11 Traffic Managementfor IP overWDM Networks

JAVIER ARACIL, DANIEL MORATO�

andMIKEL IZAL

PublicUniversityof Navarra,Spain

11.1 INTRODUCTION

The ever-increasinggrowth of the current Internet traffic volume is demandingabandwidthwhich canonly be satisfiedby meansof optical technology. However,while WDM networkswill bring anunprecedentedamountof availablebandwidth,traffic managementtechniquesbecomenecessaryin order to make this bandwidthreadily availableto the enduser. In this chapterwe provide an overview of trafficmanagementtechniquesfor IP over WDM networks. More specifically, we addressthe following issues: i) Why is traffic managementnecessaryin IP over WDMnetworks? andii) What arethe traffic managementtechniquescurrentlyproposedfor IP overWDM?

Concerningthe first issue,we provide an intuitive, ratherthan mathematicallyconcise,explanationaboutthespecificfeaturesof Internettraffic thatmakeIP trafficmanagementparticularlychallenging.Internettraffic shows self-similarity featuresand thereis a lack of practicalnetwork dimensioningrules for IP networks. Onthe other hand, the traffic demandsare highly non-stationaryand, therefore,thetraffic demandmatrix,which is theinput for any traffic groomingalgorithm,is time-dependent.Furthermore,traffic sourcescannotbeassumedtobehomogeneous,sinceveryfew usersproducealargefractionof traffic. In dynamicIP overWDM networkswe experiencethe problemof the multiservicenatureof IP traffic. Sincea myriadof new servicesandprotocolsarebeing introducedon a day-by-daybasisit turnsout that the provision of quality of serviceon demandbecomescomplicated. Weanalyzetheserviceandprotocolbreakdown of a real Internetlink in orderto showthecomplexity of thedynamicQoSallocationproblem.

Regardingexisting andproposedtraffic managementtechniqueswe distinguishbetweenstaticanddynamicWDM networks.StaticWDM networksaredimensioned

�Presentlyon leave at Universityof California,Berkeley

i

ii TRAFFIC MAN AGEMENT FOR IP OVER WDM NETWORKS

beforehandand the dynamicbandwidthallocationcapabilitiesare very restricted.Suchstaticbandwidthis not well suitedto the burstynatureof Internettraffic and,asa result,traffic peaksrequirebuffering, which is difficult to realizein theopticaldomain. Deflection routing techniques,using the sparebandwidthprovided byoverflow lightpaths,have beenproposedas a techniqueto minimize packet lossand reduceoptical buffering to a minimum. We investigatethe advantagesanddisadvantagesof suchtechniques,with emphasison the particular featuresof IPoverflow traffic. In thedynamicWDM networksscenariowestudytheOpticalBurstSwitchingparadigm[31] asa meansto incorporatecoarsepacket switchingserviceto theIP overWDM networks.

To endup the chapterwe focuson end-to-endissuesto provide QoSto the enduserapplications.Dueto theavailability of Tbpsin asinglefibertheTCPneedsto beadaptedto thishigh-speedscenario.Specifically, weanalyzetheTCPextensionsforhigh-speed,togetherwith split TCPconnectionstechniques,asa solutionto provideQoSto theenduserin a heterogeneousaccess-backbonenetwork scenario.

11.1.1 Network scenario

Along thechapterwe assumethefollowing threewavesin thedeploymentof WDMtechnology: i) First generationor static lightpathnetworks, ii) Secondgenerationor dynamiclightpath/OpticalBurstSwitchingnetworksandiii) Third generationorPhotonicPacketSwitchingnetworks.

StaticWDM backboneswill provide point-to-pointwavelengthspeedchannels(lightpaths). Suchstatic lightpathswill be linking gigabit routers,ascurrentATMor FrameRelaypermanentvirtual circuitsdo. Suchgigabit routersperformcell orpacket forwardingin theelectronicdomain.

In asecondgenerationWDM network, theWDM layerprovidesdynamicalloca-tion features,by offeringon-demandlightpathsor coarsepacketswitchingsolutions.The former providesa switchedpoint-to-pointconnectionservicewhile the latterprovides burst switching service. Precisely, Optical Burst Switching (OBS) [31]providesa transfermodewhich is halfway betweencircuit andpacket switching.In OBS,a reservationmessageis sentbeforehandsothat resourcesarereservedforthe incomingburst,which carriesseveralpacketsto thesamedestination.In doingso, a singlesignalingmessageservesto transferseveral packets, thusmaximizingtransmissionefficiency andavoidingcircuit setupoverhead.

Thethird generationof opticalnetworkswill provide photonicpacket switching,thus eliminating the electronicbottleneck. As far as the optical transmissionisconcerned,thechallengeis to provideultra-narrow opticaltransmittersandreceivers.Evenmorechallengingis thedevelopmentof all-opticalroutersthatperformpacketheaderprocessingin the optical domain. In fact, while packet headerprocessingis very likely to remainin theelectronicdomainall-opticalpacket routersbasedonopticalcodesarecurrentlyunderdevelopment[38].

TRAFFIC MAN AGEMENT IN IP NETWORKS: WHY? iii

11.2 TRAFFIC MANAGEMENT IN IP NETWORKS: WHY?

Thereis awell-establishedtheoryfor dimensioningtelephonenetworks[9] basedonthe hypothesisof the call arrival processbeingPoissonandthe call durationbeingwell-modeledasanexponentialrandomvariable. A blockingprobability objectivecanbesetand,by meansof theErlangiantheory, telephonelinescanbedimensionedto achievethetargetdegreeof service.

IPoverWDM networksand,in general,IPnetworks,lacksuchdimensioningrulesdueto atwofold reason:first thetraffic is nolongerPoissonianbut showsself-similarfeatures,non-stationarityandsourceheterogeneity. Secondly, thereis very scarcenetwork dimensioningtheoryfor self-similartraffic. In this sectionwe examinethespecificfeaturesof self-similar traffic that arenot presentin otherkinds of traffic,suchasvoice traffic. Suchfeaturesprovide the justificationfor traffic managementtechniquesin IP overWDM networks.

11.2.1 Self-similarity

In the recentpast, voice and data traffic modeling has beenprimarily basedonprocessesof independentincrements,suchasthePoissonprocess.Arrivalsin disjointtime intervals areassumedto be independentsince,intuitively, the fact that a usermakesa phonecall hasno influencein other usersmaking different calls. Moreformally, thecall arrival processhasa correlationwhich is equalto zeroin disjointtime intervals. However, the independentincrementspropertydoesnot hold forInternettraffic. Not only traffic in disjoint intervalsis correlatedbut thecorrelationdecaysslowly 1, meaningthateven if the time intervalsunderconsiderationarefarapartfrom oneanotherthetraffic is still correlated.

The traffic processis said to have long memoryor long-range dependence. Inthe specificcaseof Internettraffic, suchlong-rangedependenceis observed in thepacket countingprocesswhich representsthe numberof bytesin fixed duration( �ms)intervals[20, 29].

As a consequenceof theslow decayof theautocorrelationfunctiontheoverflowprobabilityin intermediaterouterqueuesincreasesheavily,in comparisontoaprocesswith independentincrements(Poisson).In [26] an experimentalqueueinganalysiswith long-rangedependenttraffic is presented,which comparesanoriginal Internettraffic tracewith a shuffled version,i.e. with destroyed correlations. The resultsshow a dramaticimpactin serverperformancedueto long-rangedependence.

Figure 11.1 shows the resultsfrom trace-driven simulationsof a single serverinfinite buffer systemwith self-similar traffic. The self-similar traceis shuffled inorderto show theeffect of dependencein queueingperformance.We notethat thesaturationbreakpointfor a queueingsystemwith self-similarinput occursat lowerutilization factorin comparisonto thesamesystemwith Poissonianinput.

