Optical Switching and Networking 2 (2005) 137–147www.elsevier.com/locate/osn

Keeping the packet sequence in optical packet-switched networks

F. Callegatia,1, D. Caregliob,2, W. Cerronia,∗, G. Murettoa,1, C. Raffaellia,1,

na,

particularuaranteet sequenceults

and-e

J. Soĺe Paretab,2, P. Zaffonia,1

aDipartimento di Elettronica, Informatica e Sistemistica, University of Bologna, Viale Risorgimento, 2, 40136 Bologna, Italyb Advanced Broadband Communications Center (CCABA), Universitat Polit`ecnica de Catalunya, Jordi Girona, 1-3, Modul D6, 08034 Barcelo

Spain

Received 14 January 2005; accepted 23 September 2005Available online 7 November 2005

Abstract

This paper deals with optical packet switches with limited buffer capabilities, subject toasynchronous, variable-length packetsand connection-oriented operation. The focus is put on buffer scheduling policies and queuing performance evaluation. Ina combined use of the wavelength and time domain is exploited in order to obtain contention resolution algorithms that gthe sequence preservation of packets belonging to the same connection. Four simple algorithms for strict and loose packepreservation are proposed. Their performance is studied and compared with previously proposed algorithms. Simulation ressuggest that by accepting some additional processing effort it is possible to guarantee very low packet loss probabilities whileavoiding the out-of-sequence delivery.c© 2005 Elsevier B.V. All rights reserved.

Keywords: Optical packet switching; Asynchronous and variable-length packets; Connection-oriented operation; DWDM; Packet sequence

1. Introduction

Within the research framework on the next gener-

optical packet switching (OPS) should guaranteemuch finer and flexible access to the optical bawidth [3]. At the same time, the introduction of th

ation Internet, a widely discussed issue is the inte-

a-

DWDM technology allows us to exploit the huge capac-intud-ets,

wareiza-kts,syn-esti-ofiev-

.

gration of IP with optical networks [1]. Mainly, theproposals focus on the use of the wavelength ascircuit and therefore operate with a very coarse bandwidth granularity [2]. In a longer term perspective

∗ Corresponding author. Tel.: +39 051 2093089; fax: +39 0512093053.

E-mail addresses:fcallegati@deis.unibo.it (F. Callegati),careglio@ac.upc.es (D. Careglio), wcerroni@deis.unibo.it(W. Cerroni), gmuretto@deis.unibo.it (G. Muretto),craffaelli@deis.unibo.it (C. Raffaelli), pareta@ac.upc.es(J. Solé Pareta), pzaffoni@deis.unibo.it (P. Zaffoni).

1 Tel.: +39 051 2093089; fax: +39 051 2093053.2 Tel.: +34 93 4016985; fax: +34 93 4017055.

1573-4277/$ - see front matterc© 2005 Elsevier B.V. All rights reserveddoi:10.1016/j.osn.2005.09.001

ity of optical fibers and provides effective alternativesswitch design optimization. OPS has usually been sied with reference to fixed length synchronous packresulting in easier switching matrix design [4]. In thiscase, the drawbacks are both the increase in hardcomplexity necessary to manage optical synchrontion and the limited inter-working skill with networprotocols which mainly employ variable-length packesuchas IP. For these reasons, solutions adopting achronous and variable-length packets have been invgated recently [5], showing that in this case the issuecongestion resolution becomes fundamental to aching an acceptable level of performance.

http://www.elsevier.com/locate/osn

138 F. Callegati et al. / Optical Switching and Networking 2 (2005) 137–147

Contention resolution may be achieved in the timedomain by means of optical queuing and in thewavelength domain by means of suitable wavelengthmultiplexing. Optical queuing is realized by Fiber Delay

ingareletheiblel,s

dedceofto

ofofto

S

ed.esortityrked

newcehowrt,ce

asked.l.don

eded

i-er

and

maintain the forwarding table, while the forwardingcomponent examines the headers of incoming packetsand takes the forwarding decisions. Packets comingfrom client layers are classified into a finite number of

icenalonssameheirandtedup

),a

to

theicaltputto

heeeded

nghatard

mnd,ngand

eitlyvel

calbel

h-

tiontureve-

ical

Lines (FDLs) that are used to delay packets contendfor the same output fiber in case all wavelengthsbusy. In [6] it has been shown that, by using suitabcontention resolution algorithms able to combineuse of the time and the wavelength domain, it is possto improve the performanceup to an acceptable levewith a limited number of FDLs. We refer to them aWavelength and Delay Selection(WDS) algorithms.Moreover these concepts may be effectively extento a connection-oriented network scenario, for instanbased on MPLS. In this case, a suitable designdynamic allocation WDS algorithms permits usobtain fairly good performance, by exploiting queuingbehaviors related to the connection-oriented naturethe traffic, but with significant savings in termsprocessing effort for the switch control with respectthe connectionless case [7].

