Energy Efﬁciency in Ethernet Passive Optical Networks (EPONs): … · 2013. 4. 4. · Energy Efﬁciency in Ethernet Passive Optical Networks (EPONs): Protocol Design and Performance

Energy Efficiency in Ethernet Passive OpticalNetworks (EPONs): Protocol Design and

Performance EvaluationYing Yan and Lars Dittmann

Department of Photonics Engineering, Technical University of Denmark, Kgs. Lyngby, 2800 DenmarkEmail: {yiya, ladit}@fotonik.dtu.dk

Abstract— As concerns about energy consumption grow,the power consumption of the EPON becomes a matterof increasing importance. In respect of energy efficiency,the current standard has no management protocols aimingto reduce power consumption in EPONs. In this paper,we propose an Energy Management Mechanism (EMM)for downlink EPON systems. The proposed mechanism isdesigned to enhance the standardized control scheme inEPON with the objective to increase energy efficiency whilesatisfying diverse QoS requirements. The main idea is toput an Optical Network Unit (ONU) into the sleep modeand determine a suitable wakeup time scheduler at theOptical Line Terminal (OLT). A generic EPON system isconsidered, which is composed of an OLT and several ONUsthat are EMM enabled. An energy consumption optimizationproblem aimed at saving energy is proposed and twoheuristic sleep mode scheduling policies are addressed tosolve it. The scheduling algorithms are tightly coupled withthe upstream bandwidth allocation and downstream trans-mission scheduling together through an integrated approachin which awake time in ONUs is minimized. There is atrade-off decision between maximizing the power saving andguaranteeing the network performance at the same time.Simulation results show that an EMM-based EPON withwell designed scheduling disciplines is essential to achievingsignificant energy saving while meeting the delay constraint.

Index Terms— EPON, Energy efficiency, MPCP, schedulingalgorithms.

I. INTRODUCTION

It has been widely recognized that reducing powerconsumption in data communication network becomes animportant issue to global environment and future humanlife [1]–[5]. Recent research has shown that, among allnetwork segments, broadband access networks are thehighest share of the network energy consumption, whichconsumes more than 75 percent of energy consumptionby all telecommunication equipments today. Due to therapid expansion of access network connectivity, increas-ing number of users and data rate are foreseeable in thefuture broadband access networks. As a result, energyconsumption of future broadband access networks would

Manuscript received Feb 22, 2011; accepted Mar 11, 2011.This paper is based on “Energy Efficient Ethernet Passive Optical Net-

work (EPONs) in Access Networks,” by Ying Yan and Lars Dittmann,which appeared in the conference proceedings for the 10th WSEASInternational Conference on Applied Informatics And Communications(AIC) in Taipei, Taiwan August, 2010

continue to rise, with a steady annual growth rate inthe near future. Therefore, it has raised attention in bothacademia and industry to design energy efficient networksystems. Commonly, sleep mode operation is introducedto allow that nodes (or stations) can switch to sleep whenthey are idle and wake up when they receive or transmitpackets. Such approach has been widely exploited inwireless networks, where saving battery power in mobilestations is of paramount importance due to finite powersupplies in wireless devices [6]–[11].

The EPON, a prevailing deployed broadband accesstechnology, provides cost efficiency and high data ratefor the last mile access. A typical EPON is a Point-to-Multipoint (PMP) network with a tree based topology,where an OLT connects multiple ONUs via optical links.The OLT plays a role of distributor, arbitrator and aggre-gator of traffic. In the upstream direction (from ONUs tothe OLT), multiple ONUs share a single link and trafficmay collide. The OLT distributes the fiber capacity usingan upstream bandwidth arbitration mechanism to avoidcollisions. In the downstream direction (from the OLT toONUs), data frames are broadcasted to all ONUs. ONUsfilter and accept data that are addressed to them. However,ONUs have to constantly listen and examine downstreamtraffic, which results in wasting significant energy in theONU. Shown in Fig. 1, in the traditional EPON, ONUsconsume energy to keep active when there is neitherupstream nor downstream traffic. As a result, minimizingpower consumption is a major factor driving the design ofEPON devices and of the protocols therein. An effectiveEnergy Management Mechanism (EMM) that schedulesthe sleep mode period to ONUs is a key to conservepower.

The major sources of energy waste of shared resourceEPON system include: overhearing, control packet over-head, and idle listening. Overhearing represents that anONU receives and decodes packets that are not destinedto it. Control packet overhead means the transmissionof control messages, which are necessary to coordinatetransmission between the OLT and ONUs, and unfortu-nately increases time and energy consumption. At last,idle listening means active time and energy ONUs spenton waiting for the next upstream and downstream trans-mission. It is clear that an energy efficient scheduler isdesigned aiming to eliminate these sources of useless

JOURNAL OF COMMUNICATIONS, VOL. 6, NO. 3, MAY 2011 249

© 2011 ACADEMY PUBLISHERdoi:10.4304/jcm.6.3.249-261

1:K

splitterOLT

ss1

1 2 k 1

1 1

2

k

1 k 1 2 1 2 k

Uplink

Downlink

1 1 1

2 2

k k

ONU 1

ONU 2

ONU k

ss2

ssk

Figure 1. EPON upstream (multi-point to point) and downstream (pointto multi-point) transmission.

energy consumption.To reduce power consumption, ONUs are designed to

enter sleep mode when they do not need to either receiveor send traffic. During a sleep period, ONUs turn offthe transceiver in order to save energy. In case there isincoming data for a sleeping ONU, data is queued inbuffers at the OLT. Obviously, it is better to keep an ONUto stay in sleep mode as much as possible to conserve itsenergy. However, a power management mechanism withefficient scheduling for sleep and wake-up periods amongmultiple ONUs is a challenging task due to:

- ONUs should wake up and exchange traffic with theOLT. However, arrival traffic profile is not alwaysknown by the device, because the downstream andupstream traffic can be either periodically or aperi-odically arrived. Therefore, the sleep period shouldbe carefully assigned in order to avoid missing anyincoming packets.

- During sleep period, data is buffered and presched-uled for upstream and downstream transmission inboth the OLT and ONUs. Due to QoS constraints interms of delay and latency, the sleep period shouldbe carefully scheduled, so target ONUs can wakeup and complete transmission without violating QoSrequirements.

As discussed in next Section II, several studies havebeen proposed to analyze the power consumption forEPON, while applying the approach of allowing ONUsto sleep mode. Although Different aspects have beeninvestigated on how to minimize energy consumption inPON systems, there have been few studies focusing onthe protocol design for supporting sleep mode ONUs inEPON. To implement EMM in EPON, the legacy controlscheme requires modifications and extensions. Moreover,very few have explicitly addressed the trade-off betweenenergy consumption and network performances. To thebest of our knowledge, our work is the first to proposethe assignment of sleep period to ONUs taking intoaccount of the traffic transmission in both upstream anddownstream directions.

The remainder of this paper is organized as follows.We first give a brief description of related works inSection II. The introduction of the system model isaddressed in Section III. The design of EMM based EPONwith two different downstream schedulers is introduced

in Section IV. Simulation environments and results areoutlined and discussed in Section V. Finally, conclusionsand future work are drawn in Section VI.

II. RELATED WORKS

Issues of energy efficiency have been studied in EPONsas well as in other optical access networks. In this section,we present related works and compare their models withours.

