michael mcgarry - utep college of engineeringengineering.utep.edu/docs/ath ece computer engr m...

MICHAEL MCGARRY

1009 Caxton Lane

Brunswick, OH 44212

(480) 206-0814

November 29, 2009

Search Committee

Department of Electrical and Computer Engineering

University of Texas, El Paso

El Paso, TX

Dear Search Committee,

I am writing in regard to the position of Assistant Professor of Electrical and Computer

Engineering (CompEng) advertised at Academic Keys. My current academic

appointment, as an Assistant Professor, is with the Electrical and Computer Engineering

department at the University of Akron in Ohio. My desire is to return to the Southwest and

I feel El Paso would be a perfect fit for me.

My research is in the field of computer networks. Specifically, I have made contributions

to two areas: 1) Ethernet Passive Optical Networks (EPONs), and 2) bandwidth

forecasting. My research in EPONs began during my graduate study at Arizona State

University while my research in bandwidth forecasting began last year at the University of

Akron. My most notable research accomplishments are: 1) the IEEE Communications

Society 2009 Best Tutorial Paper Award, and 2) several highly cited articles. My research

interests most closely align with Prof. Virgilio Gonzalez. I look forward to the possibility of

collaborating with him. Please see my CV and research statement for more details.

At the University of Akron, I am currently teaching a two-course track in computer

architecture. Starting in Spring 2010, I am introducing a course in computer networks. My

teaching evaluations were 3.8/4 for Fall 2008 and 3.86/4 for Spring 2009; both scores are

much higher than the average for the College of Engineering at the University of Akron.

My teaching competencies are within digital design, computer architecture, computer

networking, and software engineering. Please see my teaching statement for more

details.

I have several years of industrial experience in telecommunications. Most recently I was a

Senior Staff Scientist at ADTRAN (http://www.adtran.com), a networking equipment

company. At ADTRAN, I led an effort to conduct physical and simulation experiments to

characterize the performance of Ethernet access equipment. As a result of this

experience, I bring valuable industrial insight into the classroom that students appreciate.

I look forward to hearing from you soon. My list of references is contained in my

accompanying CV. Please feel free to contact them for information about my abilities.

Also, please feel free to call me at +1-480-206-0814 to obtain any further information

about me.

Sincerely,

Michael McGarry

Michael McGarry 1009 Caxton Lane

Brunswick, OH 44212

(480) 206-0814

[email protected]

Education

Arizona State University, Tempe, AZ. PhD in Electrical Engineering (August 2007)

Arizona State University, Tempe, AZ. MS in Electrical Engineering (December 2004)

Polytechnic University, Brooklyn, NY. BS in Computer Engineering (May 1997) Award 2009 IEEE Communications Society Best Tutorial Paper Award

Book Chapter M. P. McGarry, and M. Reisslein “Multi-channel Ethernet Passive Optical Networks” Broadband Access Networks: Technologies and Deployments, Springer, ISBN: 9780387921303 Journal Publications (Citation data obtained from Google Scholar)

J. R. Ferguson, M. Reisslein, and M. P. McGarry

“Online Excess Bandwidth Distribution for Ethernet Passive Optical Networks”, OSA Journal of

Optical Networking, vol. 8, no. 4, Apr. 2009 (JCR Impact Factor = 0.941)

M. P. McGarry, M. Reisslein, and M. Maier

“Ethernet Passive Optical Network Architectures and Dynamic Bandwidth Allocation Algorithms“,

IEEE Communications Surveys and Tutorials, Third Quarter 2008 (cited 7 times)

Won the IEEE Communications Society Best Tutorial Paper Award 2009

M. P. McGarry, M. Reisslein, C. Colbourn , M. Maier, F. Aurzada, and M. Scheutzow

“Just-in-Time Scheduling for Multichannel EPONs“, IEEE/OSA Journal of Lightwave

Technology, vol. 26, no. 10, May 2008 (JCR Impact Factor = 2.736, cited 7 times)

M. P. McGarry, M. Reisslein, M. Maier, and A. Keha

“Bandwidth Management for WDM EPONs”, OSA Journal of Optical Networking,

vol. 5, no. 9, Sept. 2006 (JCR Impact Factor = 0.941, cited 9 times)

M. P. McGarry, M. Maier, and M. Reisslein

“WDM Ethernet Passive Optical Networks“, IEEE Communications Magazine,

vol. 44, no. 2, Feb. 2006 (JCR Impact Factor = 2.799, cited 38 times)

M. P. McGarry, M. Maier, and M. Reisslein

“Ethernet PONs: A Survey of Dynamic Bandwidth Allocation (DBA) Algorithms“, IEEE

Communications Magazine, vol. 42, no. 8, Aug. 2004 (JCR Impact Factor = 2.799, cited 94 times)

M. McGarry, p. 2

Conference Publications

M. P. McGarry, and R. Haddad

“A New Approach to Video Bandwidth Prediction”, Proceedings of the

IEEE International Conference on Ultra Modern Telecommunications (ICUMT ’09), St. Petersburg,

Russia, October 12-14, 2009

J. R. Ferguson, M. P. McGarry, and M. Reisslein

“When Are Online and Offline Excess Bandwidth Distribution Useful in EPONs?”, Proceedings of the

ICST Third International Conference on Access Networks (AccessNets ’08), Las Vegas, NV, October

15-17, 2008

M. P. McGarry, M. Reisslein, C. Colbourn, and M. Maier

“Just-in-Time Online Scheduling for WDM EPONs“, Proceedings of the IEEE International

Conference on Communications (ICC ‘07), Glasgow, Scotland, June 24-27, 2007

M. P. McGarry, M. Reisslein and V. R. Syrotiuk

“Access Control in Heterogeneous Multichannel Wireless Networks“, Proceedings of the ICST First

International Conference on Integrated Internet Ad hoc and Sensor Networks (Intersense ‘06), Nice,

France, May 29-31, 2006 Academic Experience

University of Akron (August 2008-Present) Assistant Professor • Received a 3.8/4 teaching evaluation for Fall 2008 and 3.86/4 for Spring 2009. • Established a networking research laboratory. • Directing one doctoral dissertation and three masters theses. Arizona State University (August 2007-August 2008) Adjunct Professor • Research advisor to graduate students engaged in communication networking research. • Led an Ethernet Passive Optical Network bandwidth management research program. • Master’s Thesis Committee member for Amit Aneja (graduated 8/2007) and Jason Ferguson

(graduated 12/2008).

Arizona State University (January 2004-December 2006) Teaching Associate • Held lectures for digital design course (EEE120: Digital Design Fundamentals). • Helped professor develop course material for a graduate level course on Hardware Description

Languages (EEE 517: Hardware Description Languages). • Graded assignments for HDL course (EEE 517). • Responsible for assisting students with their digital design laboratory experiments (EEE 120). • Responsible for grading digital design laboratory experiments (EEE 120). • Responsible for providing digital design tutoring service to students (EEE 120).

Industrial Experience

ADTRAN (July 2007-July 2008) Sr. Staff Scientist • Analyzed network equipment performance via analytical and simulation methods. • Published internal documents to educate systems engineers on the topic of congestion management

techniques for packet switching devices. • Led an effort to analyze the performance of an Ethernet access multiplexer. • Analyzed several congestion management mechanisms via simulation experiments.

M. McGarry, p. 3 PMC-Sierra (April 1999-May 2003)

Software Design Engineer • Part of a team that developed an ATM Adaptation Layer 1 Segmentation and Reassembly (AAL1

SAR) device called the AAL1gator-32. • Part of a team that developed a SONET/SDH OC-48 channelizer device capable of transporting

ATM cells and variable length packets called the S/UNI-MACH48. • Part of a team that designed an Ethernet over SONET/SDH device called the ARROW-24xFE.

TimePlex/Cabletron (May 1998-April 1999) Software Engineer

• Part of a team that developed a Digital Subscriber Line Access Multiplexer (DSLAM).

Yurie Systems (Lucent Technologies) (August 1997-May 1998) Software Engineer

• Part of a team that developed an ATM access multiplexer that carried ATM cells and Frame Relay frames over satellite, PDH and SONET/SDH links.

Research Interests

• Optimization of medium access control (MAC) protocols

• Bandwidth forecasting

• Benchmarking fairness

• Quality of Service (QoS) guarantees

Teaching Competencies

• Communication Networks (including Network Performance Analysis)

• Digital Design (including Hardware Description Languages)

• Computer Architecture

• Software Engineering

Service

Technical Program Committee (TPC) member for:

IEEE International Conference on Communications (ICC) 2010

IEEE Globecom 2010

Peer-reviewer for:

IEEE/OSA Journal of Lightwave Technology

IEEE/ OSA Journal of Optical Communications and Networking

IEEE Transactions on Networking

U Akron ECE Dept. Graduate Policy Committee (GPC) member References

Dr. Martin Reisslein Assoc. Prof. (ASU) email: [email protected] phone: (480) 965-8593

Dr. Charles Colbourn Prof. (ASU) email: [email protected] phone: (480) 727-6631

Dr. Ahmet Keha Assistant Prof. (ASU) email: [email protected] phone: (480) 965-4055


Brunswick, OH 44212

(480) 206-0814

[email protected]

Statement of Research Interests

Improving channel utilization for Ethernet Passive Optical Networks

Ethernet passive optical networks (EPONs) can potentially suffer very poor channel utilization. This poor

channel utilization results from the type of Dynamic Bandwidth Allocation (DBA) algorithm used as well

as the distance distribution of Optical Network Units (ONUs) from the Optical Line Terminal (OLT).

The goal of this research is to increase understanding of what leads to poor channel utilization and develop

techniques to improve or maximize it.

I have developed two technologies that maximize channel utilization through transmission grant

scheduling. The first is a transmission-grant scheduling framework called Online Just-in-Time (Online JIT)

scheduling and the second is a scheduling policy called Shortest Propagation Delay (SPD) first scheduling.

I have proven that SPD can minimize the granting cycle length, which in turn maximizes channel

utilization. Through a wide set of experiments I have displayed under which conditions SPD can provide

dramatic increases in channel utilization.

I have also made a recommendation to practitioners that EPONs be designed for high channel utilization by

increasing OLT to ONU distance diversity.

This research area is continuing with an exploration of ways to provide fair access to EPON bandwidth

capacity while maximizing channel utilization. I am currently seeking sponsorship for this area of research.

Bandwidth forecasting

A core competency for emerging intelligent network protocols is the ability to predict the future bandwidth

requirements of traffic flows. With this predictive ability, intelligent network protocols will be empowered

to more efficiently manage bandwidth resources throughout the network. This more efficient bandwidth

management will lead to significant improvements for all network performance measures.

Previous efforts toward bandwidth forecasting have followed traditional time series forecasting approaches

using a multitude of existing forecast models.

My strategy for solving the bandwidth forecasting problem differs significantly from others in that I am

pursuing unconventional approaches. These unconventional forecasting approaches exploit the unique

characteristics of packet-switched networks.

I have shared some of the details of this research area with Dr. Victor Frost, a program director at NSF, and

he has indicated that this is an area of interest to the NeTS program at the NSF. A grant proposal requesting

NSF sponsorship for this research area is under preparation.


Brunswick, OH 44212

(480) 206-0814

[email protected]

Statement of Teaching Philosophy and Interests

Teaching Philosophy

Teaching is the act of knowledge dissemination. To effectively disseminate knowledge, one must be able to

clearly organize and segment knowledge. Therefore, the best teachers can clearly organize and segment

knowledge.

When preparing learning experiences for my students, I try my best to organize and segment the knowledge

they need to obtain. In further support of effective knowledge dissemination, I introduce concepts (i.e.,

segments of knowledge) one at a time and pause in the class to have students immediately utilize this new

knowledge by means of in-class problem solving exercises.

It is also important for a teacher to inspire students to be interested in the material being covered in the

classroom. To inspire students requires a little showmanship. I try my best to make my lectures highly

interactive with many opportunities for student participation as well as a number of demonstrations that

help to illustrate the concepts being introduced.

Finally, I consider homework a teaching tool, not an assessment tool. Therefore, homework is required and

will be graded but will only constitute a small portion of any course grade.

Teaching Interests

I have a very strong background in digital design, computer architecture, computer networking, and

software engineering.

At the University of Akron, I am currently teaching a two-course sequence in computer architecture.

“Computer Organization and Design: The Hardware/Software Interface” by Patterson and Hennessy is the

textbook used for the first course and “Computer Architecture: A Quantitative Approach” by Hennessy and

Patterson is the textbook used for the second course. In the first course, we begin by learning how to

convert software represented in a high level language such as C into MIPS assembly language instructions.

We learn how both instructions and data (e.g., characters, integers, and floating point numbers) are

represented as sequences of bits that are stored in memory. We continue with an exploration of data path

and control path designs for a CPU. Finally, we end the course with a discussion of computer performance

analysis techniques and hierarchical memory design. In the second course, we learn techniques to exploit

instruction level parallelism in CPU design as well as some advanced memory system optimizations.

I could teach courses in digital design from basic introductory courses through a graduate level course that

studies hardware description languages in detail. I would use “Digital Design: Principles and Practices” by

Wakerly and “Contemporary Logic Design” by Katz as the course textbooks for a two semester

undergraduate course sequence on digital design. I would use “The Verilog Hardware Description

Language” by Thomas and Moore as a course textbook for a graduate course on digital design.

In computer networking, I could teach a senior undergraduate level course in communication networks that

would cover the OSI layering model and discuss in detail some example protocols from each layer of the

OSI model. A top down approach would be the most effective. I would use “Computer Networking: A Top-

Down Approach” by Kurose and Ross as the course textbook. I could teach two graduate level courses on

advanced communication networks. In the first course, the performance analysis of computer networks

using probability theory, queueing theory and simulation experiment analysis would be covered. In the

second course, the design of communication network protocols through the use of optimization theory

would be covered. I would use “Data Networks” by Bertsekas and Gallager, “Introduction to Operations

Research” by Hillier and Lieberman, and “Discrete Event System Simulation” by Carlson, Banks, Nelson,

and Nicol for these two graduate level courses. In addition, I could teach a plethora of special topics

courses in communication networks (e.g., Multimedia Networking, Packet Switching Architectures).

1204 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 26, NO. 10, MAY 15, 2008

Just-in-Time Scheduling for Multichannel EPONsMichael P. McGarry, Member, IEEE, Martin Reisslein, Senior Member, IEEE, Charles J. Colbourn,

Martin Maier, Member, IEEE, Frank Aurzada, and Michael Scheutzow

Abstract—We investigate optical network unit (ONU) grantscheduling techniques for multichannel Ethernet passive opticalnetworks (EPONs), such as wavelength division multiplexed(WDM) EPONs. We take a scheduling theoretic approach tosolving the grant scheduling problem. We introduce a two-layerstructure of the scheduling problem and investigate techniquesto be used at both layers. We present an extensive ONU grantscheduling simulation study that provides: 1) insight into thenature of the ONU grant scheduling problem and 2) indicationof which scheduling techniques are best for certain conditions.We find that the choice of scheduling framework has typicallythe largest impact on average queueing delay and achievablechannel utilization. An offline scheduling framework is not workconserving and consequently wastes channel resources whilewaiting for all ONU REPORT messages before making accessdecisions. An online scheduling framework, although work con-serving, does not provide the best performance since schedulingdecisions are made with the information contained in a singleONU REPORT. We propose a novel online just-in-time (JIT)scheduling framework that is work conserving while increasingscheduling control by allowing the channel availability to drive thescheduling process. In online JIT, multiple ONU REPORTs can beconsidered together when making scheduling decisions, resultingin lower average queueing delay under certain conditions and amore effective service differentiation of ONUs.

Index Terms—Dynamic bandwidth allocation (DBA), Ethernetpassive optical network (EPON), media access control (MAC),scheduling, space division multiplexing (SDM), wavelength divi-sion multiplexing (WDM).

I. INTRODUCTION

CURRENT Ethernet passive optical network (EPON)standards dictate a single channel used for downstream

transmission and a single channel used for upstream trans-mission. The need for more passive optical network (PON)bandwidth capacity will drive up the utilization of multipleupstream and downstream channels. In an effort to providemore bandwidth capacity we can increase the bit-rate as well

Manuscript received June 26, 2007; revised December 14, 2007. This workwas supported in part by the DFG Research Center MATHEON “Mathematics forkey technologies” in Berlin. A preliminary overview of this work has appearedin Proceedings of IEEE International Conference on Communications (ICC),Glasgow, U.K., pp. 2174–2179, June 2007.

M. P. McGarry and M. Reisslein are with the Department of ElectricalEngineering, Arizona State University, Tempe, AZ 85287-5706 USA (e-mail:michael.mcgarry; [email protected]).

C. J. Colbourn is with the Department of Computer Science and Engi-neering, Arizona State University, Tempe, AZ 85287-8809 USA (e-mail:[email protected]).

M. Maier is with the Institut National de la Recherche Scientifique, Montreal,QC H5A 1K6, Canada (e-mail: [email protected]).

F. Aurzada and M. Scheutzow are with the Institute of Mathematics, Tech-nical University Berlin, 10623 Berlin, Germany (e-mail: [email protected]; [email protected]).

Digital Object Identifier 10.1109/JLT.2008.919366

as utilize multiple upstream and downstream channels. Thetransition from increased bit-rate to utilizing multiple channelsfor an increase in bandwidth capacity will be a function of cost.At some point the transition to multiple transmission channelsto increase bandwidth capacity will occur.

Besides an increase in bandwidth capacity there are addi-tional benefits provided by utilizing multiple upstream anddownstream channels. With multiple channels, several opticalnetwork units (ONUs) can transmit concurrently to the opticalline terminal (OLT) thereby lowering the average queueingdelay [1] experienced by the Ethernet frames queued at theONUs. Multiple channels provide a method for dynamic band-width allocation (DBA) algorithms to reserve certain channelsfor certain traffic classes [2]. The discovery and registrationprocess can be kept on a single channel which would allowtransmissions on the other channels to be uninterrupted. Fi-nally, selected channels can be used to provide all-optical OLTbypassing services [3].

Wavelength division multiplexed (WDM)-based multiplechannel PONs were first proposed in the mid-1990s [4]. Re-cently, WDM PON architectures have regained interest as theenabling technologies have become mature [5]–[10]. Spacedivision multiplexing (SDM) is another approach to channelseparation, whereby each fiber strand carries a unique channelor channels. SDM can be combined with WDM for an evenlarger number of channels and service separation.

Dynamic bandwidth allocation (DBA) in multichannelEPONs can be viewed as consisting of grant sizing and grantscheduling. Grant sizing techniques have been examined in[11]–[13]. In this paper, we suppose that the grants are sizedaccording to some existing technique and focus on the grantscheduling portion of the DBA problem for multichannelEPONs. We frame our investigation in the context of sched-uling theory [14]. We model the grant scheduling problem usingstandard scheduling theory notation. We discover solution tech-niques for this model that result in a set of possible schedulingpolicies for producing a schedule given a set of ONU grants.More specifically, we partition the scheduling problem into:1) a scheduling framework and 2) a scheduling policy operatingwithin the adopted scheduling framework. Online scheduling,where the optical line terminal (OLT) schedules grants as soonas a REPORT message is received from an ONU, and offlinescheduling where the OLT waits for REPORT messages fromall ONUs before scheduling grants are the extreme ends of acontinuum of possible scheduling frameworks. The choice ofscheduling framework, as we will explain, depends on the levelof control that is required by the scheduler. A hybrid betweenthe two extremes provides generally the best performance.

This paper is organized as follows. In Section II we review re-lated work on DBA for multichannel EPONs. In Section III wemodel the multichannel EPON ONU grant scheduling problem

0733-8724/$25.00 © 2008 IEEE

MCGARRY et al.: JUST-IN-TIME SCHEDULING FOR MULTICHANNEL EPONS 1205

using standard scheduling theory notation. We then use thismodel to find good scheduling policies and discuss schedulingframeworks. In Section IV we discuss our online just-in-time(JIT) scheduling framework and how some scheduling policiescan be adapted to be used within this framework. In Section Vwe present the results of our simulation study that analyzesthe differences between the different scheduling techniques. Weconclude in Section VI.

II. RELATED WORK

We now discuss the related work on dynamic bandwidth al-location (DBA) for multichannel or WDM EPONs.

A. WDM Variants of IPACT

In [15], Kwong et al. propose a multiple upstream wave-length version of IPACT [12], called WDM IPACT-ST, wherebyST refers to single polling table. This algorithm keeps trackof the wavelength available time of all upstream wavelengths.Upon receiving a REPORT from an ONU, the OLT schedulesthe ONU’s next transmission grant on the next available wave-length. This algorithm assumes each ONU supports all wave-lengths. This scheme is similar to the WDM IPACT we proposedin [16]. In [16] we supported differing ONU WDM architecturesby selecting the next available supported wavelength for sched-uling whereas WDM IPACT-ST does not allow for this flexi-bility. In [17], Clarke et al. propose SIPACT, another WDM ex-tension of IPACT, that is similar to the WDM IPACT in [16].

B. Dynamic Bandwidth Allocation Schemes in Hybrid

TDM/WDM Passive Optical Networks

Dhaini et al. [18] investigate WDM-PONs and dynamicwavelength and bandwidth allocation (DWBA) algorithmsfor two fundamental WDM-PON architectures they label A1and A2. In A1, the ONUs are placed into wavelength sets andcontain a fixed transmitter at the selected wavelength. In A2,the ONUs have tunable transmitters capable of transmissionon several wavelengths. In both A1 and A2, the OLT has anarray of fixed receivers for upstream reception. Dhaini et al.

proceed to investigate the problem of dynamically allocatingbandwidth in both time and wavelength. Their first approach,Static Wavelength Dynamic Time (SWDT), relies on archi-tecture A1. SWDT statically assigns a wavelength channel toall grants from each ONU, this wavelength is the wavelengthsupported by the fixed transmitter at that ONU. Time is thenmanaged using an existing single channel DBA.

Dhaini et al. then propose three variants of a dynamic wave-length and time bandwidth allocation. The first (DWBA-1)schedules ONUs after all REPORT messages have been re-ceived for a cycle. Further, DWBA-1 incorporates “fair” distri-bution of excess bandwidth. The second (DWBA-2) schedulesunderloaded ONUs upon receipt of their REPORT message andoverloaded ONUs after receiving all ONU REPORT messages.When limiting grant sizes and distributing excess from ONUsnot fully utilizing their guaranteed minimums, all REPORTsmust be received in a cycle in order to know the excess fromthat cycle. Therefore, it makes sense to have overloaded ONUswait until this information is available in order to properly sizetheir grant. Finally, the third (DWBA-3) schedules all ONUsupon receipt of their REPORT message. Since the OLT needs

to grant excess bandwidth, the authors create two grants in thisapproach. As soon as the ONU REPORT is received, one grantsthe guaranteed minimum; and after all REPORTs have beenreceived, another grants the excess assigned to that ONU. Thereare two problems with that approach: 1) each overloaded ONUreceives two grants which decreases efficiency due to moreguard times, and 2) the split between the two grants will mostlikely not occur on frame boundaries causing one frame to beunnecessarily delayed till the next cycle. Therefore, DWBA-2is more efficient.

C. Quality of Service in TDM/WDM Ethernet Passive

Optical Networks

In [2], Dhaini et al. propose two approaches for providingQoS on a TDM/WDM EPON: 1) QoS-DBA-1, which uses theOLT scheduler (DWBA-2) of [19] with a Modified DeficientWeighted Round Robin (M-DWRR) intra-ONU scheduler, and2) QoS-DBA-2 and QoS-DBA-3, which segregate constant bit-rate expedited forwarding (EF) traffic [20] from assured for-warding (AF) [21] and/or best effort (BE) traffic by wavelength.Some wavelengths are allocated for EF traffic, and some are al-located for AF and BE traffic. In QoS-DBA-2, this segregationis strict where as in QoS-DBA-3 the unused capacity on the EFwavelengths can be utilized by AF and BE traffic.

D. Summary

The various WDM IPACT variants [15]–[17] do not considerthe problem of grant scheduling. They simply apply a first-fittime and wavelength assignment. In [18], Dhaini et al. explorethe problem of grant scheduling but only in the context of excessbandwidth distribution. In this paper we examine grant sched-uling in a more general context and propose a new schedulingframework that is more efficient than either online or offlinescheduling.

III. GRANT SCHEDULING

The multichannel EPON scheduling problem can be formu-lated using the scheduling theory notation defined in [14]. The-oretical analysis of all the scheduling models discussed in thissection can be found in [14]. Scheduling theory is concernedwith scheduling a set of jobs with specific processing times tobe executed on a set of machines as efficiently as possible withrespect to an optimization criterion. We can view each ONUas representing a job, its grant size as defining its processingtime, and the channels used for transmission on the EPON asrepresenting the machines. In scheduling notation, a schedulingproblem is defined by a triple , where describes themachine environment (e.g., single machine, parallel machines,etc.), describes the processing characteristics and constraints,and describes the objective to be minimized.

For the formulation of the multichannel EPON grant sched-uling problem in the scheduling theory notation, we letdenote identical parallel machines (channels) that define ourmachine environment. Our only processing characteristic orconstraint is which refers to machine (channel) eligibilityconstraints. Specifically, is the set of machines (channels)that job (ONU) can be executed (transmitted) on. Letdenote the time at which the transmission for ONU is com-plete. With this notation we define the scheduling problem with


Fig. 1. Illustration of constituent delays of a scheduling cycle.

the objective to minimize the unweighted sum of the comple-tion times as . The processing constraint isrequired because each ONU has, in general, its own subsetof supported channels. If all ONUs supported transmissionon all wavelengths, we could remove the machine eligibilityconstraint to obtain .

Our performance objective in designing a scheduler for aWDM EPON is to lower the queueing delays experienced byframes in transit across the EPON and to increase the achiev-able resource (i.e., channel) utilization. To see how these perfor-mance objectives relate to the objectives from scheduling theorywe first explore in detail the component delays in a schedulingcycle. We start by defining cycle length, which we also referto as the GTG or GATE-to-GATE delay, as the time betweenback-to-back grants to an ONU. The component delays of ascheduling cycle are visualized in Fig. 1. GTR is the GATE-to-REPORT delay (since we append the REPORT at the end of thetransmission window, GTR is equal to the transmission time ofthe grant), and RTG is the REPORT-to-GATE delay, that in-cludes the propagation delay from ONU to OLT. Using GTRand RTG we express the cycle length as GTG GTR RTG.