The main drawback of previously proposed WDalgorithms is thatout-of-sequence deliveryof packetsbelonging to the same traffic flow cannot be avoidThe occurrence of out of order delivery raisperformance problems for the end-to-end transpprotocols and/or issues of implementation complexif re-ordering at the edges of the optical netwoshould be implemented. This will be further discussin the following and is thebasic motivation of thispaper, where we address the issue and proposeWDS algorithms that guarantee the packet sequenpreservation. The results presented in the paper sthat, at the expense of some additional processing effothe new algorithms improve the overall performanwith respect to algorithms proposed in [7].

The paper is organized as follows. InSection 2thenetworking scenario is introduced and the general tof integration between MPLS and OPS is addressIn Section 3the WDS problem is discussed in detaiIn Section 4the out-of-order problem is defined andiscussed with special reference to its effectsthe end-to-end transport control protocols.Section 5is devoted to the description of the new proposalgorithms. InSection 6numerical results are presentand conclusions are drawn inSection 7.

2. Networking scenario: MPLS/OPS

The Multi-Protocol Label Switching (MPLS) archtecture [8] is based on a partition of the network layfunctions intocontrolandforwarding. The control com-ponent uses standard routing protocols to build up

subsets, calledForwarding Equivalent Classes(FECs),based on identification address and quality of servrequirements. Each FEC is identified by an additiolabel added to the packets. Unidirectional connectithroughout the network, calledLabel Switched Path(LSPs), are set up and packets belonging to the sFEC are forwarded along these LSPs according to tlabels. On each core node, simple label matchingswapping operations are performed on a precompuLSP forwarding table, thus simplifying and speedingthe forwarding function.

On the otherhand, optical packet switching (OPSin order to be feasible and effective, requiresfurther partitioning of the forwarding component inforwarding algorithm and switching [9]. The formercorresponds to the label matching that determinesnext hop destination and the latter is the physaction of transferring a datagram to the proper ouinterface. The main goal of this separation islimit the bottleneck of electro-optical conversions: theader is converted fromoptical to electrical and thexecution of the forwarding algorithm is performin electronics, while the payload is optically switchwithout electrical conversion.

This paper considers a DWDM network integratiMPLS and OPS, which relies on optical routers texploit the best of both electronics and optics. Standrouting protocols are used as the (non-critical) routingcomponent, MPLS labels in the forwarding algorith(where strict performance limits are present) afinally, optical technologies are used in switchiand transmission, providing very high data ratethroughput.

In order to avoid scalability problems, we assumthat each LSP represents a top-level explicrouted path, formed by an aggregation of lower-leconnections including several traffic flows [10], andthat the number of LSPs managed by a single opticore router is not so high as to affect the correct laprocessing.

We also assume the availability of an optical switcing matrix able to switch variable-length packets [11].This paper is not supposed to deal with implementaissues. Therefore, a generic non-blocking architecfor an OPS node is assumed, which provides full walength conversion.

The switch is assumed to use a feed-forward optbuffering configuration [12], realized by means ofB

F. Callegati et al. / Optical Switching and Networking 2 (2005) 137–147 139

fiber delay lines. Basically, ais assumed: each waveleits own buffer. This resultsinwavelength. In principle, a

ding buffer is assigned. Thea given wavelength can besmallest integer greater thanbetween the time when the

Fig. 1. Degenerate buffer structure.

noutput queuing approachngth on each output fiber hasa logical queue per output

set of delay lines per output

packet on the corresponsmallest delay available oneasily computed using theor equal to the difference

senre

shenceputthe

illrly

bythe

d alntonthngthlarofisntheicy,ued

valon

the

tedt

urce

lthee oftputby

is

edutare

the

aess

in. It

cket

ich

onimeueisnack

wavelength could be deployed, leading to a fairly largeamount of fiber coils. However, a single pool of FDLmay be used in WDM for all the wavelengths on a givfiber or for the whole switch. The delays provided alinearly increasing with a basic delay unitD, i.e. Dj =kj D, with kj an integer number andj = 1, 2, . . . , B.For the sakeof simplicity we assume adegeneratebuffer [12] with kj = j − 1 (seeFig. 1), but the ideaspresented here are valid also fornon-degenerate buffers,i.e. with different arrangements ofkj .

3. The wavelength and delay selection problem

The electronic Switch Control Logic (SCL) takeall the decisions regarding the configuration of thardware to realize the proper switching actions. Othe forwarding component has decided to which outfiber the packet should be sent, determining alsonetwork path, the SCL:

• chooses which wavelength of the output fiber wbe used to transmit the packet, in order to propecontrol the output interface;

• decides whether the packet has to be delayedusing the FDLs or it has to be dropped sincerequired queuing resource is full.

These decisions are also routing independent anthe wavelengths of a given output fiber are equivalefor routing purposes but not from the contentiresolution point of view. The choices of wavelengand delay are actually correlated: since each wavelehas its own logical output buffer, choosing a particuwavelength is equivalent to assigning the packet onethe available delays on the corresponding buffer. This what we call theWavelength and Delay Selectio(WDS) problem. Here we assume that, oncewavelength has been chosen using a particular polalways the smallest delay available after the last que

l

wavelength will be available again and the packet arritime, divided by the buffer delay unit. This operatiprovides also the gap between the current andprevious queued packet.