Authors in [12] propose a management protocol forONUs to request entering in sleep mode. The OLT reservea minimal bandwidth for ONUs in sleep mode. During asleep period, no connection is established and no trafficis delivered. The sleep mode can be terminated by eitherthe OLT or the ONU by using an administration message.This protocol eliminates conventional requirements for acomplete activation procedure in ONUs. The authors ad-dress the impact of sleep mode protocol for re-activationprocess in GPON.

In [13], the authors present implementation challengesof sleep mode operation in PONs, such as achieving clocksynchronization and reducing the clock recovery over-head. The process of regaining network synchronizationafter waking up from sleep mode is studied in both GPONand EPON system. Two novel ONU architectures areproposed and compared with current ONU architectureson power consumption and wakeup overhead. As alreadyindicated, we focus on protocol design and do not addressthe physical implementation issues in our work. Thearchitectures as well as the overhead values proposedin [13] can be used for this purpose.

Another paper [14] proposes a power saving mecha-nism for 10 Gigabit class PON, where ONUs are im-plemented with a sleep and periodic wakeup regime. Thesleep mode is triggered based on negotiations between theOLT and the ONU, and a variable sleep time is allowed.The authors use a prediction model of traffic profiles todetermine the initiation and the duration of sleep mode.A probabilistic analysis of the presence or absence ofdownstream traffic is required in order to ensure theONU can wake up and the packet is delivered on time.This model differs from ours in that our model does notuse estimation and the energy management mechanism ismapped in EPON MAC layer.

Author in [15] presents the EPON power saving issueon the IEEE 802.3az meeting in September 2008. Thefunctions for power saving are proposed to be imple-mented in both the OLT and ONUs. The initiation andtermination of sleep mode by using handshake messagesare drafted. The attention of energy efficiency is raised inthe IEEE standardization, however, there are considerablyfew research works in saving energy consumption of opti-cal network management systems. Our work investigatesthe current control mechanism for EPON and proposescontrol protocol for sleep mode operation.

The relationship between energy and data rate in gen-eral optical IP networks was also studied in a differentcontext. For example, [16] presents power saving of

250 JOURNAL OF COMMUNICATIONS, VOL. 6, NO. 3, MAY 2011

© 2011 ACADEMY PUBLISHER

different network architectures, including all-optical net-works, first-generation networks and multi-hop networks.Power consumption of electronic versus optical compo-nents in optical WDM networks is investigated underdifferent traffic loads. In [4] this problem is generalizedby considering the power consumption in an optical IPnetwork. A new network-based model is established,which includes core, metro and access networks.

III. SYSTEM MODEL

In this paper, we consider an EPON system consistingof an OLT, 1 : K splitter and multiple ONUs. Theupstream and downstream traffic are separated in differentwavelength, typically 1310 nm for the upstream trans-mission and 1550 nm for the downstream transmission.TDMA is used in the physical layer where bandwidth isdivided in time slots. Each ONU maintains an upstreambuffer and sends upstream data to the OLT during as-signed time slots. For simplicity, we assume in this paperthat the upstream data rate is fixed for all ONUs. In thedownstream direction, when the arrival rate at the OLTexceeds the output data rate, data are queued in the OLTand transmitted when downstream bandwidth is available.

We also assume that the OLT and ONUs can regainsynchronization successfully so that ONUs can entersleep mode and return back to awake mode. The powerconsumption value in the awake state includes energyconsumed in listening to the OLT and in receiving ortransmitting data. We neglect the transition delays andthe energy consumption when an ONU changes states.We assume that the packets are able to be served withinan integer number of time slots. During the observationperiod, the number of ONUs is also fixed. Next we listsome notations (Table I) used in the rest of this work andthen formulate the energy saving scheduling problem.

Assuming that the overall transmission period is di-vided into N time slots, ONUs are either in awake mode

Notations ValueK The number of ONUs registered in the OLT.

i The index of ONUs in EPON. i ∈ K

N The total time slots within the overall observedtransmission period.

n The index of time slot. n ∈ N

sni ONU i state indication (awake modeor sleep mode) in time slot n.

Tni Period of time slot n in ONU i.

Pawake The average amount of energy (Watt)consumed during the awake period.

Psleep The average amount of energy (Watt)consumed during the sleep period.

TABLE I.NOTATIONS IN ENERGY EFFICIENCY MECHANISM

or in sleep mode during the time slot n. The total energyexpenditure model is formalized as follows:

E =K∑

i=1

N∑

n=1

Tni · [sni · Pawake + (1− sni ) · Psleep] (1)

A. Analysis of Sleep Mode OperationsWithin a period, the total energy consumption includes

energy consumed by M ONUs, which are in activestate. As expressed in our energy expenditure model(Equation 1), the total energy consumed by each ONUconsists of two parts: in either the awake or sleep mode. Inthe awake mode, there are three different states: receivingpackets, sending packets and idle listening. The idle stateis defined as the time when an ONU is active without anyreceiving or transmitting action. In the sleep state, ONUsswitch off transceiver functions and keep a timer on tocount down the wakeup time. Nodes consume much lesspower in the sleep mode than in the awake mode. Forinstance, regarding to different architectures for a sleepmode enabled ONU [4], the expected power consumptionsrequire 2.85 W for active and 750 mW - 1.28 W forsleeping. When ONUs enter sleep state, ONUs are inthe absence of traffic at the optical transmission interface.However, some functions, such as clock and data recoveryand back-end digital circuit, are still powered on. In thissense, the sleep mode differs from power down, which isusually an indication of any equipment fault or discon-nection in EPON. If an ONU is reported as deactivationafter a power down, the ONU has to register itself to theOLT again via the periodic discovery processing. On theother hand, using the sleeping mode, ONUs can wakeup, listen to, and communicate with the OLT without theregistration process, when the sleep period is finished.

As shown in Fig. 2, a timer is utilized in the sleepmode control. Upon receiving notification of enteringthe sleep mode, the timer counts down and the ONUwakes up when the timer is expired. When an ONUwakes up from the sleep mode, there is an overheadtime window (Tov) for recovering the OLT clock (Trecv)and retrieving network synchronization (Tsync). The clockrecovery time is various in different implementations ofONU receiver architectures. For synchronization process,a fixed Start Position Delimiter (SPD) is embedded inthe EPON control message header, which allows ONU togain synchronization with the start of the EPON frame.The synchronization time in EPON is up to 125 µs [13].

Following our model, it is obvious that energy depletionis determined by the time the ONU spend in awake mode.An optimal EMM solution is to schedule ONUs intosleep mode whenever there is no transmission in orderto minimize the overall power consumption. However,since there are QoS requirements in terms of delay andthroughput, it requires a scheduling policy to improveenergy efficiency without violating the QoS requirements.Therefore, this work deals with the trade-off between theallocation of wakeup frequency and the queuing delay inthe EPON.



Front-end

Analog Circuit

sleep Trecv Tsync

t

Enter sleep mode Wake up

Recover

OLT clock

Sync with

network timing

Dataup & down

ONU k

Sleep

control

Optical

Interface

Back-End

Digital

Circuit

Queue

OLTUser

Interface

ONU Receiver with sleep control

Dataup & down

Figure 2. ONU receiver with sleep control function.