1Correlationdecaysasa power law ��������� ��������������������� , � beingthetime lag.

iv TRAFFIC MAN AGEMENT FOR IP OVER WDM NETWORKS

0

5

10

15

20

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

t (s)

rho

Expected queue waiting time

Real traceShuffled trace (2s blocks)

Shuffled traze (20s blocks)

Figure 11.1 Queueingperformancewith self-similarprocess

Suchperformancedropis dueto thepresenceof burstsatany timescale.In orderto visually assesssuchphenomenon,Figure11.2shows traffic in several timescales(10, 100 and1000ms), for a real traffic traceanda Poissonprocess.We observethat while the Poissoniantraffic smoothout with the timescaletowardsthe rate �the real traffic shows burstinessat all timescales. The self-similarity propertycanbeexplainedin termsof aggregationof highly variablesources.First, themostpartof Internettraffic is dueto TCPconnectionsfrom theWWW service[4]. Secondly,both WWW objectssizeanddurationcanbe well modeledwith a Paretorandomvariable,with finite meanbut infinite variance.Theaggregation(multiplex) of suchconnectionsshows self-similarity properties[13, 33]. As a result,self-similarity isan inherentpropertyof Internettraffic, which is dueto the superpositionof a verylargenumberof connectionswith heavy-tailedduration.

11.2.2 Demandanalysis

Opticalnetworkswill provideservicetoalargenumberof usersrequiringvoice,videoanddataservices.In orderto ensurea QoSto suchusersthereis a needfor demandestimation,so that resourcesin the network canbe dimensionedappropriately. Inthetelephonenetwork it is well-known thatusersarehomogeneous, meaningthatthetraffic demandgeneratedby a numberof � usersis simply equalto thesumof theirindividual demands,namely �! "�$#%� , with �&# equalto thedemandof an individualuser.

Thehomogeneityassumptionismostconvenientsincetheoperatorwill dimensionthenetwork takingasabaseasingleuserdemandandextrapolatingto therestof thepopulation. However, the traffic demandfor the Internetradicallydiffers from thatof telephonenetworksandthehomogeneityassumptionis no longervalid.

We take a traffic tracefrom our University accesslink (seesection11.5) andshow, in Figure 11.3 (left), the total volume of traffic generatedby a group of �

TRAFFIC MAN AGEMENT IN IP NETWORKS: WHY? v

0 0.5 1 1.5 2 2.5 3x 10

4

0

50

100

150

200

250

300

350

400

time (sample number)

Kby

tes/

s

Poisson traffic Aggregation = 100ms

0 500 1000 1500 2000 2500 30000

50

100

150

200

250

300

350

400

time (sample number)

Kby

tes/

s

Poisson traffic Aggregation = 1s

0 200 400 6000

50

100

150

200

250

300

350

400

time (sample number)

Kby

tes/

s

Poisson traffic Aggregation = 4s

0 0.5 1 1.5 2 2.5 3x 10

4

0

50

100

150

200

250

300

350

400

time (sample number)

Kby

tes/

s

Real traffic Aggregation = 100ms

0 500 1000 1500 2000 2500 30000

50

100

150

200

250

300

350

400

time (sample number)

Kby

tes/

s

Real traffic Aggregation = 1s

0 200 400 6000

50

100

150

200

250

300

350

400

time (sample number)

Kby

tes/

s

Real traffic Aggregation = 4s

Figure11.2 Input traffic in severaltimescales

users,sortedaccordingto volumeof traffic. We observe thata very few numberofusersis producinga large volumeof traffic. Figure11.3 (right) shows percentageof total volume of traffic in a working day versuspercentageof usersgeneratingsuchtraffic. Lessthan10%of theusersproducemorethan90%of traffic, showingstrongnon-homogeneityin thetraffic demand.Thisresultis in agreementwith othermeasurementstakenat UC Berkeley campus[14].

3

4

5

6

7

8

9

10

11

12

0 100 200 300 400 500 600 700

GB

ytes'

Users

20

30

40

50

60

70

80

90

100

0 10 20 30 40 50 60 70 80 90 100

% B

ytes(

% Users

Figure 11.3 Traffic demandsortedby user. Left: Absolutevalue. Right: Percentageofbytesversuspercentageof users

Theresultsaboveshow thatuserpopulationcanbedividedinto threegroups:bulkusers,averageandlight users.We notethat theappearanceof a solenew bulk user

vi TRAFFIC MAN AGEMENT FOR IP OVER WDM NETWORKS

canproducea suddenincreaseof network load. For instance,the top producerisresponsiblefor the 27.56%of the traffic. Shouldher machinebe disconnectedthetraffic load would decreaseaccordingly. As a conclusion,the existenceof a groupof top producerscomplicatesmattersfor network dimensioning.For Internettraffic,theassumptionthatdemandscaleswith thenumberof usersdoesnot hold. On thecontrary, demandis highly dependentwith avery few users.

Furthermore,the dynamicbehavior of top producersdoesnot follow a generallaw. Figure11.4showstheaccumulatednumberof TCPconnectionsandbytesfromthe first andseventhtop producers.While the former is a highly regular user, thelatterproducesaburstof traffic lastingseveralminutes,andnearlynoactivity for therestof thetime.

0

0.5

1

1.5

2

2.5

3

3.5

0 5 10 15 20

GB

ytes'

Hour

0

50

100

150

200

250

0 5 10 15 20T

CP

con

nect

ions

)

Hour

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0 5 10 15 20

GB

ytes'

Hour

0

200

400

600

800

1000

1200

0 5 10 15 20

TC

P c

onne

ctio

ns

)

Hour

Figure 11.4 Demandversustime for the first (top) andseventh(bottom)producers.Left:TCPbytes.Right: TCPconnections

11.2.3 Connectionlevel analysis

Optical networks will surely evolve to providing QoS on demandat the opticallayer, in a forthcomingphotonicpacket switchingor lightpath-on-demandscenario.Nowadays,flow-switchingmechanismsarebeingincorporatedto IP routersin orderto provideQoSat theconnectionlevel. To thisend,ananalysisof theInternettrafficat theconnectionlevel is mandatory. First,we observe thatInternettraffic is mostly

TRAFFIC MAN AGEMENT IN IP NETWORKS: WHY? vii

due to TCP connections.Table11.1 shows a breakdown per protocolof the totaltraffic in ourUniversitylink. WenotethatTCPis dominant,followedatasignificantdistanceby UDP.

Table11.1 Traffic breakdown per protocol

Protocol MBytesin % Bytesin MBytesout % Bytesout

TCP 13718 99.03% 2835 89.38%UDP 109 0.8% 317 10%ICMP 13.7 0.1% 12.3 0.39%Other 10.2 0.07% 7.4 0.23%

Ontheotherhand,TCPtraffic is highly assymetricin theoutboundfrom servertoclient. Our tracesshow that82.1%of thetotal TCPtraffic correspondsto theserverto client outbound.Next, weanalyzewhatis theservicebreakdown for TCPtraffic.

11.2.3.1 Breakdown per service Figure11.5showsthenumberof bytesperTCPservice. The WWW service,either throughport 80 or variantsdue to proxiesorsecuretransactions(port 443), is the mostpopularservicein the Internet. We alsonotethatthereareanumberof serviceswhichcannotbeclassifieda-priori sincetheydo not usewell-known ports. This is anaddeddifficulty to per-flow discriminationin flow-switchingschemes.