The STG is the Schedule-to-GATE delay which is the timebetween the OLT scheduling an ONU’s next grant to the time thegrant starts being received at the OLT. The STG includes a prop-agation delay from OLT to ONU, a GATE message transmissiontime and propagation delay from ONU to OLT. The STG alongwith the grant time represents the completion time of an ONU’stransmission from the point in time it is scheduled, i.e.,STG GTR. Since, the grant time (or size) is not determinedby the scheduler, the scheduler can only work to minimize thevariable portion of completion time, i.e., the STG. Minimizing

minimizes STG. The RTS is the REPORT-to-Schedule delayand is the delay from the OLT receiving a REPORT from anONU to when the scheduling of the ONUs’ REPORT is com-pleted by the OLT. (The RTS includes the computation time forthe schedule, which we neglect in our simulations in Section V.)Thus, REPORT-to-GATE RTG delay is composed of the RTSand STG delays, i.e., RTG RTS STG.

We introduce a layered approach to scheduling (see Fig. 2).We refer to the first layer as the scheduling framework, and thesecond layer as the scheduling policy. The scheduling frame-work is a logistical framework that determines when the OLTmakes scheduling decisions, whereas the scheduling policy is amethod for the OLT to produce the schedule. The OLT can pro-duce a schedule with partial information about the ONU trans-

Fig. 2. Layered approach to scheduling.

missions to be scheduled or after waiting to receive all of the in-formation about the ONUs transmissions to be scheduled. Oncethe OLT has a set of ONU transmissions to schedule, the OLTuses a scheduling policy to create the schedule. The schedulingframework impacts the RTS delay by determining at what timeafter a REPORT message is received an ONU’s next grant isscheduled. On the other hand, the scheduling policy impacts theSTG delay by determining when an ONU grant is actually trans-mitted on a channel.

In Section III-A, we discuss scheduling frameworks in moredetail, and in Section III-B we discuss scheduling policies formultichannel EPONs.

A. Scheduling Frameworks

As mentioned above, the scheduling framework determineswhen the OLT produces a schedule. If the OLT produces aschedule as soon as any ONU REPORT is received withoutwaiting for REPORT messages from other ONUs, this isreferred to as an online scheduling framework. However, ifthe OLT were to wait for the REPORT messages from all theONUs to be received before making scheduling decisions,this is referred to as an offline scheduling framework. Thescheduling framework can be viewed as a continuum betweenthe extremes of online and offline scheduling as illustrated inFig. 3. On the online scheduling end of the continuum, theOLT only considers a single ONU REPORT in a schedulingdecision. On the offline scheduling end of the continuum, theOLT considers all ONU REPORTs in a scheduling decision.Any scheduling framework that lies between online and offlineis some form of an online scheduling framework because notall of the ONU REPORTs have been received. We will howeverreserve the term online scheduling framework to indicate thecase where the OLT considers only one ONU REPORT at atime. In Section III-A1 we will explore offline scheduling in


Fig. 3. Scheduling framework continuum (� : number of ONUs).

the context of multichannel EPONs, and in Section III-A2 wewill explore online scheduling in the context of multichannelEPONs. We will see that both schemes have very differentchannel utilization characteristics.

1) Offline Scheduling: In an offline scheduling framework,scheduling decisions are made with full knowledge of all thejobs to be scheduled including their processing times for aparticular scheduling cycle. Specifically for a multichannelEPON, an offline scheduling framework schedules ONUgrants for transmission when the OLT has received the currentMPCP REPORT messages from all ONUs. This allows theOLT to take the current bandwidth requirements of all ONUsinto consideration in the grant sizing and scheduling, i.e., thescheduling pool contains all ONUs. The scheduling policy isexecuted after the OLT receives the end of the last ONU’s gatedtransmission window. The RTS delay for the last ONU will benegligible; however the RTS may not be negligible for the otherONUs. This RTS delay introduces further queueing delays inthe ONUs because it introduces additional delay in the cyclelength GTG for an ONU. Waiting for all ONU REPORTmessages to be received results in wasted channel capacity. Thewasted capacity increases as the number of channels increases.

2) Online Scheduling: In an online scheduling framework,an ONU is scheduled for upstream transmission as soon asthe OLT receives the REPORT message from the ONU, i.e.,the scheduling pool contains one ONU. This is the schedulingframework that is at the far end of the online side of the sched-uling framework continuum depicted in Fig. 3. This approachavoids wasted channel capacity by not keeping any channelsidle while there is an ONU REPORT message for the OLT toact on.

B. Scheduling Policies

We now use our scheduling model developed in the begin-ning of Section III to find the best scheduling policies for a mul-tichannel EPON that supports an evolutionary migration fromsingle-channel EPONs to multichannel EPONs. The OLT usesthese scheduling policies once a set of ONU grants to be sched-uled has been determined by the scheduling framework.

1) Next Available Supported Channel: A simple schedulingpolicy for a multichannel EPON considers one ONU at a timeand schedules the upstream transmission for that ONU on thewavelength channel that is available the earliest among the chan-nels supported by that ONU. We refer to this scheduling policyas the next available supported channel (NASC) policy. NASCis our variation on an algorithm proposed by Graham [22] nearly

40 years ago called the List algorithm for identical parallel ma-chines. This algorithm schedules jobs one by one and assignsthem to the next available machine.

2) Parallel Machine Models and Solutions: We refer to [23]for a detailed exploration of scheduling policies for our parallelmachine environment scheduling model. Least flexible job first(LFJ) with shortest processing time (SPT) first dispatching rulewas shown to be a good heuristic for this model.

3) Unrelated Machine Models and Solutions: Another pos-sible approach to modeling the multichannel EPON grant sched-uling problem is to loosen our original parallel machine environ-ment model by recognizing that can be viewed asa special case of , where refers to an unrelated ma-chine environment where each machine executes a job at a dif-ferent speed. For machines that are in , we set the executiontime to the processing time or grant length; for machines not in

, we set the execution time to infinity. In [23] we pointed outthat a weighted bipartite matching formulation can optimallysolve .

4) Summary: Using our parallel machine model, the resultsfrom scheduling theory indicate a few dispatching rules that canprovide good scheduling policies. A dispatching rule (see [14,Secs. 14.1 and 14.2]) is a defined method of ordering jobs fordispatch in first fit fashion on available machines. Some exam-ples of general dispatching rules are: Least Flexible Job (LFJ)first, Shortest Processing Time (SPT) first, and Largest Pro-cessing Time (LPT) first. The best scheduling policies for ourmultichannel EPON grant scheduling model are the dispatchingrules discussed in Section III-B2; LFJ for machine eligibility re-strictions, SPT for minimizing the sum of the completion times,and LPT for minimizing the maximum completion time. Otherpotential dispatching rules for the multichannel EPON grantscheduling problem are: Largest Number of Frames (LNF) first,Earliest Arriving Frame (EAF) first, and Earliest Average Ar-rival (EAA) first. LNF favors ONUs with more queued Eth-ernet frames, EAF favors ONUs that have the earliest arrivinghead-of-line (HOL) Ethernet frame, and EAA favors ONUs thathave the earliest average Ethernet frame arrival time.

Dispatching rules can be used alone or grouped togetherto form composite dispatching rules. For multichannel EPONgrant scheduling, LFJ can be combined with some of the otherdispatching rules to create a composite dispatching rule thatcan provide better performance. The second dispatching rule isused to break ties from the first dispatching rule. Rather thanusing the second dispatching rule for tie breaking in the firstdispatching rule, a weight can be set for each dispatching rulein the composite.


Fig. 4. Online JIT on the Scheduling Framework Continuum (� : Number of ONUs,� : Number of Channels).

The weighted bipartite matching (WBM) formulation that isproven optimal for minimizing the sum of the completion timesfor the unrelated machine environment scheduling model oper-ates differently from the dispatching rules. Unlike the dispatchrules, the WBM scheduling policy results in a direct assign-ment of each ONU grant to a specific channel and time (po-sition). There is no need to apply NASC for the scheduling. Theoutput from the WBM scheduling policy fully characterizes theschedule. This is in contrast to the dispatching rules that specifythe order in which the ONU grants are scheduled for first fitchannel assignment according to NASC.

IV. ONLINE JIT SCHEDULING

We now present a new scheduling framework that is ahybrid between offline and online scheduling discussed inSections III-A1 and A2 respectively. We call this new sched-uling framework online just-in-time (JIT) scheduling. Thename indicates that scheduling is performed in just in timefashion. In our online JIT scheduling framework, ONUs areadded to a scheduling pool as their MPCP REPORT messagesare received by the OLT. When a wavelength becomes avail-able, the ONUs in the pool are scheduled together accordingto the selected scheduling policy across all wavelengths. TheONUs that are scheduled so that their transmissions would startshortly (i.e., close to the one-way propagation delay from OLTto the ONU into the future) after the time they are scheduled areclassified as “imminent”; the current schedule for these ONUsis considered firm. The OLT transmits GATE messages to theseONUs to inform them of their granted transmission window.The remaining ONUs are classified as “tentative” and canremain in the scheduling pool for the next scheduling round.Alternatively, all ONUs (i.e., both imminent and tentative) canalways be firmly scheduled. We refer to the case where allONUs are firmly scheduled as online JIT and the case wherethe tentative ONUs participate in future scheduling rounds asonline JIT Tentative.

The online JIT scheduling framework gives the OLT moreopportunity to make better scheduling decisions than standardonline scheduling. ONUs are scheduled at the moment rightbefore they potentially begin transmitting. To facilitate this onan EPON, we need to ensure the GATE message is transmittedby the OLT at least the one-way OLT-to-ONU propagationdelay before we intend the ONU to begin transmission. In otherwords, the GATE message must be transmitted soon enoughto accommodate the OLT-ONU-OLT round trip time (RTT)before we want to begin receiving the ONU’s transmission atthe OLT. Using the largest RTT in the EPON for this timingof the GATE message transmissions ensures that any ONU

receives the GATE message in time. Since we desire the ONUto transmit as soon as the next wavelength becomes free, weneed to schedule the ONUs in the pool at least an RTT beforethe next wavelength free time. Fig. 4 illustrates where theonline JIT scheduling framework lies on the scheduling frame-work continuum. The online JIT scheduling framework can liesomewhere from the online scheduling framework up to a pointjust short of the offline scheduling framework. Let us considerthe bounds of where the online JIT scheduling framework canlie with respect to the number of considered ONU REPORTs.

To obtain the lower bound, consider an EPON with very lowtraffic load, i.e., very few ONUs have traffic and those ONUswith traffic have small queue occupancies. In such a low trafficscenario, there are always free upstream channels. When theOLT receives an ONU REPORT, the OLT makes immediately ascheduling decision based on this one REPORT, i.e., the lowerbound is 1 ONU REPORT.

To obtain the upper bound, consider a high traffic load sce-nario where all ONUs report high queue occupancies re-sulting in grants that are larger than one RTT. In this high trafficscenario, the scheduling pool at the OLT contains all ONUs, ex-cept those that are currently transmitting (and will send theirREPORTs at the end of their transmissions). That is, withONUs in an EPON with upstream channels, there areONUs in the scheduling pool, which is the upper bound onthe number of REPORTs considered for online JIT scheduling.Thus, the online JIT scheduling framework can get very closeto emulating an offline scheduling framework, especially with asmall number of channels.

If the OLT has , ONUs inthe scheduling pool, then wavelengths are idle for one RTTto permit for the REPORTs to reach the OLT and the cor-responding grants to propagate to the ONUs. Hence, with theoffline scheduling framework, for which , the up-stream channels are not utilized for one RTT to allow for all

REPORTs to be received by the OLT and the GRANTs topropagate to the ONUs. Thus, the offline scheduling frameworkutilizes the channels less efficiently than the online and onlineJIT scheduling frameworks. On the other hand, both the on-line and online JIT scheduling frameworks are fully work-con-serving, i.e., they do not let any upstream channel go unusedwhile there is data to transmit. Importantly, note that the onlineJIT scheduling framework does not introduce any inefficienciescompared to the online scheduling framework, as further con-firmed in Section IV-C. The only added complexity with theonline JIT scheduling framework is keeping track of and sched-uling the REPORTs in the scheduling pool, which can rangefrom one to REPORTs, whereas only one REPORT isconsidered at a time in the online scheduling framework.


When using the online JIT Tentative scheduling framework,an ONU may participate in several scheduling rounds as “tenta-tive” before it becomes firmly scheduled. It is possible that cer-tain ONUs that are unfavorable to a particular scheduling policycan continuously be preempted by those that are more favorable.To prevent these ONUs from being starved of medium access,an aging mechanism is incorporated to keep these “less favor-able” ONUs (or jobs) from being starved by the scheduler. Astraightforward method to implement starvation prevention isto set a threshold at which an ONU is immediately scheduledon the next available wavelength regardless of the schedulingpolicy. This ensures that no ONU waits indefinitely for mediumaccess. This threshold can be based on the number of partici-pated scheduling rounds to adapt to changing cycle times.

A. Dispatching Rules for Online JIT

Dispatching rules or composite dispatching rules can be

used without any modification within the online JIT scheduling

framework. The dispatching rules result in an ordering of the

ONUs in the scheduling pool. This ordering is used in conjunc-

tion with NASC to schedule an ONU grant at a specific time

on a specific channel.

B. Weighted Bipartite Matching Adapted for Online JIT

The standard weighted bipartite matching (WBM) schedulingformulation [14] for minimizing the sum of the completiontimes needs to be modified to support an online schedulingframework. In any online scheduling framework, not all ma-chines are immediately available for scheduling (i.e., theymay still be processing jobs). We introduce this in the WBMformulation by setting an additive cost to a matching that isdifferent for each machine. This additive cost is related to whenthe wavelength becomes available. We refer to this cost as ,the availability cost of wavelength for ONU . Let bethe RTT delay for ONU be the time the REPORTmessage from ONU is received at the OLT, be the timewhen wavelength is free, and be the time when ONU

is ready to transmit. Given ,then . is the Euclidean distancebetween wavelength free time and ONU ready to transmit time.A weight can be used to control how much this availability costaffects the solution, we will use to represent this weight.

The following is the Integer Program that represents theWBM where is the scheduling position, is the grantprocessing time for ONU on channel (either ONU granttime for supported channel , or for nonsupported channel), is the availability cost for channel , are binary

variables representing whether or not position on machine(channel) is selected for job (ONU) , is the number ofmachines (channels), and is the number of jobs (ONUs):

(1)

subject to

(2)

(3)

The first constraint forces an ONU to be assigned to only onescheduling position. The second constraint forces each sched-uling position to be assigned to no more than one ONU. If asingle ONU supplies traffic from multiple classes, each trafficclass is treated as a separate job in the WBM formulation.

C. Stability Analysis

In this section we formally analyze the stability character-istics of the online JIT scheduling framework. Stability in thecontext of EPONs has so far been primarily examined for grantsizing techniques employing prediction of traffic newly arrivingbetween sending a REPORT and the start of the correspondingupstream transmission, e.g., see [24]–[27]. These analyses con-sider the prediction control loop and examine controllability andstability of the grant sizing for rapid fluctuations in the trafficloads; whereby the grant size prediction is considered stablewhen ONUs receive a fair bandwidth share. The scheduling ofthe grants and the resulting utilization of the upstream trans-mission channels are not explicitly considered in the existingstudies. Our stability analysis is fundamentally different fromthe existing analyses in that we examine whether the generatedlong-term traffic load can be accommodated on the upstreamtransmission channels. We consider grant sizing without predic-tion in our analysis. We remark that the online JIT schedulingframework could be used in conjunction with grant sizing tech-niques employing prediction; the analysis of such a combinedsystem is left for future work.

Recall that we consider a multichannel EPON with ONUsand channels. Let if ONU supports channel , and

if ONU does not support channel , forand . Let denote the long-run average packet(Ethernet frame) generation rate at ONU (in packets/second),and let denote the average packet length (in bits). Further, let

denote the transmission bit-rate of an upstream channel (inbits per second). We define the relative loads for

.We begin the analysis by considering the upstream transmis-

sions on a given upstream channel. We define the upstream

transmission of an ONU to consist of the upstream transmissionof Ethernet frames and the MPCP REPORT. Consider an arbi-trary upstream transmission of a given ONU. Let be a randomvariable denoting the transmission time of the payload data (Eth-ernet frames) in the upstream transmission. Let be a randomvariable denoting the “proportional” overhead in the upstreamtransmission in terms of transmission time (i.e., number of over-head bits divided by channel transmission bit-rate ). The pro-portional overhead accounts for the preamble of eight bytesfor each Ethernet frame in the upstream transmission and theinterpacket gap of 12 bytes between successive Ethernet framesin the upstream transmission. Note that is given by a con-stant times the payload transmission time, i.e.,(whereby , with denoting the fixed overhead perEthernet frame and denoting the mean of the considered Eth-ernet frame lengths). Let be a constant denoting the “fixed”overhead in the upstream transmission, i.e., the MPCP Report,plus the guard time between successive upstream transmissions.


The utilization of the upstream channel during the consideredupstream transmission is then

(1)

To examine the maximum achievable utilization, we initiallyconsider two ONUs that transmit upstream on a given upstreamchannel. (We incorporate the impact due to the possibly re-stricted set of wavelengths supported by an ONU shortly.) Withgated service and sufficiently high loads, the upstream trans-mission of an ONU is typically sufficiently long to mask theround-trip propagation REPORT-to-GATE delay for the otherONU. (For shorter upstream transmissions, we can considermore ONUs to mask the round-trip propagation delay.) Then,for any combination of scheduling framework and schedulingpolicy that ensures that there is exactly a guard time betweensuccessive upstream transmissions, the utilization on the con-sidered upstream channel is equal to the given in (1). Themaximum achievable utilization arises when the transmissiongrants become very large such that and is givenby

(2)

We now turn to the constraints imposed by the transmissioncapabilities of the individual ONUs onthe upstream channels . We immediately ob-serve that an ONU with a single transmitter can only transmiton one channel at a time. Hence, the ONU must not generatemore traffic load than it can transmit, i.e., we have to requirethat , or equivalently that , for all

.We claim that if the EPON upstream transmission system

is stable, then there are long-run average relative transmissionrates such that

for all (3)

and

for all (4)

This can be seen as follows. Note that the relative transmissionrates are obtained by normalizing long-run average transmis-sion rates (in bits/second) of ONU on channel , by thechannel bit-rate , i.e., . For any ONU , the rel-ative transmission rates over all channels that ONU supportsadd up to ; thus, (3) follows. On the other hand, in the longrun, one cannot send more over channel than permitted by themaximum utilization of the channel transmission bit-rate,which is (4).

If (3) and (4) hold, then one can construct a static periodictransmission strategy, similar to a time division multiplexing(TDM) transmission, that is indeed stable. To see this, note that(3) and (4) imply that we know how much traffic ONU can sendon channel in the long run without causing stability problems.

Namely, this is exactly . So, we allocatefraction of time of channel to ONU . This can be done in aperiodic fashion resembling a TDM strategy. Then (3) ensures

that this strategy accommodates all traffic load generated at eachONU and (4) guarantees that one does not send more on anychannel in the long run than permitted by the maximum achiev-able utilization of the channels’s transmission bit-rate, i.e., wesee that this periodic transmission strategy is stable.

Any grant sizing mechanism that bases grants on actualqueue occupancies (to avoid allocation of excess bandwidththat would then go unused), in combination with any workconserving scheduling framework and policy that ensures theminimum spacing between upstream transmissions (so thatthere is no unnecessary unused time on the upstream channels)achieves the same stability limit as the periodic transmissionstrategy. In particular, gated service grant sizing in conjunctionwith both the online and online JIT scheduling frameworkswith any scheduling policy spacing upstream transmissions ona channel by no more than the guard time are stable if and onlyif (3) and (4) hold. We remark that when all ONUs support allchannels, i.e., when for all , then (3) and (4) hold

if and only if .

We briefly remark regarding a formal analysis of the delay

performance of online JIT that online JIT scheduling always

makes scheduling decisions no earlier than online scheduling.

Hence, online JIT can make better scheduling decisions. How-

ever, it is relatively easy to find examples where online JIT and

online scheduling behave in exactly the same way. In general, it

can therefore only be proven that online JIT never makes worse

scheduling decisions than online scheduling, which is straight-

forward to verify. (Note that the converse is not true, because

online scheduling makes decisions with less information and

hence it cannot mimic online JIT.)

V. SIMULATION EXPERIMENTS

In this section we study different scheduling techniques that

can be used for grant scheduling in multichannel EPONs by

means of simulations. We developed a multichannel EPON sim-

ulation engine using the CSIM [28] simulation library. Each

wavelength supports Gbps transmission bit-rate, and the

reported load corresponds to the payload data rate .

Following common packet size models, 60% of the packets have

64 bytes, 4% have 300 bytes, 11% have 580 bytes, and 25% have

1518 bytes. The simulations were conducted using self-similar

traffic sources [29] with a Hurst parameter of 0.75. The RTTs

were uniformly distributed over s s , which corre-

sponds to distances of 2–15 km.

In our simulations we use the gated grant sizing technique

which grants each ONU its full bandwidth request. The gated

sizing technique has been demonstrated to achieve small delays

in EPONs [30]. By fixing the grant sizing technique, we are

comparing the scheduling aspects of the multichannel EPON.

A. Offline Scheduling Versus Online Scheduling

We refer the reader to our simulation study presented in [23]

for a comparison of offline versus online scheduling. Our re-

sults showed that an online scheduling framework significantly

outperformed the offline scheduling framework regardless of

scheduling policy with respect to average queueing delay. The

better scheduling decisions made by the scheduling policies


used with the offline scheduling framework were not enough

to overcome the RTS delay that dominated the cycle length.

Therefore, the average queueing delay values were significantly

higher with the offline scheduling framework. We provide fur-

ther results for the comparison of offline and online scheduling

in Section V-C.

B. Online Scheduling Versus Online JIT Scheduling

We want to study the impact of the type of scheduler and the

ONU multichannel diversity. To vary the ONU multichannel di-

versity, we simulated an EPON with ONUs and created

three sets of ONU WDM configurations: WDM Mix 1, WDM

Mix 2, and WDM Mix 3. WDM Mix 1 contains 16 ONUs that

support all wavelengths, 8 that support half of the wavelengths,

and 8 that support the other half. WDM Mix 1 provides ONU

multichannel diversity without any single channel ONUs. By

not having any single channel ONUs, the scheduling process

has increased flexibility in wavelength assignment. WDM Mix

2 contains 16 ONUs that support all wavelengths, 6 that support

half of the wavelengths, 6 that support the other half of the wave-

lengths, and 4 that only support one wavelength. WDM Mix 2

provides ONU multichannel diversity with some single channel

ONUs. Finally, WDM Mix 3 provides less multichannel diver-

sity with 8 ONUs that support all wavelengths and 24 ONUs

that support only one wavelength. Of the three ONU WDM con-

figurations, WDM Mix 3 is the most restrictive with respect to

wavelength assignment.

In the experiments for WDM Mix 1 and WDM Mix 2, all

ONUs generate the same traffic load. For the WDM Mix 3 ex-

periments, each of the eight ONUs supporting all channels gen-

erates nine times the traffic load of each of the 24 ONUs sup-

porting one channel; this load distribution gives a theoretical

stability limit of 4 Gbps for the WDM Mix 3 scenario.

We varied the scheduler for each of the ONU WDM config-

urations. Two dispatching rule-based schedulers from our par-

allel machine scheduling model (see Section III-B2): LFJ-SPT

and LFJ-LNF, and two weighted bipartite matching (WBM)-

based schedulers with weight from our unrelated ma-

chine model (see Section III-B3): WBM and WBM-LNF. These

schedulers are all compared to an online scheduler: NASC that

provides the baseline for performance.

Figs. 5, 7, and 8 show the average queueing delay plotted

against load for WDM Mix 1 with total wavelengths,

WDM Mix 2 with total wavelengths, and WDM Mix 3

with total wavelengths, respectively.

For WDM Mix 1, plots a) and b) in Fig. 5 show up to a 10%

decrease in average queueing delay at low and moderate loads

provided by the WBM-based schedulers. The plotted confidence

intervals indicate that the difference appears to be statistically

significant. The confidence intervals were obtained through the

CSIM batch means method with batches sized to minimize cor-

relations, and individual batch means representing the average

queueing delays of the Ethernet frames occurring in a given

batch. The lower average queueing delay achieved by the WBM-

based schedulers is due to more efficient wavelength scheduling

decisions made by the scheduler. The WBM-based schedulers

Fig. 5. Average queueing delay for WDM Mix 1, eight wavelength EPON.The WBM-based schedulers can provide lower queueing delays at lower loads.However, schedulers that incorporate the LNF dispatching rule provide lowerqueueing delays at higher loads. (a) Low load. (b) High load. (c) Very high load.

evaluate the costs of all possible wavelength assignments and

select the lowest cost matching, i.e., the wavelength for which

the availability time most closely matches the time when

the ONU is ready . From plots of the wavelength uti-

lization of each of the eight wavelengths for each of the sched-

uling schemes, which are not included due to space constraints,

we observed that all compared scheduling strategies achieved

equally good load balancing. The WBM-based schedulers are

simply making more efficient wavelength assignment decisions

that are resulting in shorter cycle lengths. Fig. 6 shows the cycle

length for the WBM-based schedulers compared to NASC.

At higher loads a different pattern emerges, we observe from

plot b) in Fig. 5 that at high loads the schedulers that use LNF


Fig. 6. Average cycle length for WDM Mix 1, eight wavelength EPON. Thetwo WBM-based schedulers, whose curves fall on top of each other in the figure,provide lower cycle lengths due to better wavelength assignment.

(i.e., LFJ-LNF or WBM-LNF) provide a small improvement

over the other schedulers. Plot c) in the same figure confirms this

by zooming in on the higher loads. Again, the plotted confidence

intervals appear to indicate a statistically significant difference.

This indicates that at high loads the wavelength assignment has

limited impact on average queueing delay. However, preferring

ONUs with a larger number of queued frames can lower the av-

erage queueing delay. This is largely due to a frame sampling

effect: ONUs with more frames have a larger impact on the av-

erage queueing delay measure than ONUs with fewer frames.

For WDM Mix 2, with less wavelength assignment flexibility

due to the single wavelength ONUs, we observe from Fig. 7

a smaller decrease in average queueing delay provided by the

WBM-based schedulers. The plots a) and b) in Fig. 7 indicate

about a 3% decrease in average queueing delay. The plotted

confidence intervals indicate that the difference is not statisti-

cally significant. At high loads, shown in plot b) and zoomed

in plot c), we see the same pattern as seen for WDM Mix 1: at

high loads the wavelength assignment has no impact on average

queueing delay, but time ordering does have an impact due to a

frame sampling effect.

For WDM Mix 3, with its limited ONU multichannel diver-

sity and restrictive wavelength assignment capabilities, there is

no measurable difference with respect to average queueing de-

lays between the schedulers as observed from Fig. 8.