B. Analysis of Energy Consumption and Delay Perfor-mance

When energy management mechanism is implementedin EPON, ONUs turn their optical communication in-terface on and off to minimize energy consumption.Therefore, in order for an ONU to communicate withthe OLT, it must be in active mode. As a centralizedcontrol station, the OLT schedules active duty cycles andinforms connected ONUs about its assignments. An ONUmust comply with the noticed sleep period and wakeuptime. When an ONU is in sleep mode, the downstreamtraffic destined to the ONU is queued in the OLT. Thiscommunication model imposes a clear trade-off betweenthe delay encountered by a packet and the time duringwhich ONUs are in active mode. Solutions for addressingthis trade-off depend, to a large extent, on the followingaspects:

1) The expected amount of data to be received andtransmitted.

2) The polling scheme and upstream bandwidth allo-cation scheme: for example, the sequence of polledONU and the size of bandwidth granted to theONU.

3) Whether the OLT process the packets and send toONUs with priority discrimination.

With respect to 1, if the data arrival rate is high andcontinuously delivered, data are queued and the OLT canschedule transmission. With respect to 2, if the pollingscheme and upstream bandwidth allocation are flexibleenough, it can determine the order of ONUs to commu-nication according to the expected delay constraint. Withrespect to 3, if the OLT can process packets, it may bufferand merge several packets for aggregation.

We first formulate our scheduling problem: find anoptimal scheduling discipline that minimizing the averageenergy consumption of EPONs while guaranteeing anupper bound on the queuing delay. This problem isformalized as an optimization problem that an optimalassignment of wakeup time to ONUs is an assignmentthat guarantees an upper bound Dmax on the maximumdelay while minimizing the total energy spent by theONUs in awake mode. Let K be a set of ONUs that are

awake and communicate with the OLT during the wholeN observation periods. Let Tn

i and Pawake be the awakeperiod and the average energy consumption for each ONUi, respectively. Then, the goal is to:

minimize : E =1

K

K∑

i=1

N∑

j=1

T ji · Pawake (2)

subject to:dmi < Dmax (3)

where dji is the delay experienced by every connectionsin AG i, which should deliver qualified services.

IV. ENERGY MANAGEMENT MECHANISM (EMM)

In this section, we introduce the implementation ofEMM together with EPON MAC protocol, MultipointControl Protocol (MPCP) [17]. The OLT is in full controlof the bandwidth allocation in both downstream andupstream. An ONU can associate with the OLT eitherin normal mode, referred to as Power Ignoring Scheme(PIS), or in EMM mode. When an ONU is in normalmode, it constantly stays awake and consumes power, i.e.,never goes into sleep mode. In the following, we assumethat all ONUs are EMM-enabled.

In order to minimize power consumption, ONUs re-main in a sleep mode most of the time while adheringto the assignment from the OLT. We show how beingin EMM yields considerable power saving for an ONUas compared to PIS case. In addition, we introduce howthe sleep mode scheduling scheme has an impact onthe energy and delay performance of the EMM ONUs.The sleep period and wakeup time can be calculated andassigned using either an Upstream Centric Scheduling(UCS) algorithm or a Downstream Centric Scheduling(DCS) algorithm.

A. Energy Efficient Scheduler Design

In EPONs, the downstream and the upstream trans-mission are separated, and the OLT plays a central roleto reserve bandwidth and to serve traffic transmission.In this paper, a subframe period (SP), Tsp, refers tothe time assigned to an ONU to send upstream data tothe OLT (SP-UL) or to receive downstream data fromthe OLT (SP-DL). In this work, we will analyze theenergy consumption and determine the wakeup time byconsidering both the upstream and downstream data sincethe instants of terminating sleep mode by upstream anddownstream transmission are different.

- Downstream bandwidth allocation: In the MPCPspecified in IEEE 802.3ah, the OLT broadcasts adownstream packet with destination MAC addressto all ONUs. ONUs that are awake to listen ifthe OLT has packets to deliver to them check theheader for destination identification. Since there isno retransmission mechanism in the EPON down-stream, it takes a long latency if the ONU misses



packets. Therefore, ONUs keep active all the timein order to listen to the OLT and receive packets ifthe destination identification is matched. Since thepower saving mechanism is ignored traditionally,time and energy are wasted when an ONU isactive to listening to the OLT and decoding packetsdestined to other ONUs.

- Upstream bandwidth allocation In the MPCP, theOLT maintains a polling list and a polled ONU hasthe right to transmit data within its assigned up-stream time slots. The OLT allocates upstream timeslots to each ONU in order to avoid that multipleONUs transmit to the uplink simultaneously. Theupstream bandwidth information such as the starttime and the length of granted transmission windowis carried in the GATE message. The OLT pollseach ONU by sending them GATE message andONUs transmits traffic within the assigned upstreamtime slots after receiving the GATE message. Theupstream bandwidth is distributed among all ONUsby using either a fixed allocation scheme or a dy-namic bandwidth allocation (DBA) scheme [18]. Inthe former case, each ONU is scheduled with a fixedamount of upstream transmission period. In thelatter case, the bandwidth is allocated dynamicallywith a bounded size according to ONUs’ request.A cycle (Tcycle) refers to an interval between twosuccessive polling messages to an ONU. Since thepower saving mechanism is ignored traditionally,the time is wasted when an ONU is active to waitingfor its next upstream transmission period.

The EPON supports energy management at the MAClayer, and ONUs need to negotiate with the OLT inorder to decide the power saving parameters. The powersaving parameters include the time to sleep and wake up.Theoretically, the power saving control can be initiatedeither at the OLT or at the ONUs, if they are in an idlestate, where no packet to be received or sent. However,as the ONU is an aggregation node that collects trafficfrom low level networks and relays to the OLT, it iscommonly assumed that ONUs always have queued datafor upstream transmission. In this work, we consider onlythe OLT triggered power saving mechanism. We focus onprotocol design of EMM and the scheduling schemes tomaximize power saving subject to the constraint of packetdelay.

With respect to bandwidth allocating, since the MPCPis central control protocol, the OLT has full knowledgeto assign transmission bandwidth in both upstream anddownstream directions. This implies that the order ofgranted ONUs and the amount of granted bandwidth aredetermined by the OLT. An intuitive solution for EMMscheduler is to aggregate the downstream traffic in order tominimize the time allocated to ONUs for idle states. Thegeneral design of EMM based EPON proposes followingfunctions in the OLT and ONUs:

1) OLT operation:The OLT can buffer downstream traffic and sched-

ule the appropriate transmission period to eachONU. To meet the QoS requirement of delay sensi-tive traffic, the downstream bandwidth is allocatedsuch that every packet can be served before itsdeadline. The original control message GATE ismodified with additional fields indicating the as-signed sleep time and wakeup time. To assign thewakeup time for the next upstream transmission,the OLT needs to calculate the time for an ONUto upload and download data packets. Consideringthe wakeup time to transmit the upstream packet,the OLT determines and allocates the upstreamtransmission windows for all ONUs. As for thewakeup time to receive the next downstream packet,since the number of buffered downstream data iscompletely know by the OLT, it can calculate andschedule downstream transmission.

