and anotherfor downstream transmission. Typically, a 1550 nm wave-length is used for downstream transmission and a 1310 nmwavelength is used for upstream transmission [2].

2.3. Multiple access

In the upstream direction of an EPON system, multipleONUs transmit data packets to the OLT through a commonpassive combiner and share the same optical fiber fromthe combiner to the OLT. Due to the directional propertyof a passive combiner, data packets from an ONU can onlyreach the OLT but not the other ONUs. For this reason,

J. Zheng, H.T. Mouftah / Optical Switching and Networking 6 (2009) 151–162 153

OLT

•••

ONU 1

ONU 2 3:N

star coupler

ONU N1310 nm upstream signal

1310 nm looped-back signal

Isolator

Fig. 3. EPON using looping-back star coupler.

conventional contention-based multiple access, e.g., thecarrier sense multiple access with collision detection(CSMA/CD) protocol, is difficult to implement because theONUs are unable to easily detect a collision that may occurat the OLT. Although the OLT is able to detect a collisionand inform the ONUs by sending a collision message,the transmission efficiency would be largely reducedbecause of considerable propagation delay between theOLT and the ONUs. To address this problem, an opticallooping-back technique was proposed in [6] to achievehigh channel efficiency with CSMA/CD. With this looping-back technique, a portion of the upstream signal powertransmitted by each ONU is looped back to the otherONUs at the star coupler by using a 3 × N coupler andconnecting two ports of the coupler together throughan isolator, as shown in Fig. 3. If two or more ONUstransmit data simultaneously, collisions will be detectedat each ONU and all data transmissions will be stoppedimmediately. The optical CSMA/CD protocol is applied toall upstream transmissions [7]. The OLT will receive thedata packets transmitted by each ONU and will discardthose packets with collisions. However, to implementthe optical CSMA/CD protocol, each ONU has to use anadditional receiver operating at the upstream wavelengthand a carrier sensing circuit, which would largely increasethe network cost. On the other hand, contention-basedmultiple access is unable to provide guaranteed bandwidthto each ONU and thus is difficult to support any form ofquality of service (QoS). For these reasons, contention-based multiple access is currently not a preferred solutionto the upstream multiple access.Another possible solution is to use wavelength division

multiplexing (WDM) technology and allow each ONUto operate at a different wavelength, thus avoidinginterference with the transmissions of the other ONUs. AnEPON using such WDM technology is called WDM EPON.It requires either a tunable receiver or a receiver arrayat the OLT to receive the data transmitted in multiplechannels. In particular, it also requires each ONU to usea fixed transmitter operating at a different wavelength,which would result in an inventory problem. Althoughthe inventory problem can be solved by using tunabletransmitters, such devices are costly, making the solutionnot cost-effective.Comparedwith CSMA/CD andWDM, time divisionmul-

tiplexing (TDM) on a single wavelength is more attractivefor upstream transmission in an EPON system. With TDM,each ONU is allocated a timeslot or transmission window

for data transmission, which is performed by the OLT. Eachtimeslot is capable of carrying several Ethernet packets.Packets received from one or more users are buffered inan ONU until the timeslot for that ONU arrives. Upon thearrival of its timeslot, the ONU will send out its bufferedpackets at the full transmission rate of the upstream chan-nel. Accordingly, TDM avoids data collisions from differentONUs. Moreover, it requires only a single wavelength forall ONU transmissions and a single transceiver at the OLT,which is highly cost-effective. This paper focuses on band-width allocation in TDM EPONs.

3. Bandwidth management in EPONS

In this section, we discuss the major issues that arerelated to bandwidth management in EPON systems. Asmentioned in the previous section, an EPON system typi-cally employs TDM for data transmission in the upstreamdirection. For this reason, bandwidth management be-comes critical for efficiently utilizing the bandwidth of theshared upstream wavelength. Usually, bandwidth man-agement involves twomain issues: bandwidth negotiationand bandwidth allocation, which are discussed below.

3.1. Bandwidth negotiation

Bandwidth negotiation is to exchange informationbetween the OLT and each ONU in order for each ONUto report its bandwidth demand to the OLT and for theOLT to send its bandwidth allocation decision to eachONU. For this purpose, IEEE 802.3ah defines a multipointcontrol protocol (MPCP) to support bandwidth negotiationbetween the OLT and ONUs in EPON, including two64-bytes MAC control messages: REPORT and GATE. TheREPORT message is generated by each ONU to report itsqueue status to the OLT. The OLT allocates bandwidthfor each ONU based on the queue status informationcontained in the received REPORT message, and uses theGATE message to deliver its bandwidth allocation decisionto each ONU.One way to deliver the bandwidth request or REPORT

message is to dedicate a very short timeslot in theupstream channel. This requires twice laser on/off for oneupstream transmission from each ONU. Another way isto piggyback the bandwidth request or REPORT messageat the end of a data timeslot, which reduces laser on/offtimes into one per transmission for each ONU, and thusreduces the physical-layer power overhead and the inter-frame guard [8].

3.2. Bandwidth allocation

To allocate bandwidth (or a timeslot) for each ONU, theOLT needs to perform a bandwidth allocation algorithmbased on the bandwidth requests from each ONU as wellas some allocation policy and/or service level agreement(SLA). In this context, there aremany bandwidth allocationalgorithms proposed in the literature, which can beclassified into two broad categories: static bandwidthallocation (SBA) and dynamic bandwidth allocation (DBA).With SBA, each ONU is allocated a timeslot with a fixed

154 J. Zheng, H.T. Mouftah / Optical Switching and Networking 6 (2009) 151–162

length, which does not require bandwidth negotiationand is thus simple to implement. However, due to thebursty nature of the network traffic, it may result in asituation in which some timeslots are overflowed evenunder very light load, causing packets being delayedfor several timeslots, while other timeslots are notfully used even under very heavy traffic, leading tothe upstream bandwidth being underutilized. For thisreason, static allocation is not preferred. To increasebandwidth utilization, the OLTmust dynamically allocate avariable timeslot to each ONU based on the instantaneousbandwidth demand of the ONUs. To implement DBA,polling has been widely used [9]. With polling, the OLT candynamically allocate bandwidth for each ONU and flexiblyarbitrate the transmissions of multiple ONUs, which cansignificantly increase bandwidth utilization and improvenetwork performance.

