causeddroppingofthetos-basedroutingrequirementfromtheospfspecication.

Transcription

1 ComputerScienceDepartment ImplementationandPerformanceMeasurements CollegePark,MD20742 UniversityofMaryland GeorgeApostolopoulos ofqosroutingextensionstoospf RochGuerin,SanjayKamat protocol,andonvariousperformancemeasurementsmadeonthebasisofthisimplementation Inthispaper,wereportonanimplementationofQoSroutingextensionstotheOSPF AbstractYorktownHeights,NY10598 IBMT.J.Watson toassessthecostandfeasibilityofqosroutinginipnetworks.theresultsprovideinsight ResearchCenter ofqosroutingiswellwithinthelimitsofmodernprocessors.furthermore,wealsondthat theimplementationtoresultsgeneratedbymeansofsimulations.thisprovidesacomprehensive andprotocoloverhead.moreimportant,theyestablishstrongempiricalevidencethatthecost minutefractionofthebandwidthofmostlinks.thepaperalsoexploresthesensitivityofour althoughsupportforqosroutingdoesincreasetheamountofprotocoltrac,itstillremainsa ndingstovariationsinnetworksizeandtopology,bycombiningtheinformationobtainedfrom investigationoftheoperationalcostsofqosrouting. intotherespectiveweightsofthetwomajorcomponentsofqosroutingcosts,processingcost ofqosroutingforimprovingnetworkutilizationandtheservicelevelsprovidedtorequestswith QoSrequirements.Theimprovementtotheservicereceivedbyusersisintheformofanincreased e.g.,bandwidthordelay.severalrecentresearchresults[1,2,3,4,5]havepointedoutthepotential processofselectingthepathtobeusedbythepacketsofaowbasedonitsqosrequirements, attentioninthecontextofitspossibleuseinanintegratedservicesipnetwork.qosroutingisthe 1Introduction likelihoodofndingapaththatmeetstheirqosrequirements.conversely,theimprovementto networkeciencyisusuallyintermsofincreaseinrevenue,whererevenueistypicallyafunction Becauseofitspotentialbenets,QualityofService(QoS)routinghasrecentlyreceivedsubstantial stateofnetworkresourcesthatareofrelevancetopathselection,e.g.,availablelinkbandwidth. coststhatsupportforqosroutingentails.theseaddedcostshavetwomajorcomponents:computationalcostandprotocoloverhead.theformerisduetothemoresophisticatedandmorefrequent implementingqosroutingprotocolsinipnetworks.thisisprimarilybecauseoftheadditional Suchupdatestranslateintoadditionalnetworktracandprocessing,inparticularinthecaseof ofthenumberofowsortheamountofbandwidthcarriedbythenetwork. pathselectioncomputations,whilethelatteriscausedbytheneedtodistributeupdatesonthe linkstateprotocols,onwhichmostoftheqosroutingproposalscurrentlybeingputforwardare Despitethesebenets,thereremainsmuchuncertaintyregardingthevalueandfeasibilityof based,e.g.,see[6]foranoverview.insuchacontext,itisimportanttoproperlyassessthemagnitudeofthesecosts,sothattheweightoftheassociatedbenetscanbebetterevaluated.such intoactualimplementationsandthetrade-ostheyinvolve. anassessmentshouldstrivetobothunderstandfundamentalcostcomponents,andprovideinsight 1

2 havebeeninvestigatedtoexplorethepossibilityofreducingtheprocessingcostofqospathcomputation.similarly,avarietyoflinkcostmetricsandupdatetriggeringtechniques[13,14]orpath issuesassociatedwithqosrouting.eventheverydetailedcostmodelsusedin[12,14,15],arenot selectiontechniques[14]havebeenproposedtolowertheprotocoloverheadofqosroutingwithout signicantlyaectingitsabilitytocomputeecientpaths.however,theseworksrelyprimarilyon SeveralrecentworkshaveaimedatsheddingsomelightintothecostsinherenttoQoSrouting. Inparticular,dierentvariationsofpathpre-computation[7,8,9,10]andpathcaching[11,12] simulations,andasaresultarenotabletofullycapturesomeofthemoreimplementationspecic sucienttopreciselycharacterizetheprocessingloadthatarealqoscapablerouterwillhaveto support.therefore,despitethegreaterinsightintothecostandbenetstrade-oofqosrouting, severalscenarios.thisprovidesvaluableinsightintotheoverallimplementationcomplexityofqos thedesignoftheseextensionsandtheirimplementation,wealsoevaluatetheirperformanceunder itsfeasibilityinipnetworksremainsaquestionmark,andresolvingitrequiresadditionalevidence routing,andthebehaviorofseveralofitscomponents.inparticular,usingthisimplementationwe PathFirst(OSPF)[16]routingprotocolunderthegated[17]environment.Inadditiontodescribing aqosroutingalgorithmbasedonextensionstoapopularimplementationoftheopenshortest ofacompleteimplementationofaqosroutingprotocolforipnetworks.specically,weimplement thatonlyarealimplementationcanprovide. obtainrealisticestimatesforthecostofvariousqosroutingoperationssuchaspathcomputation, Asaresult,thegoalofthispaperistollthisgapbyprovidingadetailedreportandassessment linkstateadvertisementsgenerationandreception,andcomparethecostofourqosenhanced versionofospftothatofthestandardospfprotocol.finally,thepaperalsoattemptsto combinesimulationdataandthendingsbasedonourimplementation,inordertoemulatethe operationofarouterthatispartofalargeqosenablednetworkandgetsomeinsightintothe amountofloadthatan\o-the-shelf",gatedbasedqosroutercanhandle. whichwearecurrentlypursuing. 5summarizesthendingsofthepaperandidentiesanumberofextensionsandenhancements, theqosextensionsweadded,andthegatedprogram,whichisthebaseofourimplementation. gatedbase,whilesection4reportsonperformancemeasurementsforourimplementation.section Section3givesanoverviewofourimplementationandhowwehaveaddedourQoSextensionstothe Thispaperisorganizedasfollows:InSection2,wepresentbackgroundinformationonOSPF, 2Background eachrouterandpropagatedtoallotherroutersusingreliableooding.thisoodingiscontrolledby statedatabase,whichcanberelativelyeasilyaugmentedtoincludeqosrelatedlinkmetrics.the databaseisconstructedandupdatedbymeansoflinkstateadvertisements,thataregeneratedby routingprotocolthathasbeenaninternetstandardforsometime.animportantcharacteristicof linkstateroutingprotocolsisthateachroutermaintainsthefulltopologyofthenetworkinalink TheOpenShortestPathFirst(OSPF)[16]isawellestablishedandwidelydeployedlinkstate 2.1LinkStateRoutingandOSPF havingneighboringroutersformalogicallinkbetweenthemcalledanadjacency,sothatinformation isoodedonlybetweenadjacentrouters.again,linkstateadvertisementscanthemselvesbe extendedtocarryqosinformation(see[18]). TheOSPFstandardspeciesthatroutersimplementingtheprotocolrunshortestpathDijkstra 2

