Quality of Service in Ethernet Passive Optical Networks (EPONs)

Ahmad R. Dhaini & Chadi M. Assi

Abstract Ethernet passive optical networks (EPONs) are designed to deliver multiple services and applications, suchas voice communications (VoIP), standard and high-definition video (STV and HDTV), video conferencing (interac-tive video), IPTV and data traffic access network. These Differentiated Services (DiffServ) carry strict bandwidth anddelay requirements, as well as jitter sensitivity. Supporting these bundled services in EPONs faces many challengesthat require extensive research and studies. As a result, many techniques and methods have been presented to facilitatethe latter. The process of supporting these services is so-called quality-of-service (QoS) support; which refers to theability to provide different priority to different applications, users, or data flows, or to guarantee a certain level ofperformance to a data flow.In this chapter, we overview the tools and techniques that were presented to-date aiming to provide ”fair” and ”effi-cient” support QoS in EPONs. More specifically, we overview the various intra-ONU bandwidth scheduling schemesthat have been proposed and we present our intra-ONU scheduling solution, namely Modified-DWRR (M-DWRR),that is based on the Deficit Weighted Round Robin (DWRR) scheduling mechanism. M-DWRR proved to achieveadaptive fairness among different classes of services. Next, we overview the different dynamic bandwidth allocation(DBA) schemes that enable QoS support from an OLT perspective. Here, the OLT is responsible for allocating band-width for each class of service and hence no intra-ONU scheduling is required. Finally, we discuss the quality ofservice protection issue in EPON; which is quite interesting and challenging due the Time Division Multiple Access(TDMA) nature of EPON. To resolve this issue, we present the first admission control framework in EPON, that willallow for both quality of service protection and efficient bandwidth allocation and reservation.

Ahmad R. DhainiConcordia University, 1455 De Maisonneuve Blvd. West, Montreal, Quebec, Canada, e-mail: a dhaini@ece.concordia.ca

Chadi M. AssiConcordia University, 1455 De Maisonneuve Blvd. West, Montreal, Quebec, Canada, e-mail: assi@ciise.concordia.ca

1

1 Introduction

Quality of Service (QoS) refers to the capability of a network to provide better service to selected network traffic overvarious technologies, including Frame Relay, Asynchronous Transfer Mode (ATM), Ethernet and 802.1 networks,SONET, and IP-routed networks that may use any or all of these underlying technologies.The primary goal of QoS is to provide priority including dedicated bandwidth, controlled jitter 1 and latency (requiredby some real-time and interactive traffic), and improved loss characteristics. It is also important to make sure that pro-viding priority for one or more flows does not make other flows fail. QoS technologies provide the elemental buildingblocks that will be used for future business applications in campus, WAN, and service provider networks.Broadband access providers view QoS and multimedia-capable networks as an essential ingredient to offer residen-tial customers video-on-demand, audio-on-demand, voice over IP (VoIP) and high speed Internet access. Furthermorebroadband access networks, and EPON in particular, are especially appropriate for peer to peer applications (P2P)(which permit files to be interchanged through the Internet). It was shown that P2P applications represent a high frac-tion of the upstream traffic in hybrid fiber-coax cable access network [22]. Unlike early file sharing applications (suchas Napster and Gnutella), many recent P2P applications include live media broadcasting, high bandwidth content dis-tribution and real time audio conferencing and require high performance access networks in order to deliver satisfyingQoS to the users [18].

Table 1 Services Supported by Access Networks

Service Type Downstream Bandwidth Upstream Bandwidth

Telephony Switched 4 kHz 4 kHz

ISDN Switched 144 kbps 144 kbps

Video Broadcasting Broadcast 6 MHz or 2 to 6 Mbps 0

Interactive Video Switched 6 Mbps Small

Internet Access Switched Some Mbps Small (initially)

Video-Conferencing Switched 6 Mbps 6 Mbps

Enterprise Services Switched 1.5 Mbps to 10 Gbps 1.5 Mbps to 10 Gbps

Table 1 lists the various services that are supported in access networks. As shown, different services have differentrequirements that must be respected to meet the users ”expectations”, that is, the service level agreement (SLA). TheSLA is a formally negotiated agreement between two parties. It is a contract that exists between customers and theirservice provider, client or between service providers. It records the common understanding about services, priorities,responsibilities, guarantee, and such collectively, the level of service. For example, it may specify the levels of avail-ability, serviceability, performance, operation, or other attributes of the service like billing and even penalties in thecase of violation of the SLA.Since EPON is promising to be the best wired-access cost effective network panacea, it is also supposed to supportthese services. Moreover, some of these services require a large amount of bandwidth (e.g., Video-Conferencing),hence an arbitration of the users is required in order to deny a user/node from monopolizing the EPON limited band-width (1 Gbps). In particular, the access node has a key role in providing QoS for upstream traffic. Additionally, theaccess node has a full knowledge of the access line characteristics for downstream flows and can therefore preventdownstream congestion of high priority traffic.A variety of techniques are used to optimize network load, application performance and fairness among services and

1 Jitter: The interval between successive pulses or phase of successive cycles (in this case packets).

2

applications. These services are mapped into the access nodes (here, ONUs) as ”classes”. Each class of service (CoS)is identified by the service’s traffic requirements and properties.

1.1 IEEE 802.1D Support for Classes of Service

To support CoS, Ethernet networks must be able to classify traffic into CoS and provide differentiated treatment toeach class. This was achieved by an introduction of two new standard extensions (1) P802.1p (a supplement to MACbridges): traffic class expediting and dynamic multicast filtering (later merged with P802.1D) and (2) P802.1Q Virtualbridged local area networks.P802.1Q defines a frame-format extension allowing Ethernet frames to carry a priority informationfield in their header.The standard distinguishes the following traffic classes:

1. Network control: Characterized by a ”must get there” requirement to maintain and support the network infrastruc-ture.

2. Voice: Characterized by less than 10-ms delay, and hence maximum jitter.3. Video: Characterized by less than 100-ms delay.4. Controlled load: Important business applications subject to some form of admission control, which can be eitherachieved by a preplanning of the network requirement or by bandwidth reservation per flow at the time the flowadmitted to the network.

5. Excellent effort (CEO’s best effort): As called, the best-effort-type services that an information services organi-zation would deliver to its most important customers.

6. Best effort (BE): Internet traffic as we know it today (data traffic).7. Background: Bulk transfers and other activities that are permitted on the network without affecting existing appli-cations nor monopolizing the shared bandwidth.

Note: If a bridge or a switch has less than seven queues, some of the traffic classes are grouped together. A variety oftechniques are used to optimize network load, application performance and fairness among applications.As discussed in the previous chapter, EPON engages two wavelengths in the shared channel between the ONUs andthe OLT. That is, traffic travels in both the upstream and the downstream directions. Hence, QoS operation should beconsidered for both directions.

Upstream QoS

In the upstream direction, flows/streams arrive from end-users ”undefined” as bursts at constant or variable bit rates.Hence, the following operations are conducted before transmitting these packets to the central office node.

• Traffic classification and filtering is achieved using Access Control Lists (ACL)2, which helps identifying andclassifying the traffic flows. Furthermore, filtering helps enforcing security policies by removing fraudulent trafficthat could potentially compromise the overall network behavior/performance.

• DSCP3 or p-bit (re)marking to values that can be trusted by the network elements in the aggregation network whenproviding per-flow QoS processing. The DSCP is typically used by QoS-enabled IP service routers, whereas thep-bits can be used by QoS-enabled Ethernet switches. If the DSL modem or home gateway has performed a trafficclassification process, then the access node can use the DSCP or p-bit marking to perform per-flow QoS processingas well as remarking, if required. Otherwise the access node traffic filtering process classifies the traffic into QoSsub-flows, and then (re)marks the packets accordingly.

• Ingress policing enforces the traffic contracts (SLAs), that specify how much traffic users can send to the network.A policer may apply to a full access line or to a set of QoS flows matching a certain classification. Policing may

2 Access Control List (ACL): A sequence of patterns to match packets from a traffic stream, and the associated filtering actions to be takenwhen a packet matches a certain pattern. Patterns could be based on fields such as the Ethernet header, IP header or UDP/TCP port numbers.3 Differentiated Services Code Point (DSCP): a 6-bit encoding that indicates the type of service (TOS) in the IP packet header.

3

take into account the drop precedence coded in the DSCP or p-bits; out-of-profile traffic can be filtered/dropped, orcan be remarked with a higher drop precedence.

• Traffic forwarding to the right egress interface is usually based on either the destination Media Access Control(MAC) address and/or the destination IP address.