3.3. Multipoint Control Protocol (MPCP)

MPCP [10] is a signaling protocol for facilitating dy-namic bandwidth allocation and arbitrating the transmis-sions of multiple ONUs in an EPON system. It resides at theMAC control layer and has two operation modes: normalmode and auto-discoverymode. In thenormalmode,MPCPrelies on two Ethernet controlmessages,GATE and REPORT,to allocate bandwidth to each ONU. The GATE message isused by the OLT to allocate a transmission window to anONU. The REPORT message is used by an ONU to report itslocal conditions to theOLT. In the auto-discoverymode, theprotocol relies on three control messages, REGISTER, REG-ISTER_REQUEST, and REGISTER_ACK, which are used to dis-cover and register a newly connected ONU, and to collectrelevant information about that ONU, such as the round-trip delay and MAC address.In its normal operation, MPCP in the OLT gets a request

from the higher MAC client layer to transmit a GATEmessage to a particular ONU. Upon getting such a request,MPCP will timestamp the GATE message with its localtime and then send the message to the ONU. The GATEmessage typically contains a granted start time, a grantedtransmission window, and a 4-byte timestamp, which isused to calculate the round-trip time between the OLTand the ONU. Once the ONU receives the GATE message,it programs its local register with the values containedin the GATE message. Meanwhile, it also updates its localclock to that of the timestamp extracted from the receivedGATE message in order to maintain synchronization withthe OLT. At the granted start time, the ONU will start totransmit data for up to the window size. The transmissionmay include multiple data packets, depending on thewindow size and the queue length in the ONU. No packetfragmentation is allowed during the transmission. If thenext packet cannot be transmitted in the current window,it will be deferred to the next window.A REPORT message is sent by an ONU in the allocated

transmission window together with a data packet. It canbe transmitted automatically or on demand either at thestart or at the end of a window. A REPORT is generatedat the MAC client layer and is timestamped at the MAClayer. It typically contains the bandwidth demand of an

RxTxRxTx

G1

G1

R1G2

RxTxRxTx R3

G3

R1

R2G3

R3

G2R2

ONU1

ONU2

ONU3

OLT

Guard time

Fig. 4. A flow of GATE and REPORT messages.

ONU based on the instantaneous queue length of thatONU. The ONU should also account for additional overheadin its request, including a 64-bit frame preamble and a96-bit inter-frame gap associated with each Ethernetpacket. Once a REPORT message is received by the OLT,it is passed to the MAC client layer, which is responsiblefor bandwidth allocation and recalculation of the round-trip delay to the source ONU. Fig. 4 illustrates a flow ofGATE (G) messages and REPORT (R) messages for upstreamtransmission of three ONUs.It should be pointed out that MPCP is not concerned

with any particular bandwidth allocation scheme andtransmission scheduling algorithm, and allows them tobe vendor-specific. To design an efficient polling protocolbased on MPCP, several problems must be considered,including maximum bandwidth limit, channel utilization,and packet scheduling.

3.4. Maximum bandwidth limitation

A polling protocol typically operates on a cycle-basedbasis. In each polling cycle, each ONU is polled once and isallocated a transmission window based on its bandwidthdemand. If the OLT allows each ONU to send all itsbuffered packets in one transmission, ONUs with hightraffic load may monopolize the entire bandwidth of theupstream channel. This is unfair to those ONUs with lowtraffic load. To address this problem, the OLT should limitthe maximum transmission bandwidth of each ONU. Themaximum window size can be either fixed based on somecriterion, such as a SLA, or variable based on instantaneousnetwork conditions. Under high traffic load, the maximumwindow size determines the maximum polling cycle. Ingeneral, making the maximum polling cycle too long willresult in larger delay for all packets under high traffic load,including those high-priority packets. On the other hand,making the maximum cycle too short will result in morebandwidth being wasted by inter-frame gaps (or guardtimes). Accordingly, themaximumwindow size has a greatimpact on network performance.While the maximum window size imposes a limit

on the maximum bandwidth that can be allocated toeach ONU in each polling cycle, it is also the guaranteedbandwidth available to each ONU. In fact, only whenall other ONUs use all their available bandwidth will anONU be limited to its guaranteed bandwidth. If any ONUrequests less bandwidth, it will be allocated a smallerwindow size, making the polling cycle shorter and thusincreasing the actual bandwidth available to all otherONUs.

J. Zheng, H.T. Mouftah / Optical Switching and Networking 6 (2009) 151–162 155

Rx

Tx

Rx

Tx

G1

G1

R1G2

Rx

Tx

Rx

Tx R3G3

R1

R2G3

R3

G2R2

ONU1

ONU2

ONU3

OLT

Rx

Tx

Rx

Tx

G1

G1

R1G2

Rx

Tx

Rx

Tx R3G3

R1

R2G3

R3

G2R2

ONU1

ONU2

ONU3

OLTG1

G1R1

R1G2

Rx

Tx

Rx

Tx

G1

G1

R1 G2

Rx

Tx

Rx

Tx R3G3

R1

R2 G3

R3

G2R2

ONU1

ONU2

ONU3

OLTG1

G1R1

R1

Computation time

G2

a

b

c

Fig. 5. Polling policies: (a) poll-and-stop polling; (b) interleaved polling;(c) interleaved polling with stop.

3.5. Polling strategies

A polling protocol can poll multiple ONUs for datatransmission based on different strategies [11]. A simplestrategy, called poll-and-stop polling, is to send a GATEmessage to an ONU and then stop for the data and REPORTmessage to come back from that ONU before the OLT sendsa GATE message to the next ONU, as shown in Fig. 5(a).Obviously, this protocol wastes a lot of bandwidth on theupstream channel, which would largely reduce channelutilization and increase packet delay.A more efficient strategy is to use the interleaved

polling [9], which allows the OLT to send a GATE messageto the next ONU before the data and REPORT message(s)from the previous polled ONU(s) arrive, as shown inFig. 5(b). This is feasible because the upstream channel anddownstream channel are separated, and the OLTmaintainsrelevant information about each ONU in a polling table,including the bandwidth demand of each ONU andthe round-trip time to each ONU. The results obtainedin [9] indicate that the interleaved polling protocol cansignificantly improve the network performance in termsof channel utilization and average packet delay. However,this protocol allows the OLT to allocate bandwidth onlybased on those already received bandwidth demands. TheOLT is unable to take into account the bandwidth demandsof all ONUs and make a more intelligent decision onbandwidth allocation.An effective way to overcome this shortcoming is to

use a variation of the interleaved polling, called interleavedpollingwith stop. Like the interleaved polling, this protocolallows the OLT to send a GATE message to the nextONU before the transmission and REPORT message(s) fromthe previous polled ONU(s) arrive. Unlike the interleaved

polling, the OLT does not start the next polling cycle beforethe transmissions and REPORT messages from all ONUsare received. This allows the OLT to perform bandwidthallocation based on the bandwidth demands of all ONUsat the end of each polling cycle and thus make a moreintelligent decision. However, such intelligence is obtainedat the cost of upstream channel utilization because theupstream channel is not utilized from the instant thetransmission of the last polled ONU in the previous cycle iscompleted to the instant the transmission of the first polledONU in the next cycle starts. Fig. 5(c) illustrates an exampleof the control message flows with the interleaved-polling-with-stop protocol.

3.6. Transmission scheduling

To ensure efficient transmission, a polling protocolmust schedule the transmissions of multiple ONUs ina manner that avoids data collisions from differentONUs [11]. This is not difficult to implement becausesuch scheduling is based on the granted window size andthe round-trip time to each ONU. Since the OLT knowsthe granted window size and the round-trip time to thelast polled ONU, it can calculate the transmission starttime and window size for the next ONU. Note that toallow the receiver in the OLT to prepare for receiving thetransmissions a minimum gap or guard time is usuallyrequired between the transmissions of different ONUs.On the other hand, the OLT must also be responsible

for scheduling the transmission order of different ONUs,which may have a great impact on network performance.This is not difficult to implement because the orderof the transmissions is usually determined one cycleahead by performing a scheduling algorithm. The mostwidely-used scheduling algorithm is round-robin (RR),which has been adopted by many polling protocols. RRschedules the transmissions of different ONUs in the orderof their indexes in the polling table and is simple toimplement. However, it does not take into account theinstantaneous traffic conditions at each ONU and thusmay not be able to provide the best performance interms of packet delay and data loss. To improve networkperformance, it is desirable to use an adaptive schedulingalgorithm that can dynamically schedule the order ofdifferent ONU transmissions based on the instantaneoustraffic conditions at each ONU. For example, an adaptivescheduling algorithm can schedule the ONU transmissionsin a descending order of the instantaneous queue lengthof each ONU, i.e., the longest queue first (LQF), orin an ascending order of the arrival time of the firstpacket queuing in each ONU, i.e., the earliest packet first(EPF) [12].