3 identifywherepacketsaretobesentnextbasedontheirdestination.ospfallowsmultipleequal eachdestinationisalsoderivedandusedtoconstructthecorrespondingroutingtableentry,i.e., networks.theshortestpathcomputationperformedbyeachroutercreateswhatisknownasthe SPF(shortestpathrst)tree.AspartoftheSPFtreecomputation,thenexthopinformationfor costpathstoadestinationtosupportloadbalancing. inthenetwork.verticesintheospftopologycancorrespondtorouters,transitnetworks,orstub computationontheirlocallinkstatedatabase,anddeterminetheshortestpathtoallothervertices1 databaseofonlytheareatowhichitbelongsandthesummarizedlistofreachabledestinations statedatabasemaintainedateachrouter,sinceeachrouterneedstomaintainadetailedtopology specialbackbonearea.theuseofareasreducestheroutingprotocoltracandthesizeofthelink withinotherareas. domain.theroutingdomain(alsocalledanautonomoussystemoras)canbesplitintomultipleareaswhichareallinterconnected(eitherphysicallyorlogicallyviavirtuallinks)througha completenetworklinkstatedatabase,ospfallowsforatwolevelhierarchywithintherouting Inordertohandlethescalabilityproblemsassociatedwithbothoodingandmaintaininga thetosbyteinipdatagramsisbeingrevisedbytheietfandlackofdeploymentexperience TOS-basedroutingbyallowingrouterstoadvertiseindependentcostmetricsfordierentservice typesandspecifyingthatrouterscomputeadierentspftreeforeachtypeofservice.inspiteof maximumthroughput,andminimumdelay)thatwerespeciedforipdatagrams.ospfenabled NeitherweretherehostapplicationsrequestingspecicTOSinIPdatagrams.Thesemanticsof thisprovision,thereweren'tmultipleinteroperableospfimplementationssupportingthisfeature. supportthevedierenttypesofservice(viz.,normal,minimumcost,maximumreliability, Originally,OSPFspecicationallowedforTypeofService(TOS)basedroutinginorderto causeddroppingofthetos-basedroutingrequirementfromtheospfspecication. andattempttohighlightaspectsofrelevancetoqosextensions. beendroppedfromospf.routerscanindicatetheirtoscapabilityduringadjacencyformation ThereareseveralotheraspectsoftheOSPFprotocol,thatareofimportancetotheQoSextensions weimplementedandreportoninthispaper.intherestofthissection,wereviewthemajorones aretobeusedforroutecomputationisleftunspecied.thishasprovidedanopportunityto experimentwithqosroutingasanextensionofthetosfeaturesprovidedbystandardospf. andadvertisetosspecicmetricsintheirlinkstateadvertisementsalthoughhowsuchmetrics However,inordertoavoidpotentialbackwardcompatibilityproblems,notallTOSfeatureshave distributed.thestandardospfprotocoltracconsistsofseveraltypesofpackets,andthe theamountofroutingprotocoltrac,especiallywhenadditionalqosinformationneedstobe informationaboutchangesinthetopology,andareusedtocarrymultiplelinkstateadvertisements interestedreaderisreferredto[16]fordetails.here,webrieyreviewthoseofmostsignicancein thecontextofourimplementationofqosextensionstoospf.linkstateupdatepacketscontain (LSA).Astheirnameindicates,LinkStateAcknowledgmentpacketsareusedtoacknowledgereceipt oflinkstateadvertisements.finallydatabasedescriptionandlinkstaterequestpacketsareused Oneaspectofparticularsignicanceistheoodingmechanismasithasadirectbearingon setofroutersattachedtoagivennetwork.asmentionedbefore,multiplelsascanbeincluded containinformationaboutarouterandallitsincidentinterfaces,whilenetworklsasdescribethe inagivenlinkstateupdatepacket.linkstateadvertisementareeithergeneratedperiodically andnetworklinkadvertisementsbeingthemostrelevantonestoourqosextensions.routerlsas tosynchronizethelinkstatedatabasesofneighboringrouterswhentheyformanadjacency. 1Throughoutthispaperweuseinterchangeablythetermsvertexandnode. Therearealsoseveraltypesoflinkstateadvertisements(again,see[16]fordetails),withrouter 3

4 receiptofanlsa,thereceivingrouteracknowledgesthelsawithalinkstateacknowledgment packet.multiplelsascanbeacknowledgedinasingelinkstateacknowledgmentpacket. bandwidth,ofanyofthelinks.notethatbecausearouterlsaadvertisesthecurrentstateofall linksontherouter,achangeinthestateofasinglelinkresultsinthestateofalllinksbeingupdated. WepreservedthischaracteristicinourimplementationofupdateextensionsforQoSmetrics.On QoSextensions,theywillalsobetriggeredby\signicant"changesinthemetrics,e.g.,available oraretriggeredbytopologychangessuchaslinkfailuresorrecoveries.inthecontextofour twoconsecutiveoriginationsofagivenlsabyarouter.thedefaultvalueofminlsintervalis5 seconds.thisparameterisimportantinthecontextofqosextensions,asitlimitsthefrequency relatedtotheoodingoflsas.severalofthemareofparticularrelevancetotheqosextensions weconsider.inparticular,theconstantminlsintervalspeciestheminimumtimebetweenany atwhichchangesinlinkstatecanbeadvertised.qosmetricsbeingmorevariablethanlink status,lsasaimedatupdatingqosmetricsarelikelytobemorefrequent,andtheirprocessing couldthenbeaected(see[14]foranextensivediscussiononhowtolimitthefrequencyofqos TheOSPFstandardmandatesavarietyofconstantsthatcontrolthefrequencyoftheoperations causedbytoofrequentretransmissionsinsomecongurations.itsdefaultvaluehasbeensetto1 second. andsimplydiscarded.thisconstantwasaddedinthelastversionofospfbecauseofproblems isminlsarrival,whichlimitsthefrequencyatwhichnewinstancesofagivenlsacanbeaccepted. IftwoconsecutiveinstancesofanLSAarelessthanMinLSArrivalapart,thesecondisnotprocessed updates,andtheassociatedimpactontheperformanceofqosrouting).anothersimilarconstant triptimebetweenneighboringrouters.rxmtintervalcanbesetonaperinterfacebasis,whileall quencylsrefreshtime,whichhasadefaultvalueof30minutes.finally,recallthatospfprotocol messagesaresentasrawipdatagramsandreliabilityisimplementedwithinospfthroughacknowledgementsandretransmissions.thetimebetweenlsaretransmissionsisdeterminedby BesidesLSAstriggeredbystatechanges,arouteralsoperiodicallygeneratesLSAswithfre- otherconstantsarecommontoallinterfaces. RxmtInterval,whichhasadefaultvalueof5seconds,butneedstobesettowellovertheround- sub-section,wedescribemorepreciselythenatureoftheseextensions.inparticular,weoutlinethe modiedpathselectionusedtocomputeqosroutes,andidentifythemodicationsmadetolsas tosupportadvertisementofqosmetrics. suchproblemswithoutweakeningtheprotectionprovidedbythesetwoconstants,isoneofthe issuesthatneedstobeaddressedwhenimplementingsupportforqosextensions.inthenext mostlikelytointerferewiththemorefrequentupdatesassociatedwithqosmetrics.avoiding Asmentionedearlier,oftheaboveconstants,MinLSArrivalandMinLSIntervalaretheones 2.2OSPFExtensionsforQoSSupport Wechosetoimplementpathpre-computationbecauseofitspotentiallysignicantgainsintermsof [18]identiesseveralpossiblevariationsforQoSroutingextensions,thatincludeon-demandcomputationandpre-computationofQoSroutesaswellasbothexplicitandhop-by-hoproutingmodes. Ourimplementationperformspathpre-computationandisbasedonahop-by-hoproutingmode. quirements,andasaresultlinkbandwidthistheonlymetricsextensionthathasbeenimplemented. proposals[7,10]forsupportingqosrouting.itislimitedtohandlingrequestswithbandwidthre- Ourimplementationisbasedontheapproachdescribedin[18],whichissimilartoseveralother processingload,e.g.,see[15].similarly,weoptedforahop-by-hoproutingmode,simplybecauseit canbeaccommodatedwithoutmajorchangestorsvp[19],thesignalingprotocolthatweassume isusedtorequestqosguarantees,e.g.,see[20].4