• Per-QoS class queuing and scheduling on the access node based on the DSCP or p-bits: a set of egress queues ispresent and each QoS class is mapped to the appropriate queue. The scheduling mechanisms determine way thesepackets are selected from their queues. Different schedulers (e.g. strict priority or round robin) can provide differentfunctionality on the same set of queues.

Downstream QoS

In the downstream direction, frames arrive at the access node with DSCP or p-bits that are properly marked by theapplication servers or service routers. The latter also perform ingress policing, which implies that the access node doesnot need to do this. After the forwarding decision, the following steps can be performed:

• Egress rate limiting: Similar to ingress policing, this is used to enforce traffic contracts that are associated with theservices to which the users have subscribed.

• Per-QoS class queuing and scheduling on the access lines based on the DSCP or p-bits: downstream it is crucialto use multiple QoS queues per access line and to map traffic to a specific queue based on the DSCP or pbits. Thescheduling mechanisms combined with the use of queue management features, such as (weighted) random earlydetection (WRED), determine the exact treatment received by packets in the various queues.

1.2 CoS Support in EPON

Current EPONs support diverse applications, various traffic sessions are aggregated into a limited number of classesto be serviced with differentiated services (DiffServ). These services are classified as follows: Best Effort (BE) ”data”traffic, Assured Forwarding (AF) traffic such as variable-bit-rate (VBR) video stream and Expedited Forwarding (EF)traffic used to emulate point-to-point (P2P) connections or real time services, such as Voice over IP (VoIP). The high-priority class is EF, which is delay-sensitive and requires bandwidth guarantees. The medium-priority class is AF,which is not delay-sensitive but requires bandwidth guarantees. The low priority class is BE, which is neither delay-sensitive nor bandwidth guaranteed. This information is encapsulated in the TOS of DSCP header. Note that ONUscan support up to 8 priority queues [23].To provide QoS in EPONs, bandwidth management on the upstream channel is essential for successful implementa-tions these types of networks. This is done in two ways:

• Intra-ONU scheduling: The process of selecting packets that are buffered into the various priority queues to betransmitted in the next transmission window.

• QoS-Enabled Dynamic Bandwidth Allocation: A QoS variant of the DBA presented in the previous chapter. Here,the OLT is responsible for allocating bandwidth for each class of service in each of the connected available ONUs.

In addition, to enable QoS protection and bandwidth guaranteed, Admission Control seems to be a necessity for QoSsupport in EPONs. Admission Control is the ability to judge whether the network has enough resources available toaccept the connection, and then either accepts or rejects the connection request.We discuss these three subjects in details in the following sections.

4

Fig. 1 Intra-ONU Scheduling

2 Intra-ONU Scheduling

2.1 Incoming Traffic Handling Operation

The ONU uses the traditional ”QoS classifying and scheduling” operation, that was detailed in the previous section, tohandle the users’ incoming flows. As shown in Fig.1, upon receiving traffic ”flows” from the registered subscribers, theONU performs three main operations before transmission in the upstream channel. First, it classifies every newly ar-riving packet using a ”packet-based” classifier. Next, and before placing packets in the corresponding priority queues,the ONU decides whether a packet should be admitted depending on the adopted traffic policing (admission con-trol) mechanism (e.g., Leaky Bucket). Finally, it selects packets from its queues, depending on the intra-/inter-ONUscheduling algorithm [51], and sends them to the OLT as ”flows” in the assigned transmission window (TW). Notethat ”Low Complexity” is a key design goal for intra-ONU schedulers in order to keep the ONUs’ cost at minimum.A scheduler comprises an output link, an arbiter, and a set of input class queues. Here, the arbiter runs a schedulingalgorithm to partition link capacity and deliver pre-determined (SLA) delay and throughput requirements to the inputqueues.

2.2 Existing Solutions & Schemes

Schedulers range in complexity and performance capabilities and most notable examples include priority schemes,non-work conserving schemes; e.g., stop-and-go, weighted/hierarchical round-robin (WRR/HRR), jitter earliest-due-date; and work-conserving approximations of ideal generalized processor sharing (GPS), e.g., weighted fair queuing(WFQ/WFQ2) [45], self-clocked fair queuing (SCFQ), start-time fair queuing (SFQ) [49]. Nevertheless, most sched-ulers have only been studied for intra-system roles, e.g., within ATM switches or QoS-enabled Ethernet/IP platforms.The further adaptation of these schemes in delayed, distributed EPON settings mandates more careful considerations.Moreover, there are two types of intra-ONU scheduling: strict and non-strict. priority scheduling algorithms. In strictpriority scheduling, a lower-priority queue is scheduled only if all queues with higher priority are empty. However,this may result in a starvation for low-priority traffic or as dubbed in [51], ”light-load penalty”.Non-strict priority scheduling addresses this problem by allowing reported packets (regardless of their priority) to be

5

transmitted first as long as they are transmitted in the allocated TW. In other words, here, the transmission order ofdifferent priority queues is based on their priorities. As a result, all traffic classes have access to the upstream channelwhile maintaining their priorities; which enables fairness in scheduling and statistical multiplexing (by allowing linksharing among all classes of service). Note that inter-ONU control messages for allocating bandwidth to differentONUs are transmitted via the MPCP (multi-point control protocol) access protocol presented in the previous chapter.An ideal scheduler should allow for statistical multiplexing, but guarantee a minimal portion of the available band-width to each priority queues. This is achieved by enabling link sharing among different priority queues.General Processor Sharing (GPS) achieves these goals for relatively small packets traffic (so-called fluid traffic). How-ever, GPS is not applicable in practical systems with limited-size packets since a packet monopolizes the shared linkwhile in service. This shortcoming was addressed in WFQ [45]. WFQ is basically a packet approximation of GPS thatis derived from the ideal case’s bounded by the maximum packet size. It mainly calculates the packet’s start time andrespective finish time under the ideal Global Positioning System (GPS). Consequently, packets are selected from thePQs in ascending finish time order. SCFQ and other schemes were proposed as a modified version of WFQ aiming tosimplify the time calculation. However, these schemes lacked the appromixmation accuracy of the ideal GPS used inWFQ. Start-time fair queueing (SFQ) [49] simplifies WFQ and reduces its computational complexity, by calculatingthe packet-in-service’s start-time instead of the finish time.To cope with the light-load penalty caused by applying strict priority scheduling technique, the authors of [51] pro-posed two methods. The first method involves a two-stage queueing process. Here, the incoming packets after sendingthe REPORT message are placed in the second-stage queue. Consequently, when a new GATE is received, the second-stage queue is emptied first. This however results in an increased average delay for all types of traffic. In the secondmethod, the ”after-report” incoming traffic is estimated, and thus the grant window will be large enough to accom-modate the newly arriving high priority packets. Alternatively, in [26], the intra-ONU scheduler employs priorityscheduling only on the packets that arrive before sending the REPORT message. This scheme eliminates the ”light-load penalty” and allows all services to access the shared medium.Maeda et. al [42] proposed another scheduling discipline called Priority with Insertion Scheduling (PIS), that trans-mits real-time packets when their delay bound doesn’t exceed a defined threshold (i.e., as long as the packets can bedelayed without any detriment). Kramer et. al [41] lately proposed a new hierarchical scheduler that fairly divides theexcessive bandwidth resulting from lightly loaded ONUs among priority queues (PQs) from different ONUs.On the other hand, the authors of [49] proposed a new intra-ONU scheduling scheme named ”Modified Start-Timefair queueing” (M-SFQ) that muses the performance of VBR traffic. Here, the scheduler selects for transmission thequeue with the minimal start time, derived from the head-of-line (HOL) packet in each queue, and synchronized witha Global Virtual Time similarly to SFQ. M-SFQ mainly improves the end-to-end packet delay for Assured Services atthe expense of other classes of service; which is not desirable in comparison to higher priority classes (EF). Moreover,M-SFQ tends to starve the best effort traffic to provide better QoS to AF and EF classes.The authors of [29] studied the shortcomings of M-SFQ and proposed a new intra-ONU scheduling scheme based onthe famous token bucket (TB) traffic regulator for QoS; namely Modified-TB (M-TB). TB can reshape the traffic andlimit the flow of some greedy traffic classes of service whose load are much larger than the preset limitation. M-TBalgorithm assigns the bandwidth in two stages to each queue in the grant window of each ONU at a certain cycle. Theadvantages of M-TB over M-SFQ is its low-complexity (which is a major requirement for intra-ONU schedulers). Inaddition, M-TB can guarantee both the priority and the fairness of the differentiated services while M-SFQ cannot.We have also studied the shortcomings of M-SFQ and proposed in [38, 40] a new intra-ONU scheduler, namelyM-DWRR (Modified Deficit Weighted Round Robin) that is based on the famous DWRR scheduling mechanism. M-DWRR defines weights ”!i” for each CoS queue i; where !i is adaptively set based on either the traffic requirementsor the Service Level Agreement (SLA). Moreover, M-DWRR gives the ONU enough flexibility to whether to set itsweights in each cycle time or statically; which allows the support of QoS in EPON dynamic and bendable to meettraffic requirements. In the first scheduling pass, M-DWRR offers bandwidth to each PQ i according to its !i. The ”un-needed” bandwidth remaining from the first scheduling pass is re-distributed in the second scheduling pass accordingto the same or different (depending of the traffic and queues requirements) sets of weights.M-DWRR has proven to outperform the M-SFQ scheme; as it ensure more fairness to lower priority traffic (e.g, BestEffort). In addition, the overall network throughput is improved than M-SFQ due to the elimination of best effort trafficstarvation.