80 5501 20 110 6688 443 6699 48659 25 126957

7.5

8

8.5

9

9.5

10

Service port

log1

0(B

ytes

per

ser

vice

)

Figure11.5 Breakdown perservice

11.2.3.2 Connection size and duration Sincethe mostpart of IP traffic is dueto TCPconnectionsto WWW serverswe focusour analysisin connectionsizeanddurationof suchTCP connections.Figures11.6and11.7show survival functions2 of connectionsize(bytes)andduration(seconds)in log-log scales.We notethat

2A survival functionprovidesthe probability thata randomvariable * takesvalueslarger than + , i. e., �-*/.0+1� .

viii TRAFFIC MAN AGEMENT FOR IP OVER WDM NETWORKS

the tail of the survival function fits well with that of a Paretodistribution with 2parameter. Weplot thedistributiontail leastsquaredregressionline in bothplotsandestimatevaluesof 2 of 314�365 and 31487 for durationandsizerespectively. Suchvaluesarein accordancewith previousstudiesthatreportvaluesof 1.1and1.2 [13].

1e-04

0.001

0.01

0.1

1

1 10 100 1000 10000 1000001e+06 1e+07

P(X

>x)9

File size (Bytes)

File sizePareto alpha=1.2

Figure11.6 Survival functionof file size

1e-04

0.001

0.01

0.1

1

0.01 0.1 1 10 100 1000

P(X

>x)9

Connection duration (seconds)

Connection durationPareto alpha=1.15

Figure11.7 Survival functionof connectionduration

File sizesin theInternetareheavy-taileddueto thediversenatureof postedinfor-mationwhich rangesfrom small text files to very largevideofiles [13]. Regardingduration,we noteaneven largervariability (lower 2 ), which canbedueto the dy-namicsof theTCPin presenceof congestion,which will make connectiondurationgrow largerif packet lossoccurs.

While connectionsizeanddurationareheavy-tailed, possiblycausingthe self-similarity featuresof the resultingtraffic multiplex [13, 33], we note that heavy-tailednessis a propertyof the tail of theconnectionandsizedistribution. Actually,

IP TRAFFIC MAN AGEMENT IN IP OVER WDM NETWORKS: HOW ix

themostpartof TCPconnectionsareshortin sizeandduration.Figure11.8showsa histogramof both sizeanddurationof TCP connections.We note that 75% ofthe connectionshave lessthan6Kbytesandlast lessthan11 seconds.Suchshortlastingflowscomplicatemattersfor per-flow bandwidthallocationsincetheresourcereservationoverheadis significant. As a conclusion,not only a high flexibility andgranularityin bandwidthallocationis requiredfrom the optical layer but alsoflowclassificationanddiscriminationcapabilities,in orderto distinguishwhichflowsmaybeassignedseparateresources.

100

1000

10000

100000

1e+06

0 50000 100000 150000 200000

Num

ber

of c

onne

ctio

ns

:

Bytes

0

20000

40000

60000

80000

100000

120000

140000

160000

0 5 10 15 20 25 30

Num

ber

of c

onne

ctio

ns

:

secs

Figure11.8 Histogramof TCPsize(left) andduration(right)

11.3 IP TRAFFIC MANAGEMENT IN IP OVER WDM NETWORKS:HOW

In this sectionwe focuson theimpactof IP traffic in WDM networks in first (staticlightpath)andsecondgeneration(dynamiclightpath/OBS)WDM networks. Con-cerningthe former we explain the scenarioandthe traffic groomingproblem. Wealso considerthe useof overflow bandwidthto absorbtraffic peaks. Concerningsecondgenerationnetworks we analyzethe tradeoff betweenburstinessand longrangedependenceandthe queueingperformanceimplications. Then,we considersignallingaspects,IP encapsulationoverWDM andlabelswitchingsolutions.

11.3.1 First generationWDM networks

FirstgenerationWDM networksprovidestaticlightpathsbetweennetworkendpoints.Thechallengeis to provide a virtual topology thatmaximizesthroughputandmini-mizesdelayoutof aphysicaltopologyconsistingof anetwork topologywith opticalcrossconnectslinking fiberswith limited numberof wavelengthsperfiber.

It hasbeenshown thatthegeneraloptimizationproblemfor thevirtual topologyisNP-complete[16]. Suchoptimizationproblemtakesasaparameterthetraffic matrix,which is assumedto beconstant,and,asadditionalassumptions,packet interarrival

x TRAFFIC MAN AGEMENT FOR IP OVER WDM NETWORKS

timesandpacket servicetimesat the nodesareindependentandexponentiallydis-tributed,so thatM/M/1 queueingresultscanbe appliedto eachhop. We notethattraffic stationarityandhomogeneitybecomenecessaryconditionsto assumethat atraffic matrix is constant.A non-stationarytraffic processprovidesan offeredloadwhich is time-dependent.Ontheotherhand,anon-homogeneousdemand,asshownin theprevioussection,mayinducelargefluctuationsin thetraffic flows,sinceasolebulk userproducesa significantshareof traffic. If, for instance,the top producertraffic changesdestinationsat a given time the traffic matrix is severely affected.Furthermoretheindependenceassumptionin packetarrival timesis in contrastto thelong rangedependencepropertiesof Internettraffic.

A numberof heuristic algorithmshave beenproposedto optimize the virtualtopologyof lightpaths(for further referencesee[24, part III]), assuminga constanttraffic matrix. We notethateventhoughsuchalgorithmsprovideanoptimizationofthephysicaltopologychancesarethattraffic burstscannotbeabsorbedby thestaticlightpaths. Being the buffering capabilitiesof the optical network relatively smallcomparedto theelectroniccounterparta numberof proposalsbasedon overflow ordeflectionroutinghaveappearedrecently.

WDM Cross-connects

Several routing paths

Figure11.9 WDM network

Figure11.9presentsa mostcommonscenariofor a first generationoptical net-works. IP routersusetheWDM layerasa link layerwith multipleparallelchannels,several of thosebeingusedfor protectionor overflow traffic, which leadsto a net-work designwith little buffering at the routersanda numberof alternatepathstoabsorbtraffic peaks. The samescenariois normally assumedin deflectionroutingnetworks,which arebasedon the principle of providing nearlyno buffering at thenetwork interconnectionelementsbut several alternatepathsbetweensourceanddestination,sothat thehigh-speednetwork becomesa distributedbuffering system.Theadvantageis thatbuffer requirementsat theroutersarerelaxed,thussimplifyingthe electronicdesign. In the WDM case,we notethat the backupchannelscanbeusedto provide analternatepathfor theoverflow traffic asproposedin [5]. Ratherthanhandlingthetraffic burstinessvia buffering, leadingto delayandpacket lossintheelectronicbottleneck,thebackupchannelscanbeusedto absorboverflow traffic

IP TRAFFIC MAN AGEMENT IN IP OVER WDM NETWORKS: HOW xi

bursts. Interestingly, the availability of multiple parallel channelsbetweensourceanddestinationroutersresemblesthetelephonenetwork scenario,in whichanumberof alternatepaths(tandemswitching)areavailablebetweenendoffices. Thereasonfor providing multiple pathsis not only for protectionin caseof failureof thedirectlink but alsoavailability of additionalbandwidthfor thepeakhours.

The appearanceof overflow traffic in the future Optical Internetposesa newscenariowhich hasnot beenstudiedbefore,sincestandardelectronicnetworksarebasedon buffering. While the behavior of a single-server infinite queuewith self-similar input hasbeenwell described[27, 37] thereis little literatureon the studyof overflow Internettraffic. On thecontrary, dueto therelevanceof overflow trafficin circuit switchednetworks thereis anextensive treatmentof Poissonianoverflowtraffic. Precisely, in the early stagesof deploymentof telephonenetworks A. K.Erlangfound that the overflow traffic canno longerbe regardedasPoissonian.Infact,theoverflow traffic burstinessis higherthanin thePoissonianmodel.EquivalentErlangianmodelsfor blocking probability calculationscanbe established[9], thatincorporatetheoverflow loadeffect. Theoverflow traffic canbecharacterizedby aPoissonianinput with higherintensity, in orderto accountfor the burstinessof thelatter.