5 OurimplementationcomputesQoSpathsusingthewidest-shortestpathselectioncriterion describedin[18].atarouter,thealgorithmpre-computespathsfromtherouter(thesourcevertex s)toalldestinationsinthenetwork.foreachdestination,thealgorithmcomputespathsofall possiblebandwidthvalues,andusesthemtobuildaqosroutingtablewhichiskeptseparatefrom thestandardospfroutingtable(moreonthislater).specically,foradestinationvertexd,the algorithmgeneratesthesetsd=<s1d;s2d;:::;snd>.thissetfurtherconsistsofanorderedlistof sets,wherethesetshdcontainsallpathsfromstodwithhopcounthandthemaximumbandwidth achievableforthathopcount.incaseswhenanincreaseinhopcountfromhtoh+1doesnot improvetheavailablebandwidth,thesetsh+1 discurrently2omitted.inotherwords,thesetsd maintainscandidatepathsinincreasingorderofhopcountandavailablebandwidth. TheQoSroutingtablegeneratedbythealgorithmcanthenbeconceptuallyviewedasamatrix, witheachrowassociatedwithaparticulardestination(entryintheiproutingtable),andeach columngivingtheamountofbandwidthavailableforagivenhopcount.eachentry(row-column intersection)inthetablecontains,inadditiontotheavailablebandwidthandthehopcount, informationidentifyingthenexthoponthepath(s)tothedestination3.notethatanentryinthe QoSroutingtablecancontainmorethanonenexthopincaseswhenseveralpathsexistforagiven hopcountandavailablebandwidth. TheinformationintheQoSroutingtableisusedtoidentifypathscapableofsatisfyingthebandwidthrequirementsofnewrequests.Thisisaccomplishedbycomparingtheamountofbandwidth requestedbyanewowtotheavailablebandwidthinsuccessiveentriesintherowassociatedwith theow'sdestination.thesearchstopsattherstentrywithanavailablebandwidthlargerthan therequestedvalue,atwhichpointthecorrespondingnexthopisreturnedandusedtodetermine wheretoforwardtherequestnext(see[20]fordetails).ifthereismorethanonenexthop,one ofthemmustbechosen.apossibleselectioncriterionistochoosethepathatrandombasedon theavailablebandwidthontheassociatedlocalinterface.thatis,theprobabilitypiofforwarding therequesttonexthopnireachablethroughinterfaceiwithlocalavailablebandwidthbi,isset topi=bi=pnj=1bj,wherenisthenumberofpossiblenexthopsintheroutingtableentry.this istheapproachusedinourimplementationasitfavorslocalloadbalancing.however,itshould benotedthatthismethodignorespotentialsharingofsubsequentlinksbetweenpaths,andhence mayhavelimitedeectivenessincertaintopologies.whileothermoresophisticatedmethodsare possible,inthecurrentimplementation,wechosetofocusontheabovesimplemethod. Inadditiontothechangesrequiredtoboththeroutingtableandthepathcomputation,the OSPFprotocolandcodealsoneedstobemodiedtosupportthepropagationofappropriately extendedlinkstateadvertisements.inparticular,informationaboutavailablebandwidthneedsto beaddedtothelinkstatedatabaseandupdatedthroughlinkstateadvertisements.theformatof LSAsis,therefore,extendedtocarryavailablebandwidthinformation.Thisinformationisencoded usinganewtoseld.theospfspecicationallowsavariablenumberoftos-metricstobe containedinanlsa.however,only16bitsareavailabletoadvertisethevalueofthemetric.while 16bitsaresucientforadvertisinglinkcostsforbest-eortrouting,advertisingbandwidthvalues forlinksrangingfromfewkilobitspersecondtomanyterabitspersecondrequiresmorecareful encoding.onesuchencodingschemeisdescribedin[18]andwasusedinourimplementation. Theproposalof[18]doesnotspecifyanumberofdetailssuchasthemechanismsfortriggering updatesorguidelinesonhowoftentopre-computepaths.theseissueswereinvestigatedthrough 2Wehaveexperimentedwithanumberofvariationsonthepathselectionalgorithm,thatrelaxthehopcount criterion,e.g.,keepingpathswithhigherhopcountbutthesamebandwidth,buttheyarenotincorporatedinthe currentimplementation. 3Asdescribedin[18]thiscaneasilybeextendedtoalsoprovidetheinformationneededtoconstructacomplete explicitroute. 5