3.7. Quality of service provisioning

An EPON system is expected to deliver not only best-effort data traffic, but also real-time data traffic (e.g.,voice and video) that have strict bandwidth, packetdelay, and delay jitter requirements. To meet the servicerequirements of different end users, an EPON systemmustconsider differentiated QoS provisioning.

156 J. Zheng, H.T. Mouftah / Optical Switching and Networking 6 (2009) 151–162

C

ONU

S Intra-ONU scheduler

UsersS

Traffic classifierC

Fig. 6. Priority queuing and intra-ONU scheduling.

3.7.1. Priority queuingPriority queuing is an effective way to support differ-

entiated QoS [11]. With priority queuing, network traf-fic is classified into a set of classes with diverse QoSrequirements and for each traffic class a priority queue ismaintained at each ONU. Fig. 6 illustrates an example ofpriority queuing, in which an ONU maintains three prior-ity queues that share the samememory buffer of fixed size.Data packets from end users are first classified by checkingthe type-of-service (ToS) field of the IP packets encapsu-lated in the Ethernet packets and then buffered in corre-sponding priority queues. If a higher-priority packet findsthe buffer full at the time of its arrival, it can preempta lower-priority packet. If a lower-priority packet arrivesand finds the buffer full, it will be dropped. As a result,lower-priority traffic may experience very high packet lossand even resource starvation. To address this problem, anONUshould performsomekind of traffic policing to controlthe amount of higher-priority traffic from each end user.

3.7.2. ONU schedulingTo support differentiated QoS, there are two ONU

scheduling paradigms: inter-ONU scheduling and intra-ONU scheduling [11]. Inter-ONU scheduling is responsi-ble for arbitrating the transmissions of different ONUs,and intra-ONU scheduling is responsible for arbitrating thetransmissions of different priority queues in each ONU.There are two strategies to implement these two schedul-ing paradigms. One is to allow the OLT to perform bothinter-ONU scheduling and intra-ONU scheduling. In thiscase, the OLT is the only device that arbitrates the up-stream transmissions. Each ONU can request the OLT to al-locate bandwidth for each traffic class. For this purpose,an ONU must report the status of its individual priorityqueues to the OLT through REPORT messages. MPCP speci-fies that eachONUcan report the status of up to eight prior-ity queues [10]. The OLT can then generatemultiple grants,each for a specific traffic class, to be sent to the ONUusing asingleGATEmessage. The format of the 64-byteMPCPGATEmessage can be found in [10].The other strategy is to allow the OLT to perform

inter-ONU scheduling whereas to allow each ONU toperform intra-ONU scheduling. In this case, each ONUrequests the OLT to allocate bandwidth for it basedon its buffer occupancy status. The OLT only allocatesthe requested bandwidth to each ONU. Each ONU willdivide the allocated bandwidth among different classes ofservices based on their QoS requirements and schedule

the transmissions of different priority queues within theallocated bandwidth. For intra-ONU scheduling, thereare two types of scheduling paradigms: strict priorityscheduling and non-strict priority scheduling. In strictpriority scheduling, a lower-priority queue is scheduledonly if all queueswithhigher priority are empty. Obviously,this will potentially result in infinite packet delay andhigh packet loss for low-priority traffic. In non-strictpriority scheduling, only those packets that were reportedare transmitted first as long as they can be transmittedwithin the allocated timeslot. The transmission order ofdifferent priority queues is based on their priorities. Ifthe packets that were reported are all scheduled andthe current timeslot can still accommodate more packets,those newly arriving packets that were not reported arealso transmitted based on their priorities. As a result, alltraffic classes can have access to the upstream channelwithin the allocated timeslot as reported to the OLT whiletheir priorities are maintained, which ensures fairness inscheduling.

4. Dynamic bandwidth allocation algorithms for EPONS

In this section, we present a survey of the state-of-the-art DBA algorithms that have been proposed for EPONs.Given that QoS is the main concern in EPONs, we classifythese algorithms into DBA with QoS support and DBAwithoutQoS support, anddescribe their characteristics andperformances.

4.1. DBA without differentiated QoS support

There are several DBA algorithms proposed for EPONwhich do not support differentiated services, such asIPACT [9] and BGP [13].

4.1.1. Interleaved polling with adaptive cycle time (IPACT)IPACT [9] is the first DBA algorithm proposed for EPON.

It employs a resource negotiation process to facilitatequeue report and bandwidth allocation. The OLT pollsONUs and grants timeslots to each ONU in a round-robinfashion. The timeslot granted to an ONU is determinedby the queue status reported from that ONU. Therefore,the OLT is able to know the dynamic traffic load of eachONU and allocate the upstream bandwidth in accordancewith the bandwidth demand of eachONU.Moreover, it alsoemploys the SLAs of end users to upper bound the allocatedbandwidth to each ONU.In IPACT, several bandwidth allocation schemes are

investigated, including limited allocation, constant credit,linear credit, and elastic allocation. In limited allocation,the OLT simply grants an ONU the number of bytes theONU requested, but not exceeding a maximum windowsize. This is the most conservative scheme because itassumes that no more packets arrived after the ONU sentits request. In practice, however, because of the round-trip time between the OLT and each ONU, there might bemore packets arriving between the instant an ONU sends aREPORT message and the instant the ONU receives a GATEmessage. In this case, those newly arriving packetsmay not

J. Zheng, H.T. Mouftah / Optical Switching and Networking 6 (2009) 151–162 157

be able to be transmitted in the current cycle, resulting inincreased average packet delay. To address this problem,the constant credit scheme and the linear credit schemewere proposed.In constant-credit allocation, a credit is added to the

requested window size and is considered in the grantedwindow size. The size of the credit is constant no matterhow large the requested window size is. Once an ONUreceives a GATE message, it can send packets for up tothe requested window size plus the constant credit. Thechoice of the credit size may have an impact on thenetwork performance. A too small size will not be able toimprove packet delay a lot. A too large size will reducethe bandwidth utilization of the upstream channel. Thechoice should be based on the traffic characteristics orsome empirical data.In linear-credit allocation, a similar credit is added to

the requested window size. However, the size of the linearcredit is proportional to the requested window size. Thebasis behind this scheme is that network traffic usuallyhas a certain degree of predictability. This means that if along burst of data is observed, this burst is very likely tocontinue for longer time.In elastic allocation, there is no limit imposed on the

maximum window size. The only limit is the maximumcycle time. The maximum window size Wmax is grantedin such a way that the accumulated size of last N grants(including the one being granted) does not exceed N ×Wmax, where N is the number of ONUs. In this way, if onlyone ONU has data to send, it may get a granted windowsize up to N ×Wmax.Among all the above bandwidth allocation schemes,

limited allocation exhibits the best performance [2].However, IPACT does not consider the multi-service needsof subscribers. To meet such the multi-service needs,a variety of DBA algorithms have been proposed forsupporting differentiated services.