In order to visually illustrate the burstinessincreaseof overflow traffic we plotseveral instanceof overflow versustotal traffic, for several cut-off valuesin Fig-ure 11.10. The Figureshows a significantburstinessincreasefor overflow traffic,which, consequently, requiresa differenttreatmentin termsof network androuterdimensioning[3].

0 10000 20000 30000 40000 500000

100

200

300

400

500

600

700

time (s)

Kby

tes/

s

Base Traffic (Average = 250Kbytes/s)

0 10000 20000 30000 40000 500000

50

100

150

200

250

300

time (s)

Kby

tes/

s

Overflow Traffic (Cut-off 1.5*250=375Kbytes/s)

0 10000 20000 30000 40000 500000

50

100

150

200

250

time (s)

Kby

tes/

s

Overflow Traffic (Cut-off 1.7*250=425Kbytes/s)

Figure11.10 Original input andoverflow traffic

11.3.2 SecondgenerationWDM networks

SecondgenerationWDM networkswill bring a higherdegreeof flexibility in band-width allotmentin comparisonto first generationstaticnetworks. Prior to photonicpacket switching networks, which are difficult to realizewith current technology[39], dynamiclightpathnetworksmake it possibleto reconfigurethe lightpathsvir-tual topology accordingto the varying traffic demandconditions. However, theprovision of bandwidthon demandon a per-useror per-connectionbasiscannotbeachievedwith anopticalnetwork providing lightpathson-demand,sincethey provide

xii TRAFFIC MAN AGEMENT FOR IP OVER WDM NETWORKS

a circuit switchingsolutionwith channelcapacityequalto thewavelengthcapacity.As thenext step,opticalburstswitching[31, 32] providesa transfermodewhich ishalfway betweencircuit switchingandpurepacket switching. At the edgesof theopticalnetworkpacketsareencapsulatedin anopticalburst,whichcontainsanumberof IP packetsto thesamedestination.Thereis aminimumburstsizedueto physicallimitationsin theopticalnetwork, which is unableto copewith packetsof arbitrarysize(photonicpacketswitching).

Opticalburstswitchingis basedon theprincipleof "on thefly" resourcereserva-tion. A reservationmessageis sentalongthepathfrom origin to destinationin ordertosetuptheresources(bandwidthandbuffers)for theincomingburst. A timeintervalafter thereservationmessagehasbeensent,andwithout waiting for a confirmation(circuit switching),theburstis releasedfrom theorigin node.As aresultof thelackof confirmationthereis adroppingprobabilityfor theburst. Nevertheless,resourcesarestatisticallyguaranteedandthe circuit setupoverheadis circumvented. Figure11.11shows thereservationandtransmissionprocedurefor opticalburstswitching.

S D

Burst sentbefore assignement

Reservation grantimplies delivery ACK

Accesslatency

Resourcerequests

Figure11.11 OpticalBurstSwitching

On the other hand,optical burst switching allows for differentiatedquality ofserviceby appropriatelysettingthevalueof thetime interval betweenreleaseof theresourcereservationmessageandtransmissionof theopticalburst[30, 40]. By doingso,someburstsareprioritizedover therest,and,thus,they aregrantedabetterQoS.

Even thoughthe conceptof OBS hasattractedconsiderableresearchattentionthereis scarceliteratureconcerningpracticalimplementationsandimpactin trafficengineering.A referencemodelfor anOBSedgenodeis depictedin Figure11.12.Incomingpacketsto theopticalcloudaredemultiplexedaccordingto theirdestinationin separatequeues.A timerisstartedwith thefirstpacketin aqueueand,upontimeoutexpiration, the burst is assembledandrelayedto the transmissionqueue,possiblyrequiring paddingto reachthe minimum burst size. Alternatively, a threshold-basedtriggermechanismfor burst transmissioncanbeadopted,allowing for betterthroughputfor elasticservices.

The traffic engineeringimplicationsof the referencemodel in Figure11.12canbesummarizedasfollows [23]: First, we notethat thereis anincreaseof thetrafficvariability (marginal distribution variancecoefficient) in shorttimescales,which isdueto thegroupingof packetsin opticalbursts.Furthermore,atshorttimescales,the

IP TRAFFIC MAN AGEMENT IN IP OVER WDM NETWORKS: HOW xiii

Burst assemblyqueues

Optical Link

Demux

Input Traffic

X X

Statisticalmultiplexor

.

.

.

.

.

kin

kout

Figure11.12 OBSreferencemodel

processself-similarityisdecreased,duetoburstsequencingandshuffling attheoutputof theburstassemblyqueues.Nevertheless,at longtimescaleself-similarityremainsthe same. The beneficialeffect of a self-similarity decreaseat short timescalesiscompensatedby theburstiness(traffic variability) increaseatsuchtimescales.Figure11.13shows an instanceof the input andoutputtraffic processes;=<�> and ;@?BA6C inFigure11.12,plottedin several timescales,;@<-> beinga fractionalgaussiannoise,which provesaccurateto model Internettraffic [27]. While the significanttrafficburstscanbeobservedatshorttimescaleswenotethattheprocessremainsthesameasthetimescaleincreases,thuspreservingtheself-similarityfeatures.

0

5000

10000

15000

20000

25000

30000

35000

0 50 100 150 200 250 300 350 400 450 500

Xt (

byte

s)D

n slots - t (ms)

Input traffic (1ms slots)

0

50000

100000

150000

200000

250000

0 50 100 150 200 250 300 350 400 450 500

Xt (

byte

s)D

n slots - t (x10ms)

Input traffic (10ms slots)

0

200000

400000

600000

800000

1e+06

1.2e+06

1.4e+06

1.6e+06

0 50 100 150 200 250 300 350 400 450 500

Xt (

byte

s)D

n slots - t (x100ms)

Input traffic (100ms slots)

0

10000

20000

30000

40000

50000

60000

70000

80000

90000

100000

0 50 100 150 200 250 300 350 400 450 500

Xt (

byte

s)D

n slots - t (ms)

Output traffic (1ms slots)

0

50000

100000

150000

200000

250000

300000

0 50 100 150 200 250 300 350 400 450 500

Xt (

byte

s)D

n slots - t (x10ms)

Output traffic (10ms slots)

0

200000

400000

600000

800000

1e+06

1.2e+06

1.4e+06

1.6e+06

0 50 100 150 200 250 300 350 400 450 500

Xt (

byte

s)D

n slots - t (x100ms)

Output traffic (100ms slots)

Figure11.13 Inputandoutputtraffic to OBSshaperin severaltimescales(1, 10,100ms.)

xiv TRAFFIC MAN AGEMENT FOR IP OVER WDM NETWORKS

11.3.3 Signalling

Two differentparadigmsarebeingconsideredfor integrationof IP andWDM in theforthcomingnext generationInternet. TheoverlaymodelconsidersbothIPandWDMnetworksasseparatenetworkswith differentcontrolplanes(like IP overATM). Thepeermodelconsidersthat the IP andWDM network sharethe samecontrol plane,so that IP routershave a completeview of the optical network logical topology.Figure11.14shows theoverlaymodelin comparisonto thepeermodel.Theroutersdisplayedin the Figure act as clients from the optical network standpointin theoverlaymodel,sincethey usethe optical network servicesin orderto fulfill edge-to-edgeconnectivity requirements.Therefore,an Optical User-Network Interface(UNI) becomesnecessaryat thenetwork edges,asshown in the sameFigure. TheOptical Internetworking Forum hasproduceda specificationdocumentcontaininga proposalfor implementationof an Optical-UNI which interworks with existingprotocolssuchasSONET/SDH[28].