In summary, with increased ONU multichannel diversity and

wavelength assignment flexibility, the WBM-based schedulers

are able to provide lower average queueing delays. At high

loads, the schedulers that favor ONUs with the largest number

of frames provide lower average queueing delays, because of

a frame sampling effect. Therefore, the online JIT scheduling

framework has some utility in lowering the average queueing

delay. However, its utility is much larger than this, as will be

discovered in the next section.

C. Differentiated ONU Treatment Using Online JIT Scheduling

In this set of simulation experiments we explore how the on-

line JIT scheduling framework can be used to provide differen-

tiated treatment to ONUs without using an offline scheduling

Fig. 7. Average queueing delay for WDM Mix 2, eight wavelength EPON.With less wavelength assignment flexibility as compared to WDM Mix 1, thereduction in average queueing delays achieved by the WBM-based schedulersis less pronounced. (a) Low Load. (b) High load. (c) Very high load.

mechanism. Avoiding use of an offline scheduling framework

improves channel utilization and consequently lowers queueing

delays. We have simulated the same EPON system described

above for WDM Mix 1, WDM Mix 2, and WDM Mix 3. We

now use a scheduler that always schedules two preferred ONUs,

which support all wavelengths, ahead of any of the other 30

ONUs. The other 30 ONUs are scheduled with the LFJ-LNF

dispatching rule.

Figs. 9–11 show plots of the average queueing delay experi-

enced by all ONUs (labeled “Avg ONU”) and the 2 preferred

ONUs (labeled “Pref ONU”) for the online JIT scheduling

framework, the online JIT Tentative scheduling framework


Fig. 8. Average queueing delay for WDM Mix 3, four wavelength EPON. Withvery limited wavelength assignment flexibility, the WBM-based schedulers donot provide a reduction in average queueing delay. (a) Low Load. (b) High load.

with a starvation threshold set to 32 scheduling rounds, and the

Offline scheduling framework.

Examining first the performance for the online JIT scheduling

frameworks, we observe that they are able to provide differential

treatment without using an offline scheduling framework. At

high loads, the difference in average queueing delay becomes

quite significant. For example, Fig. 10 shows that at a load of

6.8 Gbps, the preferred ONUs experience an average queueing

delay of approximately 500 s as opposed to approx. 1.6 ms

for all ONUs. Comparing the figures for the different WDM

Mixes, we see that as we move from WDM Mix 1 to WDM

Mix 2, i.e., as we increase the number of single channel ONUs

from zero to four, the average queueing delay increases but the

queueing delay experienced by the two preferred ONUs stays

the same. In Fig. 11 we see that when we increase the single

channel ONUs to 75% of all ONUs and limit the EPON to four

wavelengths, the average queueing delay increases significantly.

However, the queueing delay experienced by the two preferred

ONUs is significantly lower than the delay for the other ONUs.

At a load of 3.8 Gbps, the delay for the preferred ONUs is nearly

eight times smaller than the average queueing delay.

Comparing the online JIT scheduling framework with the on-

line JIT Tentative scheduling framework, we observe a slight

reduction in the average delay for the preferred ONUs with the

online JIT Tentative scheduling framework as compared to the

online JIT scheduling framework.

Fig. 9. Queueing delay for average ONUs versus the two ONUs given preferen-tial scheduling (WDM Mix 1). Average ONUs are scheduled using the LFJ-LNFdispatching rule. The online JIT scheduling framework is able to provide signif-icant differential treatment through scheduling. The oOffline scheduling frame-work does not provide differential treatment through scheduling. (a) Online JITscheduling framework. (b) Offline scheduling framework.

We summarize the main observations from comparing the

online JIT scheduling framework with the offline scheduling

framework as follows. First, we observe that the achievable

maximum channel utilization (stability limit) is lower for of-

fline scheduling as compared to online scheduling. Fig. 11, for

instance, indicates that for offline scheduling of WDM Mix 3

the average ONU delays shoot up to very large values for loads

around 3 Gbps, whereas we observe a similar jump in ONU

delays for online scheduling at a load of 3.75 Gbps. The lower

stability limit with offline scheduling is due to the nonwork

conserving nature of offline scheduling, which forces the OLT

to wait for all REPORT messages before making scheduling

decisions. This waiting imposes idle times on the upstream

channels which are avoided by the work conserving online

scheduling frameworks, including online JIT and online JIT

Tentative.

Second, we observe that the queueing delays are much larger

for offline scheduling than online JIT scheduling. For a 6 Gbps

load for WDM Mix 1, for instance, we observe from Fig. 9

average queueing delays around 20 ms for offline scheduling

compared to less than 0.4 ms with online JIT. The larger delays

with offline scheduling are mainly due to the increased RTS de-

lays that increase the cycle lengths and subsequently increase

the queueing delays.


Fig. 10. Queueing delay for all ONUs versus the two ONUs given preferentialscheduling (WDM Mix 2). Although the delay for the average ONU increasesbecause of the single channel ONUs, the delay for the two preferred ONUsremains the same as with WDM Mix 1. (a) Online JIT scheduling framework.(b) Offline scheduling framework.

Finally, turning to the differentiated ONU treatment, we

observe from Figs. 9–11 that in contrast to online JIT, offline

scheduling provides very little differentiation between the av-

erage ONUs and the preferred ONUs. This is mainly due to the

fact that offline scheduling forces a relatively large RTS delay

upon all ONUs, as illustrated by the delay components provided

for WDM Mix 1 in Table I. In fact, for moderate to high loads,

the RTS delay is the largest of the three delay components,

which add up to the cycle length GTG RTS STG GTR.

In addition, the grant times, i.e., GTR delays, are relatively

large compared to the STG delays, leaving little flexibility

for differentiation due to reordering of the sequence of the

upstream transmissions. In contrast, we observe from Table I

that the RTS delays are relatively small for online JIT. At

the same time, the STG delays are relatively large, compared

to both RTS and GTR delays, providing significantly more

flexibility in influencing the relative treatment of the ONUs

through scheduling of the upstream transmissions.

In summary, the online JIT scheduling framework has the po-

tential of reducing the average queueing delay experienced by

all ONUs. However, its strongest utility appears to be for sched-

ulers that provide differential treatment to ONUs. The OLT can

benefit from an increased level of scheduling control without

waiting for all ONU REPORT messages. Practical implemen-

tations may fine tune exactly when schedules are produced. It

Fig. 11. Queueing delay for all ONUs versus the two ONUs given preferen-tial scheduling (WDM Mix 3). The online JIT scheduling framework is able toprovide very significant differential treatment. (a) Online JIT scheduling frame-work. (b) Offline scheduling framework.

TABLE IAVERAGE REPORT-TO-SCHEDULE (RTS) DELAY, SCHEDULE-TO-GATE (STG)

DELAY, AND GRANT TIME (GTR) IN MICROSECONDS FOR WDM MIX 1WITH ONLINE JIT AND OFFLINE SCHEDULING FRAMEWORKS

may be advantageous to purposefully leave a channel idle while

waiting for more ONU REPORT messages to arrive at the OLT

in an effort to gain a higher level of control.

VI. CONCLUSION

We have proposed: 1) a two-layer structure of scheduling

in multichannel EPONs consisting of a scheduling frame-

work layer and a scheduling policy layer, as well as 2) online

just-in-time (JIT) scheduling, a novel work conserving sched-

uling framework. In the online JIT scheduling framework,

channel availability, rather than ONU REPORT messages,

drives the scheduling process. When a channel becomes avail-

able, the OLT makes an access decision with the information


(i.e., REPORT messages) that has accumulated since the last

channel became available. This gives the online JIT scheduling

framework the ability to make better scheduling decisions

as compared to an online scheduling framework that only

considers one ONU REPORT message. Further, the online JIT

scheduling framework is still work conserving and therefore

is more efficient than the nonwork conserving offline sched-

uling framework where the OLT waits for all ONU REPORT

messages to make access decisions.

In our simulation study, we found that with increased ONU

multichannel diversity and wavelength assignment flexibility,

the WBM-based scheduling policies used in the online JIT

scheduling framework are able to provide lower average

queueing delays. We also found that at high loads, the sched-

ulers that favor ONUs with the largest number of frames

provide lower average queueing delays, because of a frame

sampling effect. Therefore, the online JIT scheduling frame-

work has some utility in lowering the average queueing delay.

However, it has stronger utility for use with schedulers that may

provide differential treatment to ONUs. The OLT can benefit

from an increased level of scheduling control, i.e., considering

a larger scheduling pool, to differentiate ONU service. The

only trade-off is the slightly increased complexity of online

JIT, which requires the OLT to simultaneously consider and

schedule up to close to as many ONU REPORTS as there are

ONUs on the EPON; whereas, only one ONU REPORT at a

time is considered and scheduled with online scheduling.

Future research could study how grant sizing techniques are

affected by this new online JIT scheduling framework, and the

impact of schedule generation time on performance.

REFERENCES

[1] M. Marsan and D. Roffinella, “Multichannel local area network proto-cols,” IEEE J. Sel. Areas Commun., vol. 1, no. 5, pp. 885–897, Nov.1983.

[2] A. Dhaini, C. Assi, and A. Shami, “Quality of Service in TDM/WDMethernet passive optical networks (EPONs),” in Proc. IEEE ISCC, Jun.2006, pp. 616–621.

[3] M. Maier, M. Herzog, and M. Reisslein, “STARGATE: The next evo-lutionary step toward unleashing the potential of WDM EPONs,” IEEE

Commun. Mag., vol. 45, no. 5, pp. 50–56, May 2007.[4] N. Frigo, P. Iannone, P. Magill, T. Darcie, M. Downs, B. Desai, U.

Koren, T. Koch, C. Dragone, H. Presby, and G. Bodeep, “A wave-length-division multiplexed passive optical network with cost-sharedcomponents,” IEEE Photon. Technol. Lett., vol. 6, no. 11, pp.1365–1367, Nov. 1994.

[5] F. An, K. Kim, D. Gutierrez, S. Yam, E. Hu, K. Shrikhande, and L.Kazovsky, “SUCCESS: A next-generation hybrid WDM/TDM opticalaccess network architecture,” J. Lightw. Technol., vol. 22, no. 11, pp.2557–2569, Nov. 2004.

[6] Y. Hsueh, W. Shaw, L. Kazovsky, A. Agata, and S. Yamamoto, “Suc-cess PON demonstrator: Experimental exploration of next-generationoptical access networks,” IEEE Commun. Mag., vol. 43, no. 8, pp.S26–S33, Aug. 2005.

[7] F. An, D. Gutierrez, K. Kim, J. Lee, and L. Kazovsky, “SUCCESS-HPON: A next generation optical access architecture for smooth mi-gration from TDM-PON to WDM-PON,” IEEE Commun. Mag., vol.43, no. 11, pp. S40–S47, Nov. 2005.

[8] Y. Hsueh, M. Rogge, S. Yamamoto, and L. Kazovsky, “A highly flex-ible and efficient passive optical network employing dynamic wave-length allocation,” J. Lightw. Technol., vol. 23, no. 1, pp. 277–286, Jan.2005.

[9] C. Bock, J. Prat, and S. Walker, “Hybrid WDM/TDM PON usingthe AWG FSR and featuring centralized ligth generation and dy-namic bandwidth allocation,” J. Lightw. Technol., vol. 23, no. 12, pp.3981–3988, Dec. 2005.

[10] M. McGarry, M. Maier, and M. Reisslein, “WDM ethernet passive op-tical networks,” IEEE Commun. Mag., vol. 44, no. 2, pp. S18–S25, Feb.2006.

[11] C. Assi, Y. Ye, S. Dixit, and M. Ali, “Dynamic bandwidth allocation forquality-of-service over ethernet PONs,” IEEE J. Sel. Areas Commun.,vol. 21, no. 9, pp. 1467–1477, Nov. 2003.

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

[13] G. Kramer, B. Mukherjee, S. Dixt, Y. Y., and R. Hirth, “Supportingdifferentiated classes of service in Ethernet passive optical networks,”J. Opt. Networking, vol. 1, no. 8, pp. 280–298, Aug. 2002.

[14] M. Pinedo, Scheduling: Theory, Algorithms, and Systems, 2nd ed.Englewood Cliffs, NJ: Prentice-Hall, 2002.

[15] K. Kwong, D. Harle, and I. Andonovic, “Dynamic bandwidth alloca-tion algorithm for differentiated services over WDM EPONs,” in Proc.

9th Int. Conf. Communications Systems, Sep. 2004, pp. 116–120.[16] M. McGarry, “An evolutionary wavelength division multiplexing up-

grade for Ethernet passive optical networks,” Master’s thesis, ArizonaState Univ., Tempe, 2004.

[17] F. Clarke, S. Sarkar, and B. Mukherjee, “Simultaneous and interleavedpolling: An upstream protocol for WDM-PON,” in Proc. Optical Fiber

Communication Conf., Mar. 2006, p. 3.[18] A. Dhaini, C. Assi, M. Maier, and A. Shami, “Dynamic wavelength

and bandwidth allocation in hybrid TDM/WDM EPON networks,” J.

Lightw. Technol., vol. 25, no. 1, pp. 277–286, Jan. 2007.[19] A. Dhaini, C. Assi, and A. Shami, “Dynamic bandwidth allocation

schemes in hybrid TDM/WDM passive optical networks,” in Proc.

IEEE CCNC, Jan. 2006, vol. 1, pp. 30–34.[20] V. Jacobson, K. Nichols, and K. Poduri, An expedited forwarding PHB

Jun. 1999 [Online]. Available: http://www.ietf.org/rfc/rfc2598.txt,RFC 2598 (Proposed Standard) obsoleted by RFC 3246.

[21] J. Heinanen, F. Baker, W. Weiss, and J. Wroclawski, Assured for-warding PHB group June 1999 [Online]. Available: http://www.ietf.org/rfc/rfc2597.txt, RFC 2597 (Proposed Standard) updated by RFC3260.

[22] R. Graham, “Bounds for certain multiprocessing anomalies,” Bell Syst.

Tech. J. , vol. 45, pp. 1563–1581, 1966.[23] M. McGarry, M. Reisslein, M. Maier, and A. Keha, “Bandwidth man-

agement for WDM EPONs,” OSA J. Opt. Networking, vol. 5, no. 9, pp.637–654, Sep. 2006.

[24] H.-J. Byun, J.-M. Nho, and J.-T. Lim, “Dynamic bandwidth allocationalgorithm in ethernet passive optical networks,” Electron. Lett., vol. 39,no. 13, pp. 1001–1002, Jun. 2003.

[25] J. M. Joo and Y. J. Ban, “Dynamic bandwidth allocation algorithm fornext generation access network,” in Proc. OFC, Mar. 2006.

[26] S. Yin, Y. Luo, N. Ansari, and T. Wang, “Bandwidth allocation overEPONs: A controllability perspective,” in Proc. IEEE Globecom, Nov.2006.

[27] S. Yin, Y. Luo, N. Ansari, and T. Wang, “Stability of predictor-baseddynamic bandwidth allocation over EPONs,” IEEE Commun. Lett., vol.11, no. 6, pp. 549–551, Jun. 2007.

[28] Csim (Mesquite Software) [Online]. Available: http://www.mesquite.com

[29] K. Park and W. Willinger, Self-Similar Network Traffic and Perfor-

mance Evaluation. Hoboken, NJ: Wiley-Interscience, 2000.[30] G. Kramer, Ethernet Passive Optical Networks. New York: McGraw-

Hill, 2005.

Michael P. McGarry (M’98) received the B.S.degree in computer engineering from PolytechnicUniversity, Brooklyn, NY, in 1997, and the M.S.and Ph.D. degrees in electrical engineering fromArizona State University, Tempe, in 2004 and 2007,respectively.

He is currently a Senior Staff Scientist at Adtranand an Adjunct Professor at Arizona State University,Tempe. From 1997 through 2003 he was employedin industry by companies including PMC-Sierra andYurie Systems (now Lucent Technologies). His re-

search interests include congestion control and the optimization of MAC proto-cols for both optical access and mobile ad hoc networks.


Martin Reisslein (A’96–S’97–M’98–SM’03)received the Dipl.-Ing. (FH) degree from the Fach-hochschule Dieburg, Dieburg, Germany, in 1994,and the M.S.E. degree from the University of Penn-sylvania, Philadelphia, in 1996, both in electricalengineering, and the Ph.D. degree in systems engi-neering from the University of Pennsylvania in 1998.

He is currently an Associate Professor in the De-partment of Electrical Engineering at Arizona StateUniversity (ASU), Tempe. From July 1998 throughOctober 2000 he was a Scientist with the German Na-

tional Research Center for Information Technology (GMD FOKUS), Berlin andLecturer at the Technical University Berlin. From October 2000 through August2005 he was an Assistant Professor at ASU.

Dr. Reisslein served as Editor-in-Chief of the IEEE COMMUNICATIONS

SURVEYS AND TUTORIALS from January 2003 through February 2007.

Charles J. Colbourn received the M.Math. degreein 1978 from the University of Waterloo, Waterloo,ON, Canada, and the Ph.D. degree in 1980 from theUniversity of Toronto, Toronto, ON, both in computerscience.

He has held academic positions at the University ofSaskatchewan, the University of Waterloo, the Uni-versity of Vermont, and is currently a Professor ofComputer Science and Engineering at Arizona StateUniversity, Tempe. He is the author of The Combina-

torics of Network Reliability (Oxford) and Triple Sys-

tems (Oxford). He is Editor-in-Chief of the Journal of Combinatorial Designs

and serves on the editorial boards of Networks; Discrete Mathematics; Journal

of Combinatorial Theory (A); Designs, Codes and Cryptography; and others.He edited the standard reference work The CRC Handbook of Combinatorial

Designs. He is the author of more than 250 refereed journal papers focussing oncombinatorial designs and graphs with applications in networking, computing,and communications.

Dr. Colbourn was awarded the Euler Medal for Lifetime Research Achieve-ment by the Institute for Combinatorics and its Applications in 2004.

Martin Maier (M’03) received the M.Sc. andPh.D. degrees (both with distinctions) in electricalengineering from the Technical University Berlin,Berlin, Germany, in 1998 and 2003, respectively.

He is currently an Associate Professor withthe Institut National de la Recherche Scientifique(INRS), Montreal, QC, Canada. In summer 2003,he was a Postdoctoral Fellow at the MassachusettsInstitute of Technology (MIT), Cambridge. He wasa Visiting Professor at Stanford University, Stanford,CA, from October 2006 through March 2007. His

recent research activities aim at providing insights into technologies, protocols,and algorithms shaping the future of optical networks and their seamlessintegration with next-generation wireless networks. He is the author of thebooks Metropolitan Area WDM Networks—An AWG Based Approach (Kluwer,2003) and Optical Switching Network (Cambridge University Press, 2008).

Frank Aurzada studied mathematics at theFriedrich-Schiller University, Jena, Germany andthe University of York, York, U.K. He received theDipl.-Math. and Ph.D. degrees in mathematics fromFriedrich-Schiller University in 2003 and 2006,respectively.

After completing his Ph.D., he joined the DFGResearch Center Matheon at Technical UniversityBerlin. His research interests lie in the queueingtheoretic analysis of telecommunication networks,coding theory, and limit theorems in probability

theory. He is currently a Postdoctoral Researcher at the Technical UniversityBerlin, Berlin, Germany.

Michael Scheutzow received the Diploma inmathematics from the Johann-Wolfgang-GoetheUniversity, Frankfurt/Main, Germany, in 1979, thePh.D. degree in mathematics (magna cum laude)from the University of Kaiserslautern, Kaiser-slautern, Germany, in 1983, and the Habilitation inmathematics from the University of Kaiserslauternin 1988.

He is currently a Professor in the Departmentof Mathematics, Technical University (TU) Berlin,Berlin, Germany. From 1988 through 1990 he was

Project Leader with Tecmath GmbH, Kaiserslautern. Since 1990, he has beena Professor of stochastics in the Department of Mathematics, TU Berlin. From1997 through 1999 he was Associate Chair of the department. He has visitedthe University of Carbondale, Rutgers University, the University of Rochester,Warwick University, and the MSRI at UC Berkeley.

IEEE Communications Surveys & Tutorials • 3rd Quarter 200846

ver the past decade, the telecommunications infra-structure has transitioned from a copper-based plantto a fiber-based plant. The transition began with the

wide area networks (WANs) that provide connectivity betweencities and progressed through the metropolitan area networks(MANs) that provide connectivity between service providerlocations within a metropolitan area. At the same time, localarea networks (LANs) that interconnect nodes within an indi-vidual location have seen average bit rates migrate from 10Mb/s to 1 Gb/s over copper cabling. Although significantbandwidth improvements occurred in the service provider net-works (i.e., WANs and MANs), as well as at the subscriberpremises (i.e., LANs), the link between the private customernetworks and the public service provider networks did notexperience the same level of progress. This so-called accessnetwork that provides the link between the private and publicnetworks still relies on an aging copper infrastructure. ThexDSL and cable modem technology developments mademarginal improvements in bandwidth capacity but failed toopen the bottleneck that exists in access networks.

A fiber infrastructure is required in the access networks toprovide higher bit rates, as well as more flexibility. From theservice provider perspective, access network links have differ-ent revenue dynamics than links in the WAN and MAN.Whereas WAN and MAN links carry the bit streams of manyrevenue generating customers, access network links carry a

single or only a few revenue generating bit streams. For thisreason, access networks are very sensitive to cost. Cost issuesare slowing the deployment of a new physical plant in theaccess networks.

Deploying a passive optical network (PON) between ser-vice providers and customer premises can provide a cost effi-cient and flexible infrastructure that will provide the requiredbandwidth to customers for many years to come. PONs are anetwork in which a shared fiber medium is created using apassive optical splitter/combiner in the physical plant. Sharingthe fiber medium means reduced cost in the physical fiberdeployment, and using passive components in the physicalplant means reduced recurring costs by not maintainingremote facilities with power. These reduced costs make PONsan attractive choice for access networks, which are inherentlycost sensitive.

At a top level, PONs are classified by the used link-layerprotocol. Whereas an asynchronous transfer mode (ATM)PON (APON) uses ATM, an Ethernet PON (EPON) usesEthernet, and a gigabit PON (GPON) uses the GPON encap-sulation method (GEM) in addition to ATM cells to supportEthernet. The International Telecommunication Union (ITU)has generated standards for APONs: G.983 broadband PON(BPON), as well as GPONs: G.984 gigabit-capable PON(GPON). The IEEE has generated a standard for EPONs:IEEE 802.3ah Ethernet in the first mile. Given the fact that

O

S U R V E Y S

I E EE

COMMUNICAT IONS

MICHAEL P. MCGARRY, UNIVERSITY OF AKRONMARTIN REISSLEIN, ARIZONA STATE UNIVERSITY

MARTIN MAIER, INSTITUT NATIONAL DE LA RECHERCHE SCIENTIFIQUE (INRS)

ABSTRACT

We compile and classify the research work conducted for Ethernet passive optical

networks. We examine PON architectures and dynamic bandwidth allocation algo-

rithms. Our classifications provide meaningful and insightful presentations of the

prior work on EPONs. The main branches of our classification of DBA are: grant

sizing, grant scheduling, and optical network unit queue scheduling. We further

examine the topics of QoS support, as well as fair bandwidth allocation. The presen-

tation allows those interested in advancing EPON research to quickly understand

what already was investigated and what requires further investigation. We summarize

results where possible and explicitly point to future avenues of research.

ETHERNET PASSIVE OPTICAL NETWORKARCHITECTURES AND DYNAMIC BANDWIDTH

ALLOCATION ALGORITHMS

3RD QUARTER 2008, VOLUME 10, NO. 3

www.comsoc.org/pubs/surveys

1553-877X

IEEE Communications Surveys & Tutorials • 3rd Quarter 2008 47

90 percent of data traffic originates and terminates in Ether-net frames, using an EPON can reduce the adaptationrequired to move data between the LAN and the access net-work. Furthermore, ATM creates inefficiencies in data trans-port as a result of its fixed data unit that requires most datapackets to be segmented and reassembled at the end points ofthe network. This segmentation and reassembly results inhigher processing delays, as well as reduced efficiency of errorrecovery techniques. For these reasons, EPONs appear to bemore promising than APONs for data dominated networks.GPONs, on the other hand, by using GEM instead of ATM,avoid the inefficiency of segmentation and reassembly.

In this article, we review and classify the existing researchon EPONs. The focus is on EPON architectures and dynamicbandwidth allocation (DBA) for EPONs, and our classifica-tions provide insight into areas that are open for furtherinvestigation. For a survey on EPON security issues, which arenot covered in this article, see [1]. This article provides a com-prehensive and up-to-date EPON research survey as of spring2007. The status and the main directions of this research as ofearly 2004 were presented in [2]. The EPON research areahas been very active over the last few years, resulting in a dra-matically expanded and more intricate body of EPONresearch. Therefore, a fundamentally new classification andsurvey of this area is required and provided in this article.

We review the standard PON architecture and two alterna-tive architectures that were proposed. We review and classifyall of the research done on the problem of DBA for EPONs.We classify this work in a meaningful way that providesinsight to researchers currently working on EPONs and thoseconsidering working on EPONs. We discuss medium accesscontrol (MAC) protocols for the two alternative PON archi-tectures. Finally, we conclude the article with a discussion ofavenues of further investigation.

PON ARCHITECTURE

A PON generally has a physical tree topology, where oneoptical line terminal (OLT) residing at the central office ofthe service provider connects to several optical network units(ONUs) in the field. The OLT is connected to the ONUs witha feeder fiber that is subsequently split using a 1 : N optical

splitter/combiner to enable the ONUs to share the opticalfiber. This is illustrated in Fig. 1. The transmission directionfrom OLT to ONU is referred to as downstream and operatesas a broadcast medium. The transmission direction from theONUs to the OLT is referred to as upstream. The upstreamsignals propagate from ONU to OLT but are not reflectedback to each ONU; therefore, the PON is not a broadcastmedium in the upstream direction. The EPON is a multi-point-to-point [3] medium, where the ONUs cannot detecteach other’s transmission because the upstream optical signalis not received by the ONUs. However, ONUs share the samefiber; hence, their transmissions can collide, and contentionresolution must be performed.