6 simulationinseveralotherworks[14,15],andanumberofdesignchoicesintheimplementationwe protocolsonhostsrunningtheunixoperatingsystem.thedistributionofthegatedsoftwarealso gated[17]isapopular,publicdomain4programthatprovidesaplatformforimplementingrouting thisdiscussion,webrieyreviewsomeofthebasicaspectsofgatedonwhichourimplementation isbased. 2.3GateDaemon(gated) reportoninthispaperarebasedonthendingsofthoseworks.however,beforeweproceedwith servicesalsofacilitatedimplementationofsomeoftheextensionsthatwererequiredtosupport QoSrouting.Thesegatedservicesinclude: gatedenvironmentoersavarietyofservicesusefulforimplementingaroutingprotocol.these includesimplementationsofmanypopularroutingprotocols,includingtheospfprotocol.the Memorymanagement Asimpleschedulingmechanism Supportforcreationandmanagementoftimers asoneorseveraltasks.agatedtaskisacollectionofcodeassociatedwithaunixsocket.the socketisusedfortheinputandoutputofthetask.gated'smainloopconsists,amongother Routemanagement(e.g.,routeprioritizationandrouteexchangebetweenprotocols) Interfacesformanipulatingthehost'sroutingtableandaccessingthenetwork creationtimeandareinvokedbygated'smainloop.inaddition,themainloopalsoschedulesthe occurredinanyofthem.appropriatehandlersforavarietyofconditionsareregisteredattask operations,ofaselect()calloveralltasksocketstodetermineifanyread/writeorerrorconditions theschedulingmechanismsofgatedarelowprecision,andtimershaveamaximumresolutionof1 executionofexpiredtimersandlowerprioritycomputationscalledjobs.itshouldbenotedthat AllgatedprocessingisdonewithinasingleUnixprocess,androutingprotocolsareimplemented theospftask.periodicprocessingsuchasgenerationoflinkstateupdatesishandledwith makesitparticularlyimportanttoensurethatalltaskscompletereasonablyquickly.thisneeds second.furthermore,schedulingisnon-preemptiveandoperationsruntocompletion.thelatter generationoflsastoupdateqosmetrics,andre-computationoftheqosroutingtable.thisis primarilybecausewearerelyingonperiodictriggersforboththesetasks(see[14]formotivations tobetakenintoconsiderationwhenchoosingapathselectionalgorithmforqosrouting. ArrivingOSPFpacketsareprocessedbydispatchingthepacketreceptionhandlerregisteredby andevidencesupportingthesechoices). timerswhilelongerprocessingsuchasspfcomputationsarescheduledasjobs.currently,mostof ourextensionstosupportqosroutinghavebeenimplementedusingtimers.theseincludeboth Asanexample,OSPFisimplementedasasingletask,associatedwithasinglerawUnixsocket. administrativepreferencesandonlyasinglerouteisactiveperdestination.theseroutesareperiodicallydownloadedinthehost'skernelforwardingtable.ourqosroutingtableismaintainedasa tokeepmultipleentriesforagivendestinationasafunctionofhopcountandbandwidth. separatetable,primarilytoavoidconictscreatedbyprioritizationandbecauseofourrequirement 4AccesstosomeofthemorerecentversionsofthegatedisrestrictedtotheGateDconsortiummembers. gatedmaintainsasingleroutingtablethatcontainsroutesdiscoveredbyalltheactiverouting protocols.multipleroutestothesamedestinationareprioritizedaccordingtoasetofrulesand 6

7 ofthelinkstatedatabaseduringtheshortestpathcomputationeasyandecient,avoidingthe asadjacentrouters)arelinkedtogetheratthedatabaselevelwithpointers.thismakestraversal needforfulldatabaselookupsforlocatinglinkstaterecordsofneighboringnetworkelements.we takeadvantageofthesestructuresduringthecomputationoftheqosroutingtable. particularlinkstaterecord.inaddition,linkstaterecordsforneighboringnetworkelements(such Finally,theOSPFlinkstatedatabaseisimplementedusingaradixtree,forfastaccesstoa 3ImplementingtheQoSExtensionsofOSPF 3.1DesignObjectivesandLimitations QoSroutingimplementationwhilecontainingtheoverallimplementationcomplexity.Thus,our Oneofourmajordesignobjectiveswastogainsubstantialexperiencewithafunctionallycomplete importantassumptionsare: architecturewasmodularandaimedatreusingtheexistingospfcodewithonlyminimalchanges. QoSextensionswerelocalizedtospecicmodulesandtheirinteractionwithexistingOSPFcode waskepttoaminimum.besidesreducingthedevelopmentandtestingeortthisapproachalso facilitatedexperimentationwithdierentalternativesforimplementingtheqosspecicfeatures tionalitiesthatanearlyprototypeimplementationofqosroutingmustdemonstrate.someofthe suchastriggeringpoliciesforlinkstateupdatesandqosroutetablecomputation. Supportforonlyhop-by-hoprouting.ThisaectedthepathstructureintheQoSrouting Severalofthedesignchoiceswerealsoinuencedbyourassumptionsregardingthecorefunc- Supportforpathpre-computation.ThisrequiredthecreationofaseparateQoSroutingtable FullintegrationoftheQoSextensionsintothegatedframework,includingconguration beeasilyextendedtoallowconstructionofexplicitroutes,butthisiscurrentlynotsupported. tableasitonlyneedstostorenexthopinformation.asmentionedearlier,thestructurecan anditsassociatedpathstructure,andwasmotivatedbytheneedtominimizeprocessing overhead. Decouplingfromlocaltracandresourcemanagementcomponents,i.e.,packetclassiers Modularitytoallowexperimentationwithdierentapproaches,e.g.,useofdierentupdate thatcouldbeusedbyothers. thesepoliciesfromthegatedcongurationle. support,errorlogging,etc.thiswasrequiredtoensureafullyfunctionalimplementation, andschedulersandlocalcalladmission.thisissupportedbyprovidinganapibetweenqos andpre-computationtriggeringpolicieswithsupportforselectionandparameterizationof InterfacetoRSVP.TheimplementationassumesthatRSVP[19]isthemechanismusedto Futureimplementationswillbeabletospecifytheirownmechanismsforthismodule. requestrouteswithspecicqosrequirements.suchrequestsarecommunicatedthroughan routingandthelocaltracmanagementmodule,thathidesallinternaldetailsormechanisms. [18],namely Inaddition,ourimplementationalsoreliesonseveralofthesimplifyingassumptionsmadein interfacebasedon[21],andthersvpcodeusedwastheonecreatedbyisi,version4.2a2 [22]. 7

8 3.2Architecture ThescopeofQoSroutecomputationiscurrentlylimitedtoasinglearea. TheabovedesigndecisionsandassumptionsresultedinthearchitectureshowninFigure1.It AllinterfacesonarouterareassumedtobeQoScapable. AllrouterswithintheareaareassumedtorunaQoSenabledversionofOSPF,i.e.,interoperabilitywithnon-QoSawareversionsoftheOSPFprotocolisnotconsidered. consistsofthreemajorcomponents:thesignalingcomponent(rsvpinourcase);theqosrouting component;andthetracmanager.intherestofthispaperweconcentrateonthestructureand operationoftheqosroutingcomponent.ascanbeseeninfigure1,theqosroutingextensions arefurtherdividedintothefollowingmodules: Pathselectionandmanagementmoduleselectsapathforarequestwithparticular Updatetriggermoduledetermineswhentoadvertiselocallinkstateupdates. Pathpre-computationmodulecomputestheQoSroutingtablebasedontheQoSspecic Pre-computationtriggermoduledetermineswhentoperformQoSpathpre-computation. TspecmappingmodulemapsrequestrequirementsexpressedintheformofRSVPTspecs QoSroutingtablemoduleimplementstheQoSspecicroutingtable,whichismaintained linkstateinformation. QoSrequirements,andmanagesitonceselected,i.e.,reactstolinkorreservationfailures. independentlyoftheothergatedroutingtable. ever,insteadofthe\conceptual"matrixformatdescribedinthatsection,eachrowoftheqos QoSpathsarepre-computedandstoredintheQoSroutingtableasoutlinedinSection2.2.How- 3.3QoSRoutingTableandPathPre-ComputationModules Intherestofthissection,weoutlinethemainfunctionsofeachofthesemodules. andrspecsintothebandwidthrequirementsthatqosroutinguses. spondstoadierenthopcountandbandwidthvalue,arrangedinincreasingorder,i.e.,anentryfor routingtableconsistsofapathstructureintheformofalinkedlist.eachentryinthelistcorre- agivenhopcountiscreatedonlyifithasalargerbandwidththanpreviousentrieswithasmaller pathsavailablewiththesamehopcountandbandwidthvalue.asmentionedearlier,thisisused apathrequestspeciesadestinationaddressandabandwidthrequirement,andtherststepin donotcorrespondtoanincreaseinbandwidth.inadditiontothehopcountandbandwidthinformation,eachentryinthelistalsocontainsalinkedlistofnexthopsassociatedwiththedierent isthemeansforaccessingtherowassociatedwithapathrequesttoagivendestination.specically, forloadbalancingbetweenequalcostpaths. hopcount.thisavoidshavingtoallocatememoryforentriesassociatedwithhopcountvaluesthat whichthedestinationisreachable.suchavertexcanbeatransitnetwork,astubnetwork,or identifyingasuitablepathistoretrievethepathstructureassociatedwiththevertexthrough thedestinationhostitself.identifyingthisvertexisaccomplishedbyreusingthesameradixtree BesidesthelinkedliststructureofeachrowoftheQoSroutingtable,anotherimportantaspect 8