OxC

OxC

OxC

OxC

PEER MODEL

WDM Network

OVERLAY MODEL

Optical UNIOptical UNI

Telco/LAN Router Telco/LAN Router

Telco/LAN Router Telco/LAN Router

Figure11.14 Overlayversuspeermodel

In orderto integrateIP andWDM layersin thepeermodelapromisingalternativeis theuseof MultiprotocolLabelSwitching(MPLS),with enhancedcapabilitiesforoptical networks. The aim of MPLS is to provide a higherdegreeof flexibility tothe network managerby allowing the useof LabelSwitchedPaths(LSPs). Labelsin MPLS are like VPI/VCI identifiers in ATM in the sensethat they have localsignificance(links betweenrouters)and are swappedat eachhop along the LSP.They areimplementedasfixedlengthheaderswhichareattachedto theIP packet.

An LSP, thus,is functionallyequivalentto a virtual circuit. On theotherhand,aForwardEquivalenceClass(FEC)isdefinedasthesetof packetswhichareforwardedin the sameway with MPLS. The MPLS label is thus a short, fixed-lengthvaluecarriedin the packet headerto identify an FEC.Two separatefunctionalunits canbe distinguishedin an MPLS-capablerouter: control and forwarding unit. Thecontrol unit usesstandardrouting protocols(OSPF/IS-IS)to build andmaintainaforwardingtable. Whena new packet arrivesthe forwardingunit makesa routingdecisionaccordingto theforwardingtablecontents.However, thenetwork operator

IP TRAFFIC MAN AGEMENT IN IP OVER WDM NETWORKS: HOW xv

may changethe forwarding table in order to explicitly setupan LSP from originto destination. This explicit routing decisioncapability providesextensive trafficengineeringfeaturesandoffersscopefor qualityof servicedifferentiationandvirtualprivatenetworking. Figure11.15showsanexampleof LSP(routers1-2-5-6). Eventhoughthe link-stateprotocolmandates,for example,that thebestrouteis 1-4-5-6,thenetwork operatormaydecide,for loadbalancingpurposes,to divert partof thetraffic throughrouters2-5-6. On the otherhand,MPLS canwork alongsidewithstandardIP routing (longestdestinationIP addressprefix match). In our example,thepacket maycontinueits way to thedestinationhost,downstreamfrom router6,througha non-MPLScapablesubnetwork.

Routing path

2 4

4

6

5

Destination edge router

Label Switched Path

1

Source

Destination Subnetwork

Figure11.15 LabelSwitchedPathexample

Explicit routesareestablishedby meansof two differentsignallingprotocols:theResourceReservation Protocol (RSVP-TE)and Constraint-BasedRouting Label-Distributed Protocol. Both allow for strict or looseexplicit routing with qualityof serviceguarantees,sinceresourcereservationalongtheroutecanbeperformed.However, the protocolsaredifferent in a numberof ways: for instanceRSVP-TEusesTCP while CR-LDP usesIP/UDP. For an extensive discussionaboutMPLSsignallingprotocolsthereaderis referredto [15], andchapters13 and15.

MPLShasevolvedinto anew standardthatis tailoredto thespecificrequirementsof opticalnetworks: GeneralizedMPLS (GMPLS)[6]. Sinceopticalnetworksmayswitchentirefibers,wavelengthsbetweenfibersor SONET/SDHcontainersa FECmaybemappedin many differentways,notnecessarilypacketswitched.For instance,a lightpathmaybesetupin orderto carrya FEC.In this examplethereis no needtoattacha labelto thepacketsincethepacket is routedin anall-opticalfashionend-to-end. Therefore,the (non-generalized)label conceptis extendedto the generalizedlabelconcept.Thelabelvalueis implicit sincethetransportmediaidentifiestheLSP.Thus,the wavelengthvaluebecomesthe label in a wavelengthroutedLSP. Labelsmaybeprovidedto specifywavelengths,wavebands(setsof wavelengths),timeslotsor SONET/SDHchannels. A SONET/SDHlabel, for example, is a sequenceoffive numbersknown asS,U,K,L, andM. A packet comingfrom an MPLS (packetswitched)network maybetransportedin thenext hopby aSONET/SDHchannelbysimply removing the incominglabel in the GMPLS routerandrelayingthe packetto the SONET/SDHchannel. A generalizedlabel is functionally equivalentto thetimeslot location in a plesiochronousnetwork wherethe slot location identifiesachannel,with no needof additionalexplicit signalling(apacketheaderin MPLS).

xvi TRAFFIC MAN AGEMENT FOR IP OVER WDM NETWORKS

GMPLSallowsthedataplane(in theopticaldomain)toperformpacketforwardingwith thesoleinformationof thepacketlabel,andthusignoringtheIP packetheaders.By doingso,thereis no needof packetconversionfrom theopticalto theelectronicdomainandtheelectronicbottleneckis circumvented.Thesignallingfor LSPsetupis performedby meansof extensionsof bothMPLS resourcereservationprotocols(RSVP-TE/CR-LDP)in orderto allow thata LSP canbe explicitly routedthroughtheopticalcore[7, 8]. On theotherhand,modificationshavebeenmadeto GMPLSby addinga new link managementprotocoldesignedto addressthespecificfeaturesof theopticalmediaandphotonicswitches,togetherwith enhancementsto theOpenShortestPath First protocol (OSPF/IS-IS)to provide a generalizedrepresentationof the variouslink types in a network (fibers, protectionfibers, etc.), which wasnot necessaryin thepacket-switchedInternet.Furthermore,signallingmessagesarecarriedout of bandin anoverlaysignallingnetwork, in orderto optimizeresourcesandin contrastto MPLS. Sincesignallingmessagesareshortin sizeandthetrafficvolumeis not large in comparisonto the datacounterpartthe allocationof opticalresourcesfor signallingpurposesseemswasteful. Seechapters13 and15 for moredetailsonGMPLS.

As an exampleof the enhancementsprovidedby GMPLS,considerthe numberof parallelchannelsin a DWDM network, which is expectedto bemuchlargerthanin theelectroniccounterpart.Consequently, assigningoneIP addressto eachof thelinks seemswastefuldueto the scarcityof IP addressspace. The solution to thisproblemis the useof unnumberedlinks asdetailedin [19]. The former exampleillustratestheneedof adaptingexisting signallingprotocolsin theInternet(MPLS)to thenew peerIP overWDM paradigm.

As for third generationopticalnetworksbasedonphotonicpacketswitchingtherearerecentproposalsfor Optical CodeMultiprotocol Label Switching(OC-MPLS)basedon optical codecorrelations[25]. The issueis how to reada packet headerin theall-optical domain. Eachbit of thepacket headeris mappedontoa differentwavelengthin a differenttime position,forming a sequencewhich canbe decodedby correlationwith all the entriesin a codetable. While the useof photoniclabelrecognitionhasbeendemonstratedin practicethenumberof codesavailableis veryscarce(8 bits/128codes)[38], thuslimiting theapplicabilityof thetechniqueto longhaulnetworks.

11.3.4 Framing aspectsfor IP over WDM

Concerningframingtechniquesfor IP overWDM, theIP overSONET/SDHstandardis foreseenasthemostpopularsolutionin thenearfuture [10]. Therationaleof IPover SONET/SDHis to simplify the protocolstackof IP over ATM, which in turnreliesover SONET/SDHasthe physicallayer. By encapsulatingthe IP datagramson top of SONET/SDHthe bandwidthefficiency is increasedwhile the processingburdenimposedby segmentationandreassemblyproceduresdisappears.However,a link layer protocolis still neededfor packet delineation.Figure11.16shows theIP datagramin a layer 2 PDU (HDLC-framedPPP)which, in turn, is carriedin aSONET/SDHSTS-1payload.