To avoid collisions in the upstream direction, time divisionmultiplexing (TDM) or wavelength division multiplexing(WDM) can be used [4]. WDM provides a large amount ofbandwidth to each user, but requires that each ONU use aunique wavelength, which presents inventory challenges forservice providers that must stock many different ONU types.TDM allows all ONUs to share a single wavelength, thus,reducing the number of transceivers at the OLT and allowingfor a single ONU type. First generation PONs use wave-lengths to separate the upstream and downstream channelsbut use TDM to avoid upstream transmission collisionsbetween ONUs. Due to the topology of the PON, MAC pro-tocols that rely on connectivity between all nodes cannot beutilized. A PON allows for connectivity from the OLT to allONUs in the downstream and from each ONU to the OLT inthe upstream (i.e., only the OLT has connectivity to allnodes). This connectivity pattern dictates the use of a central-ized MAC protocol residing at the OLT. This leads to apolling-based MAC, where the OLT polls ONUs and grantsthem access to the shared PON medium.

BROADCAST PON

An alternative PON architecture proposed in [5, 6] requiresreflection of the upstream signal back to the ONUs, as illus-trated in Fig. 2. Splitter 1 splits the upstream signal back tothe ONUs, and splitter 2 splits the downstream signal as inthe standard PON architecture. This creates a broadcast net-work that enables a decentralized medium access control pro-tocol (e.g., carrier sense multiple access with collision

■ Figure 1. Network architecture of a PON with one optical line terminal (OLT) and N = 5 opti-cal network units (ONUs), each with a different round-trip time (RTT).

(Single subcriber)

RTT=6

RTT=10

RTT=4

RTT=18

RTT=6

RTT: Round Trip Time

1:N splitterOLT ONU3

ONU1 (Multiple subcribers)

ONU2

ONU4 (Multiple subcribers)

ONU5 (Single subcriber)

(Single subcriber)


detection [CSMA/CD]) to be employed. Unfortunately, thereare economic downsides to this approach. The ONUs becomemore expensive because they must:• Contain higher power lasers to overcome the loss

incurred by splitting their upstream signals to reflectback to other ONUs.

• Contain an extra receiver for the upstream wavelength.• Have more intelligence to participate in the medium

access arbitration.• Have an extra fiber for the reflected upstream signal.

Further, the large bandwidth-propagation delay product ofthe optical access network limits the feasibility of this type ofarchitecture.

TWO-STAGE PON

A two-stage PON architecture [7] can enable a PON toaccommodate a higher number of ONUs compared to a sin-gle-stage PON. Two-stage PONs help to increase the reach ofthe PON. In the first stage, some ONUs act as sub-OLTs forother ONUs, as illustrated in Fig. 3. These sub-OLTs regener-ate the optical signal in the upstream and downstream, as wellas aggregate the traffic of their child ONUs. This allows a sin-gle OLT in a central office to potentially reach a larger num-

ber of ONUs because the sub-OLTs actas optical switches, mitigating opticalpower budget concerns that arise whenincreasing the number of ONUs.

DYNAMIC BANDWIDTH

ALLOCATION

DBA generally is defined as the processof providing statistical multiplexing amongONUs. To understand the importance ofstatistical multiplexing in PONs, note thatthe data traffic on the individual links inthe access network is quite bursty. This isin contrast to metropolitan or wide areanetworks where the bandwidth require-ments are relatively smooth due to theaggregation of many traffic sources. In anaccess network, each link represents a sin-gle or small set of subscribers with verybursty traffic due to a small number ofbursty sources (e.g., Web data and packe-tized video). Because of this bursty traffic,the bandwidth requirements vary widelywith time. Therefore, the static allocation

of bandwidth to the individual subscribers (or sets of sub-scribers) in a PON is typically inefficient [8]. Statistical multi-plexing that adapts to instantaneous bandwidth requirements istypically more efficient. The DBA that operates at the OLT isresponsible for providing statistical multiplexing.

The OLT requires instantaneous bandwidth requirementinformation from each ONU to make access decisions. Havingthis precise information is not possible due to the non-zeropropagation delays on a PON, typically up to 100 Jsec, whichis significantly larger than the transmission time of the maxi-mum size Ethernet frame: 12.3 Jsec. The ONUs must reporttheir instantaneous queue sizes in a control frame and propa-gate this through the PON to the OLT.

A PON is a remote scheduling system [9] that suffers thefollowing problems:• Significant queue switchover overhead [9] (in the case of

PONs, this is due to the guard times between ONUtransmissions). Guard times between ONU transmissionsare required to enable the previously transmitting ONUto power off its laser to prevent spurious transmissionwhile the next ONU transmits; the next ONU to poweron its laser in preparation for transmission; and the OLTto adjust its receiver to account for power-level differ-ences in transmissions from different ONUs due to their

■ Figure 2. Broadcast PON Architecture: Downstream OLT to ONUs transmissionsare copied by splitter ”d” to all ONUs, while each upstream ONU to OLT transmissionis reflected by splitter ”ru” back to all ONUs, thus creating a broadcast network forboth upstream and downstream transmissions. The dashed lines represent the extrafibers used to carry the reflected upstream signal back to the ONUs.

(Single subscriber)OLT ONU3

(Single subscriber)ONU5

(Single subscriber)ONU2

(Multiple subscribers)ONU1

d

ru

(Multiple subscribers)ONU4

■ Figure 3. Two-Stage PON Architecture: Certain ONUs act as sub-OLTs that regenerate the opti-cal signal for ONUs in a second stage, thereby allowing for an increase in the total number ofserved ONUs.

OLT1:N splitter

ONU4 (Single subscriber)

ONU7 (Single subscriber)ONU6 (Single subscriber)



ONU1

sub-OLT1

(Multiple subscribers)

1:N splitter

sub-OLT2

1:N splitter

ONU2 (Multiplesubscribers)


different distances from the OLT. The queue switchoveroverhead should be mitigated by using a cyclic PONscheduler that issues one grant to an ONU per cycle.

• Large control plane propagation delay as a result of dis-tances between OLT and ONUs. Interleaved polling isused to mitigate the large propagation delays on PONs[3]. With interleaved polling, the next ONU to be polledis issued a message giving transmission access while theprevious ONU is still transmitting. This message, referredto as a grant, contains the start time of the transmissionwindow, as well as the length (duration) of the transmis-sion window. Figure 4 illustrates the difference betweenpolling with and without interleaving.

• Limited control plane bandwidth, which is mitigated byshort control plane messages.

A cyclic interleaved polling MAC called interleaved polling withadaptive cycle time (IPACT) [3] mitigates all of these issues.

The process of DBA consists of two parallel but potentiallyoverlapping problems: grant sizing and grant scheduling (orinter-ONU scheduling). Grant sizing determines the size of agrant, that is, the length of the transmission window assignedto an ONU for a given grant cycle. Grant scheduling deter-mines the order of ONU grants for a given cycle. Althoughthe focus of this section is on DBA for EPONs, most of theresults extend to other PONs as well. Specifically, the resultsthat are not tied to the MultiPoint Control Protocol (MPCP)or Ethernet frame can be extended beyond EPONs to BPONsand GPONs.

Figure 5 shows our taxonomy for dynamic bandwidth allo-cation. We use this taxonomy as a framework for discussing

■ Figure 4. With interleaved polling, the OLT can issue the grants on the downstream wavelengthchannel ⁄d such that successive upstream transmissions on channel ⁄u are separated in time by onlya guard time interval compared to a round trip time with standard polling.

TX OLT

6400 bytes data+ 64 bytes RPT

5120+RPT

λdRX λu

TX ONU1

λuRX λd

6400 bytes data+RPT

6400

TX ONU2

λuRX λd

5120 Standard polling

5120+RPT

TX OLT

6400 bytes data+ 64 bytes RPT

5120+RPT

λdRX λu

TX ONU1

λuRX λd

6400 bytes data+RPT

6400

TX ONU2

λuRX λd

5120 Interleaved polling

5120+RPT

■ Figure 5. Dynamic bandwidth allocation taxonomy.

Queue (intra-ONU)scheduling [25,32,33,34]

Grant (inter-ONU)scheduling [12,25,27,28]

DBA

Grant sizing

Fixed[3]

Dynamic

Gated[3]

Limited[3]

Excessdistribution(see fig. 8)

Exhaustive[17,18,19,20]


the research on dynamic bandwidth allocation for EPONs. Wediscuss the MPCP defined in the IEEE 802.3ah standard. Thisprotocol defines the control plane used in EPONs to coordi-nate medium access. We discuss two fundamentally differentproblem-solving approaches to DBA; we call these approach-es DBA frameworks. We discuss the research work done forthe problem of grant sizing. We discuss the research on grantscheduling. We discuss intra-ONU scheduling, that is, arbi-trating the different queues in a given ONU. We discuss thetopic of quality of service (QoS), and we also discuss the issueof fairness, which is touched on later. All of these sectionsassume the standard PON architecture of Fig. 1. Discussionsof protocols for broadcast and two-stage PON architecturesare deferred to later.

MULTIPOINT CONTROL PROTOCOL

To facilitate the discovery and registration of ONUs, as wellas medium access control, the IEEE 802.3ah task forcedesigned the MPCP. The MPCP consists of five messages.REGISTER REQ, REGISTER, and REGISTER ACK areused for the discovery and registration of new ONUs.REPORT and GATE are used for facilitating centralizedmedium access control. The REPORT message is used toreport the instantaneous queue occupancies at an ONU to theOLT. This REPORT message also can contain queue occu-pancies at certain threshold levels as opposed to only the fulloccupancy. This threshold queue reporting allows the OLTflexibility in determining the size of a granted transmissionwindow. The GATE message is used by the OLT to grantnon-overlapping transmission windows to the ONUs.

■ Figure 6. MPCP operation: Two-way messaging assignment of time slots for upstream transmission between ONU and OLT.

Grantedtransmissionwindow

Dataplane

REPORT msg

GATE msg

Subscriber data

DBA

OLT ONU

Controlplane

algorithm

Intra-ONUscheduler

■ Figure 7. In interleaved polling with stop, the OLT waits to receive the REPORT message fromthe last ONU in a cycle before polling the first ONU in the next cycle. This allows the OLT tomake DBA decisions based on the REPORTs from all ONUs. As a result, a walk time of at leastan RTT is incurred between grant cycles.

twalk

TXOLT

2100+RPT

2560+RPT

All REPORTsreceived

λdRXλu 2100

+RPT1518+R

2100+R

2560+R

TXONU 1

λuRXλd

2100+RPT

2100+RPT

TXONU 2

λuRXλd

2560+RPT

1518+R

TXONU 3

λuRXλd

2100+RPT

2560+RPT


Figure 6 illustrates the infrastructure for facilitating DBA.REPORT messages flow upstream to report queue occupan-cies, and GATE messages flow downstream to grant upstreamtransmission windows. Upon receiving queue occupancy infor-mation by means of REPORT messages, the OLT — using aDBA algorithm — makes MAC decisions for the next cycleand communicates these decisions to the ONUs throughGATE messages.

DBA FRAMEWORKS

With interleaved polling, the DBA granting cycles of theindividual ONUs are interleaved, and the OLT typicallymakes grant decisions based on individual ONU REPORTmessages. That is, the OLT typically does not wait untilREPORT messages are received from all ONUs before mak-ing grant decisions. This prohibits the OLT from makingDBA decisions that consider the bandwidth requirements ofall ONUs. As a result, it is very difficult for the OLT tomake fair access decisions. An alternative approach is theso-called interleaved polling with stop (Fig. 7) in which theOLT stops and waits between granting cycles for all ONUREPORT messages to be received before making DBA deci-sions. This affords the OLT the opportunity to provide a fairdistribution of bandwidth. The trade-off is that interleavedpolling with stop decreases the bandwidth utilization by forc-ing an idle period, twalk, of the one-way propagation delayfrom the last polled ONU to the OLT plus the one-waypropagation delay from the OLT to the first polled ONU, asillustrated in Fig. 7, that is, the walk time is typically equalto the average round trip time (RTT). Depending on thelength of the granting cycle, this walk time can become a sig-nificant portion of the available bandwidth. For example,with a cycle length of 1.5 msec and an RTT on the order of50 Jsec, at least 3.33 percent of the available bandwidth iswasted on the walk time. For a 750 Jsec cycle time, the walktimes would consume 6.66 percent of the available band-width.

Interleaved polling and interleaved polling with stop alsocan be compared by their problem-solving approach to DBA.Interleaved polling without stop requires an online problem-solving approach to DBA, that is, the OLT makes DBA deci-sions with incomplete knowledge of the bandwidthrequirements of all the ONUs. Whereas, interleaved pollingwith stop allows for an offline problem-solving approach toDBA, that is, the OLT makes DBA decisions with full knowl-edge of the bandwidth requirements of all the ONUs. Werefer to these DBA problem-solving approaches as the onlineand offline DBA frameworks, respectively.

GRANT SIZING

Grant sizing can be divided into four major categories:• Gated• Limited• Limited with excess distribution• Exhaustive using queue size prediction

Let Gi be the grant size for the current cycle for ONU i, Ri

the queue size reported in the most recently receivedREPORT message from ONU i, Gi

max the limit on the maxi-mum grant size for ONU i, Ei the share of the excess band-width in a cycle allocated to ONU i, and Pi the predictedqueued traffic between the time of the REPORT transmissionfrom ONU to OLT and the end of the granted transmissionwindow to ONU i for the next cycle. The general equation forgrant sizing would then be: Gi = f(Ri, Gi

max) + Ei + Pi. Wenow discuss different techniques used for this general functionf(·) and for determining Ei and Pi.

Fixed — In the fixed grant-sizing scheme, the grant size isfixed for an ONU every cycle. The function for Gi is simply Gi

= Gimax. Simulation results [3] have shown that the fixed

grant-sizing severely underperforms the dynamic grant-sizingtechniques described below. An analysis in [8] confirms thesimulation results.

Gated — In the gated grant-sizing technique, the grant sizefor an ONU is simply the queue size reported by that ONU,the function for Gi is Gi = Ri. This scheme provides low aver-age delay but does not provide adequate control to ensure fairaccess between ONUs. In-depth delay analyses of the gatedscheme can be found in [10] and [11].

Limited — In the limited grant-sizing technique [3], the grantsize is set to the reported queue size up to a maximum grantsize for that ONU. The function for Gi for the limited schemeis Gi = min(Ri, Gi

max). This grant-sizing scheme prevents anyONU from monopolizing the shared link. Simulation results[3] have shown that there is no average packet delay differ-ence between gated and limited grant sizing. However, limitedgrant sizing can assist in providing fair access between ONUsby limiting the size of the grant to Gi

max and thereby prevent-ing an ONU from monopolizing the link. Let tcycle be thelength of a grant cycle and tguard be the guard time betweengrants. Under high traffic load, tcycle

max = xi(Gimax + tguard) (i.e.,

the maximum grant cycle length is defined by the maximumgrant sizes). A large maximum grant cycle results in largerdelays, whereas a short maximum grant cycle reduces thechannel utilization due to increased guard times [12].

The limited scheme suffers from two detriments. First, thequeue is underserved if Gi

max < Ri. Second, i bandwidth is

■ Figure 8. Excess distribution taxonomy.

Excessdivision

Demanddriven excess

[13]

Equitableexcess[14]

Weightedexcess[15]

Excessallocation

Uncontrolledexcess[14]

Controlledexcess

[7,15]

Iterativeexcess[15]

Excessdistribution


wasted if the request is not fully satisfied, and the end of thegrant does not accommodate the next Ethernet frame. In thebest case, this head-of-line (HOL) Ethernet frame that doesnot fit into the remainder of the grant is only 64 bytes, whichresults in up to 63 bytes of wasted grant. The worst case is ifthe HOL packet is the maximum length 1518 bytes. Thisresults in up to 1517 bytes of wasted grant.

Limited with Excess Distribution — Limited with excessdistribution [13] augments the limited grant-sizing scheme toimprove statistical multiplexing while still preventing ONUsfrom monopolizing the link. In general, ONUs are partitionedinto two groups: underloaded ONUs and overloaded ONUs.Underloaded ONUs are those whose REPORTed queue sizeis less than or equal to the maximum grant size, that is, Ri HGi

max. Overloaded ONUs are those whose REPORTed queuesize is larger than their maximum grant size, that is, Ri >Gi

max. In the excess distribution schemes, the overloadedONUs share the unused or excess bandwidth left over fromunderloaded ONUs. The grant for an overloaded ONU thenbecomes G i = G i

max + E i. The total excess bandwidth isdefined i to be the sum of the differences between the maxi-mum grant size and the REPORTed queue size of all theunderloaded ONUs. Let Etotal be the total excess bandwidthfor a cycle, U the set of underloaded ONUs, O the set of over-loaded ONUs, and Ei the excess assigned to overloaded ONUi. The total excess is calculated as Etotal = xicu (Gi

max – Ri).The computation of the total excess bandwidth requires theOLT to wait for all ONU REPORT messages, that is, requiresthe use of interleaved polling with stop or the offline DBAframework. A hybrid DBA framework that allows underload-ed ONUs to be granted before the stop and overloadedONUs to be granted after the stop [13] can help to mitigatethe inefficiencies of the offline DBA framework. After it iscomputed, the Etotal is divided between all the overloadedONUs.

Excess distribution can be divided into excess division andexcess allocation. Excess division divides Etotal among theoverloaded ONUs and excess allocation can, if used efficient-ly, redistribute excess credits that are unused by some over-loaded ONUs. Figure 8 shows the taxonomy of excessdistribution schemes (including excess division and allocation).

One approach to excess division (referred to as DBA1 in[13]) divides the excess according to demand, that is, DBA1divides the excess according to relative request size followingthe formula:

We refer to this approach as demand-driven excess (DDE)division. Because each ONU’s share of the excess is complete-ly determined by its request size, the larger the ONU requestsize relative to the other ONUs, the more excess bandwidth it

receives. This provides statistical multiplexing but is not nec-essarily fair. The fair excess [14] or equitable excess (EE) divi-sion method divides Etotal equally among the overloadedONUs. Let M be the total number of overloaded ONUs, EEdivides the excess according to the formula:

This approach gives all ONUs an equal piece of the totalexcess, implying fairness. The weighted excess (WE) divisionmethod [15] uses ONU priority weights to divide the excessbandwidth. The total excess is divided among overloadedONUs according to their weights:

This method allows the ONUs to have differing priorities withrespect to their fair share of the excess bandwidth.

After the excess is divided among the overloaded ONUs —according to DDE, EE, or WE — it is possible for the size ofthe grant to be larger than the request, that is, Gi + Ei > Ri,which results in wasted bandwidth. The excess division can beaugmented by an excess allocation algorithm that sizes thegrant so that it does not exceed the request, as well as redis-tributes unused excess credits.

The uncontrolled excess (UE) allocation method [14]assigns overloaded ONUs their share of the excess withoutregard to their request size. Therefore, bandwidth is wasted assome overloaded ONUs receive grants that are larger thantheir requests. The controlled excess (CE) allocation method[7, 15] provides better utilization by avoiding wasted band-width under certain conditions. Specifically, if the total excessdemand from the overloaded ONUs Edemand = xjcO (Rj –Gj

max) is less than the total excess Etotal, then each ONU isgranted its full request Ri. This avoids wasted bandwidth forthe situation when the total request does not exceed the maxi-mum grant cycle size xi Gi

max.The iterative excess (IE) allocation method [15] follows an

iterative grant sizing approach to maximize the number of sat-isfied overloaded ONUs. To avoid sizing a grant larger thanthe request, the grant size for each ONU is determined as:

(1)

Initially, all of the overloaded ONUs are in a list. Ei iscomputed for each overloaded ONU according to one of theexcess division methods. One by one an overloaded ONU’sgrant size is computed according to the above formula. If an

GR R G E

G E R G Ei

i i i i

i i i i i

=H +

+ > +

Ä

ÅÈ

Ç

:

:

max

max max

if

if ÈÈ.

Ei =wi

jcOw jL· Etotal.

Ei =1

M· Etotal.

Ei =Ri

x jcO Rj·Etotal

■ Figure 9. Illustration of queue waiting times.

TXOLT

λdRXλu 6400 bytes data+

64 bytes RPT

6400 bytes data+RPT

5120+RPT

trgs

trge

TXONU 1

λuRXλd

6400 5120

5120+RPT


ONU is satisfied (i.e., Ri H Gimax + Ei), it is issued a grant and

removed from the list. Unsatisfied ONUs remain in the listand participate in future iterations. After an iteration throughthe list, Etotal is recomputed by removing the excess used bythe satisfied ONUs. This allows the excess bandwidth unusedby those satisfied ONUs to be made available to the unsatis-fied overloaded ONUs. Ei is recomputed for the overloadedONUs that remain in the list, and another iteration takesplace. The iterations continue until there are no satisfiedONUs. On this final iteration, the unsatisfied ONUs simplyare allocated their fair share of the total remaining excess(i.e., Ei). The iterative grant-sizing approach mitigates wastedbandwidth by not over-assigning bandwidth to ONUs andmaximizes the number of satisfied ONUs to lower the amountof unused slot remainder. Further, it provides a more efficientdistribution of the excess bandwidth.

Exhaustive Service System Using Queue Size Prediction— Queue size prediction is concerned with estimating thetraffic that is generated during the period between theREPORT message transmission by the ONU and the begin-ning of the gated transmission window. Let trgs denote thistime between the REPORT transmission at the ONU and thestart of the next grant for that ONU. A service system thataccommodates the traffic included in trgs is referred to as apartially gated service system [16]. Alternatively, additionallythe traffic generated during the granted transmission windowcould be predicted. Let trge denote the time between theREPORT transmission by the ONU and the end of the nextgrant at that ONU. This results in an exhaustive service sys-tem [16]. Figure 9 illustrates these time periods. Let Qi be theactual amount of traffic that queued up during trgs or trge. Thegoal of queue size prediction is to get Pi as close to Qi as pos-sible. For constant bit rate (CBR) traffic, which has a constantand therefore predictable rate of traffic generation, this israther simple. Multiplying the constant rate of the CBR trafficin bits/sec by trge [13, 17] is a sufficient predictor for CBRtraffic. For bursty variable bit rate (VBR) traffic, the queuesize prediction is more challenging.

Elementary schemes for queue size prediction for burstysources are [3]: constant credit and linear credit. In the con-stant credit scheme, the OLT adds some constant credit to thegrant size. Let † be this credit, then Pi = †. In the linearcredit scheme, the credit adapts to the size of the request. Let\ be the fraction of the request used as the credit, that is, Pi =\Ri. The idea is that the request size gives some indication asto how much traffic will arrive in the waiting period, that is,trgs or trge.

Using control theory [18] to drive the gap between predict-ed and actual queued traffic to zero, an ONU reports the dif-ference between the grant, Gi, and the actual data queued atthe start of the granted transmission window. Let Gi

prev be thegranted transmission window size, Ri

prev be the data queued atthe time of the report, and Qi

prev be the data queued duringtrgs, all for the previous grant cycle. Let ‘i be the differencereported by the ONU, that is, ‘i = Gi

prev – (Riprev + Qi

prev).Let ̃ be a control gain parameter, then: Gi = Gi

prev – ̃‘i.Control theoretic approaches are used for modifying the

control gain parameter to stabilize ‘i. Simulation results [18]show that this scheme has almost an order of magnitude lowerpacket delay compared to IPACT with gated grant sizing. Thisdifference is attributed to much better queue size prediction.The results provide limited insight because they did notexplicitly show that this control theoretic approach drove ‘i

closer to zero than gated grant sizing.A simple one step back linear predictor [19] can be used

for prediction, that is, predictions are based on the actual data

received during the previous waiting period. Let tnprev be the

time of the previous cycle; the formula used for prediction is,

This formula is identical to the linear credit scheme with

Simulation results show that a one step back linear predictorprovides lower packet delay for expedited forwarding trafficcompared to fixed bandwidth allocation and limited band-width allocation.

A higher order linear predictor [20] for predicting trafficduring the waiting period at an ONU also could be used. Thislinear predictor has its weights updated by means of the LeastMean Square (LMS) algorithm (see Section 3.5 in [21]). Thelinear predictor attempts to predict Qi using informationabout a number, L, of previous Qi values. Because it wasshown in [22] that prediction of self-similar traffic is best per-formed using short-term correlation rather than long-termcorrelation, simulations were conducted using L = 4. Analternative approach could use this value as a starting point ofa search to find the optimal value of L. Mathematical analysisproves that an increase in the accuracy of the predictor leadsto a decrease in average delay.

In the simulation results presented in [20], the higher orderlinear traffic prediction scheme is compared against fixedgrant sizing, limited grant sizing, and limited grant sizing withexcess distribution. In limited grant sizing with excess distribu-tion, the underloaded ONUs are scheduled immediately uponreceipt of the REPORT, whereas overloaded ONUs arescheduled after REPORTs are received from all ONUs. Theresults show a reduction of average packet delay, as well asloss probability. However, it is unclear how much of the dif-ference is attributable to traffic prediction and how much tothe difference in scheduling, either waiting for all REPORTsor immediately scheduling each ONU.

Summary — The gated grant-sizing technique is the simplest,adapts to changing traffic demands, and eliminates wasted por-tions of a grant. However, it does not provide partially gatedor exhaustive service, and it cannot prevent ONUs frommonopolizing the upstream channel. This lack of control limitsthe ability to provide fair access and quality of service guaran-tees. The limited grant-sizing scheme can place limits on anONU’s access to the medium but can be inefficient in utilizingthe upstream capacity. Further, the limited and any other non-gated grant-sizing scheme open up the possibility of a wastedportion of a grant due to lack of frame segmentation in Ether-net. When using a fixed-size protocol data unit (PDU), such asfor a BPON, this is not a problem. Employing excess distribu-tion can improve the efficiency of the limited scheme by allow-ing overloaded ONUs to take advantage of the bandwidth notused by underloaded ONUs in a granting cycle.

Queue size prediction can lower queuing delays by attempt-ing to predict the precise instantaneous traffic demands of anONU to allow for an exhaustive service discipline. The risk isin reduced throughput due to wasted portions of a grant thatresult from imprecise prediction. The bursty nature of datatraffic in the access network makes precise prediction difficult.A slight over prediction will reduce throughput but decreasedelay by at least accommodating all of the queued Ethernet

\ =t

t

rgs

cycleprev

.

Pt

tRi i= ·rgs

cycleprev

.


frames. Investigating non-linear neuro-computational methodsfor the queue size prediction problem potentially is an areafor further research.

GRANT SCHEDULING (INTER-ONU SCHEDULING)

Medium access control for an EPON contains two schedulingproblems. The first is concerned with scheduling grants toeach ONU, namely, inter-ONU scheduling. The second isconcerned with scheduling the individual queues of Ethernetframes at the ONU for transmission within the granted trans-mission window, that is, intra-ONU scheduling. This divisionof the scheduling often is referred to as hierarchical schedul-ing [9, 23, 24]. We discuss inter-ONU or grant scheduling inthis section and defer discussion of intra-ONU scheduling tothe next section.