9 MAPPINGFigure1:Thesoftwarearchitecture QOS ROUTE PRECOMPUTATION QOS ROUTE CORE OSPF TABLE FUNCTIONS + (ENHANCED) PRECOMPUTATION TOPOLOGY TRIGGER DATABASE FLOW PATHS TABLE structureusedtoimplementthegatedroutingtableitself.thegatedradixtreestructurecarries RECEIVE & someoverheadbecauseitneedstoallowcoexistenceofmultipleroutingprotocols,andasaresult UPDATE INTERFACE STATUS QoS LSA PATH SELECTION asimplerandmoreecientdatastructurecouldhavebeendevisedifwehadfocusedonlyonthe & MANAGEMENT BUILD & LINK STATE theimplementation(nonewcomponent)andreducethedevelopmenttime.asaresult,theexisting requirementsoftheqosroutingtable.however,wechosetoreusetheexistingstructuretosimplify SEND TRAFFIC PARAMETERS UPDATE TRIGGER QoS LSA gatedfacilityisusedtoconstructaseparateradixtreebasedontheaddressandreachability Routing Enabled OSPF informationcontainedinthelinkstatedatabase.eachleafoftheradixtreecorrespondstoa RSVP RESOURCE Thispointerisimplementedusingthegatedttsimechanism,thatallowstheinstallationofuser destinationvertex,andapointerisprovidedtotheqospathstructureassociatedwiththisvertex. MANAGER deneddatainaleafoftheradixtree. implementingthemaingatedroutingtable,thiscouldcreatepotentialconsistencyproblemsas structures.sincetheradixtreestructureisthesameastheoneusedbythemaingatedrouting availablepathstodestinationvertices,whiletheradixtreeisusedtofacilitateaccesstothesepath decidedtokeeptheqosroutingtableindependentofthegatedroutingtabletoconformtoour table,apossibleoptionwouldhavebeentosharethisstructurebetweenthetworoutingtables.we designgoalsoflocalizingchangesandminimizingimpactontheexistingospfcode.inparticular, whileitwouldbepossibletoattachtheqospathstructuresdirectlytotheleavesoftheradixtree NotethatthebulkoftheQoSroutingtableisinthepathstructuresdescribingthedierent otherroutingprotocolscontinuouslymanipulateandmodifythegatedroutingtable.inaddition, theqosroutingtablegetsre-builtindependentlyofthegatedroutingtable;afurtherreasonfor keepingthetwoseparate.finally,thegatedroutingtablecontainsroutesdiscoveredbyallactive routingprotocols,whiletheqosroutingtableisonlyassociatedwithourqosextendedversionof OSPF. thebellman-fordcomputation.duringthisconstructionphase,itisnecessarytoassociatethelink AnotherimportantaspectoftheQoSroutingtableistheoverheadinvolvedinbuildingitduring 9

10 statedatabaseentities(vertices)thatarebeingexpandedbythebellman-fordalgorithmwith containedinthelsaassociatedwiththevertexbeingexpanded(routerid'sinthecaseofarouter LSA,andnetworkid'sinthecaseofanetworkLSA).Thisprovidesageneral,albeitinecient solution,asitrequiresafulllookupintheradixtreeeachtimeavertexisexpandedinthebellman- canbeaccomplishedbyusingtheradixtreeoftheqosroutingtabletosearchfortheaddress thecorrespondingpathstructuresthatcontainthepathsdiscoveredsofarforthisvertex.this Fordcomputation.Inordertoavoidsuchapenaltyinourimplementation,weaddedapointer asmalladditionalmodicationtotheexistingvertexstructure,whichhadtobemodiedanyhow tosupporttheqosextensions. directlytothepathstructureinsidethevertexstructureinthelinkstatedatabase.thisrequired structureforthisvertex.atthistime,thepointerinthevertexstructureisupdatedtopointto insertionofthevertexintheradixtreeoftheqosroutingtableandthecreationofaninitialpath thenewlycreatedpathstructureforthevertex.inaddition,thettsieldinthenew(leaf)node structurethatpointstothevertexpathstructureisexamined.anullvalueforthispointerisan indicationthatthisisthersttimethevertexisbeingvisited.asaresult,thistriggersboththe insertedintheradixtreewiththisvertexasdestination,isalsoupdatedtopointtothesamepath expandsverticesasthehopcountincreases.whenavertexisexpanded,thepointerinthevertex WhentheBellman-Fordcomputationisinstantiated,itproceedsfromthesourcevertexand QoSroutingtablestructureisshowninFigure2. otherospfimplementations.itsmainbenetisthatitdecouplesthepathstructuresandthepath computationfromtheexactformatusedinthelinkstatedatabase.anoverviewoftheresulting structure.bothpointersarethenavailablefortherestofthepathpre-computationphase,and pointerbacktothevertexisalsoaddedinthepathstructureitself,inordertoallowdecouplingthe pathstructurefromthevertexstructure,andfacilitateportingoftheqosspecicextensionsto provideimmediateaccesstotheexistingpathstructureofanyvertexbeingexpanded.another treein-order,andde-allocatingeachnodeinthetreealongwithanypathstructurethatmay de-allocationofmemory.currently,whenanewqosroutingtableistobecomputed,both theradixtreeandthepathstructuresarefreed.thisisaccomplishedbytraversingtheradix bepointedtobythettsieldoftheleafnode.thisfullde-allocationoftheqosroutingtableis potentiallywasteful,especiallysincememoryallocationandde-allocationisanexpensiveoperation. Furthermore,becausepathpre-computationsaretypicallynottriggeredbychangesintopology, thesetofdestinationswillusuallyremainthesameandcorrespondtoanunchangedradixtree.a ThelastissueofsignicanceintheconstructionoftheQoSroutingtable,isallocationand naturaloptimizationwouldthenbetode-allocateonlythepathstructuresandmaintaintheradix tree.afurtherenhancementwouldbetomaintainthepathstructuresaswell,andattemptto 3.4UpdateandPre-computationTriggerModules incrementallyupdatethemonlywhenrequiredbecauseofadierentnumberofpathswithdistinct TheupdatetriggermoduledetermineswhenarouteroodsanewLSAtoadvertisechangesinits correctlyremovefaileddestinationsfromtheradixtreeandupdatethetreeaccordingly. andnumerouscomplexitiestoensuretheintegrityoftheoveralldatastructureatalltimes,e.g., notbeenincludedinourinitialimplementation.themainreasonisthattheyinvolvesubtle hopcountsandbandwidthvalues.however,despitethepotentialgains,theseoptimizationshave linkmetrics.thepre-computationmoduleisresponsibleforinitiatingthecomputationofanew QoSroutingtable.Inordertoallowforexperimentation,thesetwomodulessupportanumberof optionsthatcanbeconguredintheimplementation. Theupdatetriggermoduleimplements:10