2) ONU operation:After scheduling the sleep period, the OLT sendsa control message to the ONU for the permissionto transit into sleep mode. Upon receiving thismessage with parameters start time for sleep andwakeup time from sleep, the ONU enters into sleepmode. After a sleep mode, the ONU transits backto the awake mode again. Scheduled ONUs shouldwake up according to the assigned wakeup timeand check the sleep-mode GATE message (Gs).Upon receiving the Gs message, the ONU derivesthe sleep period and obeys the assigned upstreamand downstream bandwidth allocation. As for theupstream subframe period announcement, the as-signed wakeup time implies its next access timeto the upstream transmission. For a Gs messageannouncing the downstream subframe period, theONU should awake at the notified wakeup time toreceive the buffered data and then back to sleepafterwards.

B. Upstream Centric Scheduling (UCS) Algorithm

The main idea of UCS scheme is that the OLT assignsthe awake period to ONUs according to their correspond-ing upstream allocation. The OLT grants the upstreambandwidth by polling each ONU. During the grantedupstream subframe period, the ONU is awake. Once anONU transits into the awake state, the OLT with UCSalgorithm only transmits downstream packets to the awakeONU, and queues those packets destined for ’sleep’ ONUsinto a buffer.

In this subsection, we give a description of the UCSalgorithm. For simplicity of illustration, we consider asystem of an OLT and three ONUs for two upstreamtransmission periods. For more ONUs and transmissionperiods, the same logic can be applied. Fig. 3 shows theprocess of UCS protocol. In this work, the t denotesfor the instant time and T represents for a period. TheOLT maintains an entry table, which contains 1) thestart transmission time (tstartni ) and granted bandwidth(BWn

i ) for allocating upstream bandwidth; 2) the sleep



time (tsleepni ) and the wakeup time (twakeni ) for assigning

sleep period. The table is updated every cycle n forONUi. Downlink data are stored in subqueues for eachONU.

As shown in Fig. 3, the ONU1 wakes up at t0and receives sleep mode GATE message (Gs

11) at the

time t1. After learning the granted upstream subframeperiod (Tsp−ul

11), ONU1 generates REPORT message

(R1) and starts upstream transmission (dul11) complyingwith the assignment. During Tsp−ul

11, the OLT derives

downstream data (ddl11) from the subqueue for the ONU1

and transmits them to the ONU1. Notice that the data ddlrepresents either one data packet or a series of integrateddata packets. The start time of sleep mode (tsleep11)for ONU1 is assigned so that the ONU1 enters intosleep mode immediately when the upstream subframeperiod is completed at t2. The calculations of upstreamtransmission period and the time to turn into sleep modeare listed in Equation 4 and Equation 5.

Tnsp−ul,i =

BWni

Ro, i ∈ K,n ∈ N (4)

tnsleep,i = tnstart,i + Tnsp−ul,i, i ∈ K,n ∈ N (5)

where BWni is the allocated upstream bandwidth for

the ith ONU during the nth cycle. Ro is the transmis-sion rate of the optical upstream and Tg is the guardtime between two successive upstream transmissions. Thewakeup time (twakeup

11), which is same as the length of

sleep period, is calculated based on the period before theONU1 is polled again (Tcycle

11). In the Fig. 3, the next

wakeup time for the ONU1 is t3. The value of wakeuptime is computed at the OLT and assigned to each ONU.The polling cycle is computed using Equation 6.

Tncycle,i =

K∑

i=1

(Tnsp−ul,i + Tg)

=K∑

i=1

(BWn

i

Ro+ Tg), i ∈ K,n ∈ N

(6)

After a period of Tcycle, the interval between twoadjacent polling messages, an ONU is polled again at t4(t1 + Tcycle

11). The calculation of ONU wakeup time needs

to take the overhead period into account due to the timeof recovering the OLT clock and retrieving the networksynchronization. The computation of the overhead period(Tov) and the time to wake up are listed as Equation 7and Equation 8, respectively.

Tov = Trecv + Tsync (7)

tnwakeup,i = tnstart,i+Tncycle,i−Tov, i ∈ K,n ∈ N (8)

During the observation period, the total awake time foreach ONU is the sum of allocated upstream time slotsand an ONU enters the sleep mode otherwise. During thesubframe period, the ONUi transmits upstream data dul

ni

Start

Check upstream polling

timer

Time to send

GATE to ONUi

Allocate upstream bandwidth for

the ONUi based on

RTTi and BWni

Calculate the time for next

polling mesage to the ONUi

(Tcycleni)

Assign sleep period

info to the ONUi

(tsleepni and twakeup

ni)

Setup timer to poll the ONUi

after Tcycleni

Generate and transmit

sleep mode GATE

message to the ONUi

Time to send

DATA to ONUi

Check upstream transmission

period to find awake ONUi

Update OLT entry table with

tstartni

Available data

for ONUi in the buffer

Transmit DATE

message to the ONUi

Y

N

Y

Y

N

N

Figure 4. Flowchart of the UCS algorithm designed in the OLT.

and receives downstream data ddlni , within the granted

time slot n. Notice that the size of the uploaded anddownloaded data must equal or less than the upstreambandwidth, i.e. size(dulni )≤ BWn

i and size(ddlni )≤ BWni .

The total energy consumed by the ONU i is illustrated inEquation 9:

EUCS =

K∑

i=1

N∑

n=1

[Tnsp−ul,i + Tov] · Pawake

+K∑

i=1

N∑

n=1

[Tncycle,i − Tn

sp−ul,i − Tov] · Psleep

(9)

The UCS based scheduler is simple because the sleepperiod is determined based on the upstream transmission.An OLT with the UCS based scheduler uses the awakeONU, which is assigned to the upstream transmissionas destination to retrieve buffered data. The algorithmis presented in the flowchart in Fig. 4. However, theperformance of downstream data latency and bandwidthutilization may not be satisfied due to the dependencyon the upstream subframe period, the upstream pollingsequence, and the total number of active ONUs. Onedisadvantage of this scheme is that it is not suitablefor delay-sensitive traffic due to longer queuing delayexperienced in the OLT.

C. Downstream Centric Scheduling (DCS) Algorithm

The second scheduling policy presented in this paper isthat the OLT stores downstream traffic in a First-In First-Out (FIFO) buffer and the ONU has to be awake and



dul11

t1

ONU1

t

R1

t2 (tsleep 11)

t4

t3 (tstart 21)

t

ONU2

ONU3

ddl1

1Gs11

ddl21

dul2

1R1

dul12R2

ddl1

2

dul2

2R2

ddl22

dul1

3R3

ddl13

R3

Tsleep

Tsleep

Tsleep

Tsleep

Tsleep

t5

REPORT

Sleep mode GATEGs

R

t

txrx

tx

rx

txrx

ONU RTTGranted

BWtsleep

tsleep n

11

2

3

Trtt 1 BW1

Start Tx

Time

tstart n1

tstart n2

tstart n3

Trtt 2

Trtt 3

OLT Entry Table

twakeup

twakeup n

1

twakeup n2

twakeup n3

OLTtxrx

Gs1

1

1

2

3

ddl 1

OLT DL Queue

ddl 2

ddl 3

Gs2

1

dul11R1

BW11

Tg dul12R2

BW12

dul13R3

BW13

Tg

ddl11G1 ddl

12G2

Gs1

2

ONU Ddownlink

Gs12

Gs1

3

Gs21

Gs22

Gs2

3

Uplink datadul

Downlink dataddl

Tg Guard time

t

t0

BW2

BW3

tsleep n2

tsleep n3

tstart 11

Tsp-ul 11

Tsp-dl 11

Tcycle11

Tov

Figure 3. Sleep mode operation with Upstream Centric Scheduling (UCS) scheme.

receive downstream traffic whenever the OLT sends one.The awake periods are assigned to favor both upstreamand downstream traffic. The DCS scheme is illustratedin Fig. 5. The sleep mode GATE message, Gs

ni , is

used for both upstream bandwidth allocation and sleepperiod assignment. The upstream transmission period iscalculated using the same equations 4- 8 as in the UCSscheme. For example, the Gs

11 at time t1 indicates that the

next polling time is at time t8 and the ONU1 needs towake up at time t7. In the downstream, the OLT servesdata in a first-in first-out order. For instance, shown inFig. 5, during the upstream subframe, dul11, the OLT sendspackets to both ONU1 (data ddl

11) and ONU2 (data ddl

12).