The choice of the wavelength can be implemenby following different policies, producing differenprocessing loads at the SCL and different resoutilizations:

• Static. The LSP is assigned to a wavelength at LSPsetup and this assignment is kept constant over althe LSPlif etime. Therefore packets belonging tosame LSP are always carried by the same couplinput/output wavelengths. Contention on the ouwavelength can only be solved in time domainusing delay lines. In this case the WDS algorithmtrivial and requires minimum control complexity. Onthe other hand resource utilization is not optimizsince it is possible to have a wavelength of an outpfiber congested even if the other wavelengthsidle.

• Dynamic. The LSP is assigned to a wavelength atLSP setup but the wavelength may change duringLSP lifetime. Two approaches are possible:– per-packet. The wavelength is selected on

per-packet basis, similarly to the connectionlcase, choosing the most effective wavelengththe perspective of optimizing resource usagerequires some processing effort on a per-pabasis, therefore this alternative is fairly demandingin terms of processing load on the SCL, whmust be carefully dimensioned;

– per-LSP. When heavy congestion arisesthe assigned wavelength, i.e. when the tdomain is not enough to solve contention dto the lack of buffering space, the LSPtemporary movedto another wavelength. Whecongestion disappears, the LSP is switched b

140 F. Callegati et al. / Optical Switching and Networking 2 (2005) 137–147

to the original wavelength. This alternative stayssomewhat in between, aiming at realizing a trade-off between control complexity and performance

It is obvious that the per-packet alternative is theheys,

cancket

gtherity

heapketapssiblere:

thatlso

itchtionioneof

oreitha

putme

lyrenesased

on.

onn.P

thatstede),l

allocation, which has an empty queue. If such awavelength is found, the LSP is switched to it for aslong as needed, that is until congestion arises also inthe new wavelength or congestion ends on the previous

theal

d to

der-to-put

fket-tth-

grid

m-ha-n an-

orisremsanfy aableedut-ntsssing

etsy-c./or

e

theIn

thationhen

most flexible. In [5], the author observes that, since tFDL buffers are only able to provide discrete delathis creates gaps between queued packets thatbe considered equivalent to an increase of the paservice time, meaning an artificial increase in the trafficload (excess load). It has been demonstrated in [13]that a void filling algorithm that aims at minimizinthose gaps gives best performance with respect to opolicies. Nonetheless, the computational complexof such an algorithm is high, since it requires tknowledge of the length and duration of every gin the queues. Therefore, in order to keep the pacscheduling process as simple as possible, the gbetween queued packets are not considered as posbuffer places and a simplified approach is adopted hthe so-called MINGAP algorithm, proposed in [6]. It isa per-packet strategy selecting the wavelength suchthe corresponding smallest delay available provides athe smallest gap after the packet previously scheduled.

On the other hand, in [7] it has been originallyobserved that by adopting a per-LSP strategy the swperformance in general depends on the configuraof the LSP forwarding table. The basic observatis that packets following LSPs incoming on thsame input wavelength cannot overlap, becausethe serial nature of the transmission line. Therefsuch packets contend for output resources only wpackets incoming on different wavelengths. Asconsequence, if LSPs incoming on the same inwavelength are the only ones forwarded to the saoutput wavelength, contention will never arise (optimalallocation). This is just a trivial example but obviousa whole spectrum of combinations is possible. Mogenerally this effect produces a negative correlatioin the traffic profile, which is smoothed and thereforis less aggressive in terms of queuing requirementbecause it decreases the likelihood of congestion. Bon these considerations, in [7] some WDSalgorithmsare proposed that try to exploit this negative correlatiThe best performing of these algorithms is calledEmptyQueue Wavelength Selection(EQWS): it is a per-LSPstrategy trying to exploit the queuing space availableoutput wavelengths in the optimal allocation conditioIn normal conditions, packets following a given LSare transmitted on the wavelength assigned toLSP. However, as soon as a packet finds a congewavelength (i.e. a wavelength with no delays availablthe algorithm looks for a wavelength in optima

e

one. In this case the LSP is switched back tooriginal wavelength. In case a wavelength in optimallocation cannot be found, the LSP is re-assignethe wavelength with the smallest delay available.

4. Problems due to out-of-order packet delivery

As already outlined in the introduction, it is wellknown that packet losses as well as out-of-orpacket deliveries and delay variations affect endend protocols behavior and may cause throughimpairments [14,15]. In particular, the problem opacket re-sequencing is not negligible in optical pacswitched networks, especially when optical packeflows carrying traffic related to emerging, bandwiddemanding, sequence-sensitive services, such asapplications and storage services, are considered.

Whenconsidering TCP-based traffic these phenoena influence the typical congestion control mecnisms adopted by the protocol and may result ireduction of the transmission window size with cosequent bandwidth under-utilization. In particular theTCP congestion control is very affected by the losstheout-of-order delivery of bursts of segments. Thisexactly what may happen in the OPS network whetraffic is typically groomed and several IP datagra(and therefore TCP segments) are multiplexed inoptical packet, because optical packets must satisminimum length requirement to guarantee a reasonswitching efficiency. Therefore out-of-order or delaydelivery of just one optical packet may result in oof-order or delayed delivery of several TCP segmetriggering (multiple duplicate ACKS and/or timeoutthat expire) congestion control mechanisms and cauunnecessary reduction of the window size.