11 Figure2:Mappingbetweendierentdatastructures Mapping using Radix Tree the ttsi information Path... Paths to D with hop length H1... Paths to D with hop length H2... Paths to D with hop length H3... Paths to D with hop length H4 Destination D QoS Routing Table using pointer in the path Periodicupdategeneration. Avarietyoftriggeringpolicies.Examplesofsuchpoliciesare:a)Thresholdbasedpolicy:an updateistriggeredwhenthedierencebetweenthepreviouslyoodedandthecurrentvalue using a pointer Thepre-computationtriggermoduleimplements: crossed.classesmayhaveequalorexponentiallyincreasingsizesforlargervaluesofavailable bandwidth. foravailablelinkbandwidthislargerthanacongurablethreshold.b)classbasedpolicy: splitthecapacityofalinkinanumberofclassesandadvertisewhenaclassboundaryis in the vertex Vertex node D Beabletoreceivenoticationofeventsofinterest.Inparticular,thepre-computationmodule Periodicpre-computation. Triggeredpre-computationeachtimeNdistinctlinkstateadvertisementhavebeenreceived. Toimplementthisfunctionalitybothmodulesneedto: ThiscanbeaccomplishedbyinsertinghooksintotheOSPFcoderesponsibleforreceivingand processinglsas.similarly,theupdatetriggeringmoduleneedstobeinformedofchanges needstobenotiedofthearrivalofalinkstateadvertisementoratimerexpirationevent. Maintainandposttheirowntimers.Bothmodulesneedtwotypesoftimers:a)Clampdown sentifthetimerhasexpired.clamp-downtimerscanbeimplementedthroughacombination timers,thatareusedtolimitthefrequencyofmetricsupdates,i.e.,theupdatecanonlybe managertonotifytheupdatetriggeringmoduleofsuchchanges. intheavailablebandwidthonlocallinks.thisisneededtoensurethatthecorrectvalue ofastandardtimerandaag.theagissetwhenthetimerisstartedandresetwhenthe issentinthenextlsa,andinsomeinstancestodetermineifanupdateisneeded.this timerexpires.byinspectingtheagthemodulecanthendetermineiftheclampdowntimer isaccomplishedthroughtheuseofasimplemessaginginterface,thatallowstheresource hasexpiredornot.b)timersforperiodicoperation.thesearesimpletimerscoupledtothe appropriatetimerexpirationhandler.11

12 pathcomputation,itisharderforupdatetriggering.aswasalludedtointheprevioussection,this isbecausetheregularospfupdatetriggeringrulescaninterferewiththetriggeringpolicyimplementedbytheupdatetriggeringmodule.specically,theperiodic(rxmtinterval)andownclamp Ideally,theoperationoftheabovetwomodulesshouldbetransparenttotheexistingOSPF additiontothetheospfcodethathandlesinterfacestatuschanges. accomplishforthepathcomputationcomponentasitoperatesindependentlyoftheregularospf mechanismsformanagingbothpathcomputationandupdatetriggering.whilethisiseasyto Benotiedwhenalocallinkchangesstatus(up/down).Thisisaccomplishedthroughasmall onesecondisthesmallestpossiblevaluewecanspecify,anditistheonewecurrentlyuseinthe implementation.ifinsomesituations,asmallervalueisrequired,modicationstothegatedtimer morefrequentqosupdates.notethatsincethecurrentresolutionofgatedtimersisinseconds, downtimers(minlsintervalandminlsarrival)ofospfmayneedtobedisabledorbypassedin ordertoavoidinterferingwiththegenerationofqoslsas.oneoptionistodisabletheexisting codewillbeneededtoallowforanerresolution. toimplementasimilarmechanismtoensurestabilityoftheprotocolduringperiodsofoverload. Asaresult,weoptedforthesimpleapproachofdecreasingthevalueofMinLSIntervaltoallow OSPFclamp-downmechanismsinthecaseofQoSrelatedLSAs.However,itremainsnecessary infrequently(thelifetimeofadatabaseentryis1,800seconds),thecurrentimplementationallows allupdatesoriginatedbyospf.changesinthestatusofalocalinterfacecouldbeinferredfromthe originatedperiodicupdates.forsimplicityandgiventhatperiodicospfupdatesareoriginated thatalllsassentbytheospfprotocolaretrappedandroutedthroughit.asaresult,theupdate bandwidthvaluesadvertised.thisisachievedbyimplementingtheupdatetriggeringmoduleso triggeringnow\sees"alllsasandcan,therefore,controlwhichonestoblock,ifany,e.g.,theospf ofanystandardlsageneratedbyospfasthisaectsitsbook-keepingregardingthelastavailable Inadditiontotheaboveissue,itisalsoimportanttoinformtheQoSupdatetriggeringmodule onsuchanexplicitnoticationmechanism,asitlearnsaboutchangesinlinkstatusfromthe associatedlsageneratedbyospf,butagainforsimplicityweoptedforanexplicitnotication oftheupdatetriggeringmodule.notethatthepre-computationtriggermoduledoesnotrely componentresponsibleforprocessingincominglsas(anlsaisalsousedtoreectanychange ofstatusofalocallinkinthelinkstatedatabase).thisdecouplingfromlocalstateinformation alsoextendstolinkmetrics.inparticular,computationoftheqosroutingtableisdoneusing itisdiculttoaccuratelycharacterizethetransmissionpatternsoflsascorrespondingtoqos accuratemetricsinformationcouldbeobtained.thisisimportanttominimizeinconsistenciesin themetricsinformationavailableinthelinkstatedatabase,evenforlocallinksforwhichmore updates,sothatdetermininganappropriatevaluefordelayingacknowledgmentisalsodicult.in updatetrac,thiscouldcauseoverowsintheretransmissionqueuesofthesender.unfortunately, edgmentsformultiplelsas.becauseupdatesofqosmetricsarelikelytoincreasetheamount addition,thecurrentdefaultvalueisfairlylarge(about20seconds),whichwouldinterferewith forqosrouting,isthesupportfordelayedacknowledgmentsthatallowsaggregationofacknowl- theqosroutingtablescomputedatdierentrouters. QoSupdatesexceptmaybewhenusingaperiodictriggermechanismwithalargetrigger(greater AnotherexistingOSPFmechanismthathasthepotentialtointerferewiththeextensionsneeded inourcurrentimplementationtobypassthismechanismaltogetherandimmediatelyacknowledge LSAsreceivedfromneighboringrouters. than20seconds).asaresult,sinceachievingameaningfullevelofaggregationforacknowledgmentsappearstorequireadelayvaluethatwillmostcertainlyinterferewithqosupdates,wechose Anotherapproachwhichweconsideredbutdidnotimplementatthispoint,wastomakeQoS 12