As shown in Fig. 5, the admission information for bothupstream and downstream transmission is notified by thesleep mode GATE message. The Gs

11 informs ONU1

the allocated upstream transmission of dul11 and assignedsleep period during t3 to t7. After transmitting ddl

12 to

the ONU2, the OLT gets data ddl11 from the head of its

downlink queue and sends to the ONU1 together withGs

21. Since there is data ddl

21 in the buffer, the ONU1 is

notified to wake up at t4 instead of t7. The sleep periodis assigned if there is neither upstream nor downstreamtransmission scheduled. If the OLT has queued data forthe ONU, the wakeup time for the next data can becalculated. For example, the period between t3 and t4is the idle period for ONU1. Otherwise, the OLT sends amessage to inform the ONU of remaining in awake modein order to avoid missing any downstream traffic. In the

following, the assignment of sleep period is discussed indetails.

Under the condition that the downstream traffic ter-minates the sleep mode and the OLT assign the sleepperiod for the ith ONU, we distinguish four possibilitiesas shown in Fig. 6. The sleep period is determined byfavoring both the successful reception of upstream anddownstream data. Accordingly, we denote Gs as theGATE control message carrying the information of sleepperiod, such as the start time of sleep period (tsleep) andthe wakeup time (twakeup). Let Tsp−ul

ni and Tsp−dl

ni be

the nth subframe period of uplink and downlink transmis-sion, respectively. In addition, the Tov period representsan overhead time for clock recovery and synchronization.

1) Case0: GATE0 is used to inform the allocated up-stream bandwidth for the ONUi. Because the OLThas full knowledge about the upstream bandwidthallocation, the sleep interval is determined, includ-ing both the sleep time (tsleep0) and the wakeuptime (twakeup0).

2) Case1: The transmission of downstream data(Tsp−dl

1i ) is completed within the allocated up-

stream subframe period. If there is other queueddata in the OLT, such as the payload with GATE2,the sleep period is assigned as same as in Case0. Ifthe next received data is the payload with GATE3,the time to wake up is reassigned as same astwakeup2. If there is no more queued data in theOLT for the ONUi, the original assignment, tsleep0,



REPORT

Sleep mode GATEGs

R

ONU RTTGranted

BWtsleep

tsleep n11

2

3

Trtt 1 BW11

Start Tx

Time

tstart n1

tstart n2

tstart n3

Trtt 2

Trtt 3

OLT Entry Table

twakeup

twakeup n1

twakeup n2

twakeup n3

OLTtxrx

Gs1

1

OLT DL FIFO

Queue

Gs4

1

dul11R1

BW11

Tg dul12R2

BW1

2

dul13R3

BW1

3

Tg

Gs1

2

Uplink datadul

Downlink dataddl

Tg Guard time

t

BW12

BW13

tsleep n2

tsleep n3

t1 t2 t4t3 t5 t6 t7 t9t8

Gs11 ddl

12

Gs12

ddl11

Gs2

1

Gs2

2

Gs22

ddl22 Gs

31 ddl

21

Gs21 ddl

13Gs

13

Gs1

3t

t

t

Tsleep Tsleep

Tsleep

Tsleep

ddl11

dul12R2

dul13R3

ddl13 G3

Tsleep

dul23R3

dul22R2

dul21R1tx

rx

tx

rx

txrx

ONU1

ONU2

ONU3

ddl31

ddl12

ddl32

Tsleep

Gs1

1 Gs2

1 Gs4

1

Gs12

Gs51

Gs22 Gs

32

Gs13

dul11R1

ddl22

ddl21Gs

31

Gs4

1 ddl32Gs

32

ddl31

Gs51

ddl12

ddl11

ddl22

ddl21

ddl13

ddl32

Tsleep

Figure 5. Sleep mode operation with Downstream Centric Scheduling (DCS) scheme.

is removed, because ONUi should keep awake inorder to avoid missing future arrival data.

3) Case2: In this case, the downstream data (Tsp−dl2i )

can not be finished before the start time of thesleep period as assigned in Case0 and Case1. There-fore, the start time of sleep period is postponed totsleep2. The OLT examines its downstream queueand schedules the transmission of the next down-stream data (Tsp−dl

3i ) for the ith ONU. Moreover,

the wakeup time (twakeup2) is calculated in orderto ensure a successful transmission.

4) Case3: In this case, there is downstream data re-ceived during the sleep interval. In GATE3 controlmessage, the new time of entering the sleep modeis updated, tsleep3, which is calculated based on thedownstream subframe period (Tsp−dl

3i ).

5) Case4: As showing in the last GATE4 in this figure,there is downstream data arrived during the sleepinterval and lasted till the next upstream subframeperiod. Upon the sleeping period is completed, theONU transits into the awake mode and will receivethe GATE4 message with the information of up-stream bandwidth allocation, such as the start timeand length of Tsp−dl

4i . Thus, the GATE4 message

contains bandwidth assignments for both upstreamand downstream subframe period.

The flowchart of the DCS algorithm implemented inthe OLT is described in Fig. 7. Compared to the UCSalgorithm, the OLT checks the available data in the bufferand update the sleep period for the destination ONUwith a sleep mode GATE message. The wakeup time isprecisely calculated and assigned, so that the ONU canbe active to carry out the next upstream or downstream

transmission. Since the operations of the above schedulingmechanism requires the OLT to locate the next packet inthe buffer in order to calculate the wakeup time, ONUswhich have no more queued packets cannot enter the sleepmode. In this case, the OLT assigns the sleep period aszero.

In order to calculate the total energy consumption in theDCS algorithm based EMM. We define the awake periodfor upstream transmissions (Tawake−ul) and downstreamtransmissions (Tawake−dl) separately. First, we computethe awake period and sleep period in the upstream direc-tion (shown in Equation 10 and Equation 11), which issimilar to Section IV-B.

Tnawake−ul,i = Tn

sp−ul,i + Tov, i ∈ K,n ∈ N (10)

Tnsleep−ul,i = Tn

cycle,i − Tnsp−ul,i − Tov, i ∈ K,n ∈ N

(11)Next, assuming that there are M downstream trans-

missions to the ith ONU. For the downstream, the awakeperiod is calculated (in Equation 12) differently in eachcase mentioned as Fig. 6. In Case1, ONUi is in awakestate, so that the awake period for receiving downstreamdata is zero. From Case2 to Case4, if the time forwaking up and sleeping is assigned, the period that theONU has to wake up from its sleep mode is calculatedas the interval between tsleep and twakeup. Illustratedin Equation 13, the sleep period is the upstream sleepperiod minus the awake period used for the downstreamtransmission, which occurs during the upstream sleepperiod.