6

2.3 Intra-ONU Queue Management

Priority queues at the ONU are defined/set by the EPON network architect with no size restriction. However, it isimportant to avoid high packet loss rates in the Internet. When a packet is dropped before it reaches its destination,all of the resources it has consumed in transit are wasted. In extreme cases, this situation can lead to congestioncollapse. Therefore improving the congestion control and queue management algorithms seems important, howeverwas never been addressed in EPONs due the variable queue size available. Nonetheless, nowadays, network plannersand architects pay a big attention to cost reduction when designing a network. Moreover, software solutions andprotocols have always been favored over the hardware ones, due to their lower cost (price and installation). For thatreason, we will suggest a known queue management scheme that is widely used in other networks, that might beadopted in EPONs, aiming at reducing more the cost of installation of this cost-attractive solution (EPON).We investigated the Weighted Random Early Detection (WRED) [54] as a queue management scheme to improve theoverall packet delay performance in EPON. WRED states that a periodic update should be made to the average queuesize (AvQS) in order to achieve a periodic packet drop. Here, although a penalty is being paid, which is droppingpackets, but a higher gain is being achieved which is reducing the packet delay by consenting the queue to perform ona stable rate less than its maximum actual buffer size. The formula of queue update is given by :

AvQS= OAvQS! (1"2"n)+CQS!2"n (1)

where OAvQS is the Old Average Queue Size, n is an exponential weight factor, a user-configurable value, AQS isthe Actual Queue Size and CQS is the Current Queue Size. Note that This update is performed each time a packet isdequeued.In EPON, the ONU buffer size is fixed (e.g 10 Mbytes). However, the AvQS can be defined as a virtual and not anactual size. In other words, the ONU will create an ”illusion” to the incoming packet that its actual size is the updatedone. In this case, if the AvQS is ”virtually” set to full then the incoming packet will be dropped, otherwise it will beenqueued. In this case, the ONU will make sure that its buffer is not being fully utilized and hence packets will beless delayed. If the value of n gets too high, WRED will not react to congestion. That is, packets will be dropped as ifWRED was not in effect. Similarly, for low values of n, the average queue size closely tracks the current queue size.

7

Fig. 2 Conventional DBA Operation

3 QoS-enabled Dynamic Bandwidth Allocation Algorithms (DBAs)

Dynamic Bandwidth Allocation (DBA) is deployed at the OLT to assign transmission bandwidths for the differentONUs sharing the EPON network. DBA uses the services offered by the MPCP protocol to communicate assignedtransmission windows to their appropriate ONUs. In the conventional DBA operation, as shown in Fig. 2, the OLTwaits until all REPORTs from all ONUs are received4 in cycle n"1 to perform the appropriate computation. Conse-quently, the OLT broadcasts MPCP’s GATE messages to grant transmission windows for cycle n.

3.1 Existing Solutions & Schemes

Many schemes have been explored in the area to fairly allocate bandwidth for different classes of services. The authorsin [26] presented a newmethod to ”evenly” distribute the remaining excessive bandwidth overHL ONUs. This schemeresults in a remarkable improvement to the network performance for different classes of services and also allows forstatistical multiplexing traffic into unused bandwidth units. The authors also consider the option of reporting queuesize using an estimator for the occupancy of the high priority queue. Nevertheless, due to the unpredictable behaviorof HL ONUs that varies from one cycle to another where a HL ONU tends to be a ”slightly” HL one; and thus theallocated bandwidth is not being fully utilized, this uniform distribution of excessive bandwidth might not be the bestpossible solution.On the other hand, the authors of [9] proposed a new concept of DBA, where ONUs are divided into two sets, namelybandwidth guaranteed (BG) ONUs (premium subscribers according to the SLAs) and non-bandwidth guaranteed(non-BG) ONUs (subscribers with best-effort service). Here, the bandwidth guaranteed polling scheme (BGP) pro-vides guaranteed bandwidth to BG ONUs while providing best-effort services to non-BG ONUs. However, the pro-posed BGP can not be standardized with the MPCP arbitration mechanism proposed by the IEEE 803.2ah Task Forcefor the reason that in the future emerging PON technologies, each ONU must be capable of provisioning differentiatedservices for different users requirements.Alternatively, the authors of [52] proposed a newDBA to support multimedia services in EPON. Here, incoming traffic

4 Note: REPORT messages can be either sent at the end of the data transmission or at the beginning. However, if the latter is applied, theOLT might be receiving an out-of-date information from the ONUs.

8

to each ONU is buffered into one of the three priority queues (High, Medium and Low). The sizes of these queues arereported to the OLT using an ”upgraded” REPORT message. An inter-scheduler (i.e. at the OLT) is considered wherethe OLT, based on the priority queues sizes, issues grants separately. In particular, the DBA satisfies requests of flowsby priority preference (High first, Medium second and Low last). Then, if all flows are satisfied and additional band-width is still available, the remaining resources are distributed among all priority flows in the same manner. However,strict priority scheduling based on the traffic classes at the PON level may result in starvation of ONUs that have onlylow priority traffic.To overcome this problem, the authors of [10] present a new DBA for multiservice access, namely DBAM. DBAMapplies priority queuing to enqueue the EF, AF, and BE frames, and gives preference to higher-priority traffic. Priority-based scheduling is exploited to schedule the buffered frames, and the schedule interval is the time between sendingREPORT messages. DBAM also employs class-based traffic prediction to take the frames arriving during the waitingtime into account with dynamic and diverse bandwidth requests. In particular, an estimator credit ! , which is theratio of the waiting time of the ONU over the interval length, is estimated and then incorporated in the request forbandwidth of all BF, EF and AF traffic. Multiservice access for different end users is realized by means of class-basedtraffic estimation and SLA-limited bandwidth allocation.Furthermore, the authors of [11] presented a two-layer DBA where the total available bandwidth is allocated on twophases. Here, the OLT allocates bandwidth among different classes of services first, then among ONUs. This schemeprovides a higher priority to the class-level Quality of Service over the ONU-level bandwidth guarantee. However,since subscribers are practically considered non-cooperative entities, ONU-level bandwidth guarantee should be con-sidered first.In addition, the authors of [14] proposed a new GRANT ”pre-allocation” mode for EF traffic named Grant-Before-Report (GBR) and the traditional Grant-After-Report (GAR) mode for both AF and BE traffics. Here, the OLT dividesthe ONU transmission cycle into two sub-cycles; namely DBA sub-cycle (DBA-CL) reserved for EF traffic, andMPCPsub-cycle (MPCP-CL) reserved for AF and EF traffic. A new GATE scheduling mechanism is also presented in [12].This mechanism allocates GRANTs based on the traffic priority rather than the ONU classification (e.g. Round Robin).Here, all high priority traffic (from all ONUs) are granted first (in order to minimize its sensitive delay), and then lowpriority traffic second. This algorithm can be also applied with any DBA.Alternatively, Nasser and Muftah [44] presented another method to schedule CoS traffic in EPON; namely, Class-of-service Oriented Packet Scheduling (COPS). At the OLT, COPS uses two sets of ”credit pools”, where each CoSof each ONU and well as each ONU are regulated. Mainly, COPS schedules the CoS with highest priority first thenmoves to the one(s) of lower priority and henceforth. In the first scheduling round, each ONUwith traffic of the currentCoS is granted up to the number of credits stored for that CoS. To mitigate the unused bandwidth for those grants thatwere not fully satisfied, a threshold reporting scheme is used. At the end of the first round, the unused credits arepooled together and in the next round these unused credits are distributed to the CoS per ONU pairs that were not fullysatisfied. COPS provides lower delay for most CoS as compared to other famous inter-ONU scheduling schemes.Shami et. al [53], for instance, designed a new DBA, so-called the Hybrid Granting Protocol (HGP), that mainly aimsat improving the jitter performance of EF traffic. The hybrid part of HGP is that it sizes the ONU grants based onboth the REPORT messages and a queue prediction mechanism, in order to accommodate all queued traffic at thepoint the grant begins. Moreover, each scheduling cycle is divided into two sub-cycles. The first one is used for thetransmission of EF traffic and the second one for AF and BE traffic. Furthermore, to meat this cycle division, twoseparate REPORTs are sent from each ONU as well as two grants are sent from the OLT to each ONU (one for EFtraffic and the other for AF/BE traffic). Consequently, the grant of EF traffic is computed ”earlier”, and as a result,EF traffic is not only prioritized but also protected from the fluctuations of other types of traffic; hence optimizing itsjitter performance. Nonetheless, this optimization comes at the expense of more guard times per-cycle caused by theseparate grants of each ONU. To mitigate this drawback, a grant for AF/BE traffic is not sent if there is no waitingtraffic, eliminating the need for a guard time for AF/BE traffic.The authors of [55] have also tried to improve the jitter performance of EF traffic by proposing a new Hybrid SlotSize/Rate Algorithm (HSSR). HSSR was able to stabilize the EF packet delay variation by, not only fixing the cyclelength, but also fixing the position of EF traffic grants (At the beginning of the grant frame).