4.1.2. Estimation-based dynamic bandwidth allocationIn [14], Byun et al. proposed an estimation-based DBA

algorithm, which can reduce the queue length of eachONU and thus the average packet delay by estimating thepackets arrived at an ONU during the waiting time andincorporating the estimation in the grant to theONU. In thealgorithm, a control gain is used to adjust the estimationbased on the difference between the departed and arrivedpackets in the previous transmission cycle. The simulationresults show that the proposed DBA algorithm can reducethe average packet delay as compared to IPACT.

4.1.3. Interleaved polling with adaptive cycle time with grantestimation (IPACT-GE)In [15], Zhu et al. proposed another estimation-based

DBA algorithm, called interleaved polling with adaptivecycle timewith grant estimation (IPACT-GE), for efficientlysharing the upstream channel among multiple ONUS inan EPON system. With IPACT-GE, the amount of packetsarriving at an ONU between two consecutive pollingsis estimated based on the self-similarity characteristicof network traffic, and the OLT decides the granted

transmission size for the ONU based on the estimatedpacket amount as well as the amount requested in theprevious polling cycle. By estimating the amount of newarriving packets and granting an additional window size,the grant size to the ONU will be close to the real bufferoccupancy at the time when the ONU is polled. Underlight traffic load, packets are liable to be transmitted in thesame polling cyclewhen they arrivewithout a need towaituntil the next polling cycle. The simulation results showthat IPACT-GE can largely reduce the averagewaiting delayexperienced by a packet and the buffer occupancy underlight traffic load as compared to IPACT without affectingother performance in terms of average packet loss andupstream channel utilization. Moreover, when combinedwith the strict priority queue (SPQ) mechanism to supportdifferentiated services, it can greatlymitigate the light loadpenalty that is a phenomenon in using IPACT and SPQ.

4.1.4. Bandwidth guaranteed polling (BGP)Bandwidth guaranteed polling (BGP) [13] is a DBA

algorithm proposed for providing bandwidth guaranteesin EPONs. In BGP, all ONUs are divided into two groups:bandwidth guaranteed and bandwidth non-guaranteed.The OLT performs bandwidth allocation through using acouple of polling tables. The first polling table divides afixed-length polling cycle into a number of bandwidthunits and each ONU is allocated a certain number ofsuch bandwidth units. The number of bandwidth unitsallocated to an ONU is determined by the bandwidthdemand of that ONU, which is given by its SLA with aservice provider. A bandwidth guaranteed ONUwith morethan one entry in the poling table has its entries spreadthrough the table. This can reduce the average queuingdelay because theONU is polledmore frequently. However,this leads to more grants in a cycle and thus requiresmore guard times between grants, which reduces channelutilization. On the other hand, it can potentially leadto lower channel utilization because an Ethernet framecannot be fragmented in transmission. If a frame is toolarge to fit in the remainder of the current bandwidth unit,it will have to wait for the next bandwidth unit and aportion of the current bandwidth unit is thus wasted. Toaddress this problem, BGP allows an ONU to communicateto the OLT its actual use of a bandwidth unit. If theunused portion of the bandwidth unit is large enough, thisportion will be granted to a bandwidth non-guaranteedONU. Otherwise, the next bandwidth guaranteed ONU ispolled. However, this mechanism is largely limited by thepropagation delays between the OLT and ONUs.The unused portions of the bandwidth units for

the bandwidth guaranteed ONUs are distributed to thebandwidth non-guaranteed ONUs in the order of theirpositions in the second polling table. The construction ofthe second polling table is different from that of the firsttable. Each entry is dynamically created as a bandwidthnon-guaranteed ONU requests a grant. The simulationresults show that an ONU with more entries in the pollingtable has smaller queuing delay than one with fewerentries.

158 J. Zheng, H.T. Mouftah / Optical Switching and Networking 6 (2009) 151–162

4.1.5. IPACT with Smallest Available Report First (SARF)In [16], Bhatia and Bartos proposed a Smallest Available

Report First (SARF) heuristic to improve the performanceof the IPACT protocol in terms of packet delay. In thisheuristic, the OLT always grants an ONU with the smallestreported queue length first. At the same time, it treatsthe ONUs with a zero queue length differently. Under lowtraffic load, such ONUs are always served first withoutexception. However, it still requires the allocation of agrant to accommodate the next REPORT message as wellas themandatory guard-band overhead.While such grantsdo not contribute to reducing the packet delay at the ONUsthat have no packets to send, they do increase the packetdelay at the subsequent ONUs. Thus, when choosing thenext ONU to serve under higher traffic load, the ONUswithzero-length queues are treated as if they have a queuelength that is equal to the average queue length takenover all ONUs. Moreover, this average queue length isweighted by the number of times an ONU reports a zero-length queue consecutively. As a result, an ONU that hasconsecutively reported a zero-length queue many timeswill likely be served at the end of a cycle. The simulationresults show that the proposed SARF heuristic can improvethe delay performance of IPACT by about 10%–20% underthe gated allocation policy.

4.1.6. Multi-thread pollingIn [17], Song et al. proposed a multi-thread polling

algorithm to efficiently and fairly distribute the upstreambandwidth in long-reach PONs. This algorithm allows eachONU to send its REQUEST message before the previousGATE message is received from the OLT, thereby creatinga new ‘‘thread’’ of signaling between an ONU and the OLT.To implement this algorithm, the OLT must maintain apolling table and each ONU has an entry in the table, whichrecords the ONU’s Round-Trip Time (RTT) and its mostrecent requests in each thread. In each thread, the OLTperforms bandwidth allocation and distributes the GATEmessages to all ONUs. In multi-thread polling, the OLTcan make use of not only the information in the REQUESTmessages in the current thread, but also that in subsequentthreads before the OLT performs bandwidth allocation.The simulation results show that the proposed multi-thread polling algorithm, by setting the proper initialthread interval and tuning threshold, can improve theperformance of a single-thread polling algorithm in termsof average packet delay and throughput, in particular,under high traffic load.

4.2. DBA with differentiated QoS support

An EPON system is expected to deliver not only best-effort data traffic, but also real-time data traffic (e.g., voiceand video) that have strict bandwidth, packet delay, anddelay jitter requirements. In this subsection, we presentseveral DBA algorithms that can provide differentiated QoSsupport for different types of data traffic in EPON.

4.2.1. Fair sharing with dual SLAs (FSD-SLA)In [18], Banerjee et al. proposed a fair sharing with

dual SLAs (FSD-SLA) algorithm, which employs dual SLAsin IPACT to manage the fairness for both subscribers andservice providers. The primary SLA specifies those serviceswhose minimum requirements must be guaranteed witha high priority. The secondary SLA describes the servicerequirements with a lower priority. This algorithm firstallocates timeslots to those services with the primarySLA to guarantee their upstream transmissions. After theservices with the primary SLA are guaranteed, the nextround is to accommodate the secondary SLA services. If thebandwidth is not sufficient to accommodate the secondarySLA services, themax–min policy is adopted to allocate thebandwidth with fairness. If there is excessive bandwidth,FSD-SLA will allocate the bandwidth to the primary SLAentities first and then to the secondary SLA entities, bothby using max–min fair allocation.