Tsp-ul 1i Tsp-ul

2i

Tsp-dl1iGs Gs Tsp-dl

2iGs Tsp-dl

3iGs Tsp-dl

4i

Gs

GATE1

tsleep0 (tsleep1)

tsleep2 tsleep3twakeup2

twakeup0

(twakeup1)

GATE0 GATE2 GATE3 GATE4

tsleep4

tONUi

tx

rx

Tov

twakeup3

Tov Tov Tov

Figure 6. Downstream transmission in the sleep period model.

Tnawake−dl,i =

M∑

m=1

Tmawake−dl,i

=M∑

m=1

0 , if Case1tmsleep2 − tmsleep1 , if Case2tmsleep3 − tmsleep2 , if Case3tmsleep0 − tmsleep3 , if Case4

(12)

Tnsleep−dl,i = Tn

sp−ul,i−M∑

m=1

Tmawake−dl,i, i ∈ K,n ∈ N

(13)After analyzing the active behavior of an ONU in both

the upstream and downstream transmissions, the totalawake and sleep periods are concluded as Equation 14and Equation 15:

Tnawake−total,i = Tn

awake−ul,i+Tnawake−dl,i, i ∈ K,n ∈ N

(14)

Tnsleep−total,i = Tn

sleep−dl,i, i ∈ K,n ∈ N (15)

The total energy consumed is calculated in Equation 16.The DCS sleep mode scheduler transmits downstreamtraffic in a flexible way, unlike in the UCS that onlythe active ONU scheduled with an upstream transmissionperiod can receive data from the OLT. The OLT triesto schedule the queued downstream traffic in a wayto minimize the delay. The UCS is simple where thesleep period and wakeup time is determined only bythe upstream bandwidth allocation. On the other hand,the DCS scheduler requires to keeping track of bothdownstream and upstream transmission windows.

EDCS =K∑

i=1

N∑

n=1

[Tnsp−ul,i · Pawake]

+K∑

i=1

N∑

n=1

[Tnsleep−total,i · Psleep]

(16)

V. SIMULATION RESULTS

A. Simulation Scenario

In this section, we study performances and features ofour proposed energy management mechanism compared

to the fixed power control, which is represented as PowerIgnoring Scheme (PIS). Simulations have been carriedout by means of OPNET network simulator [19] to mea-sure the performance metrics. The performed simulationsaimed to show the diverse effects of system parameters,such as the upstream scheduling scheme, the number ofconnected ONUs, and the arrival packet size. Then thebenefits that can be obtained while utilizing the proposedEMM in EPON. The following parameters are monitored,such as the power consumption, the network throughput,and the average queuing delay. The average awake timein ONUs is used as a merit for energy efficiency perfor-mances. The less time an ONU is awake, the less energyan ONU consumes.

We have assumed that the OLT is connected to KONUs (K=16 or 32)., which are enabled with power sav-ing functionality. The optical link rate is 1 Gbps for bothupstream and downstream transmission. The guard timebetween two consecutive transmission slots is 5 us. Forallocating upstream bandwidth among ONUs, both fixedscheme (e.g. TDM) and dynamic scheme (e.g. IPACT)are implemented at the OLT. The maximum transmissioncycle is set to 2 ms. As the proposed power saving mecha-nism with two scheduling schemes cooperates the opticaldownstream bandwidth scheduler at the OLT, it is rea-sonable to monitor the performance of downstream traffic.Traffic is generated following Poisson distribution and thelength of packets is generated independently between 64bytes to 1200 bytes. Destinations of downstream trafficare uniformly distributed among connected K ONUs.The MPCP control messages, GATE and REPORT, areformatted according to the IEEE 802.3ah specification.The sleep mode GATE message, Gs, is an extension ofthe standard GATE message. The additional two 4 bytesfields are embedded as assigned sleep time and wakeuptime. The total size of Gs message is set to 41 bytes.

B. Validation of EMM with the fixed upstream bandwidthallocation

In Section IV, we analyze the energy efficiency mecha-nism with two proposed scheduling approaches and derivethe consumed power accordingly. The simulation scenarioconsists of 16 ONUs and the OLT grants fixed upstreambandwidth to each ONU. The results in Fig. 8 refer tothe system behavior in the presence of an fixed upstreambandwidth allocation scheme deployed at the OLT. Theyshow the energy consumption, represented by the ONU



Start

Check upstream polling

timer table

Time to send

GATE to ONUi

Transmit DATE

message to the ONUi

Y

N

Y

Y

N

NONUK is awake

Available data Dk

in the buffer

Get data destination K;

Check sleep period table;

Available next data

in the buffer for ONUK

Y


DATA to the ONUk

Generate GATE with

downstream transmission info.

(SP_DL-starttime, SP_DL-length)

Assign new sleep period info (tsleepni and twakeup

ni) according

to the current active status in the ONUi (ref to sleep period

decision figure)

Transmit GATE

message to the ONUi

Set sleep period is none

N

Allocate upstream bandwidth for

the ONUi based on

RTTi and BWni


polling mesage to the ONUi

(Tcycleni)

Assign sleep period

info to the ONUi

(tsleepni and twakeup

ni)

Setup timer to poll the ONUi

after Tcycleni

Update OLT entry table with

tstartni

Generate and transmit

sleep mode GATE

message to the ONUi

Update sleep period table

Figure 7. Flow chart of the Download Centric Scheduling (DCS)Scheme.

awake time, when the downstream traffic load changes. Inthe case of PIS, ONUs constantly stay in awake state andconsumes high energy (100% of the observation period).When the EMM is enabled in the EPON system, theawake time maintains at a low average percentage value ifthe Uplink Centric Scheduling (UCS) scheme is applied.As explained in Section IV-B, the period of being poweron is decided by allocated upstream transmission periodsin the UCS case. When each ONU is allocated with fixedamount of upstream bandwidth, the awake time is alsofixed. On the other hand, under the DCS case, ONUs wakeup and communicate with the OLT for either upstreamor downstream traffic. When arrival traffic is less thanthe optical output link rate, the awake time of EMM-DCis slightly increased due to the additional active periodfor downstream traffic. When the arrival traffic rate ishigh, the awake time increases and close to the level asin the PIS case. With respect with energy efficiency, theproposed energy management mechanism outperforms thetraditional power ignoring mechanism. Particularly, usingthe uplink centric scheduling scheme saves up to 90%power on time. The energy consumption in the DCS casevaries when the traffic load changes.

Fig. 9 shows the network throughput under the PIS andEMM cases. The downstream throughput is monitored,because the upstream throughput is not affected by theproposed power saving mechanism. When the EMMis deployed, both DCS and UCS schemes degrade thethroughput performance. In the DCS case, throughputis decreased due to the additional overhead by intro-ducing the sleep mode GATE control message. Becausethe wakeup time and sleep time are assigned to ONUswhen they receive every downstream packet, it causes aportion of control message overhead. In the UCS case,the network throughput is reduced greatly. The reasonis that the OLT can dispatch downstream traffic duringits allocated upstream transmission period. Assuming theOLT polls and grants all ONUs in a Round Robin (RR)order, each ONU can derive one over sixteen percentof total downstream link capacity. The link utilizationfactor, computed as the amount of traffic delivered to thedestination successfully over the total link capacity in thesimulation time. The higher the network throughput, thehigh link utilization. The proposed EMM receives similarresults compared to the PIS case, when the DCS schemeis applied. However, performance is highly downgradedin the UCS scheme.