Another example of how out-of-sequence packmay affect application performance is the case of delasensitive UDP-based traffic, such as real-time traffiIn fact unordered packets may arrive too late andthe delay required to reorderseveralout-of-sequencpackets may be too high with respect to the timingrequirements of the application.

These brief and simple examples make evidentneed to limit the number of unordered packets.general out-of-order delivery is caused by the factpackets belonging to the same flow of informatcan take different paths through the network and tcan experience different delays [16]. In traditional

F. Callegati et al. / Optical Switching and Networking 2 (2005) 137–147 141

connection-oriented networks, packet reordering is notan issue since packets belonging to the same connectionare supposed to follow the same virtual network pathand therefore are delivered in the correct sequence,

forhisn aey

venth,se and,

a,red

ketveork

al

ra-is

re-

es.e-penthengs isltther

0.8uttsmeby

andalso

rseorkPnce

ence

Table 1Out-of-order percentages at the input and output ports of the coreswitch, comparing the static, MINGAP, and EQWS algorithms

Algorithm Input Output

he

ase

hatlly

hisouldisbeistndwe

of-rk

emketm

mre

of

of

of

cultof

unless packet loss occurs.In an OPS network, using the wavelength domain

contention resolution (i.e. using dynamic policies), tmay notbethe case. Optical packets are transmitted ogiven wavelength depending on the flow (i.e. LSP) thbelong to. When a static WDS policy is adopted, a giflow is always transmitted on the same wavelengso the order of packets cannot be changed, becauFIFO queuing discipline is assumed. On the other hawhen a dynamic WDS policy is used, packets fromgiven flow may be transmitted on different wavelengthsexperience different delays and, eventually, be delivenot in the correct sequence.

To check the likelihood of out-of-sequence pacdelivery when the WDS algorithms mentioned aboare implemented, we set up the two hops netwscenario shown inFig. 2. Every switch is identi-cally set up with 16 wavelengths per link, an opticbuffer of B = 4 FDLs and a granularity equal to theaverage packet length. The inter-arrival packet genetion follows a Poisson model while the packet sizeexponentially distributed with an average value corsponding to a transmission time of the order of 1µs, atypical value for optical packet switching technologiWe focus on packet size only in terms of duration bcause the simulators used here are built to be indedent from the packet size in terms of bytes and frombit-rate. Since one of the benefits of optical switchiis the transparency to the bit-rate, what really matterindeed the average packet duration and the inter-arrivatime distribution. Obviously, once the optical packeduration is set, the higher the bit-rate, the higheraverage packet size in bits, leading to the need fotraffic grooming.

The input load on each wavelength is fixed toand the traffic distribution is uniform. The traffic inpat the edge switches is idealin the sense that packebelonging to the same LSP arrive in order on the sawavelength. At each switch, packets are processedthe selected WDS algorithm, sent to the next hopthe resulting amount of out-of-sequence packets isevaluated.

Table 1shows thepacket reordering distribution fothe static, MINGAP, and EQWS algorithms. Theresults, even though related to a simple netwarchitecture, are meaningful to show that MINGAand EQWS algorithms are not able to avoid sequebreaking. The percentage of packets out-of-sequ

-

Static 0 0MINGAP 3.6498 6.9948EQWS 1.6798 5.4200

of three or more locations is already not null at tinput of the core switch. By assumingn switch in seriesalong a path this percentage is expected to increaccordingly. Previous studies [14,15] confirm that justa small percentage of out-of-sequence (such as tcaused by the EQWS algorithm) may impact harmfuon the network performance.

A possible solution could be to assume that tproblem is solved at the egress edge nodes that shtake care of re-sequencing the various packet flows. Thassumption in our view is not very realistic. It canfeasible for some flow of high value traffic, but itunlikely that it will happen for all the flows of beseffort traffic, because of the amount of memory aprocessing effort that would be necessary. Thereforeargue that it becomes fundamental to control out-order delivery of packets directly in the OPS netwonodes.

5. WDS algorithms preserving the packet sequence

The way to overcome the aforementioned problis to design WDS algorithms able to preserve the pacsequence from the beginning. When facing this problein a WDM network, it is easy to realize that the ter“out-of-sequence delivery” must be defined in modetail.

In this paper we use two possible definitionsordered packets:

• Strict sequence preservation. Given a stream ofordered packets at the switch input, packetn isconsidered to be out-of-order when the first bitpacket n leaves the switch before thelast bit ofpacketn − 1.