4.2.2. Class-of-service-oriented packet scheduling (COPS)In [19], Naser andMouftah proposed a class-of-service-

oriented packet scheduling (COPS) algorithm to supportdifferentiated services. COPS uses two groups of leakybucket credit pools on the OLT side to regulate the trafficof each ONU and each class-of-service (CoS). One groupcontains k credit pools, corresponding to k CoSs in theEPON system. Each pool is used to control the average rateof certain CoS traffic from all ONUs to the OLT. The othergroup contains m credit pools, corresponding to m ONUsin the system. Each pool is used to control the usage of theupstream channel by an ONU. In allocating and grantingbandwidth or timeslots, the OLT begins with the highestCoS and ends with the lowest CoS, which is performedin two rounds. In the first round, each ONU with thetraffic of the current CoS is granted up to the number ofcredits available for that ONU. If a request is granted, thegranted bytes are subtracted from the corresponding creditpool. At the end of the first round, the unused credits arepooled together and are distributed to those ONUs whosebandwidth demands were not fully satisfied. As long asthere are credits available in the pools, a new request willbe accommodated. The simulation results show that COPShas lower average and maximum delay for all CoSs exceptthe highest-priority one as compared to IPACTwith limitedallocation.

4.2.3. Hybrid granting protocol (HGP)In [20], Shami et al. proposed a hybrid granting protocol

(HGP) to support differentiated QoS provisioning by guar-anteeing bandwidth andminimizing jitter. InHGP, traffic isclassified into three categories: Assured Forwarding (AF),Best Effort (BE), and Expedited Forwarding (EF). For the EFtraffic, HGP employs a queue predictionmechanism to sizethe grant to an ONU to accommodate all the traffic in thequeue of that ONU at the point of granting. This is becausethe EF traffic has a constant bit rate and thus can be easilypredicted. For the AF and BE traffic, it sizes the grant to anONU only based on the REPORT message from that ONU. Atransmission cycle consists of two sub-cycles: EF sub-cycleand AF/BF sub-cycle, and begins with the EF sub-cycle fol-lowed by the AF/EF sub-cycle. The EF sub-cycle carries the

J. Zheng, H.T. Mouftah / Optical Switching and Networking 6 (2009) 151–162 159

EF traffic while the AF/BF sub-cycle carries AF and BF traf-fic for each ONU. The length of the EF sub-cycle is prede-termined while that of the AF/BF sub-cycle depends on thetraffic load of each ONU. Accordingly, there are two grantsfor each ONU in every transmission cycle. The status of theAF and BF queues in an ONU is not reported until the end ofthe EF grant for that ONU, which allows the ONU to reportup-to-date queue status to the OLT. In this way, HGP guar-antees the bandwidth to the EF traffic and thus minimizesthe jitter experienced by the EF traffic, while keeping QoSsupport for theAF andBF trafficwith flexible bandwidth al-location. The simulation results show that HGP has smallerqueuing delay under higher traffic load as compared to aregular EPON scheduler. Under lower traffic load, the reg-ular EPON scheduler has smaller queuing delay because ofthe increased length of guard time per cycle.

4.2.4. Dynamic bandwidth allocation with multiple services(DBAM)In [21], Luo and Ansari proposed a dynamic bandwidth

allocation with multiple services (DBAM) algorithm to ac-commodate different types of traffic in EPONs. Instead ofproviding multiple services among ONUs and among endusers separately, DBAM incorporates both of them into theREPORT /GATE mechanism with class-based bandwidth al-location. It applies priority queuing to the EF, AF, and BEframes, and employs priority-based scheduling to schedulethe buffered frames. Moreover, DBAM uses limited band-width allocation to arbitrate bandwidth allocation amongONUs, thus prohibiting aggressive bandwidth scrambling.In addition, it employs class-based traffic prediction to takeinto account the traffic that arrives during the waiting pe-riod, which ranges from sending the queue status report tosending the traffic buffered in each ONU. Such a predictionis based on the actual traffic received in the previous wait-ing period. The OLT serves all ONUs in a fixed round-robinfashion in order to facilitate traffic prediction. The simula-tion results show that such prediction can provide smallerpacket delay for the EF traffic compared to fixed bandwidthallocation and limited bandwidth allocation.

4.2.5. Limited sharing with traffic prediction (LSTP)In [22], Luo and Ansari proposed a limited sharing

with traffic prediction (LSTP) algorithm, which employsan adaptive filter to predict the traffic that arrivesduring the waiting period and thus more accurately grantbandwidth to each ONU. For each class of traffic, LSTPestimates the data that arrive during the waiting periodbased on the data of this class that actually arrived inprevious transmission cycles by using a linear predictor.The bandwidth demand of an ONU is thus the reportedqueue length plus the estimation. The OLT arbitrates theupstream bandwidth using this estimation and reserves aportion of the upstream bandwidth for transmitting theestimated data in the earliest transmission cycle, thusreducing packet delay and loss. In addition, LSTP facilitatesservice differentiation by using different SLA parametersto restrict different classes of traffic. The simulation resultsshow that it improves the network performance in termsof average packet delay and loss as compared to fixedallocation, limited allocation, and limited allocation withexcess distribution.

4.2.6. Two-layer bandwidth allocationIn [23], Xie et al. proposed a two-layer bandwidth

allocation (TLBA) algorithm for supporting differentiatedservices in EPONs. TLBA is a hierarchical allocationalgorithm that allocates bandwidth in two layers. In thefirst layer, the transmission cycle is partitioned or theupstream bandwidth is allocated among differentiatedservice classes, which is called class-layer allocation. Inthe second layer, the partition or bandwidth allocated toeach class is distributed to all ONUs within the same classbased on a max–min fairness policy, which is called ONU-layer allocation. The OLT allocates bandwidth based on theinstantaneous demand of each ONU and does not limitthe size of each demand. Accordingly, an ONU is allowedto report the lengths of all its queues to the OLT and theOLT allocates the upstream bandwidth to meet all thedemands as much as possible. To avoid any class frommonopolizing the available bandwidth in a cycle, a per-class threshold is introduced. The bandwidth thresholdguarantees a minimum bandwidth for a class under hightraffic load. Any excessive bandwidth from the classes thatneed less than their thresholds is distributed among theclasses that needmore than their thresholds. The excessivebandwidth distribution is performed based on the weightsthat are assigned to each class. The simulation results showthat under lower traffic load, TLBA can achieve efficientutilization of the upstream channel.