The average queueing delay for analyzing performancedifferences under different power saving settings is shownin Fig. 10. The average delay of the EMM DC schemeis better than that of the UCS scheme since, with UCSscheme, downstream packets have to wait for their awakeperiod in the OLT buffer. We observe that when thetraffic intensity is low, average delay remains constantsince the transmission rate is high enough to serve undera range of values of traffic load. However, when theincoming traffic intensity reaches a certain point andcauses congestion, average delay increases rapidly. Asexpected, the average queueing delay is consistent withthe throughput performance, where the EMM DC schemeachieves better performance.

C. Impact of Upstream Bandwidth Allocation Schemes

Now we examine the impact of upstream bandwidthallocation schemes on the proposed energy managementmechanism with the uplink centric scheme. In this sce-nario, only upstream traffic from ONUs to the OLT isconsidered to highlight the results differed from differentupstream bandwidth allocation schemes. If the down-stream traffic is involved, then additional time is requiredfor transmitting downstream packets.

Fig. 11 illustrates a simple scenario that shows therelationship between the upstream bandwidth allocationalgorithm, TDM and DBA, with the ONU awake time.Because the downstream traffic is ignored, the sleep pe-riod is assigned to each ONU only based on the allocatedupstream bandwidth, as the EMM UCS performs. Inthe TDM case, each ONU is assigned with a fixed slottime. In the DBA scheme case, the OLT grants ONUsaccording to their bandwidth requests. Under low trafficload, buffered data at ONUs are small, and therefore,



0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0

0

10

20

30

40

50

60

70

80

90

100

Ave

rag

y p

erc

en

tag

e o

f O

NU

aw

ake

tim

e (

%)

ONU Downstream Transmission Rate (Gbps)

EMM-UC (TDM)

EMM-DC (TDM)

PIS

Figure 8. Average ONU awake time in PIS, EMM-UC, and EMM-DCalgorithms.

0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0

0

100

200

300

400

500

600

Th

rou

gh

pu

t T

x to

ON

U.m

ea

n (

Mb

ps)


EMM-UC (TDM)

EMM-DC (TDM)

PIS

Figure 9. Network throughput in PIS, EMM-UC, and EMM-DCalgorithms.

small amount of requested time slots. The percentageof awake time is increased along with the increasedtraffic rate. Because the maximum upstream slot time isconstrained as 2 ms, the DBA based approach has similarresults under the heavy traffic load.

Fig. 12 and Fig. 13 show the network throughput andqueuing delay for the downstream traffic, respectively.The volume of downstream traffic is fixed at 1 Gbpsand destination is uniformly distributed among 16 ONUs.Consistent with the curve of awake time, the downstreamthroughput increases when the ONU receives high inputtraffic and requests for large timeslots. On the other hand,when the required upstream transmission period is small,ONUs are turned off most of the observation time anddownstream data are queued at the OLT. Thus the queuingdelay is high under light traffic load in the dynamicbandwidth allocation case.

From previous studies [18], dynamic bandwidth alloca-tion schemes improve EPON channel utilization and pro-vide QoS guarantee. Our simulations over the upstreamchannel have demonstrated that the dynamic upstreambandwidth allocation scheme gains system power effi-ciency. However, the delay performance is downgradedwhen the upstream traffic load is low.

0,0 0,2 0,4 0,6 0,8 1,0 1,2 1,4 1,6 1,8 2,0

0

20

40

60

80

100

120

140

160

180

200

220

Avg

Qu

eu

e D

ela

y in

ON

U.m

ea

n (

ms)


EMM-UC (TDM)

EMM-DC (TDM)

PIS

Figure 10. Average queueing delay in PIS, EMM-UC, and EMM-DCalgorithms.

0 10 20 30 40 50 60 70 80 90 100

0

10

20

30

40

50

60

70

80

Ave

rag

y p

erc

en

tag

e o

f O

NU

aw

ake

tim

e (

%)

ONU Upstream Transmission Rate (Mbps)

EMM-UC (TDM)

EMM-UC (DBA)

Figure 11. Awake time compared in different upstream bandwidthallocation schemes.

D. Impact of the sleep mode GATE message overhead

In this study, simulations show impacts of the sleepmode GATE control message. The MPCP control mes-sages, GATE and REPORT, are formatted according tothe IEEE 802.3ah specification. The sleep mode GATEmessage, Gs, is an extension of the standard GATEmessage. The additional two 4 bytes fields are embeddedas assigned sleep time and wakeup time. The total sizeof Gs message is set to 41 bytes. In the UCS case,the sleep mode GATE message is used to poll ONUs,grant upstream bandwidth, and inform assigned sleepperiod. The Gs is generated and sent to ONUs in thesame means as the original MPCP case. However, inthe DCS case, Gs message is transmitted together witheach downstream packet, which results in amount ofcontrol message overhead. Therefore, in this test case,we simulate a network scenario with 16 ONUs and focuson EMM DCS scheme.

Fig. 14 shows the number of transmitted Gs as afunction of incoming packet sizes. In all cases, the totalnumber of Gs is increased at the downstream data rateincreases. The size of data packets are varied between400 bits and 9400 bits. Under a certain input data rate,the larger the data size, the less transmitted Gs controlmessage. With respect to bandwidth utilization, increasednumber of Gs message introduces high value of control



0 10 20 30 40 50 60 70 80 90 100

0

10

20

30

40

50

60

70T

hro

ug

hp

ut T

x to

ON

U.m

ea

n (

Mb

ps)


EMM-UC (TDM)

EMM-UC (DBA)

Figure 12. Network throughput compared in different upstream band-width allocation schemes.

0 10 20 30 40 50 60 70 80 90 100

0,0

0,5

1,0

1,5

2,0

Avg

Qu

eu

e D

ela

y in

ON

U.m

ea

n (

s)


EMM-UC (TDM)

EMM-UC (DBA)

Figure 13. Average queuing delay compared in different upstreambandwidth allocation schemes.

overhead, therefore, results in low bandwidth utilization.The delay performance is illustrated by Fig. 15. The

delay difference is hardly to tell when the downstreamtraffic load is low. When traffic load becomes high, thelarge packet size case produces better delay performance.This is for the reason that higher channel utilization canbe achieved by reducing the ratio of the overhead of Gs

message and the size of payload.

E. Performance comparison with multiple ONUs

We now consider a scenario where 32 ONUs areconnected to the OLT. With this network structure, anONU is allocated less upstream transmission periods.Fig. 16 shows the awake periods are reduced in both UCSand DCS schemes compared to the 16 ONUs structure.Consequently, the queuing delay is increased (shown inFig. 17). When the traffic load becomes high enough,the downstream traffic takes all transmission time andconsumes the same power as in the PIS case.

VI. CONCLUSION

In this paper, we focus on the energy consumptionin the EPON system. We proposed and investigated ananalytical model for the energy expenditure in EPON.We developed an energy management mechanism foroptimal assignments of sleep periods to ONUs. The delayconstraints that need to be satisfied is analyzed in the

0,0 0,2 0,4 0,6 0,8 1,0 1,2

0

400000

800000

1200000

To

tal n

um

be

r o

f tr

an

sm

itte

d s

lee

p m

od

e G

AT

E m

essa

ge

s


400 bits

1400 bits

2400 bits

3400 bits

4400 bits

5400 bits

6400 bits

7400 bits

8400 bits

9400 bits

Figure 14. the number of transmitted Gs as a function of incomingpacket sizes.