9

Fig. 3 Proposed Cycle Framework

4 Quality of Service Protection and Admission Control in EPON

4.1 Preliminaries

In order to provide sustainable QoS in the access network, bandwidthmanagement on the upstream channel is essential.In order to support and protect the QoS of real time traffic streams, one needs, in addition to bandwidth allocation andservice differentiation, an admission control algorithm which makes decision on whether or not to admit a real-timetraffic stream based on its requirements and the upstream channel usage condition. The problem of QoS protection issignificant because the bandwidth allocated by the OLT to one ONU can only be guaranteed for one cycle. Furthermore,appropriately controlling the admission of real time traffic will preventmalicious users frommanipulating the upstreamchannel by sending traffic into or requesting bandwidth from the networkmore than their SLA. Accordingly, admissioncontrol helps in protecting the QoS of existing traffic and admit new flows only if their QoS requirements can beguaranteed. Note that admission control has never been addressed in EPONs. However, the authors of [43] haveproposed a DBA that is based on the BGP presented in the previous section, where admission control is conducted onan ONU-basis rather than a flow/stream-basis. In other words, an ONU is either accepted as a bandwidth guaranteed(BG) ONU or as a best effort non-BG ONU.In current EPON networks, the bandwidth of the upstream channel is shared among different ONUs using a TDMAscheme; the OLT allocates a transmission bandwidth for every ONU either equals to its bandwidth request fromthe previous cycle, or equals to the minimum bandwidth guaranteed (B min), or equals to the minimum bandwidthguaranteed plus a surplus bandwidth that may remain unused in the cycle. Clearly, the bandwidth of one ONU cannotbe guaranteed and may vary from one cycle to another according to the load at other ONUs.Bandwidth reservation resolves the uncertainty in allocating enough bandwidth resulting from the load variations atdifferent ONUs. Hence, each ONU is required to reserve bandwidth for its real time streams in order to satisfy theirQoS requirements. Once this bandwidth is reserved, the OLT can no longer allocate it to other ONUs. Every ONUis guaranteed a new minimum bandwidth (Bmin) and could be allocated up to a maximum bandwidth (Bmax) in orderto allow other ONUs to receive their share of the channel. Best effort (BE) traffic shares a fraction of the total cycle(Tcycle, Tcycle # 2ms in EPON networks [23]), e.g., ! ! Tcycle where ! < 1. When ! = 0, all the bandwidth of theupstream channel is used to transmit bandwidth guaranteed traffic.The new cycle ((1"!)!Tcycle) is used to provide services for bandwidth guaranteed traffic. This new cycle in turnis divided into two sub-cycles (T1, T2); the OLT computes the minimum bandwidth guaranteed (Bmin) for each ONU

10

using T1 (Bmin = (Tcycle"N!Tg)!"8!N , where " is the transmission speed of the PON in Mb/s, N is the number of ONUs

and Tg is the guard time that separates the TW for every ONUn and ONUn+1) and the ONU has total control overthis bandwidth, while the bandwidth of the second sub-cycle is under the control of the OLT (please refer to Fig.3 for a graphical elaboration, with N = 4). This new system enables us to implement a two-step admission control(AC); the first is a local AC at the ONU and the second is a global AC at the OLT (as explained later). Note that,although the minimum guaranteed bandwidth is under the control of the ONU, the scheduling of various ONUs isstill done centrally at the OLT in order to achieve a collision free access to the upstream channel. The two sub-cyclesare selected of equal length; however, if T1 < T2, then the OLT will have more control over the bandwidth with lessbandwidth guaranteed per ONU. Conversely, the ONU is guaranteed more bandwidth, which may be un-utilized if theload at a particular ONU is not high. Under our assumption of equal lengths for the sub-cycles, we set the maximumbandwidth that a highly loaded ONU can be allocated, Bmax = #! Bmin. For example, when # = 3, a highly loadedONU may or could be assigned a maximum of 2! Bmin from the second half cycle and hence a total of 3! Bmin percycle.For real-time applications, QoS metrics can be predefined in a policy control unit (PCU) and various thresholds couldbe specified/defined. For example, if the expected drop rate or the delay requirement for a certain flow/applicationcannot be be respected, the flow should not be admitted. Admitting such a stream will not only experience a degradedlevel of service, but it will also degrade the QoS of existing streams. Alternatively, best-effort traffic is never rejected.and is always guaranteed a minimal bandwidth (BminBE ). Hence, to achieve these goals, the following two rules shouldnot be violated before and after admitting a new real-time flow:

1. The QoS of each real-time stream (existing or new) should be guaranteed.2. The BE traffic throughput (BEThroughput) $ BminBE

In every cycle, the ONU reports (using the MPCP protocol) to the OLT the BE buffer occupancy for bandwidthallocation in the next cycle; for real-time streams that the ONU has already admitted, the OLT will schedule only theirtransmission since the bandwidth of each stream has already been pre-determined and reserved and it is guaranteedper cycle for the rest of the lifetime of each stream.

Fig. 4 Dual Token Bucket Filter Framework

4.2 Traffic Characteristics and QoS Requirements

Clearly, an admission decision for a new flow arrival should be made according to both admission policies and QoSrequirements supplied often by the application layer at the end users. The set of parameters that characterize the trafficstream vary from one traffic class to another. For example, CBR traffic is non bursty and characterized by its mean datarate (µ), which makes it quite predictable. With respect to QoS, CBR traffic requires stringent packet delays and delayvariations (jitter). Alternatively, VBR traffic is quite bursty and may be characterized by the following parameters [19]:

• Mean Data Rate (µ) in bits per second (bps).• Peak Arrival Data Rate ($ ) in bits per second (bps).• Maximum Burst Size (%) in bits.

11

• Delay Bound (& ) which is the maximum amount of time in units of microseconds allowed to transport a trafficstream (flow) measured between the arrival of the flow to the MAC layer and the start of transmission in thenetwork.