• Loose sequence preservation. Given a stream ofordered packets at the switch input, packetn isconsidered to be out-of-order when the first bitpacketn leaves the switch before thefirst bit ofpacketn − 1.Indeed the strict sequence preservation is the ideal

case. The loose sequence preservation is more diffito control, especially when considering a cascade

142 F. Callegati et al. / Optical Switching and Networking 2 (2005) 137–147

ets.

switches and, most of aof-sequence arrivals atdue to the fact that an oaggregate of several IP d

rategy: packets following aon the wavelength assigned

e corresponding buffer is not

Fig. 2. Network scenario to evaluate the amount of out-of-order pack

ll, may still cause limited out-the edge nodes. This may bepticalpacket is, in general, an

the following per-LSP stgiven LSP are transmittedto that LSP as long as th

atagrams and hence the relativewoomenceDS

heource

the

t

t

ricte

apvet

full. When congestion arises, the algorithm looks for aketare

withthePS-

ithm

scketat

the

ce is

yual

ere

t

position of subsequent IP packets included in tsubsequent optical packets cannot be controlled if soverlapping is permitted. Nonetheless looser sequepreservation imposes weaker constraints on the Walgorithm, which is expected to perform better. In tfollowing subsections we describe the proposal of fsimple algorithms for strict and loose packet sequenpreservation.

To describe the algorithms, let us introducefollowing notations:

• tn: the arrival time of the generic packetn;• dn: thedelay assigned to the generic packetn;• ln: theduration of the generic packetn;• an: the time at which the first bit of the generic packe

n is scheduled to leave the switch;• bn: the time at which the last bit of the generic packe

n is scheduled to leave the switch. It is obvious thatbn = an + ln andan = tn + dn.

5.1. Strict packet sequence

The WDS algorithms designed to keep a stpacket sequence are calledStrict Packet Sequencwith Minimum Length wavelength selection(SPS-ML) and Strict Packet Sequence with Minimum Gwavelength selection(SPS-MG). These algorithms habeen originally proposed in [17]. Basically, they adop

wavelength that can provide a delay such that the pacsequence is not broken. In case multiple choicesavailable, the options are to choose the wavelengththe smallest delay available (SPS-ML) or providingsmallest gap after the previous queued packet (SMG).

In particular, when packetn arrives at timetn, inorder to preserve the packet sequence, the algorselects a queue depending on the timebn−1 at whichthe last bit of packetn − 1, following the same LSP, ischeduled to leave the node. Assuming ordered pastreams arriving at the switch input, it is obvious thtn ≥ tn−1 + ln−1, but if packetn − 1 is queued, thendn−1 > 0 and there aretwo possible alternatives:

1. tn < bn−1: packetn − 1 wasqueued and packen overlaps withit. There is the chance to break tpacket sequence.

2. tn ≥ bn−1: packetn arrives when or after packetn−1has completely left the node. The packet sequenalways guaranteed.

In the first case, packetn has to be delayed ban amount of time at least as long as the residtransmission time of packetn − 1. Due to the discretnumber of delays provided by the FDLs, which amultiples of the basic unitD, the minimum delay tha

F. Callegati et al. / Optical Switching and Networking 2 (2005) 137–147 143

may be assigned to packetn is:

Dmin =⌈

bn−1 − tnD

⌉D. (1)

ondak

ehishiserebe

ensseetheend

ationfor

red

otenuessenter

rewith

PSof

,lyevether’s

sede.

. In

ms

lay

ior.

lledave-ce

gver-to

ets

the

woketket

,

he

On the other hand, the minimum delay for the seccase isDmin = 0 because there is no possibility to brethe correct packet sequence.

OnceDmin is determined, the algorithm verifies if thwavelength assigned to the LSP is able to provide tdelay. If not, it searches for another wavelength. Tsearch may give multiple results, meaning that thare several wavelengths on which the packet maytransmitted with the required delay. When this happthe WDS algorithm must choose one among thequeues. This choice is independent of the sequencissue that has already been solved by choosingproper delay. Therefore the choice of the queue to sthe packet can be made based on general optimizprocedures similar to what has already been studiedthe pure connectionless case [6]. Here we follow thesame approach and two algorithms are then considethat adopt different choices:

• SPS-MG selects the queue among those nfull which introduces the minimum gap betwesubsequent queued packets. If two or more queprovide the same gap, the shortest one is choi.e. the queue providing the smallest delay greathanDmin.

• SPS-ML selects the shortest queue. If two or moqueues have the same minimum length, the onethe smallest gap is chosen.

It has to be underlined that to execute the Salgorithms the SCL needs to store the last valuebn for each LSP in the forwarding table. Nonethelesswith respect to previous algorithms, the extra cost onstays in the memory requirements, because whatalgorithm is used, the SCL must always calculatevalue ofbn for each queued packet to know the buffeoccupancy.

Fig. 3 illustrates an example of the behavior of thealgorithms. InFig. 3(a), packetn arrives at the node anpacketn − 1 on the same LSP is still in the queuTherefore, packetn overlaps withit and theminimumassignable delay isDmin = 3. Both algorithms selectthe second queue because it provides this delayFig. 3(b), packetn arrives when packetn − 1 hascompletely left the node. In this case, the algorithbehave differently: with SPS-MG, packetn is put inthe second queue (minimum gap) with delay 2D; whilewith SPS-ML, the third queue (shortest one) and deD is selected.