4.2.7. Limited allocation with excess distributionIn [24], Assi et al. proposed an efficient DBA algorithm

to support QoS in EPONs. This algorithm is based onlimited allocation. Due to the bursty nature of Ethernettraffic, in each transmission cycle, some ONUs (also calledlightly-loaded ONUs) may have less traffic to transmitand thus need smaller bandwidth than the minimumguaranteed bandwidth, while other ONUs (also calledheavily-loaded ONUs) may have more traffic to transmitand need larger bandwidth than the minimum guaranteedbandwidth. For this reason, those lightly-loaded ONUsmay result in an excessive bandwidth. Obviously, thisexcessive bandwidth can be exploited to meet thebandwidth demands of the heavily-loaded ONUs. For thispurpose, an allocation schemewas proposed to allocate theexcessive bandwidth among the heavily-loaded ONUs inproportional to their bandwidth demands. To implementthis allocation scheme, the OLT must collect the REPORTmessages from all ONUs before it performs computationfor bandwidth allocation and then schedule (i.e., sendsGATE messages to) the ONUs for the next transmissioncycle. However, this would result in an idle period inwhich the upstream channel is unutilized. To make useof this idle period and improve bandwidth utilization,an early allocation mechanism was proposed, whichschedules a lightly-loaded ONU instantaneously withoutany delay, whereas schedules those heavily-loaded ONUsafter the OLT receives all REPORT messages and performscomputation for bandwidth allocation. However, thisscheduling mechanism may not be able to make sufficientuse of the idle period in many cases. While it cansignificantly improve bandwidth utilization under lowand medium traffic load, it is unable to make use of

160 J. Zheng, H.T. Mouftah / Optical Switching and Networking 6 (2009) 151–162

the idle period under high traffic load [25]. To furtherimprove bandwidth utilization under high traffic load, anew scheduling control mechanism is proposed to addressthe idle period problem in [25]. In this mechanism, the OLTstill employs an early allocationmechanism that schedulesa lightly-loaded ONU instantaneously without any delay.At the same time, it accumulates the excessive bandwidthcontributed by each lightly-loaded ONU. For the heavily-loaded ONUs, the OLT normally waits until after all REPORTmessages are received to perform the computation forbandwidth allocation and send GATE messages to theONUs. To ensure that the idle period is not wasted, theOLT maintains a tracker that records the ending time ofthe timeslot for the last scheduled ONU and updates thetracker every time the next ONU is scheduled. The OLTwill schedule a heavily-loaded ONU under some specialconditions depending on the value of the tracker. Thesimulation results show that the DBA algorithm proposedin [25] can effectively improve the network performancein terms of packet delay and throughput under high trafficload compared with the algorithm proposed in [24].

4.2.8. Queue-based dynamic bandwidth allocationIn [26], Choudhury and Saengudomlert proposed an

OLT-centricDBA schemebased on individual requests fromthe service queues in ONUs for a QoS-aware EPON. Toprevent misbehaving high-priority traffic from completelystarving all lower-priority traffic, the proposed DBAalgorithm employs queue scheduling and makes use ofthe excess bandwidth of lightly-loaded queues to meetthe bandwidth demand of heavily-loaded queues. For thispurpose, it incorporates an efficient polling mechanismto solve the idle period problem and uses a noveldifferent-cycle policy to reduce the scheduling overhead,which selectively allocates bandwidths to different serviceclasses based on their delay bounds. The simulation resultsshow that the proposed DBA algorithm outperforms theIPACT algorithm that uses a strict priority policy in termsof average packet delay and bandwidth utilization.

4.2.9. QoS-aware dynamic bandwidth allocationIn [27], Miyoshiet al. proposed a QoS-aware DBA

algorithm, called Dynamic Credit Distribution (D-CRED),for gigabit EPONs and argued that in order to achievehigher bandwidth utilization unused slot remainders mustbe eliminated. To eliminate the unused slot remainders,D-CRED introduces a dynamic queue threshold technique,which allows each ONU to have only one threshold andthe OLT to dynamically change the threshold value (alsocalled credit here). This dynamic threshold technique caneliminate unused slot remainders and keep the cyclelonger, which in turn leads to higher throughput efficiencyand more precise bandwidth allocation. The simulationresults show that D-CRED can achieve a 6.4% higherbandwidth utilization compared to IPACT with limitedservice, and 99% of the theoretical maximum utilization.D-CRED can also be extended to QoS support. For thispurpose, a concept of fairness is defined and basedon the defined concept D-CRED calculates a degree ofsatisfaction of an ONU regarding credit allocation in each

cycle and keeps track of the degree during the entire busyperiods. If the ONU receives a full credit as requested,D-CRED considers that the ONU is satisfied with the creditallocation. Otherwise, the ONU is not satisfied with thecredit allocation. Therefore, D-CRED tries to make everyONU satisfied in its credit allocation.

4.2.10. Intra-ONU bandwidth schedulingIn [28], Ghani et al. proposed a modified start-time fair

queuing (M-SFQ) virtual time scheduler for decentralizedintra-ONU bandwidth allocation in EPONs. Unlike othervirtual time schedulers which timestamp all packets,the M-SFQ scheduler only maintains timestamps for thehead-of-line (HOL) queue packets, yielding much lowercomplexity. It features low implementation complexityand can be incorporated with any OLT-ONU (inter-ONU)DBA algorithm. The simulation results confirm that the M-SFQ scheduler can achieve a very fine degree of bandwidthallocation and good delay performance.

4.2.11. Intra-ONU bandwidth allocationIn [29], Chen et al. proposed a novel modified to-

ken bucket (M-TB) algorithm for decentralized intra-ONUbandwidth allocation in EPONs. Compared with the purestrict priority (SP) algorithm and the M-SFQ algorithm,which have a complexity of O(k) and O(k logM), respec-tively, this algorithm has a complexity of O(k), where kisthe total number of packets that can be sent in one grantwindow andM is the number of input queues. The simula-tion results confirm that theM-TB algorithm can guaranteeboth the priority and the fairness of differentiated serviceswhile the SP and M-SFQ algorithms cannot. In addition, itcan also be incorporated with any OLT-ONU (inter-ONU)DBA algorithm.

4.2.12. Fine schedulingIn [30], Chen et al. proposed a fine scheduling algorithm

for upstream bandwidth allocation, which consists ofan inter-ONU scheduler at the OLT and an intra-ONUscheduler at each ONU. In the inter-ONU scheduling, anovel DBA algorithm is proposed to fairly allocate thebandwidth for upstream data transmission and optimizethe upstream bandwidth utilization. Unlike most otherDBA algorithms which only report the buffered data sizeat each ONU, the proposed DBA algorithm allows eachONU to send two requests in its REPORT messages: amaximum window size and a minimum (or guaranteed)window size. After receiving all the REPORT messages ina cycle, it performs bandwidth allocation and allocates abandwidth to each ONU based on the following fairnesscriteria: (1) The minimum bandwidth demand is alwaysguaranteed; (2) The excess bandwidth is allocated toa unit (an ONU or a queue) according to its weight;(3) No unit is allocated a bandwidth more than itsmaximum bandwidth demand. Furthermore, it introducesa novel approach to eliminate unused slot remainderswithout causing transmission delay by sending the REPORTmessage ahead of the data stream, and informing the OLTof the requested maximum and minimum bandwidths forthe next cycle as well as the total actual bandwidth of each

J. Zheng, H.T. Mouftah / Optical Switching and Networking 6 (2009) 151–162 161

ONU for this cycle. In addition to the inter-ONU scheduler,a novel hierarchical intra-ONU scheduler is also proposedto realize fine granularity scheduling to support QoS fortraffic of each individual user by combining the M-TBalgorithm for inter-class [29] and the modified start-timefair queuing (for intra-class) [28]. The simulation resultsshow that the proposed overall scheduling algorithm canmeet the performance requirements in terms of packetdelay and throughput for the transmission of multimediatraffic of each end user.