Finally, BE traffic is bursty and requires neither delay requirements nor bandwidth guaranteed (note that network op-erators may set a certain minimum bandwidth that should be guaranteed for BE traffic; e.g., by appropriately adjustingthe value of !).When these parameters are specified by the end-user, the problem left for the admission control unit (ACU, whichis either at the ONU or OLT) is simply to determine whether a new stream should be admitted and whether its QoSrequirements can be guaranteed while the QoS requirements for already admitted streams can be protected. For CBRtraffic, the admission decision is straight forward; if the mean data rate can be supported, then the stream is admitted.Hence, enough bandwidth per cycle should be reserved to guarantee the stream data rate. Here, the average delay ofCBR traffic is guaranteed to be bounded by the length of cycle. For VBR traffic, the ACU may decide to admit astream only if its peak rate can be supported (for the best QoS) or may admit the stream as long as the mean data rateis available [19]. The former approach ends up admitting few streams and the latter approach barely supports QoSfor bursty streams. Therefore, a guaranteed bandwidth based on the traffic parameters could be derived and we usea dual-token bucket for traffic regulation; this dual-token bucket is situated at the entrance of the MAC buffer and isassociated with each stream. Fig. 4 shows the dual token bucket where the bucket size is calculated:

B= %! (1" µ/$) (2)

Accordingly, one can easily determine the arrival process of the stream passing through the filter [19]:

A(t,t+ ') =min($' ,B+ µ') (3)

Where A(t, t + ') is the cumulative number of arrivals during (t,t+ '). The arrival rate curve could be constructedfrom the above equation and is shown in Fig. 5. Therefore, the guaranteed rate for every real-time flow i can be easilyderived from Fig. 5 using the distance formula [19]:

gi =%i

&i+ %i$i

(4)

Since CBR traffic is deterministic and its peak rate is equivalent to its mean rate, therefore its bandwidth guaranteedwill be:

gi = µi (5)

Consequently, a conventional rate based admission control [20] can be used to determine whether a new stream can beadmitted or not. For example, if STWj is the bandwidth (bps) allocated and reserved for ONU j, then a new flow i+1could be admitted if:

g ji+1+h j

(i=1

g ji # STWj (6)

Where h j is the number of real time streams (CBR or VBR) at ONU j. Now, the difficulty stems from the fact that inEPON the bandwidth assigned per ONU is not guaranteed, as mentioned earlier. Hence, we next propose a two stepadmission control scheme that will provide bandwidth guaranteed for each CoS stream.

4.3 Local Admission Control (LAC)

As we discussed earlier, each ONU is guaranteed a minimum bandwidth per cycle, Bmin. Hence, the ONU can lo-cally perform rate-based admission control according to the bandwidth requirement of the new arriving flow and thebandwidth availability. For example, if g jf is the guaranteed rate for the new flow, f , arriving at ONU j, then the band-width requirement (in bytes) per cycle for the new flow is: R j

f = g jf !Tcycle. Therefore, this new flow will be admitted

12

Fig. 5 Guarantee Bandwidth Derivation Graph [19]

according to the following condition:

Rjf +

h j

(i=1

Rjfi # Bmin (7)

Where h j is the total number of flows already admitted by the ONU; R jfi is the bandwidth requirement for a flow f i,

Rjfi = g jfi !Tcycle, and g jfi is the guaranteed rate (bps) for the flow computed according to either equation (4) or (5).

The scheme classifies the arriving flow into BE traffic or real-time traffic. If it is BE, then the traffic is admitted. Oth-erwise, the ONU will derive the guaranteed rate and check equation (7). If (7) holds, then the ONU will conditionallyadmit the flow and monitor its QoS for a predefined number of cycles (e.g., for 20 ms). If the QoS requirements of thenewly admitted flow are satisfied and the QoS of existing flows remain intact, then the flow is admitted. Otherwise,the flow is dropped.

4.4 Global Admission Control (GAC)

When a flow f cannot be admitted locally at the ONU (due to bandwidth insufficiency), the ONU reports the arrivalof a new flow to the OLT. The OLT may admit this new flow only if there is bandwidth available in the second sub-cycle (T2) and if the ONU sending the request has not been allocated more than Bmax. Hence, the OLT maintains avariable for every ONU designating the bandwidth allocated so far to this ONU, B j

alloc = (h ji=1R

ji , where R

ji denotes

the bandwidth guaranteed for already admitted h j flows for ONU j. The OLT maintains as well another variable thatindicates the bandwidth that is still available, Bavail , (i.e., not committed yet) in T2. The new flow may be admitted ifthe following two conditions (8a, 8b) hold simultaneously:

Rjf +

h j

(i=1

Rji # Bmax (8a)

Rjf # Bavail (8b)

Upon admitting a new flow, the OLT will reserve additional bandwidth for ONU j and update accordingly the totalavailable bandwidth: Bavail = Bavail"Rj

f .Similarly, the OLT performs the above algorithm for every admission request of a new flow at any ONU. A flow will

13

be rejected if at least one of the above two conditions is not satisfied. If both conditions are satisfied, then the OLT willconditionally admit the new flow and monitor its QoS parameters for the subsequent n cycles in order to determinewhether it finally should admit the flow. When a flow leaves the network, the ONU reports to the OLT and the latterwill update the available bandwidth accordingly: Bavail = Bavail +Rj

f .

4.5 Issues and Solutions

In the proposed AC scheme, every real time stream is provided a guaranteed bandwidth that is computed based onthe guaranteed rate of the flow and is reserved and fixed per cycle. The OLT then allocates a transmission windowthat encompasses all the guaranteed bandwidth for every ONU per cycle. A subtle issue which may arise is due to thestatistical nature of real-time traffic and hence guaranteeing bandwidth per flow per cycle may ultimately waste thebandwidth. In other words, if one ONU is being reserved bandwidth for a particular flow per cycle and has no trafficfrom this flow to transmit, then this bandwidth is not utilized and wasted. This issue arises because the allocationbecame static (i.e., reservation) and not dynamic as in traditional EPON systems, where the bandwidth is allocatedon demand. Moreover, if a flow had more bytes to be sent than the reserved ones (i.e., guaranteed), then our purposeon providing bandwidth guaranteed in every cycle will be unsuccessful. This is because estimating the bandwidthrequirement for a flow based on its guaranteed rate does not accurately reflect the real nature of the traffic, especiallywith respect to the arrival of its packets in a short period of time (i.e., the short length of the cycle) and hence theinefficiency of the bandwidth prediction and reservation.To resolve the above problems, we propose a two-branch solution. In the first branch, the OLT selects a super-cycle(Tsc = ) !Tcycle, where ) is a constant) instead, and every admitted real-time flow is now guaranteed a bandwidth perTsc. The purpose of this proposal is to mitigate the inefficiency of the bandwidth reservation caused by the short-timeprediction, and thus a more accurate bandwidth estimation will take place. Here as before, the period (1"!)!T sc isdivided into two periods, T1 and T2. Each ONU is now guaranteed a bandwidth of Bnewmin which is computed based onT1. The OLT controls the remaining bandwidth of the super-cycle. Upon the arrival of a new flow f at ONU j withbandwidth guaranteed B f

g , the flow is either admitted/rejected locally at the ONU or globally by the OLT, as describedearlier.In the second branch, we ensure that the reservation does not waste any bandwidth. Here, we apply a crediting systemwhere each flow’s estimated bandwidth is saved as credits at the OLT. In other words, every time a flow is admitted,the OLT will be informed and it will compute/estimate a total credit (number of bytes available per Tsc for this flow)Cjfi = Bfi

g !Trsc, where T rsc is the period between the arrival of the flow and the end of the current super-cycle. The OLTmaintains as well a total credit per type of traffic (C j

CBR for CBR and CjVBR for VBR) per ONU; for example,C

jCBR =

(Nji=1C

jfi where Nj is the number of CBR flows at ONU j. Now, in every cycle, the OLT deducts the requested/allocated

bandwidth of this flow from its reserved credits until the time of a new super-cycle. At this point, the credits are resetto the estimated ones. Next, we will explain how this solution will help in designing a DBA with effective reservationscheme.

4.6 Admission Control-enabled Dynamic Bandwidth Allocation Scheme (AC-DBA)

To apply the solutions proposed in the pervious Section, we propose a new hybrid DBA that will perform both band-width allocation and reservation at the same time. As any conventional DBA, the ONU reports to the OLT, in everycycle, its buffer occupancy (QCBR(n" 1), QVBR(n" 1), and QBE(n" 1), where n is the cycle number) and requeststransmission bandwidth accordingly. However here, the OLT will allocate bandwidth to each CoS at each ONU ac-cording to its available credit in the current super-cycle, as well as based on the requests received from other ONUs.Let Aj

CBR(n), AjVBR(n), A

jBE(n) be the bandwidth allocated for ONU j; then we have:

14

N

(j=1

(AjCBR(n)+Aj

VBR(n)) # Bcycle"Ttgt" (N!BminBE ) (9)

N

(j=1

AjBE(n) # N!BminBE (10)

where Bcycle is the total bandwidth available in a Tcycle and T tgt is the total guard time (in bytes) between ONUstransmissions and BminBE is the minimum bandwidth guaranteed (in bytes) for Best Effort traffic computed as follows:

BminBE =Tcycle! !!Tsc

N8!Tsc

! " =Tcycle!!8!N

! " (11)

where " is the PON speed (1Gbps).Every time the OLT allocates bandwidth to one ONU, it will adjust the available credit for every CoS accordingly:CjCBR(n) =C j