13 MakingQoSLSAsunreliablewouldbeareasonabledesignchoicebecauseoftheirmorefrequent LSAsunreliable,i.e.,eliminatetheiracknowledgments,soastoavoidanypotentialinterference. theproperoperationofthebaseospfprotocolitself.onesuchinstanceistheimpactofqos implementingandexperimentingwithsuchanoptioninthefuture. Anequallyandevenmoreimportantaspect,istoensurethatQoSextensionsdonotinterferewith operationofospf,andonlyreducethequalityofpathsdiscoveredbyqosrouting.weplanon LSAsonthestandardSPFcomputation.BecausesuchLSAsdonotcorrespondtochangesinlink transmission,butmoreimportantbecausethelossofaqoslsawillnotinterferewiththebase status,anewspfcomputationistypicallynotrequiredandshould,thereforebeavoided.our thathavethepotentialtointerferewiththeoperationoftheextensionsneededforqosrouting. implementationaddressesthisissuethroughtheuseofaagthatidentiesqoslsasandisused tobypassthespfcomputation. Finally,itshouldbenotedthattheabovediscussionhasfocusedonexistingOSPFmechanisms request,tosearchthroughtheradixtreeoftheqosroutingtable.thissearchidentiesthepath 3.5PathSelectionandManagementModule Thismodulehastwoparts,ofwhichonlythepathselectionparthasbeenimplemented.Path selectionisresponsibleforhandlingincomingrequestsforqosroutes,e.g.,triggeredbythereceipt ofanrsvppathmessage.thisisdonebyrstusingthedestinationinformationprovidedinthe structureassociatedwiththedestination.recallthatthepathstructureforadestinationconsists ofalistofnexthopscorrespondingtolonger,buthigherbandwidthpathstothedestination. Oncetheappropriatepathstructurehasbeenidentied,itisscanneduntiltheminimumhoppath withanavailablebandwidthlargerthanorequaltotheamountrequestedisfound.thenexthop oneisselectedatrandomwithprobabilityproportionaltotheamountofavailablebandwidthon information,i.e.,theassociatedinterface,oroneofthenexthopsifmorethanoneispresent,is thenreturnedandusedtoforwardtherequestonward.whenmorethanonenexthopispresent, theinterfaceasdiscussedinsection2.2.ifthesearchofthepathstructureterminateswithout timeofpathselection. (i.e.,thecorrespondingnexthop)tothedestination,i.e.,wedonotperformcalladmissionatthe ndingapathwithsucientbandwidth,wecurrentlyreturn(oneof)thewidestavailablepaths 3.6TspecMappingModule ThismodulesimplyextractsinformationfromtheRSVPTspecthatdescribestheQoSrequirements lastfeatureisnotyetsupportedinthecurrentimplementation. withtherequest.themanagementofthisentryinthepresenceofpathorreservationfailures(see [20]fordetails)istheresponsibilityofthepathmanagementpartofthemodule.However,this Onceapathhasbeenidentied,acorrespondingentryiscreatedinaowtable,andisassociated RSVP,version4.2a2. therestofthesystem.theintegratedservicesdatastructuresarethesameonesusedbytheisi wesupportonlyasimplemappingwherethetokenrateoftherequestisusedasthebandwidth requirementoftherequest.othermoresophisticatedmappingscanbeaddedlaterwithoutaecting ofarequest,andmapsittotheqosmodelsupportedbythesystem,i.e.,bandwidth.currently, 13

14 4PerformanceEvaluation 4.1Methodology Inthissection,weattempttoevaluatethecostofQoSrouting,whenusingourimplementation. Asdiscussedintheintroduction,weassumethatcomparedtoregularBest-Eortrouting,QoS constructarealnetworkwithroutersrunningourimplementation,andobserveandmeasuretheir costofqosrouting.consequently,wedonotdiscussroutingperformanceissuesorcomparerouting a)processingcost,b)messagegenerationandreceptioncost,andc)memoryrequirements.for processingcost,wefurthersubdivideitintopathpre-computationandpathselectioncosts. performanceofqosandnon-qosrouting(see[14]forsuchadiscussion),andconcentrateinstead onimplementationcostcomparisons.weexplorethreedierentdimensionsinourcomparisons: routinghasbenetsforthenetworkandtheuser,andtheprimaryissueisthepotentiallygreater operationalperformance.unfortunately,thisapproachisnotpracticalforanynon-trivialsize networkduetothehighcostofbuildingareallargescalenetwork.asaresult,weinvestigate stand-aloneoperations.forexample,thetimeneededforasinglepathpre-computation,orthe stepinthatdirectionistonoticethatmostoftheabovecostscanbemeasuredindividuallyoras measurementmethodsthatareaccuratewithoutrequiringalargeequipmentinvestment.therst sizeoftheqosroutingtablecanbeestimatedbasedonasinglerouter,whoselinkstatetopology ThemostcomprehensiveandaccuratemethodformeasuringthecostofQoSroutingisto timeittakestoselectapath.eventhecostofreceivingororiginatinglsascanbemeasured databasehasbeenpopulatedusingsomeexternalmechanism.thesameholdsformeasuringthe ciesandinteractionsthattakeplaceinarealoperationalenvironment.theseaectperformance, estimatestheintrinsiccostofqosrelatedoperations,itdoesnotfullycapturethemanydependenmentoftheimpactofqosroutingonarouter'soperation.thisisbecause,whilethisaccurately measurementsas\stand-alone"evaluationmode. reasonablyaccuratelybyusingonlytworoutersaftertheyformanadjacencyandstartexchanging ifonlybecausetheydeterminethefrequencyandtimingofmanyofthoseoperations,andthese LSAs.Thus,itispossible,withaminimumamountofequipment,toobtaingoodatomicestimatesofthecostofallindividualoperationsofinterest.Werefertothistypeofperformance parametersarediculttoestimatewithoutafullscalenetworkenvironment.inordertoaddress simulationsthatweusetocreatetheappearanceofalargenetwork. thisshortcomingofthestand-alonemeasurementmode,weproposetocombineitsresultswith Stand-aloneperformanceresultsaloneare,however,notsucienttoprovideacompleteassess- correspondtorepresentativeandrealisticoperationalconditions.onceagivensetofparameters estimateoftheoperationalbehaviorofaqosroutingenabledrouterinthesimulatednetwork. hasbeenxed,weselectoneofthesimulatedroutersasa\test-node"anduseittoobtainan parametersis\tuned"basedonourpreviousexperiencewiththissimulationenvironment,soasto requests,arrivalratesanddistributionofrequestsourcesanddestinations,andc)choiceofpath pre-computationandlinkstateupdategenerationtriggerpoliciesintherouters.eachoftheabove workwiththefollowingparameters:a)networktopology,b)traccharacteristicssuchassizeof Specically,wedeneasimulationenvironmentthatallowsustospecifyanoperationalnet- Thisisaccomplishedasfollows: thefollowingoperationsoccurredatthetestnode:a)generationofanlsa,b)receptionofan LSA,c)initiationofpathpre-computationand,d)initiationofpathselection.Theinformation usingeitheroneofthefollowingtwomethods.therstmethodusestheindividualoperations gatheredinthesimulationlogsisthenusedtoderiveoperationalcostsofthetest-noderouter Whileperformingasimulationrun,wegeneratealogthatcontainsthetimeatwhicheachof 14