Since grant scheduling works at the inter-ONU level and iscoupled with the process of grant sizing, it is performed at theOLT. Typically, to change the scheduling order from roundrobin, the OLT must wait for all REPORTed queue sizesfrom the ONUs and then determine the grant order. Thisrequires the use of interleaved polling with stop or the offlineDBA framework, as illustrated in Fig. 7.

ONU transmissions ordered longest queue first (LQF) [12,25] or earliest packet first (EPF) [12] have been examined.LQF allows ONUs with the largest grant size to transmit first.This is the same as largest processing time (LPT) first schedul-ing in standard scheduling theory notation [26]. EPF allowsONUs with the earliest arriving HOL Ethernet frame totransmit first. To avoid the walk time between cycles when theOLT waits for all ONUs to REPORT before making anyscheduling decisions, the scheduling can leave out the last orlast few ONUs when scheduling. Simulation results [25, 27]using Poisson traffic show that both LQF and EPF providelower average delay at medium loads compared to a roundrobin scheduler. At low and high loads, the average delay wasthe same as a round robin scheduler.

Implementing the limited with excess distribution grant siz-ing technique requires special consideration with respect togrant scheduling. To accumulate and distribute the totalexcess bandwidth, the OLT must wait to receive REPORTmessages from all ONUs before issuing grants for the nextcycle. This means a walk time, twalk, is wasted between cycles.To mitigate this, lightly loaded ONUs can be scheduled imme-diately upon receiving their REPORT messages because theywill not receive any excess [13]. However, as the traffic loadincreases to the point where all ONUs become overloaded,twalk still is wasted between cycles. To avoid wasting twalk inbetween cycles, an overloaded ONU can be scheduled imme-diately if there is no underloaded ONU available for schedul-ing when the channel becomes available [28].

Summary — Examining the performance of these existingscheduling schemes for self-similar traffic is an important

topic for future research. Another avenue for future researchis to employ scheduling theory to find a better scheduler.

ONU QUEUE SCHEDULING (INTRA-ONU SCHEDULING)

Intra-ONU scheduling is concerned with scheduling the multi-ple queues of Ethernet frames at an ONU, for transmissionwithin the ONU’s granted transmission window. If the numberof queues in an ONU is relatively small, this intra-ONUscheduling can be performed at the OLT. However, as thenumber of queues increases, scheduling is typically made hier-archical [9] with the inter-ONU scheduling at the root of thehierarchy in the OLT and one level of branches. The ONUcontains the branch (i.e., intra-ONU) schedulers.

Low complexity is a key design goal for intra-ONU sched-ulers so that the cost of the ONUs is kept at a minimum.There are typically two classes of scheduling that service mul-tiple queues of differing priority:• Strict priority (SP) scheduling, which can be unfair• Weighted fair queuing (WFQ) schedulingSP scheduling creates unfairness when starving lower prioritytraffic due to unrestricted preemption. The ideal schedulershould allow statistical multiplexing, but guarantee a minimalportion of the available bandwidth to each priority queue (i.e.,provide link sharing). Generalized processor sharing (GPS)[29] achieves these goals for the fluid traffic model, wherepackets are infinitesimally small. Unfortunately, in practicalsystems with finite-size packets, the ideal GPS link sharing isnot directly applicable because a packet must monopolize theserver (i.e., transmission link) while in service. WFQ [30] is apacket approximation of GPS whose deviation from the idealcase is bounded by the maximum packet size. WFQ calculatesthe start time of a packet under the ideal GPS system andbased on this start time, computes the finish time under idealGPS. Then, packets are transmitted in the order of the calcu-lated finish time. The calculations of the ideal GPS times canbe computationally intensive for ONUs. A few schemes wereproposed to simplify these calculations at the expense ofapproximation accuracy to the ideal GPS.

Start-time fair queuing (SFQ) [31] is one simplified versionof WFQ. In SFQ, the calculation of server virtual time that isused to calculate the start time of a packet is reduced to thestart time of the packet currently in service, greatly reducingthe computational complexity. In contrast to WFQ, SFQ sortspackets by start time rather than finish time. An intra-ONUframe scheduler that employs a modified start-time fair queu-ing (M-SFQ) algorithm [32, 33] can be used as a low complex-ity alternative to SFQ. M-SFQ further simplifies thescheduling by calculating the start time only for the HOLpackets.

Simulation results comparing M-SFQ to strict priorityscheduling indicate that M-SFQ provides the same averagedelay for the expedited forwarding class (delay sensitive traf-fic), higher average delay for assured services class 1 (high-

■ Figure 10. Quality of service taxonomy.

Differentiatedservice

[13,23,35,36,37]

Bandwidthguarantees

[38,39]

Delayguarantees

[40]

Delay jitterguarantees

[17,41]

Admissioncontrol

[44,45,46]

Quality of service


speed video), and lower average delay for assured servicesclass 2 (low-speed pre-recorded video). Further, the averagedelay for class 2 is lower than for class 1, which seems unde-sirable, but is not commented on in the simulation study. Onecan conclude that M-SFQ provides improved delay andthroughput for assured services class 2 at the expense of worseperformance of assured services class 1. It is not clear howthis displays the strengths of M-SFQ over strict priorityscheduling. M-SFQ achieves better inter-class isolation [33]but also treats class 2 traffic better than class 1 traffic, whichis undesirable.

An additional shortcoming of M-SFQ is that it tends tostarve best-effort traffic to provide better QoS to the assuredservices and expedited forwarding classes [34]. A modifiedversion of deficit weighted round robin (M-DWRR) was pro-posed and examined in [34] to address this shortcoming. M-DWRR maintains a credit deficit counter for each class andconsiders only the HOL packets, ensuring low computationalcomplexity. In a first scheduling pass, M-DWRR offers band-width to each class queue according to the class weight (whichis reflected in the deficit credit counter). The bandwidth thatis not required by the individual queues is redistributed in asecond scheduling pass, which has some resemblance to theexcess distribution mentioned earlier, but is conducted inter-nally by the ONU. Simulation results indicate that M-DWRRensures fairness according to the chosen weights for the dif-ferent service classes, including best-effort traffic. Also, over-all throughput with M-DWRR is about 10 percent higher thanwith M-SFQ due to eliminating best-effort traffic starvation.

A non-work-conserving scheduling discipline, called priori-ty with insertion scheduling (PIS) [25], that transmits real-time packets when their delay-bound will be exceeded is yetanother ONU queue scheduling approach. PIS allows non-real-time traffic to gain access to the medium, as long as thereal-time traffic can be delayed without detriment.

Summary — To keep ONU costs low, the complexity of theONUs should be kept low. Therefore, the ideal intra-ONUscheduler provides quality of service guarantees through linksharing with low complexity. Alternatively, the intra-ONUscheduling can be performed at the OLT. This can result inpotential scalability problems as the number of queuesincreases [9]. However, allowing the OLT to perform theONU queue scheduling reduces ONU complexity concerns.

QUALITY OF SERVICE

EPONs are intended not only to carry best-effort data traffic,they also are expected to carry packetized voice and videothat have strict bandwidth and delay requirements, as well asdelay jitter sensitivity. We present in Fig. 10 the taxonomy fororganizing the research work on quality of service guaranteesfor EPONs.

Differentiated Service — The simplest way to facilitate QoSis to provide differentiation of traffic and different service toeach differentiated traffic class. The ONUs classify and sepa-rately buffer ingress traffic and can perform strict priorityscheduling between the classes when deciding which frames tosend during a gated transmission window. The use of strictpriority scheduling is required for compliance with 802.1dbridging [23]. Standard strict priority scheduling results in aphenomenon referred to as the light load penalty [13, 23].The individual queue sizes are REPORTed at the end of agrant. During the period trge, more high-priority traffic canarrive at the ONU, which can preempt the lower-priority traf-fic that was accounted for in the REPORT. If the grant sizing

predicts this higher priority traffic, newly arriving during trge,then the grant will accommodate this traffic; otherwise, it willunfairly preempt the lower priority traffic that was accountedfor in the REPORT. This problem occurs at low loads whenthe grants are typically small and are more likely used up bythe newly arriving high-priority traffic. To alleviate this prob-lem, a two-stage buffering scheme [23] should be used at theONU. The two-stage buffering moves the frames that wereaccounted for in the REPORT scheduled into a single sec-ond-stage queue that is emptied first during the next grant.This scheme effectively enforces a strict priority schedulingperformed at REPORT time, as opposed to the time of thegrant, which will have a queue occupancy that is potentiallydifferent than that REPORTed. Simulation results [13, 23]demonstrate the existence of the light-load penalty and indi-cate that the two-stage buffering eliminates the light-loadpenalty at the expense of higher delay of high-priority traffic.

Strict priority scheduling can be extended to the PON level[35]. A DBA algorithm that uses a fixed-cycle length anddivides this cycle between three priority classes on a strict pri-ority basis is one approach. The OLT would send a separategrant for each priority class [13, 35].

A two-layer DBA (TLBA) scheme [36] for differentiatedservices also could be used. In the first layer, the OLT decidesthe cycle partitioning between the classes of service (class-layer allocation), and in the second layer, the partition foreach class is further divided for each ONU (ONU-layer allo-cation). Within a class, all ONUs share the bandwidth accord-ing to a max-min fairness policy. To keep any class frommonopolizing the available bandwidth in a frame, a per-classbandwidth threshold is enforced. The bandwidth thresholdguarantees a minimum bandwidth for a class under heavyload. Any remaining bandwidth from classes that request lessthan their threshold is divided among classes that requestmore than their threshold. The division of that remainingbandwidth is handled through weights that are assigned toeach class. For ONU buffer management, weighted randomearly detection (RED) can be used. Simulation results [36]indicate that TLBA can divide the bandwidth as set throughclass weights when under heavy load. This allows for band-width guarantees for each class. The results also show thatunder lower loads, TLBA allows for effective utilization of themedium.

Class-of-service oriented packet scheduling (COPS) [37] isanother method to provide differentiated services. COPS reg-ulates the traffic of each ONU, as well as each class-of-service(CoS) using two sets of credit pools, one per ONU and oneper CoS. Granting begins with the highest CoS and ends withthe lowest CoS. In the first round of granting, each ONU withtraffic for the current CoS is granted up to the number ofcredits stored for that ONU, as well as that CoS. To mitigatethe unused slot remainder for those grants that cannot befully satisfied, a threshold queue-reporting scheme is used. Atthe end of the first round, the unused credits are pooledtogether and in the second round, these unused credits aredistributed to the CoS-ONU pairs that were not fully satisfied.Simulation results [37] show that COPS can provide loweraverage and maximum delay for all but the highest CoS ascompared to IPACT with limited grant sizing (IPACT-LS).The highest CoS experiences slightly higher average delayunder COPS as compared to IPACT-LS.

All of the above schemes suggest differentiating traffic atthe ONU into classes, separately reporting queue sizes ofeach class, and allowing the OLT to provide individual grantsto each class. The schemes differ in how they determine thegrant sizes for each class. It is also apparent that two-stagebuffering is required to keep higher priority traffic arriving


during trge from preempting the lower priority traffic that wasaccounted for in the REPORT.

Bandwidth Guarantees — A DBA algorithm for EPONscalled Bandwidth Guaranteed Polling (BGP) [38] can be usedfor providing bandwidth guarantees. The BGP algorithm actsas a compromise between fixed TDM and statistical multiplex-ing. In BGP, ONUs are divided into two disjoint sets:• Bandwidth guaranteed ONUs• Non-bandwidth guaranteed ONUsThe algorithm maintains two polling tables.

The first polling table divides a fixed length polling cycleinto a number of bandwidth units. The required bandwidth ofan ONU, as dictated by a service level agreement (SLA) witha service provider, determines the number of bandwidth unitsallocated in the polling table to that ONU. A bandwidth-guar-anteed ONU with more than one entry in the polling tablehas these entries spreading through the table rather thanappearing contiguously. This lowers the average queuing delaybecause these ONUs are polled more frequently. However,the increased polling frequency results in more grants percycle and hence, more guard times between grants, and thuslower channel utilization. Further, fragmenting the grants canpotentially lead to lower grant utilization because Ethernetframes cannot be fragmented to be transmitted across grantboundaries. Therefore, a frame that is too large to fit in theremainder of a bandwidth unit must wait for the next band-width unit, and a portion of the current bandwidth unit iswasted. Lower grant utilization further reduces the channelutilization. BGP employs a method to mitigate the grant uti-lization problem by allowing an ONU to communicate itsactual use of a bandwidth unit to an OLT. If the unused por-tion of the bandwidth unit is sufficiently large, this portion isgranted to a non-bandwidth guaranteed ONU. Otherwise, thenext bandwidth-guaranteed ONU is polled. However, thisapproach is severely limited by the propagation delays (i.e.,walk times) required for message exchange on an EPON.

In BGP, unused bandwidth units in the first polling tableare given to non-bandwidth-guaranteed ONUs in their orderof appearance in the second polling table. The second pollingtable, is constructed differently than the first. Entries in thesecond polling table are created as non-bandwidth-guaranteedONUs request grants. This is in contrast to the first pollingtable, which represents a division of time on the upstreamchannel of the EPON. Simulation results presented in [38]show that, as one expects, ONUs with more entries in thepolling table have lower queuing delay than those with fewerentries, and IPACT lies somewhere in the middle.

BGP can be augmented to include admission control fornew bandwidth-guaranteed ONUs [39]. This admission con-trol uses standard parameter-based admission control. Thereare two parameters that describe the resource requirements ofONUs:• Bandwidth requirement (peak rate)• Delay bound requirementUsing these parameters, the admission control determineswhether the ONU is accepted as a bandwidth-guaranteed(BG) ONU or as a best-effort (non-BG) ONU.

Delay Guarantees — A DBA algorithm called Dual DEB-GPS Scheduler [40] potentially can help provide delay guaran-tees. This DBA uses deterministic effective bandwidth (DEB),to determine the scheduling weights used in a generalizedprocessor sharing (GPS) scheduler. The scheduling is per-formed in two layers, hence the name Dual. The first layerperforms class-level multiplexing at the OLT. The secondlayer performs source level multiplexing at the ONU.

Traffic arriving at the ONU is regulated by a leaky bucketmechanism. This leaky bucket enforces a traffic profile char-acterized by: burst size P, peak rate {, and average rate J.These leaky bucket parameters are used to determine theDEB for a source. This DEB is directly used as a weight todetermine the portion of the upstream bandwidth assigned tothis traffic source. The DEB guarantees a particular delaybound. Let U be the desired delay bound, and Beff (U) be theeffective bandwidth to guarantee the delay bound. Then,

(2)

The OLT divides each bandwidth cycle according to theBeff (U) weights; the remaining bandwidth in a cycle is dividedequally among all best-effort sources (i.e., those that do notrequire any delay bounds). The OLT generates the grants perONU based on the weights and the information about whichsources belong to each ONU. The ONU is then responsiblefor performing the intra-ONU scheduling that determineshow the different sources fill the granted transmission win-dow. DEB attempts to guarantee delay bounds by guarantee-ing a certain bandwidth (i.e., Beff).

Simulation results are presented [40] with ONUs havingtwo QoS-aware traffic sources (QoS1 and QoS2) and twobest-effort sources (BE1 and BE2). QoS1 has a delay boundof 1 msec and jitter bound of 0.2 msec. QoS2 has a delaybound of 2 msec and jitter bound of 0.4 msec. According tothe presented results, QoS1 traffic experiences roughly a 2 to2.5-millisecond delay at a load of 0.7 and higher (bound is 1millisecond). QoS2 traffic experiences roughly a 3-milliseconddelay at a load of 0.7 and higher (bound is 2 milliseconds).From the presented results, it seems Dual DEB-GPS provideslow delay and jitter for traffic sources that require QoS butdoes not provide a guaranteed delay bound at higher loads.

Delay Jitter Guarantees — A scheme called the Hybrid SlotSize/Rate algorithm (HSSR) [41] can stabilize packet delayvariation in EPONs for jitter-sensitive high-priority traffic.HSSR not only uses a fixed cycle length but also fixes theposition of jitter-sensitive high-priority traffic grants (fixed tothe beginning of the frame). The lower priority traffic fromONUs occupies the remainder of the frame. HSSR causesmore than one grant per cycle to an ONU, which reduces effi-ciency due to extra guard times. However, the reduced effi-ciency allows for guaranteeing packet-delay variation boundsfor certain traffic. A portion of the fixed grant cycle is parti-tioned for jitter-sensitive traffic. Quasi non-intrusive rangingkeeps the ranging and registration process of new ONUs fromdisturbing high-priority traffic. The ranging and registrationresponses from new ONUs are scheduled to occur during thebest-effort portion of the fixed frame. The fixed frame is largeenough to provide ample time for this process.

Simulation results show that HSSR offers lower averagedelay and packet-delay variation than the conventionalscheme (i.e., no fixed position of high-priority traffic). Fur-ther, the packet-delay variation for HSSR is due solely toqueuing delay and not from the scheduling by the DBA.

The Hybrid Granting Protocol (HGP) [17] can ensure QoSthrough minimizing jitter and guaranteeing bandwidth. It is ahybrid of two approaches to sizing grants, one uses the

Beff

if

if

( )

:

:

U

P

UP{

U P{

U P

=+

¾ÀÏ

ÁÂÑ

H H [¾

ÀÏÁ

ÂÑ

>

01 1

J

J11 1

J[

¾

ÀÏÁ

ÂÑ

Ä

Å

ÈÈÈ

Ç

ÈÈÈ {

.


REPORT message to size the grant; the other uses a queueprediction mechanism to size the grant to accommodate allqueued traffic at the point the grant begins (i.e., accommo-dates traffic in trgs). The first approach is used for assured for-warding (AF) [42] and best-effort (BE) services, whereas thelatter approach is used for expedited-forwarding (EF) [43]services that are assumed to have a constant bit rate andtherefore can be estimated easily.

A scheduling cycle is then divided into two subcycles: EFsubcycle and AF/BE subcycle. The EF subcycle carries the EFservices for each ONU and the AF/BE subcycle carries theAF and BE services for each ONU. Hence, every schedulingcycle there are two grants for each ONU. REPORTing of theAF and BE queues in an ONU is delayed until the end of theEF grant for that ONU. This allows the OLT to obtain moreup-to-date queue occupancy for the ONU because the DBAcomputation is performed after the EF subcycle.

By fixing the scheduling cycle size and fixing the positionof the EF grant to each ONU, HGP can guarantee bandwidthto EF traffic and minimize the jitter experienced by the EFtraffic. This QoS comes at the expense of more guard timesper cycle because of the separate grant to each ONU for itsEF traffic. To mitigate this inefficiency, a grant for AF/BEtraffic is not sent if there is no pending traffic, eliminating theneed for a guard time for the AF/BE grant.

HGP and HSSR [41] share the same frame structure anddivision. The novelty of HGP is the way in which theREPORTs are generated. Simulation results [17] show thatHGP provides lower queuing delay at higher loads as com-pared to a regular EPON scheduler. At lower loads, the regu-lar EPON scheduler provides lower queuing delay, which isattributed to the increased number of guard times per cyclewith HGP.

Admission Control — As reviewed, a variety of inter- andintra-ONU scheduling solutions exist to provide QoS inEPON networks. These solutions should be effective not onlyin supporting QoS but also must be designed carefully to pro-tect the requirements of already admitted traffic, as specifiedin the associated SLAs. Toward this end, admission controlmay become necessary to both support and protect the QoSrequirements in EPON networks.

Research on admission control for EPON began only veryrecently [44, 45]. An admission control framework togetherwith an appropriate DBA algorithm that is capable of sup-porting and protecting QoS of real-time traffic while guaran-teeing a minimum bandwidth for best-effort traffic wasintroduced and studied in [46]. The examined admission con-trol algorithm determines whether or not to admit a new real-time traffic stream based on its requirements and theutilization of the upstream wavelength channel. To achievethis, each polling cycle is divided into two subcycles. In thefirst subcycle, each ONU is assigned a guaranteed minimumbandwidth to support the respective QoS requirements of itsstreams. The second subcycle is used by the OLT to dynami-cally assign transmission windows to best-effort traffic of allONUs. The proposed admission control proceeds in twosteps. First, by using its assigned guaranteed bandwidth, eachONU performs local rate-based admission control accordingto the bandwidth requirements of newly arriving flows andcurrent bandwidth availability. Second, flows that could not belocally admitted are reported to the OLT, which in turn triesto admit them in the second subcycle, provided sufficientunused bandwidth is available.

The performance of the proposed admission control wasstudied by means of simulation using a strict priority and adeficit-weighted round-robin based intra-ONU scheduling

algorithm for real-time voice and video streams, as well asbest-effort data traffic. The obtained results show that theconsidered admission control is able to satisfy the QoSrequirements in terms of delay bound and throughput.

Summary — Differentiated services require differentiatedqueuing and reporting and different grant sizing and schedul-ing treatment for each class at the OLT. As the number ofqueues increases, scalability becomes an issue and a hierarchi-cal approach should be followed. Grant reservations arerequired for providing bandwidth guarantees. A fixed andproperly-sized cycle length with fixed position of delay and jit-ter-sensitive traffic can provide delay and jitter guarantees.More research must be conducted in the area of providingbandwidth, delay, and jitter guarantees on an EPON.

FAIRNESS

EPONs carry traffic from a diverse group of non-cooperativesubscribers. This non-cooperation requires fairness mecha-nisms to ensure that all nodes receive their “fair” share of thenetwork resources. Fairness is most effectively tackled by thegrant-sizing process and requires the use of the offline DBAframework. The limited and limited with excess distributiongrant-sizing schemes offer a method to provide fairness on anEPON. We now discuss other approaches to fairness.

A sibling fair scheduler [9], in the case of EPONs, is ascheduler that ensures fairness for queues within an ONU. Acousin fair scheduler [9] can guarantee fairness among allleaves (i.e., queues) regardless of grouping. An EPONrequires a scheduler that is fair among all queues and there-fore requires a cousin fair scheduler. Fair Queuing with Ser-vice Envelopes (FQSE) [9] is a hierarchical schedulingalgorithm that is cousin fair.

A service envelope (SE) is a grant to a node in the schedul-ing hierarchy. A service envelope for a leaf node is a piece-wise linear function of how satisfied a node is, wherebysatisfaction is measured in terms of a satisfiability parameter(SP). The SP begins at 0 when the SE is at its minimum guar-anteed bandwidth and increases linearly until the request iscompletely satisfied. The slope of the linear increase is deter-mined by the weight of the node.

The SE of non-leaf nodes is calculated as an approxima-tion of the sum of the SEs of all their children. The approxi-mation is necessary to keep the number of knee points in thepiecewise linear function from exceeding a prescribed maxi-mum. Limiting the number of knee points is necessary to keepthe request message that conveys the piecewise linear functionwithin a fixed length.

A deficit round robin approach within an ONU can beused to allow queues to contend for the cumulative slotremainder. This guarantees that there is only one slot remain-der per ONU as opposed to a slot remainder for each queue.As a result, unused slot remainders are reduced. Simulationand mathematical analysis [9] show that FQSE is capable ofmaking bandwidth guarantees, as well as providing fairnessbetween all queues regardless of which ONU they reside in(i.e., cousin fairness).

For open access EPONs, in which multiple service pro-viders share a single EPON, fairness must be maintainedamong service providers, as well as subscribers. Dual servicelevel agreements (Dual-SLAs) [47] can be used to manage thefairness for both subscribers and service providers. A Dual-SLA manages two sets of SLAs, one for the subscribers andone for the service providers. One of the sets of SLAs isselected as the primary. The primary SLA set is given priorityover the secondary.


MAC PROTOCOLS FOR

ALTERNATIVE PON ARCHITECTURES

MAC PROTOCOLS FOR A BROADCAST PON

Full-utilization local-loop request contention multiple access(FULL-RCMA) [5] is an extension of RCMA that allows fortransmission interleaving to tolerate the walk times on anEPON and provides a bounded cycle time for support of CBRtraffic. RCMA is a hybrid between a token-passing and con-tention MAC protocol. A cycle of time is divided into a con-tention-based request period and a token-passing data period.During the request period, ONUs contend through requestmessages to be added to the token-passing list for the upcom-ing data period. One of the ONUs is designated the masterand generates the token-passing list for the upcoming dataperiod. ONUs, according to the order determined by the mas-ter, pick up the token to gain access to the medium for theirtransmission. If ONUs remain backlogged after receiving theirtransmission window, they set a “more data” bit to signify thatthey automatically should be added to the token-passing listfor the next cycle. This helps to reduce the probability ofrequest collisions. ONUs that are added to the token-passinglist through request contention are given priority over ONUsthat are added through setting the “more data” bit in the pre-vious cycle. This allows new ONUs to gain prompt access tothe network. Performance analysis shows that FULL-RCMAcan provide higher upstream link utilization compared toIPACT or RCMA. However, a comparison of average packetdelay between the schemes is not available.

For the MAC protocol proposed in [6], cycles are dividedinto a control period and a data period separated by a waitingtime that is used to collect all control messages and to pro-duce a schedule. The control period is divided using fixedtime division multiple access (TDMA), one slot for eachONU. The ONUs independently compute the same transmis-sion schedule given the control information, and the dataperiod is divided between the ONUs according to this sched-ule. Because the control messages and data messages are sep-arated, there are 2 · N guard intervals per scheduling cycle asopposed to N guard intervals for a typical EPON, in which thecontrol information is appended to the end of the data trans-missions. This increased number of guard intervals degradesthe channel utilization.

MAC PROTOCOLS FOR TWO-STAGE PONS

A DBA scheme for a two-stage PON called EPON DynamicScheduling Algorithm (EDSA-2) [7] takes advantage of somepredictability of the aggregated traffic from the sub-OLTs.Because sub-OLTs aggregate traffic from several burstysources, the sub-OLT traffic tends to be less bursty and hence,more predictable. Specifically, EDSA-2 predicts the traffic dur-ing trgs by assuming short-term rate stationarity from the aggre-gated ONUs. Differentiated services are provided by having aset of class queues for the local sub-OLT traffic, as well as aseparate set of class queues for the aggregated ONU traffic.The aggregate ONU class queues are given priority over thecorresponding class queues for the local sub-OLT traffic.

Simulation results that compare EDSA-2 to a QoS DBAthat performs no traffic prediction (EDSA-1) show thatEDSA-2 provides lower average packet delay for all classes oftraffic from both standard ONUs and the sub-OLT ONUs.The lower delays for the sub-OLT traffic can be explained bythe accommodation of traffic received during the grantingperiod (i.e., trgs) through traffic prediction. However, it is notclear why the standard ONUs would experience lower delays.