CBR(n"1)"A jCBR(n). The credit for VBR traffic is updated similarly. If the ONU has run out of credits,

then the OLT does not allocate any bandwidth for this CoS at this ONU during this super cycle.As for the computation of the available bandwidth for each CoS, the OLT waits until all requests (R(Q j

CBR(n" 1)+QjVBR(n"1)+Qj

BE(n"1)) are received from all ONUs. If (Nj=1(QjCBR(n"1)+Qj

VBR(n"1))# Bcycle"Ttgt"N!BminBE ,then A j

CBR(n) = min(QjCBR(n" 1),C

jCBR(n" 1)); similarly for VBR traffic, A

jVBR(n) = min(Qj

VBR(n" 1),CjVBR(n"

1)) and their credits (for both CBR and VBR) are updated accordingly. Otherwise, the OLT will compute the totalguaranteed bandwidth, B j, for each ONU j as follows:

Bj(n"1) =Rj(n"1)! (Bcycle"Ttgt" (N!BminBE ))

(Nj=1R j(n"1)

(12)

where R j(n"1) = QjCBR(n"1)+Qj

VBR(n"1). Then the OLT allocates bandwidth as follows:

AjCBR(n) = min(Qj

CBR(n"1),CjCBR(n"1)) (13a)

AjVBR(n) = min(B j(n"1)"Qj

CBR(n"1),CjVBR(n"1)) (13b)

Next, the OLT will allocate bandwidth to BE traffic based on the requests received from the ONUs. The total BEbandwidth per cycle is BBE = N!BminBE , which is shared by all ONUs. Note, however, if (

Nj=1(A

jCBR(n)+Aj

VBR(n)) #Bcycle"Ttgt"N!BminBE , then the total bandwidth available for BE traffic becomes:

BBE = N!BminBE +(Bcycle"Ttgt"N

(j=1

(AjCBR(n)+Aj

VBR(n))) (14)

If QjBE(n" 1) # BminBE , then A

jBE(n) = Qj

BE(n" 1). Otherwise, the OLT will allocate to the ONU requesting less thanBminBE and will compute the excess bandwidth from these ONUs to distribute them to other ONUs requesting more BEtraffic. Accordingly, if Q j

BE(n" 1) > BminBE , then AjBE(n) = BminBE + * j, where * j is the excess bandwidth allocated for

ONU j:

* j =! j!BremBE (n)

!t(15)

where ! j = QjBE(n"1)"BminBE , !t = (Nj=1! j, and BremBE (n) is the remaining bandwidth in the cycle n after allocating

all ONUs bandwidth for their BE traffic such that :

BremBE = BBE " (N"L)!BminBE "L

(j=1

QjBE(n"1) (16)

15

Now, in order to prevent the waste of bandwidth and control the allocation of surplus to various ONUs, the excessbandwidth allocated for the BE traffic at a highly loaded ONU (* j) is computed as follows:

* j =min(* j,! j) (17)

0 1000 2000 3000 4000 5000 6000 7000 8000 9000 100000

50

100

150

200

250

300

Time (ms)

Num

ber o

f Flo

ws

Simulations Traffic Model

CBRVBRBE

Fig. 6 Traffic Model Used for the AC framework

4.7 Performance Evaluation

The evaluation of the AC framework is based on the second simulation model that is, to the best of our knowledge, thefirst EPON simulation model that truly models the real Internet traffic (that arrives to the ONUs as flows or streams);and that enables the support of AC in EPON.We begin by testing the behavior of our admission control by showing in Fig. 7 the number of admitted real-timetraffic streams. As shown, our system reaches saturation (i.e., no more real-time flows can be admitted in the network)at time 7000 ms. As we continue generating real-time flows until 7500 ms, all the real-time flows arriving afterwardsare rejected. However, this does not mean that no flows were rejected earlier since conditions (12) or (13a) and (13b)need to be respected to admit a new arriving real-time flow otherwise a flow is rejected. The figure shows that starting450 ms as they arrive are rejected.Next, we study the performance of real-time traffic by measuring the instantaneous average packet delays. To reducethe measurements complexity, we choose the sampling period T = Tsc = 500 ms. Figs. 8 and 9 show these mea-surements, with admission control (i.e., AC-DBA) and without admission control (using M-DWRR (set 1) and StrictPriority (SP) schedulers). Clearly, using M-DWRR and SP schedulers, CBR traffic shows the optimal performancewhere its average packet delay remains under 2 ms even when the load continuously increases (i.e., as the simulationtime continues to increase). This shows the advantage of M-DWRR; that is, although it divides the cycle among theCoS queues based on their assigned weights, it also provides an optimal performance for CBR traffic. This is due tothe fact that the assigned weights are adaptively set based on the QoS requirements. On the other hand, using the StrictPriority scheduler that always selects packets from higher priority queue until satisfied (i.e., until it is empty), CBRtraffic will accordingly exhibit the best performance. As for AC-DBA, it makes sure to satisfy the QoS requirementsdefined previously (in terms of delay and throughput) by crediting every real-time traffic the appropriate bandwidthand reserving it in every super-cycle/cycle, since a CBR flow is admitted only if its guaranteed bandwidth is assured

16

0 1 2 3 4 5 6 7 8 9 100

50

100

150

200

250

300

350

400

450

500

Time (seconds)

Num

ber o

f Rea

l!Ti

me

Adm

itted

Flo

ws

Real!Time Admitted Flows vs. Time

With Admission ControlWithout Admission Control

Fig. 7 Admission Control Behavior

in every cycle. Hence, AC-DBA maintains a CBR average packet delay of 2%4 ms with a noticeable slight decreasepattern that repeats every super-cycle. This is due to the fact that the credits assigned in one super-cycle n were con-sumed just before the ”credit refilling” for super-cycle n+1, which is due to the statistical multiplexing nature of CBRtraffic; and thus, the delay decreases after the latter operation.As for the VBR traffic, as shown in Fig. 9, AC-DBAmaintains its delay performance to meet the specified QoS require-ments of the stream (i.e., 25%30 ms) while the delay witnesses an exponential increase under both adopted schedulers(Figs. 9(b) and 9(c)), i.e., a system that does not deploy any admission control. This behavior highlights the needfor the application of admission control in EPON, because when the system reaches saturation (as described earlier)and all the arriving streams are admitted, the performance is no more maintained; more specifically, no bandwidth isguaranteed for all types of traffic and the QoS requirements are no longer met (not only for new application but forexisting applications as well). On the other hand, the deployment of AC in EPON allows for a bandwidth guaranteedservice with guaranteed protected QoS.We further investigate our AC framework by measuring/monitoring the throughput of one flow from each CoS (i.e.,CBR, VBR) with AC (i.e., AC-DBA) and with no AC (i.e. M-DWRR and SP) in Fig. 10. As shown and expected,the selected CBR flow exhibits the same performance with and without AC, while the selected VBR flow shows adifferent behavior. Here, the VBR flow with AC, maintains its derived 4 Mbps throughput through out the simulation,even after the system reaches saturation. On the other hand, when no AC is applied, the VBR flow does not show astable throughput behavior. Moreover, when the system reaches saturation, the throughput of the VBR flow starts de-creasing. This is due to the fact that when more real-time flows are admitted and no AC is applied, the bandwidth thatwas guaranteed for the already admitted flows (before saturation) is now shared by more flows. Hence, the guaranteedbandwidth is no longer guaranteed for the already admitted flows and for the newly admitted ones. This again showsthe need for admission control in EPON to stabilize and guarantee the throughput for all admitted flows and reject theflows that will break this theme. This, in real and practical settings, will deny all malicious users from monopolizingthe bandwidth provided; and at the same time, it will allow for bandwidth protection to the bandwidth assigned forother well-behaved users.As for the BE traffic, our concern is to guarantee a minimum total throughput that meets rule (2) in the AC scheme. Forthat reason, we measure its total throughput rather than the per-flow throughput as we did for CBR and VBR traffic.Here, the BE throughput increases to reach a total of & 400 Mbps under all schemes (i.e., with AC and with no AC)when the load is low and decreases when more flows are admitted into the network. However, when the system reachessaturation, AC-DBA makes sure to preserve the minimum pre-defined throughput; while with M-DWRR and StrictPriority schedulers, the throughput is not guaranteed and hence the pre-defined throughput is no longer respected. Nev-ertheless, M-DWRR still provides a minimum throughput (which is one of the advantages of M-DWRR) by forcing

17

the weight policy, while it reaches a very low one (& 0Mbps) with SP; a phenomenon known as BE traffic starvation.