4.2.13. Priority-based dynamic bandwidth allocation foremergency handlingIn [31], Moon studied the DBA problem in network

emergencies and argued that priority-based DBA shouldonly be used when a network emergency occurs. Thenetwork state is classified into two different states: normalstate and emergency state. For the normal state, a DBAalgorithm proposed in [32] is adopted, which is slightlyextended to consider the weight of each ONU. For theemergency state, three priority-based DBA algorithms areproposed,which differ in theway they protect non-priorityONUs from starving. In the first algorithm, there is noprotection of non-priority ONUs from starving. That is,non-priority ONUs can be served only after the requestsof all priority ONUs are served. In the second algorithm,a counter is used to protect non-priority ONUs fromstarving. That is, non-priority ONUs can be served afterpriority ONUs are served over a certain number, say αtimes. Also, when non-priority ONUs are served over acertain time, say β times, the service is handed over topriority ONUs. In the third algorithm, the OLT maintains apolling sequence table to protect non-priority ONUs fromstarving. The simulation and analytical results show thatthe proposed DBA algorithms perform efficiently in termsof packet delay and bandwidth utilization, and can meetmore stringent QoS requirements than the conventionalweighted round-robin-based algorithms [33].

4.2.14. Fair bandwidth allocation using effective multicasttraffic shareIn [34], Kim et al. proposed a fair bandwidth allocation

mechanism, called share-based proportional bandwidthallocation (S-PBA), to effectively supportmulticast servicesin the downstream direction of a TDM-PON. In order toprovide anONUwith a fair amount of downlink bandwidthand high throughput independent of the traffic type,the S-PBA mechanism arbitrates the amount of unicasttimeslot by using effective multicast traffic share, whichis determined based on multicast traffic load distributionand traffic-sharing density. In the S-PBA mechanism, atransmission cycle consists of three types of timeslots:a static timeslot that provides the ONUs with a staticbandwidth for real-time constant bit rate (CBR) trafficor management signaling in the manner of unsolicitedpolling; a multicast timeslot for transmission of multicasttraffic; and unicast timeslots allocated to the ONU queuesfor transmission of unicast traffic. In each cycle, the controlagent in the OLT allocates timeslot bandwidths to the staticservice queue and the single copy broadcast (SCB port),respectively. Then a remaining unicast timeslot bandwidth

is distributed over the ONU queues to guarantee eachONU queue a minimum fair bandwidth. A restriction onthe amount of bandwidth for multicast traffic is neededin the case of traffic congestion in order to preventa bandwidth monopolization of multicast applications,while guaranteeing a strict QoS to important unicastapplications. The analytical and simulation results validatethe effectiveness of the proposed S-PBA mechanism andthe mechanism is applicable to multicast video delivery ormulticast traffic transmission in general.

4.2.15. Admission control for QoS protectionIn [35,36], Dhaini and Assi et al. proposed a per-stream

QoS protection in EPONs using a two-stage admissioncontrol system. In this admission control system, a totalcycle is divided into two main cycles, one for real-time traffic and the other for best-effort traffic. For thereal-time traffic, each ONU is guaranteed a minimumbandwidth and could be allocated up to a maximumbandwidth. Thus, the corresponding cycle can further bedivided into two sub-cycles: T1 and T2. The OLT allocatesthe minimum guaranteed bandwidth for each ONU inT1, which is under the control each ONU, whereas thebandwidth of T2 is under the control of the OLT. The systememploys a two-stage admission control mechanism: alocal admission control (LAC) at each ONU and a globaladmission control (GAC) at the OLT. The first stageallows each ONU to perform flow admission controllocally based on its bandwidth availability. At the firststage, an ONU can locally perform flow admission controlbased on the bandwidth demand of a new arriving flowand its bandwidth availability, and best-effort traffic isalways admitted. For a real-time traffic flow, it willconditionally admit the flow based on the bandwidthusage condition, and will monitor its QoS status for apredefined number of cycles. If the QoS of the newlyadmitted flow are satisfied without affecting the QoS ofthe existing flows, the flow will be admitted. Otherwise,the flow will be dropped. If a flow cannot be admittedat the ONU due to bandwidth insufficiency, the ONUwill report the arrival of a new flow to the OLT. TheOLT will admit this new flow only if there is bandwidthavailable in T2 and the ONU has not been allocated abandwidth more than the maximum bandwidth. Althoughthe minimum guaranteed bandwidth is under the controlof each ONU, the scheduling of different ONUs is stillcentrally done at the OLT in order to avoid collisionsin the upstream channel. To support the admissioncontrol mechanism, a hybrid DBA algorithm is alsoproposed, which performs both bandwidth allocation andreservation simultaneously. The simulation results showthat the proposed admission control system can achieve agood performance in providing QoS for both real-time andbest-effort traffic.

5. Conclusions

Bandwidth allocation is a critical issue in the design ofan EPON system. Since multiple ONUs share a commonupstream channel, an EPON systemmust efficiently utilizethe limited upstream bandwidth in order to meet the

162 J. Zheng, H.T. Mouftah / Optical Switching and Networking 6 (2009) 151–162

bandwidth demands and quality of service requirementsof end users. For this purpose, DBA is highly preferred. Inthis chapter, we discussed themajor issues that are relatedto bandwidth allocation in an EPON system and presenteda survey of the state-of-the-art DBA algorithms for TDMEPON systems. Although many DBA algorithms have beenproposed to provide differentiated QoS in EPONs, furtherresearch efforts are still needed in order to explore bettersolutions to this critical issue.On the other hand, recent advances in enabling

technologies havemade optical devices that are previouslycost-prohibited become more affordable, which makes iteconomically viable to use multiple upstream channels inan EPON system. This has presented new challenges in thedesign of DBA algorithms for multi-channel WDM EPONs,which has become a hot research topic in recent years. Dueto the limitation of space, the reader is referred to [37–40]for the current research status on this topic.

References

[1] J. Zheng,H.T.Mouftah, OpticalWDMNetworks: Concepts andDesignPrinciples, Wiley-IEEE Press, Hoboken, NJ, 2004.

[2] G. Kramer, G. Pesavento, Ethernet passive optical network (EPON):Building a next-generation optical access network, IEEE Communi-cations Magazine 40 (2) (2002) 66–73.

[3] C. Foh, L. Andrew, E. Wong, M. Zukerman, FULL-RCMA: A highutilization EPON, IEEE Journal on Selected Areas in Communications22 (8) (2004) 1514–1524.

[4] S. Sherif, A. Hadjiantonis, G. Ellinas, C. Assi, M. Ali, A novel decentral-ized Ethernet-based PON access architecture for provisioning dif-ferentiated QoS, IEEE/OSA Journal of Lightwave Technology 22 (11)(2004) 2479–2483.

[5] A. Shami, X. Bai, N. Ghani, C. Assi, H. Mouftah, QoS control schemesfor two-stage Ethernet passive optical access networks, IEEE Journalon Selected Areas in Communications 23 (8) (2005) 1467–1478.

[6] B.N. Nesai, N.J. Frigo, A. Smiljanic, K.C. Reichmann, P.P. Iannone,R.S. Roman, An optical implementation of a packet-based (Ethernet)MAC in a WDM passive optical network overlay, in: Proc. of OFC’01,vol. 3, 2001, pp. WN5-1–WN5-3.

[7] C.-J. Chae, E. Wong, R.S. Tucker, Optical CSMA/CD media accessscheme for Ethernet over passive optical network, IEEE PhotonicsTechnology Letters 14 (5) (2002) 711–713.