,

r

Fig. 3. Example of the SPS-ML and SPS-MG algorithms’ behav(a) Packetn − 1 is still in the queue. (b) Packetn − 1 hasalready leftthe queue.

5.2. Loose packet sequence

The algorithms proposed for this case are caLoose Packet Sequence with Minimum Length wlength selection(LPS-ML) andLoose Packet Sequenwith Minimum Gap wavelength selection(LPS-MG)and have been originally proposed in [18]. Theirbehavior is the same as SPSpolicies, with the onlydifference that the delay requirements for not breakinthe sequence are less strict and allow partial packet olapping. LPS-ML and LPS-MG are therefore similarSPS-ML and SPS-MG.

In particular, by considering the stream of packbelonging to the same LSP, now packetn is consideredto be out-of-sequence when its first bit is sent onoutput wavelength before the first bit of packetn − 1.This definition allows a partial overlapping between tconsecutive packets. In practice if the first bit of pacn − 1 has been already sent when the first bit of pacn arrives, then the SCL schedules packetn withoutbothering about the sequence problem.

Following the same notation of the previous sectionagain we have two possibilities:

• tn < an−1: then there is the chance to break tpacket sequence definition.

144 F. Callegati et al. / Optical Switching and Networking 2 (2005) 137–147

or.

ket

lay

n

ogth

avetsg aikeescts.

d bhasS

algorithms. The simulator implements a 4× 4 switchwith 16 wavelengths per fiber, which results in a 64×64 optical switching matrix. Each input wavelengthis supposed to carry 3 different LSPs, for a total of

ithing

heTwocalelaral

anlue

r ofallyto

ncence

heend

icfor

all,annce

ra

etens.gh

icheith

rvecaltchket

ave

oth

m

Fig. 4. Example of the LPS-ML and LPS-MG algorithms’ behavi(a) Packetn − 1 is still in the queue. (b) Packetn − 1 is leaving thequeue.

• tn ≥ an−1: packetn arrives when packetn − 1 isleaving or has already left the node. Then the pacsequence is guaranteed.

In the first case, the amount of time needed to depacketn in order to keep the sequence is:

Dmin =⌈

an−1 − tnD

⌉D. (2)

Fig. 4 shows how LPS-ML and LPS-MG work. Ithe first case (Fig. 4(a)) packetn− 1 is still in the queueand packetn has to be delayed by three granularities tbe brought back in order. Only the second wavelencan provide this delay and both algorithms behsimilarly. In the second case, the first bit of packen − 1 is leaving the switch. In this case the algorithmdo not have constraints apart from that of choosinfeasible delay. For this choice LPS-MG behaves lSPS-MG and looks for the wavelength that minimizthe gap while LPS-ML resembles SPS-ML and selethe wavelength that minimizes the length of the queue

6. Numerical results

The results presented here have been obtainemeans of an ad hoc event-driven simulator, whichbeen widely tested in the past with different WD

y

192 incoming LSPs. As already outlined inSection 4,the simulator is bit-rate transparent and works wpacket transmission and inter-arrival times only, makit scalable to different bit-rates and packet sizes.

The input traffic is generated according to tappropriate statistics on a per wavelength basis.types of input traffic are considered: a classiuniform traffic with Poisson distribution of thinter-arrival times and a more realistic self-simimodel implemented with 32 multiplexed point arrivprocesses having Pareto distribution [19]. We assumeasynchronous, variable-length optical packets, withexponential distributed packet size with average vacorresponding to a transmission time of the orde1 µs.The input load on each wavelength is 0.8, equdivided among the LSPs and the traffic distributiontheoutputs of the switching matrix is uniform.

In this simulation study we compare the performaof the new algorithms oriented to the packet sequewith the static, the EQWS [7], and the MINGAPalgorithms [6]. It is reasonable to forecast that tnew algorithms will show intermediate performancbetween the connectionless case (i.e. MINGAP) athe worst case (i.e. static). This is because the statalgorithm does not use the wavelength domainstatistical multiplexing of contending packets atwhile the new ones do, but with less freedom thMINGAP, due to the constraints on the packet sequepreservation.

Fig. 5 plots the packet loss probability (PLP) foSPS-ML, SPS-MG and other WDS policies asfunction of D normalized to the average packduration, withB = 4 FDLs. Confidence intervals arnot shown for all the curves for readability reasoHowever, simulations have been carried on long enouin order to typically generate a billion packets, whmakes loss measures up to 10−6 quite accurate. To provthis, the curve related to SPS-ML has been plotted wconfidence intervals. It should be noticed that cufluctuations are not caused by insufficient statistidata, but are most probably due to the mismabetween FDL buffer granularity and average pacsize.