END-TO-END ISSUES xvii

(1 byte)Path Overhead

STS-1

Payload Envelope (87 bytes)

Flag(0x7E)

Protocol Data(0xFF)

Address Control(0x03)

Flag(0x7E)

1 1 1 1 or 2 Variable 2 or 4

FCS

1Bytes:

STS-1 Frame

Transport Overhead(3 bytes)

and scramblingByte stuffing

9 rows

90 bytes

Figure11.16 IP over SONET/SDH

Thecurrentstandardsfor layer2 framingproposetheuseof HDLC-framedPPPasdescribedin RFC1662/2615[34, 35]. However, dueto scramblingandreliabilityproblemsthescalabilityis compromisedbeyondOC-48[21]. In responseto theneedof higherspeedsotherproposalsfor IP over SONET/SDHhave appearedsuchasthe PPPover SDL (Simplified DataLink) standard(RFC 2823)[11]. In PPPoverSDL thelink synchronizationis achievedwith analgorithmwhich is similar to I.432ATM HEC delineation. Insteadof searchingfor a flag (Ox7E),asis donein POS,the receiver calculatesthe CRC over a variablenumberof bytesuntil it “locks” toa frame. Then the receiver entersthe SYNC stateandpacket delineationis thusachieved. In addition,datascramblingis performedwith aself-synchronousEGFIHKJ�3scrambleror anoptionalset-resetscramblerindependentof userdata,which makesit impossiblefor themalicioususerto breakSONET/SDHsecurity.

Wenotethattheperformanceof IP overSONET/SDH,bothin POSandPPPoverSDL is highly dependenton theIP packetsize.Assumingnobytestuffing (escapingOx7E flags)and16 bits framechecksequence,the POSoverheadis 7 bytes. Fora SONET/SDHlayerofferedrateof 2404Mbps(OC-48)theuser-perceivedrateontop of TCP/UDPis 2035Mbpsfor anIP packetsizeof 300bytes.

Finally, therearealsoongoingeffortsto providelightweightframingof IP packetsover DWDM using10 GigabitEthernet(IEEE 802.3aetaskforce) [36] andDigitalWrappers(ITU G.709)[1].

11.4 END-TO-END ISSUES

The successof the next generationoptical Internetwill not only dependon opticaltechnology, but alsoon the setof protocolsthat translatethe availability of gigabit

xviii TRAFFIC MAN AGEMENT FOR IP OVER WDM NETWORKS

bandwidthinto user-perceived quality of service. Figure11.17shows a referencemodelfor WDM network architecture.TheWDM wide/metropolitanareanetworkservesasan Internetbackbonefor thedifferentaccessnetworks. Thegeographicalspanof theWDM ring canbemetropolitanor regional,coveringareasupto thousandmiles. TheWDM network input traffic comesfrom themultiplex of a largenumberof users(in thethousands)at eachaccessnetwork. Examplesof accessnetworksinour architecturearecampusnetworksor Internetserviceprovidernetworks. AccessandWDM backbonewill be linked by a domainbordergateway that will performthenecessaryinternetworkingfunctions.Suchdomainbordergatewaywill typicallyconsistof a high-speedIP router.

Access network

Access network

WDM backbone

Server

Client

Optical burst

Figure11.17 WDM network referencemodel

The scenarioshown in Figure 11.17 is the most likely network configurationfor future all-optical backbones,that will surelyhave to coexist with a numberofsignificantlydifferenttechnologiesin theaccessnetwork suchasEthernets,wireless,HFCandxDSL networks. Thedeploymentof fiberopticsto theend-usersitecanbeincompatiblewith otheruserrequirements,suchasfor instancemobility, eventhoughthereis a trendtowardsproviding high-speedaccessin the residentialaccesses.Inany case,theaccessnetworkwill besubjecttopacketlossanddelayduetocongestionor physicallayer conditionswhich arenot likely to happenin the optical domain.On the contrary, the high speedoptical backboneswill provide channelswith hightransmissionrates(in the10 Gbps)andextremelylow bit errorrates(in the 3ML�NPOBQ ).

Bridging thegapbetweenaccessandbackbonenetwork becomesanopenissue.A flat architecturebasedon a userend-to-endTCP/IPconnection,althoughsimpleandstraightforward,maynot bea practicalsolution. TheTCPslow startalgorithmseverely constrainsthe useof the very large bandwidthavailable in the lightpath

END-TO-END ISSUES xix

until thesteady-stateis reached.Evenin suchsteady-stateregimetheuser’s socketbuffer maynot belargeenoughto provide thestoragecapacityneededfor thehugebandwidth-delayproductof thelightpath.Furthermore,anempiricalstudyconductedby theauthorsin aUniversitynetwork [4] showedthatnearly30%of thetransactionlatency wasduetoTCPconnectionsetuptime,whichposestheburdenof theroundtriptime in a three-wayhandshake.

However, TCPprovidescongestionandflow controlfeaturesneededin theaccessnetwork,whichlackstheidealtransmissionconditionsthat areprovidedby theopticalsegment.We mustnoticethatheterogeneousnetworksalsoexist in otherscenarios,such as mobile and satellite communications. In order to adaptto the specificcharacteristicsof eachof the network segmentssplit TCP connectionmodels, anevolutionaryapproachof the TCP end-to-endmodel,have beenrecentlyproposed[2, 12]. We notethatTCPsplitting is not anefficient solutionfor opticalnetworks,sincedue to the wavelengthspeed,in the order of Gbps, the useof TCP in theopticalsegmentcanbequestioned.For example,consideringa 10 Gbpswavelengthbandwidthand 10 ms propagationdelay in the optical backbone(2000 km) thebandwidthdelayproductequals25 MBytes. File sizesin the Internetare clearlysmaller than suchbandwidth-delayproduct [4, 22]. As a result, the connectionis always slow-starting,unlessthe initial window size is very large [18]. On theotherhand,sincethe pathsfrom gateway to gateway in the optical backbonehavedifferentroundtripdelayswe notethat thebandwidthdelayproductis not constant.For example,a1 msdeviation in roundtriptimemakesthebandwidth-delayproductincreaseto 1.25Mbytes. Therefore,it becomesdifficult to optimizeTCPwindowsto truly achieve transmissionefficiency in this scenario.Furthermore,theextremelylow lossratein theopticalnetwork makesretransmissionsvery unlikely to happenandsincethenetwork canoperatein a burst-switchedmodein theoptical layerwenotethat thereareno intermediatequeuesin which overflow occurs,thusmakingmostof theTCPfeaturesnot necessary.

11.4.1 TCP for high-speedand split TCP connections

As a first approachto solvingtheadaptationproblembetweenaccessandbackbonethe TCP connectioncanbe split in the optical backboneedges. As a result,each(separate)TCP connectioncan be provided with TCP extensionstailored to thespecificrequirementsof bothaccessandbackbonenetwork. More specifically, thebackboneTCPconnectionusesTCPextensionsfor speed,asdescribedin [18]. SuchTCPextensionsconsistof largertransmissionwindow andnoslow start.

A simulationmodel for the referencearchitectureof Figure 11.17 is shown inFigure11.18. The �SR 3 simulatoris selectedasa simulationtool sincean accurateTCPimplementationis available. We choosea simplenetwork topologyconsistingof an optical channel(1 Gbps)which connectsa coupleof accessrouterslocatedat the boundariesof the opticalnetwork, asshown in Figure11.18. Regardingthe

3http://www-mash.cs.berkeley.edu/ns/

xx TRAFFIC MAN AGEMENT FOR IP OVER WDM NETWORKS

accessnetwork two accesslinks provide connectivity betweenthe accessroutersand client and server respectively. We simulatea numberof network conditionswith varyingnetwork parameters,namelylink capacities,propagationdelayandlossprobability. Theobjectiveis to evaluatecandidatetransfermodesbasedonTCPwiththeperformancemetricbeingconnectionthroughput.