CONCLUSION

In this article, we summarized and classified the existingresearch on EPONs. We introduced a meaningful frameworkthat allows those interested in advancing EPON research toquickly understand the state-of-the-art and to identify areasrequiring further study. We outlined the standard physicalPON architecture, as well as two alternative architectures,broadcast PON and two-stage PON. We also examined andprovided a meaningful taxonomy for dynamic bandwidth allo-cation. Using this taxonomy, we presented the existing workon dynamic bandwidth allocation. The major branches of thetaxonomy are• Grant sizing• Grant scheduling• Queue schedulingWe also surveyed the existing approaches for supporting qual-ity of service and fairness. Finally, we presented a discussionof protocols for the alternative physical PON architectures.We conclude by outlining areas that we believe are in urgentneed of further research.

The problem of ONU grant sizing has received significantattention from the research community, but there are stillimportant open questions. Providing an exhaustive service dis-cipline for EPONs can lower queuing delays by up to a cycletime. However, due to the nature of EPONs as a remotescheduling system, an exhaustive service discipline is not pos-sible without queue-size prediction. Prediction of queuedCBR traffic is straightforward as a result of its constant rate.VBR traffic, on the other hand, is difficult to predict. Withthe proliferation of VBR video through IPTV services, itwould be worthwhile to explore schemes to predict its short-term bandwidth requirements. These schemes can be used forqueue-size prediction to help lower the queuing delay for thisdelay-sensitive traffic. Data traffic is typically delay insensitive,so there is less of a need to reduce the queuing delays for thistype of traffic.

The problem of distributing excess bandwidth has beenexplored in the context of the offline DBA framework (i.e.,interleaved polling with stop). It would be of value to explorethe possibility of providing fair excess bandwidth distributionin the online DBA framework (i.e., a purely interleaved grant-ing system).

The topic of ONU grant scheduling has received someattention from the research community. However, we feelthis topic can be explored further to uncover the best grantscheduler for an EPON. The topic of ONU grant schedulinganchored in scheduling theory has been studied extensivelyin [48, 49] within the context of multiple-channel EPONs.Similar exposition should be extended to single-channelEPONs.

With respect to providing QoS on EPONs, providing dif-ferentiated services has received significant attention from theresearch community. However, providing bandwidth, delay,and delay variation (i.e., jitter) guarantees requires furtherstudy. Providing guaranteed service across an EPON will becritical because the access network is required to be an inte-grated services network carrying packetized voice and videoalong with data traffic. The voice and video services willrequire some level of bandwidth, delay, and jitter guaranteesfor successful operation.

Emerging from the work on single-channel EPONs,researchers are beginning to extend the DBA problem toEPONs that employ more than one upstream and/or down-stream channel [14, 50, 51, 52]. DBA for multi-wavelengthEPONs represents a broad area for future research.


REFERENCES[1] M. Hajduczenia et al., “On EPON Security Issues,” IEEE Com-

mun. Surveys and Tutorials, vol. 9, no. 1, 1st Quarter 2007,pp. 68–83.

[2] M. McGarry, M. Maier, and M. Reisslein, “Ethernet PONs: ASurvey of Dynamic Bandwidth Allocation (DBA) Algorithms,”IEEE Commun. Mag., vol. 42, no. 8, Aug. 2004, pp. S8–S15.

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

[4] G. Kramer, B. Mukherjee, and G. Pesavento, “Ethernet PON(ePON): Design and Analysis of an Optical Access Network,”Photonic Network Commun., vol. 3, no. 3, July 2001, pp.307–19.

[5] C. Foh et al., “FULL-RCMA: A High Utilization EPON,” IEEE

JSAC, vol. 22, no. 8, Oct. 2004, pp. 1514–24.[6] S. Sherif et al., “A Novel Decentralized Ethernet-Based PON

Access Architecture for Provisioning Differentiated QoS,”IEEE/OSA J. Lightwave Tech., vol. 22, no. 11, Nov. 2004, pp.2483–97.

[7] A. Shami et al., “QoS Control Schemes for Two-Stage EthernetPassive Optical Access Networks,” IEEE JSAC, vol. 23, no. 8,Aug. 2005, pp. 1467–78.

[8] T. Holmberg, “Analysis of EPONs under the Static PriorityScheduling Scheme with Fixed Transmission Times,” Proc. IEEE

Conf. Next Generation Internet Design and Engineering (NGI),Apr. 2006, pp. 192–99.

[9] G. Kramer et al., “Fair Queuing with Service Envelopes (FQSE):A Cousin-Fair Hierarchical Scheduler For Subscriber Access Net-works,” IEEE JSAC, vol. 22, no. 8, Oct. 2004, pp. 1497–1513.

[10] S. Bhatia, D. Garbuzov, and R. Bartos, “Analysis of the GatedIPACT Scheme for EPONs,” Proc. IEEE ICC, June 2006, pp.2693–98.

[11] F. Aurzada et al., “Delay Analysis of Ethernet Passive OpticalNetworks with Gated Service,” Arizona State University Techni-cal Report, Mar. 2007.

[12] J. Zheng and H. Mouftah, “Media Access Control for EthernetPassive Optical Networks: An Overview,” IEEE Commun. Mag.,vol. 43, no. 2, Feb. 2005, pp. 145–50.

[13] C. Assi et al., “Dynamic Bandwidth Allocation for Quality-of-Service over Ethernet PONs,” IEEE JSAC, vol. 21, no. 9, Nov.2003, pp. 1467–77.

[14] A. Dhaini et al., “Dynamic Wavelength and Bandwidth Alloca-tion in Hybrid TDM/WDM EPON Networks,” IEEE/OSA J. Light-

wave Tech., vol. 25, no. 1, Jan. 2007, pp. 277–86.[15] X. Bai, A. Shami, and C. Assi, “On the Fairness of Dynamic

Bandwidth Allocation Schemes in Ethernet Passive Optical Net-works,” Computer Commun., vol. 29, no. 11, July 2006, pp.2123–35.

[16] D. Bertsekas and R. Gallager, Data Networks, 2nd Ed., Pren-tice Hall, 1991.

[17] A. Shami et al., “Jitter Performance in Ethernet Passive Opti-cal Networks,” IEEE/OSA J. Lightwave Tech., vol. 23, no. 4,Apr. 2005, pp. 1745–53.

[18] H.-J. Byun, J.-M. Nho, and J.-T. Lim, “Dynamic BandwidthAllocation Algorithm in Ethernet Passive Optical Networks,”Electronics Letters, vol. 39, no. 13, June 2003, pp. 1001–2.

[19] Y. Luo and N. Ansari, “Bandwidth Allocation for MultiserviceAccess on EPONs,” IEEE Commun. Mag., vol. 43, no. 2, Feb.2005, pp. S16–S21.

[20] Y. Luo and N. Ansari, “Limited Sharing with Traffic Predictionfor Dynamic Bandwidth Allocation and QoS Provisioning overEthernet Passive Optical Networks,” OSA J. Opt. Net., vol. 4,no. 9, Sept. 2005, pp. 561–72.

[21] S. Haykin, Neural Networks: A Comprehensive Foundation,

2nd Ed., Prentice Hall, 1999.[22] S. Ostring and H. Sirisena, “The Influence of Long-Range

Dependence on Traffic Prediction,” Proc. IEEE Int’l. Conf. Com-

munications (ICC), vol. 4, June 2001, pp. 1000–1005.[23] G. Kramer et al., “Supporting Differentiated Classes of Service

in Ethernet Passive Optical Networks,” OSA J. Opt. Net., vol. 1,no. 8, Aug. 2002, pp. 280–98.

[24] Y. Zhu, M. Ma, and T. Cheng, “Hierarchical Scheduling toSupport Differentiated Services in Ethernet Passive Optical Net-

works,” Computer Networks, vol. 50, no. 3, Feb. 2006, pp.350–66.

[25] M. Ma, L. Liu, and T. H. Cheng, “Adaptive Scheduling for Dif-ferentiated Services in the Ethernet Passive Optical Networks,”Proc. 9th Int’l. Conf. Commun. Systems, Sept. 2004, pp.102–6.

[26] M. Pinedo, Scheduling: Theory, Algorithms, and Systems, 2nd

Ed., Prentice Hall, 2002.[27] J. Zheng and H. Mouftah, “Adaptive Scheduling Algorithms

for Ethernet Passive Optical Networks,” IEE Proc. Commun.,vol. 152, no. 5, Oct. 2005, pp. 643–47.

[28] J. Zheng, “Efficient Bandwidth Allocation Algorithm for Eth-ernet Passive Optical Networks,” IEE Proc. Commun., vol. 153,no. 3, June 2006, pp. 464–68.

[29] A. Parekh and R. Gallager, “A Generalized Processor SharingApproach to Flow Control in Integrated Services Networks: TheSingle-Node Case,” IEEE/ACM Trans. Net., vol. 1, no. 3, June1993, pp. 344–57.

[30] A. Demers, S. Keshav, and S. Shenker, “Analysis and Simula-tion of a Fair Queuing Algorithm,” Internetworking: Research

and Experience, vol. 1, no. 1, Sept. 1990, pp. 3–26.[31] P. Goyal, H. M. Vin, and H. Cheng, “Start-Time Fair Queuing:

A Scheduling Algorithm for Integrated Services Packet Switch-ing Networks,” IEEE/ACM Trans. Net., vol. 5, no. 5, Oct. 1997,pp. 690–704.

[32] N. Ghani et al., “Quality of Service in Ethernet Passive OpticalNetworks,” Proc. 2004 IEEE/Sarnoff Symp. Advances in Wired

and Wireless Commun., Apr. 2004, pp. 161–65.[33] N. Ghani et al., “Intra-ONU Bandwidth Scheduling in Ethernet

Passive Optical Networks,” IEEE Commun. Letters, vol. 8, no.11, Nov. 2004, pp. 683–85.

[34] A. Dhaini et al., “Adaptive Fairness through Intra-ONUScheduling for Ethernet Passive Optical Networks,” Proc. IEEE

Int. Conf. Commun. (ICC), June 2006, pp. 2687–92.[35] S.-I. Choi and J.-D. Huh, “Dynamic Bandwidth Allocation

Algorithm for Multimedia Services over Ethernet PONs,” ETRI

Journal, vol. 24, no. 6, Dec. 2002, pp. 465–68.[36] J. Xie, S. Jiang, and Y. Jiang, “A Dynamic Bandwidth Alloca-

tion Scheme for Differentiated Services in EPONs,” IEEE Com-

mun. Mag., vol. 42, no. 8, Aug. 2004, pp. S32–S39.[37] H. Naser and H. Mouftah, “A Joint-ONU Interval-Based

Dynamic Scheduling Algorithm for Ethernet Passive OpticalNetworks,” IEEE/ACM Trans. Net., vol. 14, no. 4, Aug. 2006,pp. 889–99.

[38] M. Ma, Y. Zhu, and T. Cheng, “A Bandwidth GuaranteedPolling MAC Protocol for Ethernet Passive Optical Networks,”Proc. IEEE INFOCOM, vol. 1, Mar. 2003, San Francisco, pp.22–31.

[39] M. Ma, Y. Zhu, and T. Cheng, “A Systematic Scheme for Mul-tiple Access in Ethernet Passive Optical Access Networks,”IEEE/OSA J. Lightwave Technology, vol. 23, no. 11, Nov. 2005,pp. 3671–82.

[40] L. Zhang et al., “Dual DEB-GPS Scheduler for Delay-ConstraintApplications in Ethernet Passive Optical Networks,” IEICE

Trans. Commun., vol. E86-B, no. 5, May 2003, pp. 1575–84.[41] F. An et al., “A New Dynamic Bandwidth Allocation Protocol

with Quality of Service in Ethernet-Based Passive Optical Net-works,” Proc. IASTED Int’l. Conf. Wireless and Optical Commu-

nications (WOC 2003), vol. 3, July 2003, pp. 165–69.[42] J. Heinanen et al., “Assured Forwarding PHB Group,” RFC

2597 (Proposed Standard), June 1999, updated by RFC 3260;http://www.ietf.org/rfc/rfc2597.txt

[43] V. Jacobson, K. Nichols, and K. Poduri, “An Expedited For-warding PHB,” RFC 2598 (Proposed Standard), June 1999,obsoleted by RFC 3246; http://www.ietf.org/rfc/rfc2598.txt

[44] A. Dhaini et al., “Admission Control in Ethernet Passive Opti-cal Networks (EPONs),” Proc. IEEE Int’l. Conf. Communications

(ICC), Glasgow, Scotland, June 2007.[45] A. Dhaini et al., “Per-Stream QoS and Admission Control in

Ethernet Passive Optical Networks (EPONs),” IEEE/OSA J. Light-

wave Tech., vol. 25, no. 7, July 2007, pp. 1659–69.[46] C. Assi, M. Maier, and A. Shami, “Toward Quality-of-Service

Protection in Ethernet Passive Optical Networks: Challengesand Solutions,” IEEE Network, vol. 21, no. 5, Sept.–Oct. 2007.

[47] A. Banerjee, G. Kramer, and B. Mukherjee, “Fair Sharing


Using Dual Service-Level Agreements to Achieve Open Accessin a Passive Optical Network,” IEEE JSAC, vol. 24, no. 8, Aug.2006, pp. 32–44.

[48] M. McGarry et al., “Bandwidth Management for WDMEPONs,” OSA J. Opt. Net., vol. 5, no. 9, Sept. 2006, pp.637–54.

[49] M. McGarry et al., “Just-in-Time Online Scheduling for WDMEPONs,” Proc. IEEE ICC 2007, June 2007.

[50] K. Kwong, D. Harle, and I. Andonovic, “Dynamic BandwidthAllocation Algorithm for Differentiated Services over WDMEPONs,” Proc. 9th IEEE Int’l. Conf. Commun. Systems (ICCS),Sept. 2004, pp. 116–20.

[51] M. McGarry, M. Maier, and M. Reisslein, “WDM Ethernet Pas-sive Optical Networks,” IEEE Commun. Mag., vol. 44, no. 2,Feb. 2006, pp. S18–S25.

[52] A. Dhaini, C. Assi, and A. Shami, “Quality of Service inTDM/WDM Ethernet Passive Optical Networks (EPONs),” Proc.

IEEE ISCC 2006, June 2006, pp. 616–21.

BIOGRAPHIESMICHAEL MCGARRY ([email protected]) received his BS inComputer Engineering from Polytechnic University, Brooklyn, NYin 1997. He received his M.S. and Ph.D. in Electrical Engineeringfrom Arizona State University, Tempe, AZ in 2004 and 2007respectively. He is an Assistant Professor at the University ofAkron, Akron, OH. From 2007 through 2008 he was a Senior StaffScientist at ADTRAN and an Adjunct Professor at Arizona State

University. From 1997 through 2003 he was employed in industryby companies including PMC-Sierra and Yurie Systems (now partof Alcatel-Lucent). His research interests include congestion con-trol and the optimization of MAC protocols for both opticalaccess and mobile ad hoc networks.

MARTIN REISSLEIN ([email protected]) received his Ph.D. in systemsengineering from the University of Pennsylvania, Philadelphia, in1998. He is an associate professor in the Department of ElectricalEngineering at Arizona State University, Tempe. From July 1998through October 2000 he was a scientist with the GermanNational Research Center for Information Technology (GMDFOKUS), Berlin, and lecturer at the Technical University Berlin. Hemaintains an extensive library of video traces for network perfor-mance evaluation, including frame size traces of MPEG-4 andH.264 encoded video, at http://trace.eas.asu.edu.

MARTIN MAIER ([email protected]) received his M.Sc. and Ph.D.degrees (both with distinction) in electrical engineering from theTechnical University Berlin, Berlin, Germany, in 1998 and 2003,respectively. In the summer of 2003, he was a Postdoctoral Fellowat the Massachusetts Institute of Technology (MIT), Cambridge.He is an associate professor with the Institut National de laRecherche Scientifique (INRS), Montreal, Canada. His recentresearch activities aim at providing insights into technologies,protocols, and algorithms shaping the future of optical networksand their seamless integration with next-generation wireless net-works. He was a visiting professor at Stanford University, Califor-nia, October 2006 through March 2007.

1

Shortest Propagation Delay (SPD) First Scheduling

for EPONs with Heterogeneous Propagation DelaysMichael P. McGarry, Martin Reisslein, Frank Aurzada, and Michael Scheutzow

Abstract—Due to the geographic distribution of its subscribers,Ethernet Passive Optical Networks (EPONs) have typicallyvarying propagation delays between the Optical Network Units(ONUs) and the Optical Line Terminal (OLT). In this paper, weconsider EPONs with an offline scheduling framework, whichenables Quality-of-Service mechanisms by collecting bandwidthrequests from all ONUs before the OLT makes dynamic band-width allocations for transmissions on the shared ONUs-to-OLT upstream channel. We propose and evaluate the ShortestPropagation Delay (SPD) first scheduling policy which sequencesthe ONUs’ upstream transmissions in increasing order of theONUs’ propagation delays, i.e., the upstream transmission ofthe ONU with the smallest propagation delay is scheduled first.We formally analyze the competitiveness of SPD first schedulingand find that it achieves very close to optimal performance.We characterize the stability limit for Gated and Limited grantsizing in conjunction with SPD grant scheduling. We evaluatethe cycle length and packet delay with SPD scheduling throughprobabilistic analysis and simulations and find significant reduc-tions in packet delay with SPD first scheduling in EPONs withheterogeneous propagation delays, especially when Limited grantsizing is employed.

Index Terms—Ethernet Passive Optical Network, Grantscheduling, Packet delay, Propagation delay.

I. INTRODUCTION

Passive Optical Networks (PONs) have emerged as an

attractive technology for high-speed access networks [1]–[4].

In particular, the combination of PON technologies with the

ubiquitous Ethernet networking technologies has made Ether-

net PON (EPON) a promising access network choice [5]–[14].

In Ethernet Passive Optical Networks (EPONs), an Optical

Network Unit (ONU) provides high-speed network access to

an individual subscriber or a group of subscribers. Several

ONUs connect to a single Optical Line Terminal (OLT),

typically in the form of a tree topology rooted at the OLT.

Due to the geographic distribution of the served subscribers,

the individual ONUs have typically different distances, and

thus different propagation delays from the OLT. With the

emergence of long reach and next-generation PONs [15]–

[19] covering larger geographic areas with spans of 100 km

or higher (i.e., one-way ONU-to-OLT propagation delays of

M. McGarry is with the Dept. of Electrical and Computer Eng., Universityof Akron, Akron, OH 44325, Email: [email protected], Phone: (330)972-7168, Fax: (330) 972-6487.

M. Reisslein is with the School of Electrical, Computer, and EnergyEngineering, Arizona State University, Tempe, AZ 85287-5706, Email:[email protected], Phone: (480) 965-8593, Fax: (480) 965-8325.

F. Aurzada and M. Scheutzow are with the Institute of Mathematics,Technical University Berlin, Berlin, Germany, Email: {aurzada, ms}@tu-berlin.de.

0.5 ms or higher), the disparities of the propagation delays are

likely becoming more pronounced.

EPONs avoid collisions on the shared upstream (ONUs-to-

OLT) channel through a polling based medium access control

protocol. The ONUs signal their bandwidth demands with

REPORT messages to the OLT, while the OLT dynamically

allocates bandwidth and schedules the upstream transmis-

sions so as to avoid collisions. The OLT signals the ONUs

with GATE messages their upstream transmission windows

(grants). A key challenge for efficient sharing of the upstream

channel is the masking of the round trip propagation delay

between OLT and ONUs. One of the first approaches for

masking the propagation delays has been the Interleaved

Polling with Adaptive Cycle Time (IPACT) approach [9], [20]

which interleaves the REPORT-GATE cycles of the individual

ONUs so that they can mask each others propagation delays.

The basic IPACT approach implements the online scheduling

framework in that the OLT considers a single ONU REPORT

when making bandwidth allocation and scheduling decisions.

Quality of Service (QoS) control generally requires that the

OLT considers and trades off requests from several ONUs

when making bandwidth allocation and scheduling decisions.

With the offline scheduling framework [21], which is referred

to as interleaved polling with stop in [14], [22], the OLT

collects REPORTs from all ONUs before making bandwidth

allocation and scheduling decisions. The offline scheduling

framework thus enables the wide variety of QoS mechanisms,

see for instance, [23]–[28], which consider jointly all RE-

PORTs in their bandwidth allocation and scheduling decisions.

On the downside, the offline scheduling framework imposes

an idle period on the upstream channel between cycles due to

the OLT schedule computation time, and transmission time of

the first GATE message and the round trip propagation delay

to the first scheduled ONU of a new cycle, as described in

more detail in Section III-B. Further idle periods are possible

if an upstream transmission is not long enough to mask the

propagation delay to the next ONU in the schedule.

A few studies have pursued strategies that combine online

scheduling and offline scheduling. For instance, the stud-

ies [29]–[32] schedule ONUs with small bandwidth requests

immediately (i.e., in online fashion), while ONUs with large

bandwidth requests are only scheduled after REPORTs from

all ONUs have been collected (i.e., in offline fashion) and more

informed decisions are possible. When the ONU propagation

delays are fairly homogeneous, scheduling small bandwidth

request (which have a small impact on the QoS and fairness

properties of the schedule) right away can indeed be a good

strategy to mask propagation delays for the larger requests,

2

which require more careful informed decisions. However,

when the propagation delays of the ONUs are vastly different

then scheduling a small grant for a far-away ONU can result

in large idle times. Thus, for EPONs with heterogeneous

propagation delays, the scheduling decisions need to take the

propagation delays into consideration.

In this paper, we examine, to the best of our knowledge, for

the first time the problem of efficiently masking heterogeneous

propagation delays in EPONs with offline scheduling. We

propose and evaluate the Shortest Propagation Delay (SPD)

first scheduling policy. The SPD first policy strives to mask

the long round trip propagation delays to far-away ONUs

by first scheduling the upstream transmissions of near-by

ONUs. We prove that the SPD first policy minimizes the

cycle length to within a small time period (number of ONUs

times transmission delay of GATE message) of an optimal

scheduling policy. We characterize the cycle length and packet

delay of SPD first scheduling for Gated grant sizing in low

load and high load regimes through probabilistic analysis and

derive stability limits for Limited grant sizing [9]. We conduct

extensive simulations to verify our analysis and to broadly

assess the reductions in cycle length and packet delay as well

as the increase in channel utilization achieved with SPD first

scheduling.

Importantly, by including a sufficient number of close-

by ONUs with small propagation an EPON using SPD first

scheduling can be engineered to allow for offline scheduling

with a very small imposed idle time between scheduling

cycles. Further, SPD first scheduling is very simple in that

a given set of served ONUs needs to be sorted only once in

increasing propagation delays.

This article is organized as follows. In Section II we

provide background on scheduling in EPONs and review

related work. In Section III we formally model the grant

scheduling problem with heterogeneous propagation delays

and characterize the competitiveness of SPD first scheduling.

We also derive approximations of the mean cycle length and

packet delay. In Section IV we present numerical results from

our analytical cycle length and delay evaluation and provide

extensive simulation results for SPD first scheduling. Finally,

in Section V we summarize our findings.

II. BACKGROUND AND RELATED WORK

In this section we briefly provide background on the dy-

namic bandwidth allocation in EPONs and review related

research on scheduling in EPONs. The dynamic bandwidth

allocation (DBA) in EPONs can be divided into: 1) the

sizing of the upstream transmission windows (grants), and

2) the scheduling of the grants on the upstream wavelength

channel [11]. Widely used grant sizing methods are Gated

grant sizing, where the OLT sets the grant size equal to the

ONU request and Limited grant sizing, where the OLT sets

the grant size equal to the ONU request up to a maximum

grant size; if request exceeds the maximum grant size, then

the maximum grant size is allocated [9], [20].

As noted in the Introduction, the basic online scheduling

framework considers and schedules one ONU request at a

time, whereas the offline scheduling framework collects re-

ports from all ONUs before making scheduling decisions. A

few studies have examined a just-in-time scheduling frame-

work, where ONU requests are collected and scheduling

decisions are made when the channel is about to become

idle [21], [33], [34]. In [35] the ONUs are split into two groups

whereby each group is scheduled in offline fashion and the

cycles of the two groups are interleaved to mask the idle time

in between cycles. We also note that a few studies have sought

to improve on the REPORT-GATE traffic signaling through

traffic prediction, see e.g., [36]–[38]. In this first study on grant

scheduling in EPONs with heterogeneous propagation delays

we focus on the offline scheduling framework as efficient

offline scheduling in EPONs with heterogeneous propagation

delays can serve as a basis for future studies combining QoS

mechanisms, such as [23]–[28], with efficient scheduling for

heterogeneous propagation delays. We leave the study of grant

scheduling for heterogeneous propagation delays in EPONs

with just-in-time scheduling, a combination of online and

offline scheduling, or traffic prediction for future research.

Primarily for the offline scheduling framework, where all

ONU requests are considered in scheduling decisions, but also

for the approaches combining online and offline scheduling

and for the just-in-time approaches, where a subset of the

ONUs are considered, a wide variety of specific scheduling

policies have been examined. A number of studies have

examined scheduling policies that provide prescribed QoS

differentiation or fairness properties, e.g., [23]–[32]. We focus

in this first study on heterogeneous propagation delays on min-

imizing the average packet delay; considering heterogeneous

propagation delays in conjunction with QoS differentiation

and fairness mechanisms are important directions for future

research. Existing scheduling policies for minimizing the

average packet delay include:

• Earliest Arrival First (EAF) scheduling [14], [39] which

orders ONUs by the arrival time of the head of line

packet.

• Shortest Processing Time (SPT) first scheduling [21], [33]

which orders ONUs by their grant size.

• Largest Processing Time (LPT) first scheduling [14],

[25] which orders ONUs by their grant size (descending

order).

• Largest Number of Frames (LNF) first scheduling [21]

which orders ONUs by the number of frames queued

(descending order).

A recent comparison found that LNF provided slightly smaller

or the same average queueing delays [21] than the other

policies and we consider therefore LNF as a benchmark in

our performance evaluation in Section IV.

We also note that efforts to mask propagation and other

system delays (such as laser tuning times) have been examined

for medium access control and scheduling in WDM star

networks, e.g., [40], [41]. These WDM star networks provide

all-to-all connectivity and are thus fundamentally different

from the EPON tree network, where only the OLT can reach

all ONUs.

3

Γ

2

stall

schedt

tR tR tRG1 G2 G3

tR

tR

tR

G1

G2

G3

tG tG tG

τ (3)

τ (3)τ (3)τ (3)

τ (2)

τ (2)

τ (1)

τ (1)

tR

tR

tstall

1

OLT

t 3

stall=0

t=0

t tgg

ONU 1

ONU 2

ONU 3

t

Fig. 1. Illustration of offline scheduling with three ONUs. The upstream channel experiences an idle (stall) period due to the scheduling computation timetsched, transmission time of first GATE message tG, and round trip propagation delay 2τ(1). There are further stall periods if the upstream transmission ofan ONU does not mask the round trip propagation delay of the next ONU.