As expected the PLP shows a typical concbehavior as a function ofD, with an optimal valuethat depends on the algorithm. It can be seen that bSPS algorithmssignificantly improve theperformanceof the static allocation, which is the only other algorith

F. Callegati et al. / Optical Switching and Networking 2 (2005) 137–147 145

eythend

theeue

evetionanda

tomor

thatPS-s a

tleuchtedncehmsthiswo

aseforesesch

nce

at,therngterssgth

heSPSy asding.itch

nceed

Fig. 5. Packet loss probability as a function ofD, comparing the SPSalgorithms with other algorithms under Poisson traffic model.

that avoids the out-of-sequence delivery. Moreover, thalso improve the overall performance with respect toEQWS, being closer to MINGAP. While SPS-MG aSPS-ML show the same behavior for small values ofD,SPS-ML performs slightly better than SPS-MG whenDis large. This is because by choosing the queue withsmallest gap, SPS-MG may buffer a packet in a quthat will become empty later (higherDmin) than usingSPS-ML (as in the case ofFig. 3). This means that thfollowing packet belonging to the same LSP will halessqueuing resources available due to the preservaof the packet sequence constraint. On the other hSPS-ML, by choosing the shortest queue, providesbetter utilization of the buffering space since it trieskeep the queues as short as possible, thus leavingroom for following packets.

Fig. 6 compares the performance of the stricsequence algorithms (SPS-ML and SPS-MG) with tof the loose packet sequence ones (LPS-ML and LMG). Again the graph shows the performance afunction of the granularityD. We see that by allowinga partial overlapping between the packets a litfurther improvement can be obtained. Nonetheless simprovement is rather small, especially when weighagainst the possible drawbacks of this solution. Hewe conclude that the loose packet sequence algoritare less appealing than SPS-ML and SPS-MG. Forreason we focus the following results on these talgorithms only.

Fig. 7 shows that a significant improvement of theperformance may be obtained with a small increof the number FDLs in the buffer. For instance,SPS algorithmsadding 2–4 FDLs, the PLP decreasby two orders of magnitude. In this figure the valuof D are those providing the lowest PLP for ea

e

Fig. 6. Packet loss probability as a function ofD, comparing the strictsequence definition (SPS-ML and SPS-MG) with the loose sequedefinition (LPS-ML and LPS-MG) under Poisson traffic model.

Fig. 7. Packet loss probability as a function ofB, comparing the SPSalgorithms with other algorithms under Poisson traffic model.

algorithm. It is interesting to see from this graph thas expected, the SPS algorithms outperform the oconnection-oriented algorithms in spite of providimore functionality. The only case that provides a betPLP is the MINGAP case that, due to its connectionlebehavior, has more freedom to exploit the wavelendomain for statistical multiplexing.

Table 2 compares the algorithms in terms of tpercentage of strict out-of-sequence packets.algorithms guarantee the correct sequence deliverwell as the static allocation, while both MINGAP anEQWS are not able to avoid the sequence breakMoreover, here we have considered a single swscenario; by assumingn switches in series along a path,the percentages can increase accordingly.

The price to pay to keep the correct sequedelivery is the additional processing effort need

146 F. Callegati et al. / Optical Switching and Networking 2 (2005) 137–147

Table 2Out-of-sequence percentages

Algorithm Out-of-sequence (%)

onr

eoficy.retket

c

inghall

annong

--egth

not

l,

S-etsed.rictLSPcketheied

sedgolehehe

tchms

nrk

ishct

,

Static 0EQWS 3.347MINGAP 4.018SPS-ML 0SPS-MG 0

Fig. 8. Percentage of LSP-to-wavelength reassignments as a functiof D, comparing the SPS algorithmswith other algorithms undePoisson traffic model.

by SPS algorithms that reallocate the LSPs fairlymore often than EQWS. In this sense, the percentagof LSP-to-wavelength reallocations is a measurethe processing effort that the switch control logmust deal with when applying a given WDS policFig. 8 shows that the static algorithm does not requiany reassignments, while the SPS algorithms presenintermediate values between MINGAP (a per-pacstrategy) and EQWS.

Finally, in Fig. 9 the PLP as a function ofD withB = 4 is plotted considering a self-similar traffimodel with Hurst parameterH = 0.9. This figurecompares the SPS algorithms with the best performalgorithm (MINGAP) and the EQWS algorithm whicneeds the lowest processing effort. As expected,algorithms perform worse under self-similar traffic thusing the Poisson model;the PLP increases by aorder of magnitude. Nevertheless, the relation amthe algorithms remains the same as inFig. 5.

7. Conclusions

In this paper we have analyzed the contention resolution problem in optical packet switches with MPLSlike connection-oriented operation. In particular, whave pointed out that previously proposed wavelen

Fig. 9. Packet loss probability as a function ofD, comparing the SPSalgorithms with other algorithms under self-similar traffic model.

and delay selection algorithms (such as EQWS) domaintain the sequence of thepacket flows neither at aglobal switch level nor at the single traffic flow leveexcept for the static allocation solution.

Four novel dynamic algorithms (SPS-ML, SPMG, LPS-ML, and LPS-MG) that maintain the packsequence within the same LSP have been propoTwo different strategies have been pursued: stpacket sequence where packets following the samecannot overlap among themselves, and loose pasequence where partial overlapping is allowed. Tperformance of these algorithms has been studand compared with respect to previously propoalgorithms. Simulation results show that by acceptinsome additional processing effort it is possible tguarantee very low packet loss probabilities whiavoiding the undesired out-of-sequence delivery. Tresults also show that the new algorithms improve tperformance compared to EQWS.