0 10 20 30 40 50 60 70 80 90

1

2

3

4

5

6

7

8

9

x 10!3

Time (x0.1 seconds)

Pack

et D

elay

(sec

onds

)

CBR Average Delay vs. Time

CBR Average Delay every "T"Curve Fitting

(a) AC-DBA

0 20 40 60 80 100 120 140 160 180 2000

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5x 10!3

Sampled Time (seconds)

CB

R P

acke

t D

elay

(se

con

ds)

Average Sampled Packet CBR Delay vs. Time

CBR Avereage Delay every "T"Curve Fitting

(b) NO-AC using M-DWRR

0 20 40 60 80 100 120 140 160 180 2000

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5x 10!3

Sampled Time (seconds)

CB

R A

vera

ge

Pac

ket

Del

ay (

seco

nd

s)

Sampled CBR Average Packet Delay vs. Time

CBR Average Delay every "T"Curve Fitting

(c) NO-AC using Strict Priority

Fig. 8 CBR Packet Delay

0 10 20 30 40 50 60 70 800

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

Time (x0.1 seconds)

Pack

et D

elay

(sec

onds

)

VBR Average Delay vs. Time

VBR Average Delay every "T"Curve Fitting

(a) AC-DBA

0 20 40 60 80 100 120 140 160 180!0.05

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

Sampled Time (seconds)

Aver

age

Pack

et V

BR D

elay

(sec

onds

)

Average Sampled VBR Packet Delay vs. Time

VBR Average Delay every "T"Curve Fitting

(b) NO-AC using M-DWRR

0 20 40 60 80 100 120 140 160 180!0.05

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

Sample Time (seconds)

VB

R A

vera

ge

Pac

ket

Del

ay (

seco

nd

s)

VBR Sampled Average Packet Delay vs. Time

VBR Average Delay every "T"Curve Fitting

(c) NO-AC using Strict Priority

Fig. 9 VBR Packet Delay

18

0 1 2 3 4 5 6 7 8 9 100

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

Time (seconds)

CB

R F

low

Th

rou

gh

pu

t (M

bp

s)

CBR Flow Throughput vs. Time

AC!DBANO AC ! M!DWRRNO AC ! Strict Priority

(a) CBR Flow Throughput

0 1 2 3 4 5 6 7 8 9 10!1

0

1

2

3

4

5

6

7

8

Time (seconds)

VB

R F

low

Th

rou

gh

pu

t (M

bp

s)

VBR Flow Throughput vs. Time

AC!DBANO AC ! M!DWRRNO AC ! Strict Priority

(b) VBR Flow Throughput

0 1 2 3 4 5 6 7 8 9 100

50

100

150

200

250

300

350

400

450

Simulation Time (seconds)

Th

rou

gh

pu

t (M

pb

s)

BE traffic Throughput vs. Time

AC!DBANO AC ! M!DWRRNO AC ! Strict Priority

(c) BE Total Throughput

Fig. 10 CoS Traffic Throughput

19

5 Conclusion

Ethernet Passive Optical Networks (EPONs) have emerged as the best solution for the last mile bottleneck. EPONsnot only provide high speed bandwidth for the emerging QoS applications, but also offer high reliability, maintenance,low cost and most importantly an easy spatial upgrade that can meet the continuous Internet growth in terms of usersand bandwidth demand.Although standardized, EPON carries many ”yet-to-be-solved” problems such as, efficient bandwidth allocation (inter-and intra-ONU schedulers), fairness, QoS protection and many more. From this point, EPON stands as an interestingto-be-investigated technology and is still exposed to intensive research from both the industry and academia. In thischapter, we addressed these problems and overviewed various solutions that aim to improve the overall performancein the access network. Moreover, we presented a novel decentralized bandwidth intra-ONU scheduler (M-DWRR) thatallows for a unique ONU-gripping to QoS traffic by adaptively setting weights for the different CoS.We also addressed the QoS protection problem in EPONs, and motivated the need for Admission Control in order to”respect” QoS requirements and enable an efficient bandwidth allocation. To implement this solution, we presented thefirst and complete EPON framework that supports the application of admission control (AC). This framework resolvesthe guaranteed bandwidth issue for the QoS applications and protects the performance of on-going admitted traffic.The AC framework implements a two-stage admission control, namely locally at the ONU and globally at the OLT,with all its rules. Moreover, we have supported this framework with the first hybrid AC-enabled DBA that performsboth bandwidth allocation and reservation. We have also presented the first simulation model that is designed to testthis framework. Our AC framework showed that the application of admission control in EPON is becoming crucial forproviding bandwidth guaranteed for the emerging QoS applications, and their protection against the malicious usersthat aim on monopolizing the bandwidth provided.

20

References

1. OECD Organisation For Economic Co-Operation and Development, ”OECD Communications Outlook”, June 2003.2. F. J. Effenberger, H. Ichibangase, and H. Yamashita, ”Advances in broadband passive optical networking technologies”, IEEE Com-

munications Magazine, vol. 39, no. 12, December 2001, pp.118-124.3. S. Hardy, ”Verizon staffers find fiber-to-the-home cheaper than copper”, Lightwave, PennWell, vol. 17, no. 134 pp. 1, December 2000.4. Glen Kramer, ”Ethernet Passive Optical Networks”, Addison-Wesley, 2005.5. H. Shinohara, ”Broadband Access in Japan: Rapidly Growing FTTH Market”, IEEE Communication Magazine, September 2005, pp.72-78.

6. Dror Salee, ”EPON Ushers In a New Generation of Bandwidth-Intensive Applications”, EPON Tutorial, Passave Inc.7. J. Zheng, H. T. Mouftah, ”An Adaptive MAC Polling Protocol for Ethernet Passive Optical Networks”, Proc. IEEE ICC 05, May 2005.8. Su-Il Choi and Jae-Doo Huh., ”Dynamic bandwidth allocation algorithm for multimedia services over ethernet PONs”, ETRI Journal,24(6):465468, December 2002.

9. Maode Ma, Yongqing Zhu, and Tee Hiang Cheng, ”A bandwidth guaranteed polling MAC protocol for ethernet passive optical net-works”, Proc. of IEEE INFOCOM, volume 1, pages 2231, March 2003. San Francisco, CA.

10. Yuanqiu Luo, and Nirwan Ansari, ”Bandwidth allocation for multiservice access on EPONs”, IEEE Communications Magazine, vol.40, Issue 2, Feb. 2002.

11. Jing Xie, Shengming Jiang, and Yuming Jiang, ”A Dynamic bandwidth allocation scheme for Differentiated Services in EPONs”,IEEE Communications Magazine, vol.42 Issue 8, Aug. 2004.

12. Ahmed E. Kamal and Brian F. Blietz, ”A Priority Mechanism for the IEEE 802.3ah EPON”, IEEE ICC’05, Seoul, Korea, May 2005.13. Hee-Jung Byun, Ji-Myung Nho, and Jong-Tae Lim, ”Dynamic bandwidth allocation algorithm in ethernet passive optical networks”,

Electronics Letters, 39(13):10011002, June 2003.14. Xiaofeng Bai, A. Shami, and C. Assi, ”A Hybrid Granting Algorithm for QoS Support in Ethernet Passive Optical Networks”, IEEE

International Conference on Communications (ICC’05), Seoul, Korea, May 2005.15. Yongquing Zhu, Maode Ma, and Tee Hiang Cheng, ”A novel multiple access scheme for Ethernet Passive Optical Networks”, GLOBE-

COM 2003 - IEEE Global Telecommunications Conference, vol. 22, no. 1, Dec. 2003, pp. 2649-2653.16. A. Shami, X. Bai, C. Assi, and N. Ghani, ”New dynamic bandwidth allocation scheme in ethernet passive optical access networks”,

Proc. IEEE ICPP 2004, Aug. 2004, pp 371-378.17. ”Virtual Bridged Local Area Networks”, IEEE Standard 802. Iq., 1998.18. Jin Liang, Klara Nahrstedt, ”RandPeer: Membership Management for QoS Sensitive Peer-to-Peer Applications”, Proc. IEEE INFO-

COM, Barcelona, April 2006.19. Chun-Ting. Chou, Sai Shankar N., Kang G. Shin, ”Achieving Per-Stream QoS with Distributed Airtime Allocation and Admission

Control in IEEE 802.11e Wireless LANs”, Proc. IEEE INFOCOM, April 2005.20. Sugih Jamin, Peter Danzig, Scott Shenker, and Lixia Zhang ”A Measurement-based Admission Control Algorithm for Integrated