III. PERFORMANCE ANALYSIS

A. Model Notation

We consider the EPON reporting and granting cycle with the

offline scheduling framework, which is illustrated in Fig. 1 for

N = 3 ONUs. We denote tsched for the schedule computation

time, i.e., the time duration from the instant when all REPORT

messages have been received at the OLT to the instant the

transmission of the first GATE message commences. We

denote tG for the fixed transmission time [in seconds] of

an MPCP GATE message, tg for the fixed guard time [in

seconds] required between ONU transmission windows, and

tR for the fixed transmission time of a MPCP REPORT

message. We let the constants τi, i = 1, . . . , N , denote the

one-way propagation delays [in seconds] between OLT and

ONU i (which we consider to be equal to the ONU i to OLT

propagation delay). We let τ(i), i = 1, . . . , N , denote the prop-

agation delays sorted in ascending order, i.e., τ(1) = mini τi

and τ(N) = maxi τi. For a given cycle, we let Ri be a

random variable denoting the reported queue depth (in units of

seconds of upstream transmission time) and Gi [in seconds]

be a random variable denoting the duration of the upstream

transmission window (grant) of ONU i, i = 1, . . . , N . We

suppose that Gi includes all “per-Ethernet frame” overheads,

such as Preamble and Inter Packet Gap (IPG).

For a given cycle, we define the cycle length Γ as the

time period from the instant the scheduling commences to the

instant the upstream transmissions of the cycle are completely

received. We define the upstream channel utilization η as the

ratio of the sum of upstream transmission windows to the cycle

length, i.e., η =∑N

i=1 Gi/Γ

B. Problem Overview

Generally, in order to minimize packet delays and maximize

the utilization on the EPON upstream channel, idle periods

on the upstream channel should be minimized. In turn, with

minimal idle periods, the cycle length Γ is minimized. With

the offline scheduling framework, there is an idle period

(stall time) between the instant the end of the last upstream

transmission of the preceding cycle arrives at the OLT and the

instant the beginning of the first upstream transmission of the

current cycle arrives at the OLT. Clearly, this first stall time

is minimized by sending the first GATE message to the ONU

with the shortest propagation delay, which results in

tstall1 = max(tsched + tG + 2τ(1), tg). (1)

For illustration of the problem suppose that next the GATE

messages and upstream transmissions of ONUs 2 and 3 follow,

see Fig. 1. If the first upstream transmission is too short to

mask the round-trip propagation delay to ONU 2, a stall time

tstall2 occurs between the end of the first upstream transmission

and the beginning of the reception of the second upstream

transmission. More specifically, if

tstall1 + G1 + tR + tg < tsched + 2tG + 2τ2, (2)

then a non-zero stall time tstall2 occurs. On the other hand, as

illustrated for ONU 3 in Fig. 1, if the round-trip propagation

delay is masked by a preceding upstream transmission, then

there is no stalling.

Note that in the illustration in Fig. 1, the sequence of the

GATE message transmissions is equal to the sequence of up-

stream transmissions reaching the OLT. These two sequences

do not necessarily need to be the same. In fact, one can

4

relatively easily construct examples where first sending the

GATE message to a far-away ONU, followed by sending the

GATE messages and receiving the upstream transmissions of

near-by ONUs, followed by the reception of the upstream

transmission from the far-away ONU minimizes idle periods,

and thus the cycle length. We also note that [20] briefly

mentioned that GATE messages to far-away ONUs may need

to be sent before GATE messages to near-by ONUs to achieve

close to continuous utilization of the upstream channel, but

did not analyze in any detail the scheduling of the upstream

transmission windows.

C. Solution Strategy

The scheduling of the upstream transmissions can be viewed

as a generalized version of the scheduling problem with release

times, i.e., times when a given job becomes eligible for

execution. Even for fixed known release times, the problem of

minimizing the total completion time is strongly NP-hard [42].

Our problem is more general in that the release times, i.e., the

times when upstream transmissions can at the very earliest

arrive at the OLT depend on the sequencing of the GATE

message transmissions.

To the best of our knowledge, the combined problem of

scheduling the sequence of GATE message transmissions and

upstream transmissions so as to minimize the cycle time is

mathematically intractable. Our solution strategy is to consider

two restrictions: (R1) We suppose that the GATE message

transmission time tG is equal to the REPORT message trans-

mission time tR, This is reasonable as both of these MPCP

messages are sent in minimum-length Ethernet frames. Noting

that every upstream transmission must at least contain a

REPORT to obtain the current queue depth of the ONU,

tG = tR implies that Gi ≥ tG ∀i. (R2) We initially suppose

that the sequence of the GATE message transmissions is equal

to the sequence of the upstream transmissions arriving at

the OLT. Note that even with this restriction, the problem

of scheduling the upstream transmissions is a generalized

version of the scheduling with release times in that the release

times of the jobs are not fixed, but rather depend on their

position in the schedule (i.e., the number of tG delays in

the release time varies according to the sequence of the

upstream transmissions). Given these two restrictions, we show

in Section III-D that the shortest propagation delay (SPD) first

scheduling policy minimizes the cycle time. Subsequently, in

Section III-E we relax the restriction R2 on the sequence of

the GATE messages and characterize the competitiveness of

SPD scheduling.

D. Shortest Propagation Delay (SPD) First Optimality

Theorem 1. If Gi ≥ tG ∀i and the GATE message trans-

mission sequence is equal to the sequence of the upstream

transmissions, then Shortest Propagation Delay (SPD) first

scheduling of the upstream transmissions minimizes the cycle

duration Γ.

Proof: Without loss of generality, we neglect the schedul-

ing time tsched and the guard times tg in the following

Case (A): Cycle length governed by τ2:

tG tG¡

¡¡

¡¡@@

@@@

τ1 τ1

G1

@@

@@

@@

τ2

¡¡

¡¡

¡¡

τ2

G2

Γ1,2 = 2tG + 2τ2 + G2

Case (B): Cycle length governed by τ1:

tG tG¡

¡¡

¡¡

¡¡@@

@@

@@@

τ1 τ1

G1 G2

@@@τ2 ¡

¡¡τ2

Γ1,2 = tG + 2τ1 + G1 + G2

Fig. 2. Illustration of cases for evaluation of cycle length Γ1,2.

as they are not affected by the scheduling of the upstream

transmissions.

a) Comparison of two ONUs: Consider two ONUs 1 and

2. Let Γ1,2 denote the cycle length when ONU 1 is scheduled

before ONU 2. As illustrated in Fig. 2, there are two cases for

the evaluation of Γ1,2: (A) the propagation delay τ2 governs

the cycle length, and (B) the propagation delay τ1 governs the

cycle length. Clearly, the actual cycle length Γ1,2 is obtained

as the maximum of the two cases:

Γ1,2 = max (2tG + 2τ2, tG + 2τ1 + G1) + G2. (3)

Analogously, we obtain by symmetry (which only exchanges

the roles of 1 and 2) the cycle length when ONU 2 is scheduled

first, followed by ONU 1:

Γ2,1 = max (2tG + 2τ1, tG + 2τ2 + G2) + G1. (4)

We want to show that if Gi ≥ tG then we cannot have

τ1 > τ2 and Γ1,2 < Γ2,1, (5)

i.e., it cannot be that scheduling the ONU with the longer

propagation delay (no. 1 in this case) leads to a shorter cycle

length.

We proceed to show that (5) leads to a contradiction. We

distinguish all possible cases according to where the maximum

in the definition of Γ1,2 and Γ2,1, respectively, is attained.

Case 1: 2tG + 2τ2 ≤ tG + 2τ1 + G1 and 2tG + 2τ1 ≤

tG +2τ2 +G2. In this case, we have (the first condition comes

from Γ1,2 < Γ2,1, the other two from the condition for the

case).

tG + 2τ1 + G1 + G2 < tG + 2τ2 + G2 + G1 (6)

2tG + 2τ2 ≤ tG + 2τ1 + G1 (7)

2tG + 2τ1 ≤ tG + 2τ2 + G2. (8)

Note that equation (6) is a contradiction to τ1 > τ2.

5

Case 2: 2tG + 2τ2 ≥ tG + 2τ1 + G1 and 2tG + 2τ1 ≤

tG + 2τ2 + G2. Here, we get

2tG + 2τ2 + G2 < tG + 2τ2 + G2 + G1 (9)

2tG + 2τ2 ≥ tG + 2τ1 + G1 (10)

2tG + 2τ1 ≤ tG + 2τ2 + G2. (11)

Note that (10) is a contradiction to τ1 > τ2 and tG ≤ G1.

Case 3: 2tG + 2τ2 ≤ tG + 2τ1 + G1 and 2tG + 2τ1 ≥


tG + 2τ1 + G1 + G2 < 2tG + 2τ1 + G1 (12)

2tG + 2τ2 ≤ tG + 2τ1 + G1 (13)

2tG + 2τ1 ≥ tG + 2τ2 + G2. (14)

Note that (12) is a contradiction to tG ≤ G2.

Case 4: 2tG + 2τ2 ≥ tG + 2τ1 + G1 and 2tG + 2τ1 ≥


2tG + 2τ2 + G2 < 2tG + 2τ1 + G1 (15)

2tG + 2τ2 ≥ tG + 2τ1 + G1 (16)

2tG + 2τ1 ≥ tG + 2τ2 + G2. (17)

Note that (16) is a contradiction to τ1 > τ2 and tG ≤ G1.

b) General case: Now, consider N ONUs with propaga-

tion delays τ1, . . . , τN , respectively. We show that it is optimal

to schedule them in SPD first manner.

Assume that they are scheduled in the natural order (first

ONU 1 with τ1, etc.). Consider ONUs i and i + 1. We have

shown above that it is better (in the sense that the overall

cycle time is shorter (at least not longer)) when scheduling

the upstream transmission of ONU i before the upstream

transmission of ONU i+1 if τi ≤ τi+1. This is not influenced

by the fact that the transmissions of ONUs i and i + 1 may

be delayed by preceding transmissions of other ONUs. In any

case, it cannot be a loss to schedule the ONU with the shortest

propagation delay first. This implies the assertion, since i was

arbitrary.

Note that an immediate corollary of Theorem 1 is that SPD

first scheduling maximizes the upstream channel utilization.

As examined in detail in the next section, the bound in

Theorem 1 shows that SPD first scheduling is very close to

optimal, since the GATE transmission time tG is typically

small compared to traffic.

E. Competitiveness of SPD Scheduling

In this section, we examine the competitiveness of SPD

first scheduling in comparison to an optimal schedule that

minimizes the cycle length. The optimal schedule may have

different sequences of GATE transmissions and upstream

transmissions, i.e., does not need to meet restriction R2. We

still require that the optimal schedule meets restriction R1

that Gi ≥ tG since the EPON polling requires that each

upstream transmission contains at least a REPORT. We first

characterize the absolute difference of the cycle length with

SPD ΓSPD compared to the minimal cycle length Γopt of an

optimal schedule. Next, we examine the competitive ratio of

SPD scheduling, i.e., the bound on the ratio of the cycle length

with SPD scheduling ΓSPD to the cycle length of an optimal

schedule Γopt.

Theorem 2. The cycle length with SPD first scheduling

exceeds the minimal cycle length by no more than (N −

1)tG, i.e.,ΓSPD ≤ Γopt + (N − 1)tG.

Proof: For SPD first scheduling, let tstarti , i = 1, . . . , N ,

denote the instant when the upstream transmission of ONU ibegins to arrive at the OLT. We define for convenience tstart0 :=0 and G0 = 0 and note that

tstarti = max(itG + 2τ(i), tstarti−1 + Gi−1 + tg). (18)

The cycle length is

ΓSPD = tstartN + GN . (19)

Now, consider a imaginary EPON where the grants to all

ONUs are communicated with one GATE message requiring

only one transmission time tG. This imaginary EPON serves

as a comparison for the optimal scheduling in the real EPON.

Let Γopt,c denote the cycle time in the imaginary EPON (the

subscript c is for “comparison strategy”). Clearly, SPD first

scheduling of the upstream transmissions is optimal in the

imaginary EPON.

We proceed to show that

ΓSPD − (N − 1)tG ≤ Γopt,c ≤ Γopt ≤ ΓSPD, (20)

where the last two inequalities are trivially satisfied. We prove

the first inequality by induction. In the imaginary EPON let

sstarti denote the instant when the upstream transmission of

ONU i begins to arrive at the OLT. Denote sstart0 := 0 and

note that

sstarti = max(tG + 2τ(i), sstart

i−1 + Gi−1 + tg),

for i = 1, . . . , N. (21)

We obtain by induction that

tstarti ≤ sstarti + (i − 1)tG. (22)

The case i = 1 is trivial. In general, we get

tstarti = max(itG + 2τ(i), tstarti−1 + Gi−1 + tg) (23)

≤ max(itG + 2τ(i),

sstarti−1 + (i − 2)tG + Gi−1 + tg) (24)

≤ max((i − 1)tG + tG + 2τ(i),

sstarti−1 + (i − 1)tG + Gi−1 + tg) (25)

= (i − 1)tG + max(tG + 2τ(i),

sstarti−1 + Gi−1 + tg) (26)

= (i − 1)tG + sstarti . (27)

The assertion follows from (22), since

ΓSPD = tstartN + GN (28)

≤ sstartN + GN + (N − 1)tG (29)

= Γopt,c + (N − 1)tG. (30)

We remark that the bound in Theorem 2 is attained for an

example scenario with τ1 = . . . = τN−1 = 0, 2τN = NtG,

6

and G1 = . . . = GN = tG. For this example, the cycle length

with optimal scheduling is Γopt = (N + 2)tG, whereas the

cycle length with SPD scheduling is ΓSPD = (2N + 1)tG.

That is, the difference ΓSPD−Γopt is exactly (N−1)tG in this

example, and thus the bound in Thm 2 cannot be improved.

Proposition 1. The competitive ratio for the cycle length with

Smallest Propagation Delay (SPD) first scheduling is

ΓSPD

Γopt≤ min

{

tG + 2maxi τi +∑

i Gi

tG + 2mini τi +∑

i Gi

,

tG + 2maxi τi +∑

i Gi

tG + maxi(2τi + Gi)

}

. (31)

Proof: First, note that the shortest possible cycle length

must satisfy

Γopt ≥ tG + 2τ(1) +

N∑

i=1

Gi (32)

since at least the first GATE needs to be transmitted and at

least the round trip propagation delay to the nearest ONU is

incurred before all the upstream transmissions (with aggregate

duration∑N

i=1 Gi) can arrive at the OLT. This bound is

attained when the ONU with the shortest propagation delay

has a very large upstream transmission.

Second, note that cycle must be long enough to accommo-

date the GATE message transmission, round-trip propagation

delay, and upstream transmission of each individual ONU i,i.e.,

Γopt ≥ tG + 2τi + Gi (33)

for each ONU i, i = 1, . . . , N . Since this holds for all i we

can take the maximum over all ONUs yielding

Γopt ≥ tG + maxi

(2τi + Gi). (34)

This bound is attained if one ONU with a large propagation

delay has a large upstream transmission and all other ONUs

have small propagation delays and upstream transmissions.

Thirdly, for the cycle time with SPD first scheduling

ΓSPD ≤ tG + 2τ(1) +

N∑

i=1

Gi +

N−1∑

i=1

δi,i+1 (35)

where we define δi,i+1 as the difference between the (i +1)th smallest and the ith smallest round trip propagation delay,

i.e., δi,i+1 = 2τ(i+1) − 2τ(i) for i = 1, . . . , N − 1. Note that∑N−1

i=1 δi,i+1 is the worst case for the stall times in between

the upstream transmissions arriving at the OLT. This worst-

case occurs if each ONU has only a REPORT message to

send upstream, i.e., Gi = tG. The δ-sum is telescoping and

we get

N−1∑

i=1

δi,i+1 = 2τ(N) − 2τ(1). (36)

Thus,

ΓSPD ≤ tG + 2τ(N) +

N−1∑

i=1

Gi. (37)

Combining (32) and (37) we get

ΓSPD

Γopt≤

tG + 2τ(N) +∑N

i=1 Gi

tG + 2τ(1) +∑N

i=1 Gi

. (38)

From (34) and (37) we get

ΓSPD

Γopt≤

tG + 2τ(N) +∑N

i=1 Gi

tG + maxi(2τi + Gi). (39)

The bounds in Proposition 1 show that the SPD first

scheduling policy has a good competitive ratio in most cases.

Indeed, the first bound is a good competitive ratio, i.e., is

close to one, if∑N

i=1 Gi is large, i.e., if there is heavy traffic.

The second bound is a good competitive ratio if τ(N) is large

compared to∑N

i=1 Gi, i.e., typically if there is light traffic.

From the derivations of the preceding bounds we obtain

the following insights into SPD first scheduling and potential

heuristics to improve on SPD first scheduling: a) If there is

heavy traffic at ONUs with short propagation delays, then send

them their GATE messages as soon as possible following the

SPD first policy. With this strategy the traffic from the ONUs

with short propagation delays can mask the GATE transmis-

sions and round trip propagations to the far-away ONUs. b)

If there is little traffic at the ONUs with short propagation

delays, then GATE messages could first be sent to the ONUs

with large propagation delays, even though their upstream

transmission windows should, of course, be scheduled after the

windows of the ONUs with short propagation delays. With this

strategy, the ONUs with short propagation delays can finish

their transmissions by the time the upstream transmissions

from the ONUs with long propagation delays arrive at the

OLT.

F. Packet Delay Analysis

In this section, we analyze the cycle length and packet delay

with SPD first scheduling for Gated grant sizing. We consider

approximations for light traffic and heavy traffic. We define

the Poissonian packet traffic load ρi, i = 1, . . . , N of ONU ias the ratio of the average traffic bit rate (including Ethernet

frames plus “per-Ethernet frame” overhead, i.e., preamble and

IPG) generated at ONU i to the upstream channel bit rate

C [bit/s]. Unlike for the analysis in the preceding sections,

we do not consider the REPORT message with transmission

time tR = tG as part of a grant Gi in this section, but rather

account for the REPORT transmission times separately from

the grant duration. We denote ρt =∑N

i=1 ρi for the total load.

Noting that with Gated grant sizing the “per-grant” overhead

(REPORT message, guard time) becomes negligible as the

grants grow very large in heavy traffic, the stability condition

(maximum throughput) with gated grant sizing is ρt < 1(which holds for arbitrary packet traffic patterns, including

self-similar traffic). We let L̄ and σL [in bit] denote the mean

and standard deviation of the packet size (plus Preamble and

IPG). We define Poi[ω] for a random variable with Poisson

distribution with parameter ω.

7

1) Low load scenario: In a low load situation with SPD

scheduling, the cycle time is determined by the last (longest

prop. delay) ONU. Namely, the cycle time in cycle n + 1satisfies

ΓLL(n + 1) = NtG + 2τ(N) +L̄

CPoi

[

ρ(N)C

L̄ΓLL(n)

]

+ tG,

(40)

where ρ(N) is the load parameter of the ONU with the

largest propagation delay τ(N). The cycle consists of N GATE

message transmission and the data packet transmissions plus

REPORT message transmission of the last ONU. The other

transmissions are masked by the long propagation delay of

the last ONU (due to the low load assumption).

Taking expectations we get

EΓLL = NtG + 2τ(N) + ρ(N)EΓLL + tG. (41)

This yields

EΓLL =(N + 1)tG + 2τ(N)

1 − ρ(N). (42)

Since we need the second moment of the cycle time for the

delay analysis, we obtain in the same way from (40):

EΓ2LL = VΓLL + (EΓLL)2 (43)

= ρ(N)L̄

CEΓLL + (EΓLL)2 (44)

= EΓLL(ρ(N)L̄

C+ EΓLL). (45)

The delay of a packet generated at ONU i consists of the

backward recurrence time of the cycle length [43, Ch. 5.5]EΓ2

LL

2EΓLL

, i.e., the time until the generated packet is included in

a REPORT, the time period from the instant the transmission

of the REPORT is complete to the instant the next upstream

transmission commences, which is itG + 2τ(i) due to the low

load assumption, and the time needed within the grant until

the considered packet commences transmission ρiEΓ2

LL

2EΓLL

:

D(i)LL =

EΓ2LL

2EΓLL

+ itG + 2τ(i) + ρi

EΓ2LL

2EΓLL

+τ(i) +L̄

C. (46)

= (1 + ρi)ρ(N)

L̄C

+ EΓLL

2+ itG

+3τ(i) +L̄

C(47)

The overall delay is then given by

DLL =1

ρt

N∑

i=1

ρ(i)D(i)LL, (48)

where ρt :=∑N

i=1 ρi and ρ(i) is the load parameter of the

ONU with the i-th smallest propagation delay.

2) Heavy load scenario: In a heavy load scenario, all ONUs

have data to send, which masks the round-trip delays between

OLT and the second to last ONUs in the SPD first schedule.

In this case, the cycle length satisfies

EΓHL = tG +2τ(1) +

N∑

i=1

ρiEΓHL +NtG +(N −1)tg, (49)

because a cycle consists of the GATE transmission plus round

trip propagation delay to the first ONU (the one with the

shortest propagation delay) and all the traffic that is sent

(which consists of the accumulated data traffic and the NREPORT messages). This gives

EΓHL =(N + 1)tG + (N − 1)tg + 2τ(1)

1 −∑N

i=1 ρi

. (50)

This scenario resembles the case of a single ONU for which

the packet delay was analyzed in [44]. Inserting τHL :=N+1

2 tG + N−12 tg + τ(1) and ρt :=

∑N

i=1 ρi in Eqn. (39)

in [44], namely

DHL = 2τHL

2 − ρt

1 − ρt

+ρt

2C(1 − ρt)

(

σ2L

L̄+ L̄

)

+L̄

C(51)

gives an approximation of the packet delay.

G. Stability of Limited grant sizing

For Limited grant sizing with SPD scheduling we evaluate

the maximum throughput (stability limit) as follows. First,

for arbitrary packet traffic, including self-similar traffic, we

calculate the maximal cycle time by tstart,max0 := 0

tstart,maxi = max(itG + 2τ(i), tstart,max

i−1 + Gmaxi−1 + tg), (52)

where Gmaxi are the given maximal grant sizes (Gmax

0 = 0).

Then,

ΓmaxSPD = tstart,max

N + GmaxN (53)

is the maximum cycle length (even in the heaviest traffic, no

cycle will be longer than this). The stability condition for SPD

is then

ρiΓmaxSPD < Gmax

i (54)

for all i. (The stability limit for Largest Propagation Delay

(LPD) first scheduling is obtained analogously by considering

the τi in decreasing order in (52).) For homogeneous ONU

loads ρ1 = ρ2 = · · · = ρN and maximum grant sizes Gmax1 =

· · · = GmaxN , the total load is ρt = Nρi, resulting in the

stability condition

ρmaxt,SPD <

NGmaxi

ΓmaxSPD

. (55)

8

0.0 sec

200.0 usec

400.0 usec

600.0 usec

800.0 usec

1.0 msec

1.2 msec

1.4 msec

1.6 msec

1.8 msec

2.0 msec

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

cycl

e le

ngth

(in

sec

onds

)

load (in Gbps)

Self−similar, simulationPoisson, simulationPoisson, analysis high loadPoisson, analysis low load

a) 50 µs max. prop. delay (i.e., up to 10 km)

200.0 usec

400.0 usec

600.0 usec

800.0 usec

1.0 msec

1.2 msec

1.4 msec

1.6 msec

1.8 msec

2.0 msec

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

cycl

e le

ngth

(in

sec

onds

)

load (in Gbps)


b) 250 µs max. prop. delay (i.e., up to 50 km)

200.0 usec

400.0 usec

600.0 usec

800.0 usec

1.0 msec

1.2 msec

1.4 msec

1.6 msec

1.8 msec

2.0 msec

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

cycl

e le

ngth

(in

sec

onds

)

load (in Gbps)


c) 500 µs max. prop. delay (i.e., up to 100 km)

Fig. 3. Average cycle length for SPD grant scheduling and Gated grantsizing.

IV. EXPERIMENTAL PERFORMANCE ANALYSIS

We conducted a set of simulation experiments to: 1) validate

our analytical models presented in Section III, and 2) quantify

the improvements to cycle length and packet delay achieved

by SPD grant scheduling under different operating conditions.

We use an EPON simulator that we have developed using the

CSIM discrete event simulation library [45]. We simulated an

EPON with 32 ONUs and varied the maximum propagation

delay to represent different EPON reaches.

In the absence of any diversity in propagation delay, the

impact of minimizing cycle length will be minor. In our

experiments, we use a continuous (uniform) distribution for

one-way (OLT-to-ONU) propagation delays with a minimum

value of 6.68 µs and different maximum one-way propagation

delay values. We modify this distribution slightly by forcing

0.0 sec

500.0 usec

1.0 msec

1.5 msec

2.0 msec

2.5 msec

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

cycl

e le

ngth

(in

sec

onds

)

load (in Gbps)

LPDLNFSPD


0.0 sec

1.0 msec

2.0 msec

3.0 msec

4.0 msec

5.0 msec

6.0 msec

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9cy

cle

leng

th (

in s

econ

ds)

load (in Gbps)

LPDLNFSPD


0.0 sec

2.0 msec

4.0 msec

6.0 msec

8.0 msec

10.0 msec

12.0 msec

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

cycl

e le

ngth

(in

sec

onds

)

load (in Gbps)

LPDLNFSPD


Fig. 4. Average cycle length with Gated grant sizing for self-similar traffic.

one ONU to have a minimum propagation delay value and

another to have a maximum propagation delay value. The

propagation delays for the other 30 ONUs are continuously

distributed over the range.

We compare SPD scheduling to Largest Number of

Frames (LNF) first scheduling which was previously demon-

strated [21] to provide low average queueing delay compared

to other scheduling policies [39]. We also present some results

for Largest Propagation Delay (LPD) first scheduling which is

the opposite of SPD to illustrate the range of possibilities for

cycle length and packet delay. The average cycle length and

average packet delay values presented in this section represent

the mean of several independent runs that were constructed

using the batch means feature in the CSIM discrete event

simulation library. The resulting statistical confidence intervals

for Poisson traffic are smaller than the point marks in the plots.