0,0 0,2 0,4 0,6 0,8 1,0 1,2

0,0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

Avg

Qu

eu

e D

ela

y in

ON

U.m

ea

n (

s)


400 bits

1400 bits

2400 bits

3400 bits

4400 bits

5400 bits

6400 bits

7400 bits

8400 bits

9400 bits

Figure 15. Average queuing delay performance as an impact of thesleep mode GATE message overhead.

scheduling problem. We solve the problem by proposingtwo heuristic scheduling algorithms: the upstream centricscheduling algorithm and the downstream centric schedul-ing algorithm. An EPON system with energy efficiencybased scheduling schemes are simulated in the OPNETsimulator. We verified the solution quality with respect toa set of chosen metrics such as awake time and queuingdelay. Specifically, we compared the energy performanceand the network performance under different traffic loadand different upstream bandwidth allocation schemes. Wefound that the energy management mechanism reduces thepower consumption, and the DCS algorithm outperformsthe UCS algorithm in the delay performance.

REFERENCES

[1] C. Lange and A. Gladisch, “Energy consumption oftelecommunication networks: A network operator’s view,”in Work shop on Optical Fiber Communication Confence(OFC) ”Energy Footprint of ICT: Forecasts and networksolutions”, 2009.

[2] C. Lange, D. Kosiankowski, C. Gerlach, F. Westphal, andA. Gladisch, “Energy consumption of telecommunicationnetworks,” in 35th European Conference on Optical Com-munication (ECOC), 2009, pp. 1–2.

[3] M. Gupta and S. Singh, “Greening of the internet,” ACMSIGCOMM, pp. 19–26, 2003.

[4] J. Baliga, R. Ayre, K. Hinton, W. V. Sorin, and R. S.Tucker, “Energy consumption in optical IP networks,”IEEE/OSA Journal of Lightwave Technology (JLT), vol. 27,pp. 2391–2403, 2009.



0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0

0

10

20

30

40

50

60

70

80

90

100A

ve

rag

y p

erc

en

tag

e o

f O

NU

aw

ake

tim

e (

%)


16 ONU: EMM-UC (TDM)

16 ONU: EMM-DC (TDM)



PIS

Figure 16. Awake time compared in scenarios with different connectedAGs.

0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0

0

20

40

60

80

100

120

140

160

180

200

220

240

Avg

Qu

eu

e D

ela

y in

ON

U.m

ea

n (

ms)






PIS

Figure 17. Network throughput compared in scenarios with differentconnected AGs.

[5] K. Christensen, C. Gunaratne, B. Nordman, and A. George,“The next frontier for communications networks: Powermanagement,” Computer Communications, vol. 27, pp.1758–1770, 2004.

[6] C. E. Jones, K. M. Sivalingam, P. Agrawal, and J. C.Chen, “A survey of energy efficient network protocols forwireless networks,” Wireless Networks, vol. 7, pp. 343–358, 2001.

[7] J. Lee, C. Rosenberg, and E. Chong, “Energy efficientschedulers in wireless networks: Design and optimization,”Mobile Networks and Applications, vol. 11, pp. 377–389,2006.

[8] Y. Zhang, “Performance modeling of energy managementmechanism in IEEE 802.16e mobile WiMAX,” in WirelessCommunications and Networking Conference (WCNC),2007, pp. 3205–320.

[9] R. Cohen and B. Kapchits, “An optimal wake-up schedul-ing algorithm for minimizing energy consumption whilelimiting maximum delay in a mesh sensor network,”IEEE/ACM Transactions on Networking, vol. 17, pp. 570–581, 2009.

[10] J. Hsieh, T. Lee, and Y. Kuo, “Energy-efficient multi-polling scheme for wireless LANs,” IEEE Transactions onWireless Communications, vol. 8, pp. 1532–1541, 2009.

[11] Y. Chen and S. Tsao, “Energy-efficient sleep-mode oper-ations for broadband wireless access systems,” in IEEEVehicular Technology Conference, 2006, pp. 1112–1116.

[12] T. Smith, R. S. Tucker, K. Hinton, and A. V. Tran,“Implications of sleep mode on activation and rangingprotocols in PONs,” 21st Annual Meeting of the IEEELasers and Electro-Optics Society (LEOS), vol. 11, pp.604–605, 2008.

[13] S. Wong, L. Valcarenghi, S. Yen, D. Campelo, S. Ya-mashita, and L. Kazovsky, “Sleep mode for energy sav-

ing PONs: Advantages and drawbacks,” in Globlecom’09Second International workshop on Green Communication,2009.

[14] R. Kubo, J. Kani, Y. Fujimoto, N. Yoshimoto, and K. Ku-mozaki, “Proposal and performance analysis of a power-saving mechanism for 10 gigabit class passive opticalnetwork systems,” in NOC, vol. 1, 2009, pp. 87–94.

[15] EPON Powersaving via Sleep Mode, IEEE 802.3az Meet-ing, 2008.

[16] I. Cerutti, L. Valcarenghi, and P. Castoldi, “Power savingarchitectures for unidirectional WDM rings,” in OpticalFiber Communication (OFC), 2009, pp. 1–3.

[17] http://www.ieee802.org/3/efm/, IEEE 802.3ah Ethernet inthe First Mile Task Force.

[18] K. Glen, M. Biswanath, and P. Gerry, “IPACT: a dynamicprotocol for an Ethernet PON (EPON),” IEEE Communi-cation Magazine, vol. 40, pp. 74–80, 2002.

[19] http://www.opnet.com/, OPNET Modeler 14.5, 2005.

Ying Yan received the B.Eng. degree in electrical engineeringfrom the Beijing University of Technology, China, in 2002and the M.S. degree in electronics engineering in 2004 fromthe Technical University of Denmark. After completing theMaster degree, Ying Yan began to pursuit her Ph.D. studies ontelecommunication engineering in 2005. During 2006-2007, sheworked as a research scientist at the department of communi-cation platforms in the Technical Research Centre of Finland(VTT), Finland. While pursuing the Ph.D. study, Ying Yanparticipated in European projects (the IST-MUPBED projectand the ICT-ALPHA project) and a Danish national project (theHIPT project). Currently she is an Assistant Professor at theTechnical University of Denmark. Her research interests includescheduling algorithms, resource management mechanism, qual-ity of service, and energy efficiency in fixed wireless networksand Carrier Ethernet transport network.

Lars Dittmann was born in 1962 and received the M.Sc.EE and Ph.D. from the Technical University of Denmark in1988 and 1994, and is currently professor at the universitywithin the area of integrated networks. He has since January1999 been heading the network competence area (coveringboth optical and electrical networks) within DTU Fotonik atthe Technical University of Denmark and was prior to thatresponsible for electronic switching and ATM networks at theCenter for Broadband Telecommunication. Main research areais network architectures and performance.



Documents

Energy Efﬁciency in Ethernet Passive Optical Networks (EPONs): … · 2013. 4. 4. · Energy Efﬁciency in Ethernet Passive Optical Networks (EPONs): Protocol Design and Performance