In this paper we have considered a single swiscenario, letting the case of evaluating the algorithin a whole network context for future studies.

Acknowledgments

This work was partially funded by the Commissioof the European Union through the IST-FP6 Netwoof Excellence e-Photon/ONe and by MCYT (SpanMinistry of Science and Technology) under contraFEDER-TIC2002-04344-C02-02.

References

[1] N. Ghani, S. Dixit, T.-S. Wang, On IP-over-WDM integrationIEEE Communications Magazine 38 (3) (2000) 72–84.

F. Callegati et al. / Optical Switching and Networking 2 (2005) 137–147 147

[2] E. Mannie (Ed.), Generalized Multi-Protocol Label Switching(GMPLS) Architecture,IETF, RFC3945, October 2004.

[3] T.S. El-Bawab, J.-D. Shin, Optical packet switching in corenetworks: between vision and reality, IEEE CommunicationMagazine 40 (9) (2002) 60–65.

orkEE

98)

E

ofl

ices,

el

thernal

t,

ousit-e,

[12] D.K. Hunter, M.C. Chia, I. Andonovic, Buffering in opticalpacket switches, IEEE/OSA Journal of Lightwave Technology16 (12) (1998) 2081–2094.

[13] L. Tancěvski, S. Yegnanarayanan, G. Castanon, L. Tamil,F. Masetti, T. McDermott, Optical routing of asynchronous,

in

ey,in a

a5)

t a

et-n04,

et

3,

ffic,ndpp.

[4] P. Gambiniet al., Transparent optical packet switching: netwarchitecture and demonstrators in the KEOPS project, IEJournal on Selected Areas in Communications 16 (7) (191245–1259.

[5] F. Callegati, Optical buffers for variable length packets, IEECommunications Letters 4 (9) (2000) 292–294.

[6] F. Callegati, W. Cerroni, G. Corazza, Optimizationwavelength allocation in WDM optical buffers, OpticaNetworks Magazine 2 (6) (2001) 66–72.

[7] F. Callegati, W. Cerroni, C. Raffaelli, P. Zaffoni, Dynamwavelength assignment in MPLS optical packet switchOptical Network Magazine 5 (5) (2003) 41–51.

[8] E. Rosen, A. Viswanathan, R. Callon, Multiprotocol LabSwitching Architecture, IETF, RFC 3031, January 2001.

[9] C. Guillemot et al., Transparent optical packet switching:European ACTS KEOPS project approach, IEEE/OSA Jouof Lightwave Technology 16 (12) (1998) 2117–2134.

[10] K. Kompella, Y. Rekhter, LSP Hierarchy with GeneralizedMPLS TE, draft-ietf-mpls-lsp-hierarchy-08.txt, IETF drafSeptember 2002.

[11] D. Chiaroni et al., First demonstration of an asynchronoptical packet switching matrix prototype for multiterabclass routers/switches, in: Proc. of 27th European Conferencon Optical Communications, ECOC 2001, October 2001Amsterdam, The Netherlands, vol. 6, 2001, pp. 60–61.

variable length packets, IEEE Journal on Selected AreasCommunications 18 (10) (2000) 2084–2093.

[14] S. Jaiswal, G. Iannacone, C. Diot, J. Kurose, D. TowslMeasurement and classification of out-of-sequence packetstier-1 IP backbone, in: Proc. of INFOCOM 2003, March 2003,San Francisco, CA, vol. 2, 2003, pp. 1199–1209.

[15] M. Laor, L. Gendel, The effect of packet reordering inbackbone link on application throughput, IEEE Network 16 ((2002) 28–36.

[16] J.C.R. Bennett, C. Patridge, Packet reordering is nopathological network behavior, IEEE/ACM Transactions onNetworking 7 (6) (1999) 789–798.

[17] F. Callegati, D. Careglio, W. Cerroni, J. Sol´e-Pareta, C. Raffaelli,P. Zaffoni, Keeping the packet sequence in optical packswitched networks,in: Proc. of 9th European Conference oNetworks and Optical Communications, NOC 2004, June 20Eindhoven, The Netherlands, 2004, pp. 443–450.

[18] G. Muretto, C. Raffaelli, P. Zaffoni, Evaluation of packreordering in optical packet networks with dynamic algorithms,in: Proc. of Photonics in Switching, PS 2003, September 200Paris, France, 2003, pp. 77–79.

[19] J.J. Gordon, Long range correlation in multiplexed Pareto train: Proc. ofInternational IFIP-IEEE Conference on BroadbaCommunications, April 1996, Montreal, Canada, 1996,28–39.

Keeping the packet sequence in optical packet-switched networksIntroductionNetworking scenario: MPLS/OPSThe wavelength and delay selection problemProblems due to out-of-order packet deliveryWDS algorithms preserving the packet sequenceStrict packet sequenceLoose packet sequence

Numerical resultsConclusionsAcknowledgmentsReferences