[8] Y. Luo, S. Yin, N. Ansari, T. Wang, Resource management forbroadband access over time-division multiplexed passive opticalnetworks, IEEE Network 21 (5) (2007) 20–27.

[9] G. Krammer, B. Mukherjee, G. Pesavento, IPACT: A dynamic protocolfor an Ethernet PON (EPON), IEEE Communications Magazine 40 (2)(2002) 74–80.

[10] IEEE 802 3ah Ethernet in the First Mile Task Force. http://www.ieee802.org/3/efm/index.html.

[11] J. Zheng, H.T. Mouftah, Media access control for Ethernet passiveoptical networks (EPONs), an overview, IEEE CommunicationsMagazine 43 (2) (2005) 145–150.

[12] J. Zheng, H.T. Mouftah, Adaptive scheduling algorithms for Ethernetpassive optical networks, IEE Proceedings-Communications 152 (5)(2005) 643–647.

[13] M. Ma, Y. Zhu, T. Cheng, A bandwidth guaranteed polling MACprotocol for Ethernet passive optical networks, in: Proc. of IEEEINFOCOM’03, vol. 1, Mar.–Apr. 2003, pp. 22–31.

[14] H.-J. Byun, J.-M. Nho, J.-T. Lim, Dynamic bandwidth allocationalgorithm in Ethernet passive optical networks, IEEE ElectronicsLetter 39 (13) (2003) 1001–1002.

[15] Y. Zhu, M. Ma, IPACT with grant estimation (IPACT-GE) scheme forEthernet passive optical networks, IEEE/OSA Journal of LightwaveTechnology 26 (14) (2008) 2055–2063.

[16] S. Bhatia, R. Bartos, IPACT with smallest available report first: Anew DBA algorithm for EPON, in: Proc. of IEEE ICC’07, Glasgow, UK,June 2007, pp. 2168–2173.

[17] H. Song, B.-W. Kim, B. Mukherjee, Multi-thread polling: A dynamicbandwidth distribution scheme in long-reach PON, IEEE Journal onSelected Areas in Communications 27 (2) (2009) 134–142.

[18] A. Banerjee, G. Kramer, B. Mukherjee, Fair sharing using dualservice-level agreements to achieve open access in a passive opticalnetwork, IEEE Journal on Selected Areas in Communications 24 (8)(2006) 32–44.

[19] H. Naser, H. Mouftah, A joint-ONU interval-based dynamic schedul-ing algorithm for Ethernet passive optical networks, IEEE/ACMTransactions on Networking 14 (4) (2006) 889–899.

[20] A. Shami, X. Bai, C. Assi, N. Ghani, Jitter performance in Ethernetpassive optical networks, IEEE/OSA Journal of Lightwave Technology23 (4) (2005) 1745–1753.

[21] Y. Luo, N. Ansari, Bandwidth allocation for multiservice accesses onEPONs, IEEE Communications Magazine 43 (2) (2005) S16–S21.

[22] Y. Luo, N. Ansari, Limited sharing with traffic prediction for dynamicbandwidth allocation and QoS provisioning over EPONs, OSA Journalof Optical Networking 4 (9) (2005) 561–572.

[23] J. Xie, S. Jiang, Y. Jiang, A dynamic bandwidth allocation scheme fordifferentiated services in EPONs, IEEE CommunicationsMagazine 42(8) (2004) S32–S39.

[24] C.M. Assi, Y. Ye, S. Dixit, M.A. Ali, Dynamic bandwidth allocationfor quality-of-service over Ethernet PONs, IEEE Journal on SelectedAreas in Communications 21 (9) (2003) 1467–1477.

[25] J. Zheng, H.T. Mouftah, Efficient bandwidth allocation algorithm forEthernet passive optical networks, IEE Proceedings – Communica-tions 153 (3) (2006) 464–468.

[26] P.K. Choudhury, P. Saengudomlert, Efficient queue based dynamicbandwidth allocation scheme for Ethernet PONs, in: Proc. of IEEEGLOBECOM’07, Washington DC, Nov. 2007, pp. 2183–2187.

[27] H. Miyoshi, T. Inoue, K. Yamashita, QoS-aware dynamic bandwidthallocation scheme in Gigabit-Ethernet passive optical networks,in: Proc. of IEEE ICC’04, Paris, France, June 2004, pp. 90–94.

[28] N. Ghani, A. Shami, C. Assi, M.Y.A. Raja, Intra-ONU bandwidthscheduling in Ethernet passive optical networks, IEEE Communica-tion Letters 8 (11) (2004) 683–685.

[29] J. Chen, B. Chen, S. He, A novel algorithm for intra-ONU bandwidthallocation in Ethernet passive optical networks, IEEE Communica-tion Letters 9 (9) (2005) 850–852.

[30] B. Chen, J. Chen, S. He, Efficient and fine scheduling algorithm forbandwidth allocation in Ethernet passive optical networks, IEEEJournal of Selected Topics in Quantum Electronics 12 (4) (2006)653–660.

[31] B. Moon, Emergency handling in Ethernet passive optical networksusing priority-based dynamic bandwidth allocation, in: Proc. of IEEEINFOCOM’08, Phoenix, AZ, Apr. 2008, pp. 1319–1327.

[32] K. Son, H. Ryu, T. Yoo, S. Chong, Dynamic bandwidth allocationschemes to improve utilization under non-uniform traffic inEthernet passive optical networks, in: Proc. of IEEE ICC’04, Paris,France, June 2004, pp. 1766–1770.

[33] N. Ghani, A. Shami, C. Assi, M.Y.A. Raja, Quality of service in Ethernetpassive optical networks, in: Proc. of 2004 IEEE Sarnoff Symposium,Princeton, NJ, Apr. 2004, pp. 161–165.

[34] N. Kim, H.-S. Lim, M. Kang, Fair bandwidth allocation using effectivemulticast traffic share in TDM-PONs, IEEE/OSA Journal of LightwaveTechnology 26 (7) (2008) 756–767.

[35] A.R. Dhaini, C. Assi, M. Maier, A. Shami, Per-stream QoS andadmission control in Ethernet passive optical networks (EPONs),IEEE/OSA Journal of Lightwave Technology 25 (7) (2007) 1659–1669.

[36] C. Assi, M. Maier, A. Shami, Toward quality-of-service protectionin Ethernet passive optical networks, IEEE Network 21 (5) (2007)12–19.

[37] M. McGarry, M. Maier, M. Reisslein, WDM Ethernet passive opticalnetworks, IEEE Communications Magazine 44 (2) (2006) S18–S25.

[38] M.P. McGarry, M. Reisslein, M. Maier, A. Keha, Bandwidth manage-ment for WDM EPONs, OSA Journal of Optical Networking 5 (9)(2006) 637–654.

[39] A.R. Dhaini, C.M. Assi, M. Maier, A. Shami, Dynamic wavelengthand bandwidth allocation in hybrid TDM/WDM EPON networks,IEEE/OSA Journal of Lightwave Technology 25 (1) (2007) 277–286.

[40] M.P. McGarry, M. Reisslein, C.J. Colbourn, M. Maier, F. Au-rzada, M. Scheutzow, Just-in-time scheduling for multichannelEPONs, IEEE/OSA Journal of Lightwave Technology 26 (10) (2008)1204–1216.