TCPClient

OpticalGateway

OpticalGateway

TCPServer

Accesslink

Accesslink

Backbonelink

Figure11.18 nsmodel

Table 11.2 Summary of simulation parametersParameter Value

BW of backbonelink 3MT UWVXRBackbonelink propagationdelay LZY\[1L1]^RBW of accesslink _�4 `$a"U�VbR&c�[1`4 `$a"UWVXR1cM3ML&L&adUWVbRAccesslink propagationdelay 715e]^R&cf5eLe]�R&cg36L1Le]�R

Figure 11.19shows throughputversusfile size for split-TCP (STCP-PPS)andend-to-endTCP(EE-PPS)connectionwith theaccessnetwork parametersshown intable11.2.Ontheotherhand,resultsfromaproposalof atransferprotocolusingOBS[17], FilesoverLightpaths(FOL), arealsopresented.In FOL, filesareencapsulatedin anopticalburstin orderto betransmittedacrosstheopticalnetwork. Wenotethatthroughputgrowswith file sizetowardsa valuewhich is independentof accessBW,andthelessRTT themoresteady-statethroughput.Forsmallfile sizestheconnectiondurationis dominatedby setuptimeandslow start,whichdoesnotallow thewindowsizeto reacha steady-statevalue. For large files the TCP reachessteady-stateandthe throughputis equalto window sizedivided by round-triptime. Suchbehavioris expectedin a large bandwidth-delayproductnetwork, in which connectionsareRTT-limited ratherthanbandwidth-limited.

Wenotethattheuseof splitTCPprovidesasignificantperformanceimprovement.Thus,weexpectthattheforthcomingWDM networkswill incorporatea connectionadaptationmechanismin theedgerouterssothat theWDM bandwidthcanbefullyexploited.

END-TO-END ISSUES xxi

0

500000

1e+06

1.5e+06

2e+06

2.5e+06

0 1e+06 2e+06 3e+06 4e+06 5e+06

thro

ughp

ut(b

ps)

h

transfer size(bytes)

Access BW=8.448Mbps

EE-PPS 25 msSTCP-PPS 25 ms

FOL 25 msEE-PPS 50 ms

STCP-PPS 50 msFOL 50 ms

EE-PPS 100 msSTCP-PPS 100 ms

FOL 100 ms

0

500000

1e+06

1.5e+06

2e+06

2.5e+06

0 1e+06 2e+06 3e+06 4e+06 5e+06

thro

ughp

ut(b

ps)

h

transfer size(bytes)

Access BW=34.368Mbps

EE-PPS 25 msSTCP-PPS 25 ms

FOL 25 msEE-PPS 50 ms

STCP-PPS 50 msFOL 50 ms

EE-PPS 100 msSTCP-PPS 100 ms

FOL 100 ms

0

500000

1e+06

1.5e+06

2e+06

2.5e+06

0 1e+06 2e+06 3e+06 4e+06 5e+06

thro

ughp

ut(b

ps)

h

transfer size(bytes)

Access BW=100Mbps

EE-PPS 25 msSTCP-PPS 25 ms

FOL 25 msEE-PPS 50 ms

STCP-PPS 50 msFOL 50 ms

EE-PPS 100 msSTCP-PPS 100 ms

FOL 100 ms

Figure11.19 Averagetransferthroughput

11.4.2 Performanceevaluation of file transfer (WWW) servicesover WDMnetworks

The resultsshown in the previous sectionprovide a comparisonin error free con-ditions, for instancein a first generationWDM network (static lightpath betweenrouters).However, it turnsout thatsecondgenerationWDM networkssuffer block-ing probability as a consequenceof the limited numberof wavelengthsandburstdroppingdueto limited queueingspacein opticalburst or photonicpacket switch-es. In suchconditionssplit TCP becomesinefficient andalternateprotocoldesignbecomenecessary.

xxii TRAFFIC MAN AGEMENT FOR IP OVER WDM NETWORKS

In FOL [17], filesareencapsulatedin opticalburstswhicharereleasedthroughtheopticalbackboneusinga simplestopandwait protocolfor errorcontrol. Assumingthat the setupof an optical burst takes RTT/2, Figure 11.20 shows the achievedthroughputwith error probabilitieslower than 0.1, for both FOL (different burstsizes)andSplit TCP (STCP).We observe thatTCP congestionavoidanceseverelylimits transferefficiency. If lossprobabilty is equalto 0.01thethroughputobtainedwithTCPishalf thethroughputobtainedwithasimplestopandwaitprotocolin FOL.This servesto illustratethat thethroughputpenaltyimposedby theTCPcongestioncontrolmechanismsis rathersignificant.

0

20

40

60

80

100

120

140

1e-07 1e-06 1e-05 0.0001 0.001 0.01 0.1

Thr

ough

put (

Mbp

s)

P(error)

Average throughput in a TCP transfer with errors

TCP TheoreticalBurst Switching L=500KBBurst Switching L=250KB

Burst Switching L=100KBTCP Simulated

Figure11.20 Throughputcomparison

The main differencebetweena simpleFOL protocolandTCP is the way bothprotocolsinterpretcongestion.While TCPconsidersthatlossis producedby queue-ing overflow FOL is awarethatlossis dueto blocking. In a losssituation,TCPwilllower the transmissionwindow, which resultsin no effectat all sincecongestionisdueto blocking, andthe moreis the blocking probability the larger is the numberof accessesto the opticalnetwork. Furthermore,sincethe ikjmlonqprp productisextremelylarge the TCP window sizetakeson a very high value. As a result, theslow startor congestionavoidancephasewhich follow apacket losstake thelongesttime to complete.

11.5 CONCLUSIONS

In this chapter, we have presentedthe motivation, solutionsand openchallengesfor the emerging field of traffic managementin all-optical networks. At the timeof the writing of the chapterthereare a large numberof open issues,includingtraffic managementfor IP traffic in dynamicWDM networks(dynamiclightpathandburst switching), designof an efficient integratedcontrol planefor IP andWDMandproposalof transportprotocolthattranslatetheavailability of opticalbandwidthinto user-perceivedquality of service.We believe thattheresearcheffort in thenextgenerationoptical Internetwill be focusednot only in theprovision of a very largebandwidthbut alsoin theflexibility , easeof useandefficiency of opticalnetworks.

xxiii

Appendix: Measurementscenario

Our traffic tracesare obtainedfrom the network configurationdepictedin Figure11.A.1. Themeasurementspresentedin thispaperareperformedat theATM Perma-nentVirtual Circuit (PVC) thatlinks PublicUniversityof Navarrato thecorerouterof the Spanishacademicnetwork (RedIris4) in Madrid. Rediristopologyis a starof PVCswhich connecttheUniversitiesaroundthecountryto thecentralintercon-nectionpoint in Madrid. FromthecentralRedIrisfacilities in Madrid a numberofinternationallinks connecttheSpanishATM academicnetwork to theoutsideInter-net. ThemeasuredPVCusesbridgedencapsulationandit is terminatedatbothsidesby IP routers.ThePeakCell Rate(PCR)of thecircuit is limited to 4 Mbpsandthetransmissionratein theopticalfiber is 155Mbps.

Table 11.A.1 Trace characteristics

Startdate Mon 14/02/200000:00Enddate Mon 14/02/200024:00TCPconnectionsrecorded 957053IP packetsanalyzed 16375793

We notethatthescenariounderanalysisis a representativeexampleof a numberof very commonnetwork configurations.For example,the mostSpanishInternetServiceProviders(ISPs)hire ATM PVC links to the operatorsin order to providecustomerswith accessto theInternet. Thesamesituationariseswith corporateandacademicnetworks,thatarelinkedto theInternetthroughsuchIP over ATM links.On the other hand,measurementsare not constrainedby a predeterminedset ofdestinationsbut representa real exampleof a very largesampleof usersaccessingrandomdestinationsin theInternet.Table11.A.1summarizesthemaincharacteristicsof thetraffic tracepresentedin this chapter.