Services Packet Networks”, IEEE/ACM Transactions on Networking, 5(1):56-70. Feb. 199721. Chuan H. Foh, Lachlan L. Andrew, Moshe Zukerman, and Elaine Wong, ”Full-RCMA: a high utilization EPON”, In Proceedings of

OFC, volume 1, pages 282284, March 2003. Atlanta, GA.22. M. Garcia, D. F. Garcia, V. G. Garcia, and R. Bonis, ”Analysis and Modeling of Traffic on a Hybrid Fiber-Coax Network”, IEEE JSAC,

Vol. 22, No. 9, Nov. 2004.23. G. Kramer, B. Mukherjee, and G. Pesavento, ”Ethernet PON (ePON): Design and Analysis of an Optical Access Network”, Photonic

Network Communications, vol. 3 (2001) pp. 307-319.24. G. Kramer, B. Mukherjee, and G. Pesavento, ”Ethernet Passive Optical Network (EPON): Building a Next-Generation Optical Access

Network”, IEEE Communications Magazine, Feb. 2002, pp. 66-7325. M. P. McGarry, M. Maier, and M. Reisslein, ”Ethernet PONs: A survey of dynamic bandiwdth allocation (DBA) algorithms”, IEEE

Communications Magazine, 42(8):S8S15, August 2004.26. Chadi M. Assi, Yinghua Ye, Sudhir Dixit, and Mohamed A. Ali, ”Dynamic bandwidth allocation for Quality-of-Service over Ethernet

PONs”, IEEE JSAC, vol. 21 (2003) pp. 1467-1477.27. IEEE 802.3ah, Ethernet in the First Mile Task Force, http://www.ieee802.org/3/efm/index.html.28. R D. Feldman, E. E. Harstead, S. Jiang, T H.Wood, and M. Zirngibl, ”An evaluation of architectures incorporating wavelength division

multiplexing broad-band fiber access”, IEEE Journal of Lightwave Technology, pp. 1546-1558, 1998.29. J. Chen, B. Chen, and S. He, ”A Novel Algorithm for Intra-ONU Bandwidth Allocation in Ethernet Passive Optical Networks”, IEEE

Communications Letters, vol. 9, no. 9, Sep. 2005.30. M. Zirngibl, C. H. Joyner, L. Stulz, C. Dragone, H. Presby, and I. Kaminow, ”LARNET: a local access router network”, IEEE Photonics

Tech. Letters, pp. 215-217, 1995.31. N. J.Frigo, et.al., ”RITENet: a passive optical network architecture based on the remote interrogation of terminal equipment”,

http://ieeexplore.ieee.org, 1994.32. G. Mayer, M. Martinelli, A. Pattavina, and E. Salvadori, ”Design and cost performance of the multistage WDMPON access networks”,

Journal of Lightwave technology, IEEE, pp. 121-142, 2000.33. J.D. Angelopoulos, N. Lepidas, E. Fragoulopoulos, and I. Venieris ”TDMA Multiplexing of ATM cells in a residential access super-

PON”, IEEE JSAC, pp. 1123-1133, 1998.34. I. Van de Voorde, C. M. Martin, J. Vandewege, X. Z. Oiu, ”The superPON demonstrator: an exploration of possible evolutionpaths for

optical access networks”, IEEE Communications Magazine, Feb 2000 Volume: 38, Issue: 2 On page(s): 74-82

21

35. G. Kramer, B. Mukherjee, and G. Pesavento, ”IPACT: A Dynamic Protocol for an Ethernet PON (EPON)”, IEEE CommunicationsMagazine, vol. 40, Issue 2, Feb. 2002.

36. Ahmad R. Dhaini, Chadi M. Assi, Martin Maier, and Abdallah Shami, ”Dynamic Wavelength and Bandwidth Allocation in HybridTDM/WDM Ethernet Passive Optical Networks (EPONs)”, IEEE/OSA Journal of Lightwave technology, under review.

37. Ahmad R. Dhaini, Chadi M. Assi, and Abdallah Shami, ”Dynamic Bandwidth Allocation Schemes in Hybrid TDM/WDM PassiveOptical Networks”, In Proc. IEEE Consumer Communications and Networking Conference (CCNC2006), Las Vegas, Jan. 2006.

38. Ahmad R. Dhaini, Chadi M. Assi, Abdallah Shami, and Nasir Ghani, ”Adaptive Fairness through intra-ONU scheduling in EthernetPassive Optical Networks”, In Proc. IEEE International Communications Conference (ICC2006), Istanbul, June 2006.

39. Ahmad R. Dhaini, Chadi M. Assi, and Abdallah Shami, ”Quality of Service in TDM/WDM Passive Optical Networks (EPONs)”,IEEE Symposium on Computers and Communications (ISCC06), Sardinia, Italy, June 2006.

40. Ahmad R. Dhaini, Chadi M. Assi, Martin Maier, and Abdallah Shami, ”Per-Stream QoS and Admission Control in Ethernet PassiveOptical Networks (EPONs)”, IEEE Journal of Lightwave Technology (JLT), 25, 1659-1669 (2007).

41. G. Kramer et al., ”Fair Queuing with Service Envelopes (FQSE): A Cousin-Fair Hierarchical Scheduler for Subscriber Access Net-works”, IEEE JSAC, pp. 1497-1513, Oct. 2004.

42. M. Ma, L Liu, and H. Cheng, ”Adaptive scheduling for differentiated services in the ethernet passive optical networks”, IEEE ICCS2004, pp. 102-106.

43. M. Ma, L Liu, and H. Cheng, ”A systematic scheme for multiple access in ethernet passive optical access network”, IEEE/OSA JLT,vol. 23, no. 11, pp. 3671-3682, Nov. 2005.

44. H. Nasser and H. Muftah, ”A joint-ONU interval-based dynamic scheduling algorithm for ethernet passive optical networks”,IEEE/ACM Transactions on Networking, vol. 14, no. 4, pp 889-899, Aug. 2006.

45. A. Demers, S Keshav, and S. Shenker, ”Analysis and simulation for a fair queueing algorithm”, Internetworking: Research and Expe-rience, vol.1, no.1, Sep. 1990.

46. J. Zheng, H. T. Mouftah, ”Media Access Control for Ethernet Passive Optical Networks: An Overview”, IEEE CommunicationsMagazine, Feb. 2005, pp. 145-150.

47. W. Leland, M. Taqqu, W Willingler, and D. Wilson, ”On the Self-similar Nature of Ethernet Traffic (Extended Version)”, IEEE/ACMTransactions on Networking, pp. 1-15 Feb. 1994.

48. Glen Kramer, ”Synthetic traffic generation”, C++ source code version 1, http://wwwcsif.cs.ucdavis.edu/ kramer/research.html.49. N. Ghani, A. Shami, C. Assi, and Y. Raja, ”Intra-ONU Bandwidth Scheduling in Ethernet Passive Optical Networks”, IEEE Commu-

nication Letters, November 2004.50. Chuck Semeria, ”Supporting Differentiated Service Classes: Queue Scheduling Disciplines”, white paper, Juniper Networks, January

2002.51. G. Kramer, B. Mukherjee, S. Dixit, Y. Ye, and R. Hirth, ”Supporting differentiated classes of service in Ethernet passive optical

networks”, OSA Journal of Optical Networking, vol. 1, no. 8/9, pp. 280-298, August 2002.52. Su-Il Choi and Jae-Doo Huh., ”Dynamic bandwidth allocation algorithm for multimedia services over ethernet PONs”, ETRI Journal,

24(6):465468, December 2002.53. A. Shami, X. Bai, C. Assi, and N. Ghani, ”Jitter Performance in Ethernet Passive Optical Networks,” IEEE Journal of Lightwave

Technology, Vol. 23, No. 4, pp. 1745-1753, April 2005.54. R. Stankiewicz and A. Jajszczyk, ”Analytical Models for Multi-RED Queues Serving as Droppers in DiffServ Networks”, IEEE

GLOBECOM 2007, pp. 2667-2671, Nov. 2007.55. F. An, Y.L.K. Lim, L. White, and L. Kazovsky, ”A new dynamic bandwidth allocation with quality of service in ethernet passive

pptical networks,” IEEE WOC 2003, Vol. 3, pp. 165-169, July 2003.56. Virtual Bridged Local Area Networks, IEEE Standard 802. Iq., 1998.57. Chuan H. Foh, Lachlan L. Andrew, Moshe Zukerman, and Elaine Wong, ”Full-RCMA: a high utilization EPON”, In Proceedings of

OFC, volume 1, pages 282284, March 2003. Atlanta, GA.58. ITU-T Recommendation G.1010 (2001), ”End-user multimedia QoS categories”.

22