9

1.0 msec

2.0 msec

3.0 msec

4.0 msec

5.0 msec

6.0 msec

7.0 msec

8.0 msec

0.9 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98

cycl

e le

ngth

(in

sec

onds

)

load (in Gbps)

LPDLNFSPD


0.0 sec

2.0 msec

4.0 msec

6.0 msec

8.0 msec

10.0 msec

12.0 msec

14.0 msec

16.0 msec

18.0 msec

20.0 msec

0.9 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98

cycl

e le

ngth

(in

sec

onds

)

load (in Gbps)

LPDLNFSPD


0.0 sec

5.0 msec

10.0 msec

15.0 msec

20.0 msec

25.0 msec

30.0 msec

35.0 msec

0.9 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98

cycl

e le

ngth

(in

sec

onds

)

load (in Gbps)

LPDLNFSPD


Fig. 5. Average cycle length with Gated grant sizing for self-similar trafficas the load approaches the channel capacity.

A. Gated grant sizing

In this section we present our results of the experiments we

conducted using Gated grant sizing.

1) Cycle length: Figure 3 shows the average cycle length

as a function of the total load for SPD grant scheduling

by means of analysis using Eqs. (42) and (50), as well as

simulation experiments using Poisson traffic sources and self-

similar traffic sources. We observe from this figure that Eqs.

(42) and (50) provide an excellent fit to the average cycle

length measured in the simulation experiments.

Figure 4 shows the average cycle length for LPD, LNF,

and SPD grant scheduling and varying propagation delay

configurations. Figure 5 shows the same for load values

approaching the channel limit. We observe from these figures

that SPD always provides a lower average cycle length than

0.0 sec

2.0 msec

4.0 msec

6.0 msec

8.0 msec

10.0 msec

12.0 msec

14.0 msec

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

dela

y (in

sec

onds

)

load (in Gbps)



0.0 sec

2.0 msec

4.0 msec

6.0 msec

8.0 msec

10.0 msec

12.0 msec

14.0 msec

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1de

lay

(in s

econ

ds)

load (in Gbps)



0.0 sec

2.0 msec

4.0 msec

6.0 msec

8.0 msec

10.0 msec

12.0 msec

14.0 msec

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

dela

y (in

sec

onds

)

load (in Gbps)



Fig. 6. Average packet delay for SPD grant scheduling and Gated grantsizing.

LPD or LNF. The difference increases significantly with an

increasing load and increasing maximum propagation delay.

As an example, with a 500 µsec maximum propagation delay

and load value of 0.9 the average cycle length was 2.0

milliseconds using SPD and 6.0 milliseconds using LNF.

2) Packet delay: Figure 6 shows the average packet delay

for SPD grant scheduling by means of analysis using Eqs. (48)

and (51), as well as simulation experiments using Poisson traf-

fic sources and self-similar traffic sources. The measurements

for average packet delay with the self-similar traffic sources

are, as expected, much higher despite average cycle lengths

similar to those observed with Poisson traffic sources. This is

a result of a few very long cycles, whose lengths are observed

once, that have many packets whose associated large delay is

observed once for each of these packets.

Figure 7 shows the average packet delay for LPD, LNF,

10

0.0 sec

5.0 msec

10.0 msec

15.0 msec

20.0 msec

25.0 msec

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

dela

y (in

sec

onds

)

load (in Gbps)

LPDLNFSPD


0.0 sec

5.0 msec

10.0 msec

15.0 msec

20.0 msec

25.0 msec

30.0 msec

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

dela

y (in

sec

onds

)

load (in Gbps)

LPDLNFSPD


0.0 sec

5.0 msec

10.0 msec

15.0 msec

20.0 msec

25.0 msec

30.0 msec

35.0 msec

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

dela

y (in

sec

onds

)

load (in Gbps)

LPDLNFSPD


Fig. 7. Average packet delay with Gated grant sizing for self-similar traffic.

and SPD grant scheduling and varying propagation delay

configurations for self-similar traffic. Figure 8 shows the same

for load values approaching the channel limit. We observe

from Figures 7 and 8 that SPD provides lower average packet

delay for the experiments with 250 µs and 500 µs maximum

propagation delays. The difference with LNF becomes more

pronounced at high load values. As an example, with a 500

µs maximum propagation delay and load value of 0.9 Gbps

the average packet delay was 21.3 milliseconds using SPD

and 26.5 milliseconds using LNF. At 50 µs, the difference

between LNF and SPD is rather insignificant. However, it is

worth noting that SPD clearly provides a lower average cycle

length for all load values (see Figures 4 and 5) which leads

to lower average packet delay.

15.0 msec

20.0 msec

25.0 msec

30.0 msec

35.0 msec

40.0 msec

45.0 msec

50.0 msec

55.0 msec

60.0 msec

65.0 msec

70.0 msec

0.9 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98

dela

y (in

sec

onds

)

load (in Gbps)

LPDLNFSPD


10.0 msec

20.0 msec

30.0 msec

40.0 msec

50.0 msec

60.0 msec

70.0 msec

80.0 msec

90.0 msec

0.9 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98

dela

y (in

sec

onds

)

load (in Gbps)

LPDLNFSPD


10.0 msec

20.0 msec

30.0 msec

40.0 msec

50.0 msec

60.0 msec

70.0 msec

80.0 msec

90.0 msec

100.0 msec

110.0 msec

0.9 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98

dela

y (in

sec

onds

)

load (in Gbps)

LPDLNFSPD


Fig. 8. Average packet delay with Gated grant sizing for self-similar trafficas the load approaches the channel capacity.

One-way prop. delay [µs] ΓmaxSPD

Eq. (53) [ms] ΓmaxSPD

sim. [ms]

50 (up to 10 km) 2.058 2.03250 (up to 50 km) 2.058 2.03500 (up to 100 km) 2.058 2.031000 (up to 200 km) 2.104 2.11

TABLE IMAXIMUM AVERAGE CYCLE LENGTH (IN MILLISECONDS) FOR SPD FOR

LIMITED GRANT SIZING WITH Gmaxi = 7188 BYTES,∀i FOR

SELF-SIMILAR TRAFFIC.

B. Limited grant sizing

In this section we present our results of the experiments

we conducted using Limited grant sizing with Gmaxi = 7188

bytes, ∀i.

1) Cycle length: Figure 9 shows the average cycle length

for LPD, LNF, and SPD grant scheduling for varying propa-

11

200.0 usec

400.0 usec

600.0 usec

800.0 usec

1.0 msec

1.2 msec

1.4 msec

1.6 msec

1.8 msec

2.0 msec

2.2 msec

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

cycl

e le

ngth

(in

sec

onds

)

load (in Gbps)

LPDLNFSPD


400.0 usec

600.0 usec

800.0 usec

1.0 msec

1.2 msec

1.4 msec

1.6 msec

1.8 msec

2.0 msec

2.2 msec

2.4 msec

2.6 msec

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

cycl

e le

ngth

(in

sec

onds

)

load (in Gbps)

LPDLNFSPD


1.0 msec

1.2 msec

1.4 msec

1.6 msec

1.8 msec

2.0 msec

2.2 msec

2.4 msec

2.6 msec

2.8 msec

3.0 msec

3.2 msec

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

cycl

e le

ngth

(in

sec

onds

)

load (in Gbps)

LPDLNFSPD


2.0 msec

2.2 msec

2.4 msec

2.6 msec

2.8 msec

3.0 msec

3.2 msec

3.4 msec

3.6 msec

3.8 msec

4.0 msec

4.2 msec

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

cycl

e le

ngth

(in

sec

onds

)

load (in Gbps)

LPDLNFSPD

d) 1000 µs max. prop. delay (i.e., up to 200 km)

Fig. 9. Average cycle length for different grant scheduling policies withLimited grant sizing for self-similar traffic.

gation delay configurations. We make two observations: 1) in

all plots SPD provides lower average cycle length for all load

One-way prop. delay [µs] ρmaxt,SPD

Eq. (55) ρmaxt,SPD

sim.

50 (up to 10 km) 0.894 0.90250 (up to 50 km) 0.894 0.90500 (up to 100 km) 0.894 0.901000 (up to 200 km) 0.874 0.87

TABLE IILOAD AT WHICH THE AVERAGE CYCLE LENGTH APPROACHES THE

MAXIMUM CYCLE LENGTH, WHICH IS EQUAL TO THE STABILITY LIMIT,FOR SPD WITH LIMITED GRANT SIZING WITH Gmax

i = 7188 BYTES,∀i

FOR SELF-SIMILAR TRAFFIC.

values, and 2) as the maximum propagation delay is increased

from 50 µs to 1000 µs, SPD is able to maintain a maximum

average cycle length around 2 milliseconds.

Exploring the second observation further, we see that the

maximum average cycle length for LNF increases from around

2 milliseconds at 50 µs to close to 3.6 milliseconds at 1000

µs. With a 1000 µsec maximum propagation delay, the load at

which the cycle length reaches its maximum is 0.87 Gbps for

SPD and approximately only 0.46 Gbps for LNF. In Table I

we compare the maximum average cycle length using Eq. (53)

with the results from our simulation experiments. The data in

this table indicates that the equation is within 1.4% of the

experimental data.

Our experimental data and Eq. (53) indicate that SPD is able

to keep the maximum cycle length near 2 milliseconds as the

maximum propagation delay is increased. Whereas, LNF is

unable to do the same. Eq. (53) illustrates that the maximum

cycle length is a function of the order of the ONUs as well

as their propagation delays. LNF constantly changes the order

of ONUs with respect to their propagation delays resulting in

the variations that can be seen in its maximum average cycle

length. Further, SPD is able to minimize the cycle length by

ordering the ONUs in increasing order of propagation delay.

LNF does not order the ONUs according to their propagation

delays and, as a result, does not minimize the cycle length.

2) Stability limit: As the load increases and the grant sizes

approach the prescribed maximum grant size Gmaxi , the cycle

length approaches the maximum cycle length. That is, the

average cycle length levels out at the maximum cycle length

plotted in Fig. 9 and tabulated for SPD in Table I. Further

increases in the load can not increase the cycle length, but

result in infinite queue build-up in the ONUs, i.e., instability.

The load value at which the average cycle length levels

out to the maximum cycle length in Fig. 9 thus represents

the stability limit, which is tabulated for SPD in Table II.

We observe from Table II that Eqn. (55) very accurately

characterizes the stability limit.

Turning to Fig. 9, we observe that the stability limit dif-

ference between LNF and SPD increases significantly as the

maximum propagation delay is increased. As an example, with

a 500 µs maximum propagation delay, the stability limit is

approximately 0.89 Gbps using SPD (the average cycle length

converges to the maximum cycle length very slowly between

a load of 0.85 and 0.89) and approximately 0.62 Gbps using

LNF.

By scheduling the grants to the close-by ONUs first, SPD

masks the long round-trip delays to the ONUs that are further

12

0.0 sec

50.0 msec

100.0 msec

150.0 msec

200.0 msec

250.0 msec

300.0 msec

350.0 msec

400.0 msec

450.0 msec

500.0 msec

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

dela

y (in

sec

onds

)

load (in Gbps)

LPDLNFSPD


0.0 sec

50.0 msec

100.0 msec

150.0 msec

200.0 msec

250.0 msec

300.0 msec

350.0 msec

400.0 msec

450.0 msec

500.0 msec

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

dela

y (in

sec

onds

)

load (in Gbps)

LPDLNFSPD


50.0 msec

100.0 msec

150.0 msec

200.0 msec

250.0 msec

300.0 msec

350.0 msec

400.0 msec

450.0 msec

500.0 msec

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

dela

y (in

sec

onds

)

load (in Gbps)

LPDLNFSPD


100.0 msec

200.0 msec

300.0 msec

400.0 msec

500.0 msec

600.0 msec

700.0 msec

800.0 msec

900.0 msec

1.0 sec

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

dela

y (in

sec

onds

)

load (in Gbps)

LPDLNFSPD

d) 1000 µs max. prop. delay (i.e., up to 200 km)

Fig. 10. Average packet delay for different grant scheduling policies withLimited grant sizing for self-similar traffic.

away. This more efficient utilization of the upstream channel

increases the stability limit substantially as the propagation

0.0 sec

100.0 msec

200.0 msec

300.0 msec

400.0 msec

500.0 msec

600.0 msec

700.0 msec

800.0 msec

900.0 msec

1.0 sec

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9load (in Gbps)

dela

y (in

sec

onds

)

LNF, 0SPD, 0LNF, 4SPD, 4LNF, 8SPD, 8LNF, 16SPD, 16

Fig. 11. Average packet delay for EPON with 4 ONUs between 50 km and100 km from OLT with varying number of close-by ONUs (i.e., 0.5 km to 5km from OLT) added.

delays become more diverse.

3) Packet delay: Figure 10 shows the average packet delay

for LPD, LNF, and SPD grant scheduling and varying prop-

agation delay configurations. We observe that SPD provides

lower average packet delay for all load values and reconfirm

the higher stability limit for SPD. It is instructive to compare

the packet delay reductions with SPD compared to LNF grant

scheduling for Gated grant sizing in Figs. 7 and 8 to the

corresponding delay reductions for Limited grant sizing in

Fig. 10. Clearly, for Limited grant sizing we observe substan-

tially larger delay reductions. With Gated grant sizing, any

reported queue size is served in one upstream transmission.

In contrast, with Limited grant sizing, a large reported queue

requires several maximum sized grants of size Gmaxi , and

thus several cycles to transmit the traffic to the OLT. The

cycle length minimizing SPD scheduling policy leads hence

to significantly more pronounced delay reductions for Limited

grant sizing than for Gated grant sizing.

C. Engineering EPONs for better channel utilization

In this final set of simulation experiments we wish to illus-

trate the utility of SPD grant scheduling in allowing close-by

ONUs to be added to an EPON without taking bandwidth from

existing ONUs. Essentially, the maximum channel utilization

is significantly improved when adding close-by ONUs and

utilizing SPD scheduling. We consider an EPON with 4 ONUs

with ONU-to-OLT propagation delays continuously distributed

between 250 µs and 500 µs (i.e., ONU-to-OLT distances of

50 km to 100 km). We then added a varying number of close-

by ONUs with ONU-to-OLT propagation delays continuously

distributed between 2.5 µs and 25 µs (i.e., ONU-to-OLT

distances of 0.5 km to 5 km).

Figure 11 shows the average packet delay for different

numbers of added close-by ONUs for both LNF and SPD. We

observe that adding close-by ONUs increases the maximum

achievable channel utilization (stability limit). The shorter

propagation delays of the close-by ONUs allow these ONUs to

be serviced while the GATE messages are propagating to the

ONUs with the larger propagation delays and their upstream

transmissions are propagating up to the OLT. With LNF

13

scheduling the ONUs are ordered by their grant size, irrespec-

tive of their propagation delays leading to poor exploitation

of this ability. SPD, on the other hand, always services the

ONUs with shorter propagation delays first allowing them

to mask the round-trip time to the ONUs with the larger

propagation delays, and thus achieving substantially higher

channel utilization than LNF.

V. CONCLUSION

In conclusion, we introduced a new EPON grant scheduling

technique called Shortest Propagation Delay (SPD) first grant

scheduling to exploit heterogenous propagation delays. We

proved that SPD minimizes granting cycle length to within

a small time period (number of ONUs times GATE message

transmission time) and maximizes channel utilization. We

analytically characterized the stability limit (maximum packet

throughput or equivalently maximum channel utilization) for

both Gated and Limited grant sizing for arbitrary traffic and

characterized the cycle length and packet delay for Gated grant

sizing for Poisson traffic. We have illustrated the utility of SPD

through a set of simulation experiments. Specifically, we found

that SPD can improve performance measures when using

Gated grant sizing as well as Limited grant sizing. The most

significant improvements came from its use with Limited grant

sizing and long reach EPONs. In those circumstances, packet

delay and channel utilization were significantly improved.

Another significant finding is the potential channel uti-

lization improvement that is possible when using SPD grant

scheduling in conjunction with certain EPON design princi-

ples. Generally, EPONs should be engineered to have prop-

agation delay diversity to promote higher channel utilization.

ONUs that are within a short range of the OLT can fill the

idle times in which the OLT waits for data from the ONUs

that are further away.

An interesting avenue for future research arises from the

convergence of fiber-based and wireless access networks, see

e.g., [46], [47], which will potentially cover large geographic

areas and thus have highly heterogeneous propagation delays.

Further, the integration of medium access control on the fiber

and wireless media may lump the propagation delay on the

fiber and the wireless medium access delay together to lead to

additional diversification of the round-trip delays experienced

by the OLT.

REFERENCES

[1] F. Effenberger, “An introduction to PON technologies,” IEEE Commu-

nications Magazine, vol. 45, no. 3, pp. S17–S25, Mar. 2007.[2] P. Green, Fiber to the Home: The New Empowerment. Wiley-

Interscience, 2005.[3] G. Keiser, FTTX Concepts and Applications. Wiley-IEEE Press, 2006.[4] C. Lam, Passive Optical Networks: Principles and Practice. Academic

Press, 2007.[5] M. De Andrade, L. Gutierrez, and S. Sallent, “A distributed scheduling

approach for Ethernet-based passive optical networks,” in Proceedings

of IEEE Conference on Local Computer Networks, Oct. 2007, pp. 205–206.

[6] C. Foh, L. Andrew, E. Wong, and M. Zukerman, “FULL-RCMA: a highutilization EPON,” IEEE Journal on Selected Areas in Communications,vol. 22, no. 8, pp. 1514–1524, Oct. 2004.

[7] M. Hajduczenia, H. J. A. da Silva, and P. P. Monteiro, “EPON versusAPON and GPON: a detailed performance comparison,” OSA Journal

of Optical Networking, vol. 5, no. 4, pp. 298–319, Apr. 2006.

[8] G. Kramer, B. Mukherjee, and G. Pesavento, “Ethernet PON (ePON):Design and Analysis of an Optical Access Network,” Photonic Network

Communications, vol. 3, no. 3, pp. 307–319, July 2001.[9] ——, “IPACT: A dynamic protocol for an Ethernet PON (EPON),” IEEE

Communications Magazine, vol. 40, no. 2, pp. 74–80, February 2002.[10] M. Ma, Y. Zhu, and T. Cheng, “A systematic scheme for multiple

access in Ethernet passive optical access networks,” IEEE/OSA Journal

of Lightwave Technology, vol. 23, no. 11, pp. 3671–3682, November2005.

[11] M. McGarry, M. Reisslein, and M. Maier, “Ethernet Passive OpticalNetwork Architectures and Dynamic Bandwidth Allocation Algorithms,”IEEE Communication Surveys and Tutorials, vol. 10, no. 3, pp. 46–60,October 2008.

[12] M. R. Radivojevic and P. S. Matavulj, “Implementation of intra-ONUscheduling for Quality of Service support in Ethernet passive opticalnetworks,” IEEE/OSA Journal of Lightwave Technology, vol. 27, no. 18,pp. 4055–4062, Sept. 2009.

[13] B. Skubic, J. Chen, J. Ahmed, L. Wosinska, and B. Mukherjee, “Acomparison of dynamic bandwidth allocation for EPON, GPON, andnext-generation TDM PON,” IEEE Communications Magazin, vol. 47,no. 3, pp. S40–S48, Mar. 2009.

[14] J. Zheng and H. Mouftah, “Media access control for Ethernet passive op-tical networks: an overview,” IEEE Communications Magazine, vol. 43,no. 2, pp. 145–150, February 2005.

[15] L. G. Kazovsky, W.-T. Shaw, D. Gutierrez, N. Cheng, and S.-W. Wong,“Next-generation optical access networks,” IEEE/OSA J. of Lightwave

Tech., vol. 25, no. 11, pp. 3428–3442, Nov. 2007.[16] J. Lazaro, J. Prat, P. Chanclou, G. T. Beleffi, A. Teixeira, I. Tomkos,

R. Soila, and V. Koratzinos, “Scalable extended reach PON,” in Proc.

OFC, Feb. 2008.[17] C. Michie, T. Kelly, I. Andonovic, and J. McGeough, “Reach extension

of passive optical networks using semiconductor optical amplifiers,” inProceedings of IEEE Int. Conf. on Transparent Opitcal Networks, June2008, pp. 194–197.

[18] H. Song, B. Kim, and B. Mukherjee, “Multi-thread polling: a dynamicbandwidth distribution scheme in long-reach PON,” IEEE Journal on

Selected Areas in Communications, vol. 27, no. 2, pp. 134–142, Feb.2009.

[19] G. Talli and P. Townsend, “Hybrid DWDM-TDM long-reach PONfor next-generation optical access,” IEEE/OSA Journal of Lightwave

Technology, vol. 24, no. 7, pp. 2827–2834, July 2006.[20] G. Kramer, B. Mukherjee, and G. Pesavento, “Interleaved polling with

adaptive cycle time (IPACT): A dynamic bandwidth distribution schemein an optical access network,” Photonic Network Communications,vol. 4, no. 1, pp. 89–107, Jan. 2002.

[21] M. McGarry, M. Reisslein, C. Colbourn, M. Maier, F. Aurzada, andM. Scheutzow, “Just-in-Time scheduling for multichannel EPONs,”IEEE/OSA Journal of Lightwave Technology, vol. 26, no. 10, pp. 1204–1216, May 2008.

[22] J. Zheng and H. Mouftah, “A survey of dynamic bandwidth allocationalgorithms for Ethernet Passive Optical Networks,” Optical Switching

and Networking, vol. 6, no. 3, pp. 151–162, July 2009.[23] A. Banerjee, G. Kramer, and B. Mukherjee, “Fair sharing using dual

service-level agreements to achieve open access in a passive opticalnetwork,” IEEE Journal on Selected Areas in Communications, Part

Supplement, vol. 24, no. 8, pp. 32–44, Aug. 2006.[24] B. Chen, J. Chen, and S. He, “Efficient and fine scheduling algorithm

for bandwidth allocation in ethernet passive optical networks,” IEEE

Journal on Selected Topics in Quantum Electronics, vol. 12, no. 4, pp.653–660, July-August 2006.

[25] M. Ma, L. Liu, and T. Cheng, “Adaptive scheduling for differentiatedservices in an Ethernet passive optical network,” OSA Journal of Optical

Networking, vol. 4, no. 10, pp. 661–67, 2005.[26] H. Naser and H. Mouftah, “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.[27] F. Melo Pereira, N. L. S. Fonseca, and D. S. Arantes, “A fair scheduling

discipline for Ethernet passive optical networks,” Computer Networks,vol. 53, no. 11, pp. 1859–1878, July 2009.

[28] A. Shami, X. Bai, N. Ghani, C. Assi, and H. Mouftah, “QoS controlschemes for two-stage Ethernet passive optical access networks,” IEEE

Journal on Selected Areas in Communications, vol. 23, no. 8, pp. 1467–1478, Nov. 2005.

[29] C. Assi, Y. Ye, S. Dixit, and M. Ali, “Dynamic bandwidth allocationfor Quality-of-Service over Ethernet PONs,” IEEE Journal on Selected

Areas in Communications, vol. 21, no. 9, pp. 1467–1477, November2003.

14

[30] P. Choudhury and P. Saengudomlert, “Efficient queue based dynamicbandwidth allocation scheme for ethernet PONs,” in Proceedings of

IEEE Globecom, Nov. 2007, pp. 2183–2187.[31] A. Shami, X. Bai, C. Assi, and N. Ghani, “Jitter performance in Ethernet

passive optical networks,” IEEE/OSA Journal of Lightwave Technology,vol. 23, no. 4, pp. 1745–1753, Apr. 2005.

[32] J. Zheng, “Efficient bandwidth allocation algorithm for ethernet passiveoptical networks,” IEE Proceedings Communications, vol. 153, no. 3,pp. 464–468, June 2006.

[33] S. Bhatia and R. Bartos, “IPACT with smallest available report first:A new DBA algorithm for EPON,” in Proc. of IEEE International

Conference on Communications, June 2007, pp. 2168–2173.[34] A. Kamal and B. F. Blietz, “A priority mechanism for the IEEE 802.3ah

EPON,” in Proceedings of IEEE ICC, May 2005, pp. 1879–1883.[35] G. Kramer, A. Banerjee, N. Singhal, B. Mukherjee, S. Dixit, and Y. Ye,

“Fair queueing with service envelopes (FQSE): a cousin-fair hierarchicalscheduler for subscriber access networks,” IEEE Journal on Selected

Areas in Communications, vol. 22, no. 8, pp. 1497–1513, Oct. 2004.[36] Y. Luo and N. Ansari, “Bandwidth allocation for multiservice access on

EPONs,” IEEE Communications Magazine, vol. 43, no. 2, pp. S16–S21,Feb. 2005.

[37] ——, “Limited sharing with traffic prediction for dynamic bandwidthallocation and QoS provioning over EPONs,” OSA Journal of Optical

Networking, vol. 4, no. 9, pp. 561–572, 2005.[38] Y. Zhu and M. Ma, “IPACT with grant estimation (IPACT-GE) scheme

for Ethernet passive optical networks,” IEEE/OSA Journal of Lightwave

Technology, vol. 26, no. 14, pp. 2055–2063, July 2008.[39] J. Zheng and H. Mouftah, “Adaptive scheduling algorithms for ethernet

passive optical networks,” IEE Proceedings Communications, vol. 152,no. 6, pp. 643–647, October 2005.

[40] P. Sarigiannidis, G. I. Papadimitriou, and A. S. Pomportsis, “A high-throughput scheduling technique with idle timeslot elimination mecha-nism,” IEEE/OSA Journal of Lightwave Technology, vol. 24, no. 12, pp.4811–4827, Dec. 2006.

[41] V. Sivaraman and G. N. Rouskas, “HiPeR-l: A high performancereservation protocol with look-ahead for broadcast WDM networks,”in Proceedings of IEEE Infocom, 1997, pp. 1270–1277.

[42] M. Pinedo, Scheduling: Theory, Algorithms, and Systems, 3rd ed.Springer, 2008.

[43] D. P. Heyman and M. J. Sobel, Stochastic Models in Operations Re-

search: Volume 1: Stochastic Processes and Operating Characteristics.Courier Dover, 2003.

[44] F. Aurzada, M. Scheutzow, M. Herzog, M. Maier, and M. Reisslein,“Delay analysis of Ethernet passive optical networks with gated service,”OSA Journal of Optical Networking, vol. 7, no. 1, pp. 25–41, Jan. 2008.

[45] “Csim (mesquite software),” http://www.mesquite.com.[46] P. L. Tien, Y.-M. Lin, and M. C. Yuang, “A novel OFDMA-PON

architecture toward seamless broadband and wireless integration,” inProceedings of OFC, Mar. 2009, pp. OMV2–1–OMV2–3.

[47] Z. Zheng and J. Wang, “A study of network throughput gain in optical-wireless (FiWi) networks subject to peer-to-peer communications,” inProceedings of IEEE ICC, June 2009.

michael mcgarry - utep college of engineeringengineering.utep.edu/docs/ath ece computer engr m...

Documents