4http://www.rediris.es

xxiv MEASUREMENT SCENARIO

Israel

Japan

Cyprus

USA

Canarian Islands

RedIRIS/TEN-155 May 1999 Topology155 Mbps34/45 Mbps10 Mbps

2 MbpsPlanned during 1999Various speeds

Measurement Point

Figure 11.A.1 Network measurementscenario

REFERENCES

1. ANSIT1X1.5/2001-064Draft ITU-T Rec. G. 709. Network nodeinterfacefortheopticaltransportnetwork (OTN), March2001.

2. E. Amir, H. Balakrishnan,S. SeshanandR. Katz. Improving TCP/IPperfor-manceoverwirelessnetworks. In ACM MOBICOM’95, Berkeley, CA, 1995.

3. J. Aracil, M. Izal, andD. Morato. Internettraffic shapingfor IP over WDMlinks. Optical NetworksMagazine, 2(1),January/February2001.

4. J.Aracil, D. Morato,andM. Izal. Analysisof internetservicesfor IP overATMlinks. IEEE CommunicationsMagazine, December1999.

5. B. Arnaud.Architecturalandengineeringissuesfor building anopticalinternet.In ProceedingsofSPIEInternationalSymposiumonVoice,Video,andDataCom-munications– All-Optical Networking:Architecture, Control, andManagementIssues, Boston,MA, November1998.

6. P. Ashwood-Smithet al. GeneralizedMPLS-SIgnalingFunctionalDescription.Internet-Draft,work in progress,Nov 2000.

xxv

7. A. Banerjee,J. Drake, J. Lang, B. Turner, D. Awduche,, L. Berger, K. Kom-pella,andY. Rekhter. Generalizedmultiprotocollabelswitching: An overviewof signalling enhancementsand recovery techniques. IEEE CommunicationsMagazine, July2001.

8. A. Banerjee,J.Drake,J.Lang,B. Turner, K. Kompella,andY. Rekhter. Gener-alizedmultiprotocollabelswitching: An overview of routingandmanagementenhancements.IEEE CommunicationsMagazine, January2001.

9. J.Bellamy. Digital telephony. JohnWiley & Sons,New York, 1991.

10. P. Bonenfant and A. Rodriguez-Moral. Framingtechniquesfor ip over fiber.IEEE Network, July/August2001.

11. J. Carlson,P. Langner, E. Hernandez-Valencia,andJ. Manchester. PPPoversimpledatalink (SDL) usingSONET/SDHwith ATM-lik e framing. RFC2823(Experimental),May 2000.

12. R. Cohenand I. Minei. High-speedinternet accessthrough unidirectionalgeostationarychannels. IEEE Journal on SelectedAreasin Communications,17(2):345–359,February1999.

13. M. E. Crovella and A. Bestavros. Self-Similarity in World Wide Web Traf-fic: EvidenceandPossibleCauses. IEEE/ACM Transactionson Networking,5(6):835–846,December1997.

14. R. Edell,N. McKeown, andP. Varaiya.Billing usersandpricing for TCP.IEEEJournalOnSelectedAreasIn Communication, 13(7),September1995.

15. E. W. Gray. MPLSImplementingtheTechnology. Addison-Wesley, 2001.

16. A. Ganz,I. ChlamtacandG.Karmi. Lightnets:Topologiesfor highspeedopticalnetworks. IEEE/OSAJournalof LightwaveTechnology, 11,May/June1993.

17. M. Izal andJ.Aracil. IP overWDM dynamiclink layer: challenges,openissuesand comparisonof files-over-lighpathsversusphotonicpacket switching. InProceedingsof SPIEOptiComm2001, Denver, CO,August2001.

18. V. JacobsonandR.Braden. TCP extensionsfor long-delaypaths. RFC 1072,October1998.

19. K. KompellaandY.Rekhter. Signallingunnumberedlinks in RSVP-TE.Internet-Draft, work in progress,Sept2000.

20. W. E. Leland,M. S.Taqqu,W. Willinger, andD. V. Wilson. On theself-similarnatureof Ethernettraffic. IEEE/ACM Transactionson Networking, 2(1):1–15,January1994.

21. J. Manchester, J.Anderson,B.Doshi, and S. Davida. IP over SONET. IEEECommunicationsMagazine, May 1998.

xxvi MEASUREMENT SCENARIO

22. G. Miller, K. ThompsonandR. Wilder. Wide-areainternettraffic patternsandcharacteristics.IEEENetwork, pages10–23,November/December1997.

23. D. Morato, J. Aracil, M. Izal, E. Maga na, and L. Diez-Marca. Explainingthe impactof opticalburstswitchingin traffic self-similarity. TechnicalReportPublicUniversityof Navarra.

24. B. Mukherjee.Optical CommunicationNetwoks. McGrawHill, 1997.

25. M. Murata and K. Kitayama. A perspective on photonicmultiprotocol labelswitching. IEEENetwork, July/August2001.

26. O. Narayan,A. Erramilli and W. Willinger. ExperimentalQueueingAnaly-sis with Long-RangeDependentPacket Traffic. IEEE/ACM TransactionsonNetworking, 4(2):209–223,April 1996.

27. I. Norros. On the Useof FractionalBrownian Motion in the Theoryof Con-nectionlessNetworks. IEEE Journal on SelectedAreas in Communications,13(6):953–962,August1995.

28. Optical Internetworking Forum(http://www.oiforum.com).OIF UNI 1.0- con-trolling opticalnetworks,2001.

29. V. PaxsonandS. Floyd. Wide areatraffic: The failure of Poissonmodeling.IEEE/ACM Transactionson Networking, 4(2):226–244,April 1996.

30. C. Qiao. Labeledoptical burst switching for ip-over-wdm integration. IEEECommunicationsMagazine, 38:104–114,September2000.

31. C. Qiao andM. Yoo. Optical burst switching(OBS) - A new paradigmfor anopticalInternet.Journalof High-SpeedNetworks, 8(1),1999.

32. C. Qiao andM. Yoo. featuresand issuesin optical burst switching. OpticalNetworksMagazine, 1:36–44,April 2000.

33. R. Sherman,W. Willinger, M. S. TaqquandDaniel V. Wilson. Self-SimilarityThroughHigh-Variability: StatisticalAnalysisof EthernetLAN Traffic at theSourceLevel. IEEE/ACM Transactionson Networking, 5(1),February1997.

34. W. Simpson,A. Malis. PPPin HDLC-like framing. RFC1662,July1994.

35. W. Simpson,A. Malis. PPPoverSONET/SDH.RFC2615,June1999.

36. D. H. Su. Standards:TheIEEE P802.3aeprojectfor 10 Gb/sEthernet.OpticalNetworksMagazine, 1(4),October2000.

37. B. Tsybakov and N. D. Georganas. On self-similar traffic in ATM queues:Definitions,overflow probabilityboundandcell delaydistribution. IEEE/ACMTransactionsonNetworking, 5(3):397–409,June1997.

xxvii

38. N. WadaandK. Kitayama. PhotonicIP routing usingoptical codes:10 gbit/sopticalpacket transferexperiment. In Proceedingsof Optical Fiber Communi-cationsConference2000, Baltimore,MD, March2000.

39. S.Yao,B. MukherjeeandS.Dixit. Advancesin photonicpacketswitching: Anoverview. IEEECommunicationsMagazine, 38(2):84–94,February2000.

40. M. Yoo,C.Qiao,andS.Dixit. Opticalburstswitchingfor servicedifferentiationin thenext generationopticalinternet.IEEE CommunicationsMagazine, 39(2),February2001.