15 cumulativecostisobtainedbyaddingindividualcostsbasedonthesimulationlogthatidentiesall thedierentoperationsperformedbytherouter.thecumulativecostcanthenbedividedbythe totalsimulationtimetoyieldanestimateoftherouterload.thesecondmethodusesthetiming andoperationinformationcontainedinthesimulationlogtodrivearealrouterandobserveits costsderivedfromthestand-aloneexperimenttocomputeacumulativecostatthetest-node.this themandthegrowthofoperationalcosts.ontheotherhand,thesecondmethodfocuseson Inparticular,therstmethodallowsustoestimatethenumberofoperationsperformedasvarious providesadierentestimateofperformancecosts,andasaresultallowsustocross-validateresults. parameterssuchasnetworksize,tracload,etc.,vary,butitassumesanadditiverelationbetween thenmeasuredandusedtoprovideanotherestimateoftheimpactofqosrouting.eachmethod behavior.thisisactuallyaccomplishedusingtworouters.oneusedtogeneratelsasaccording capturingpossibleinteractionsthatcantakeplacewithinarouterastheprocessingloadtriggered usedtoperformalltheotheroperationsspeciedinthetestlog.theloadonthistestrouteris toinformationinthesimulationlogonreceptionoflsas.thesecondisourtestrouter,andis dierenceisthatospfimplementstwotypesoflsas,routerandnetworklsas,andthesimulator First,thesimulator(amodiedversionofMaRS[23]builtforpreviousworks)usedtoderivethe byexternaleventsvaries. operationsandtiminglogs,doesnotexactlymimicthebehavioroftheospfprotocol.themain dier,theimpactofignoringnetworklsasshouldbeminimalasupdatesoflinkbandwidth assumesthatonlyrouterlsasaresent.however,whilehandlingofrouterandnetworklsasdo simulationbasedmethodwejustdescribedremainsanapproximation,albeitareasonableone. informationareprovidedonlythroughrouterlsas.suchlsasrepresent,therefore,thebulkof Beforeweproceedwithourperformanceevaluation,itisimportanttopointoutthatthe combinationofmultiplelsasintoasingleospfnetworkpacket,whichlowerstheirtransmission overhead.theseeectsarenotcapturedinthesimulator,whichdoesnotimposeanyminimum senttoaneighboringrouter.thisaectsthearrivalpatternsoflsasandisalsolikelytocausethe thequeuingofupdatesgeneratedatafasterrate,sothattheyareproperlyspacedbeforebeing ingof1secondweimposebetweenthetransmissionoftwoconsecutivelsas.thisconstraintforces LSAswhenoperatingaQoSroutingenableddomain,withnetworkLSAsbeingoriginatedonlyin spacingbetweenconsecutivelsas,andfurtherassumesthateachlsaistransmittedinitsown caseoftopologychangeswhichweanyhowdonotconsiderinoursimulation. OSPFpacket.Thiscanleadtoslightlylargerestimatesforoperationalcostssincethereceptionof Anotherdiscrepancybetweenthesimulatorandtherealimplementation,istheminimumspac- However,wedon'texpectthistohaveamajorimpact.Finally,themachineusedastestrouterisan machine.asaresult,whenmeasuringtheutilizationofthetestrouter,allthetrac,whichinthe simulatednodecameandleftthroughmultipleinterfaces,isnowpassingthroughthesameinterface. ordinaryunixbasedhostrunninggatedwithoutanyofthesoftwareandhardwareoptimizations, secondevaluationmethod.specically,onlyasinglenetworkinterfacewasavailableforthetest containingmultiplelsas. multipleindividuallsasislikelytobemoreexpensivethanthereceptionofasingleospfpacket thatmostcommercialroutersareequippedwith.nevertheless,despitetheseshortcomings,we believethatthisstudyprovidesrealisticinsightintotheprocessingrequirementsof\typical"qos Hardwarelimitationsintroduceanothersourceofdiscrepancy,whichismorerelevanttothe gated3.6a.2software.theethernetadaptersusedinthetestsare10mbit/secondsmcpci. processor,64mbytesofrealmemory,3.4gbytesofdisk,runningfreebsd2.2.5-releaseand routingloads. Inallexperiments,thetestsystemsusedareIBMIntellistationswithaPentiumPro200MHz 15

16 computingtheqosroutingtabletothatoftheospfshortestpathcomputation,andthememory 4.2Stand-AloneCost requirementsoftheqosroutingtabletothoseofthestandardospfroutingtable.pathselection onesinanon-qosroutingenabledimplementation.inparticular,wecomparethecostofpre- WeappropriatelyinstrumenttheimplementationinordertomeasurethecostofeachoftheQoS specicoperationsweidentiedearlier.inparticular,wemeasurethetimeneededtocompute Therstofthesetwocomponentsiscommontobest-eortroutingwhilethesecondisuniqueto costinqosroutinghastwocomponents:theradixtreesearchbasedondestinationaddressand thelinearscanningofthepathstructurelistforafeasiblepathbasedonbandwidthrequirement. theqosroutingtable,theamountofmemorythistableneeds,andthetimeittakestoselecta QoSrouting. pathfromthistable.afterobtainingthesemeasurements,wecontrastthemwithcorresponding numberofpathsimpliesahigherpathcomputationtimeaswellasabiggerqosroutingtable.in links.thisisbecausepathpre-computationmaintainsalternatepathswithbottleneckbandwidth largerthanthatofminimumhoppaths.asaresult,thenumberofdistinctpathstoagiven ordertoreportrepresentativeguresfortheoverheadofqosrouting,wethereforeneedtoconsider asisthecaseforthestandardospfprotocol,butalsoonthedistributionofavailablebandwidthon destinationvariesaccordingtothedistributionofavailablebandwidthonnetworklinks.alarger theimpactoflinkbandwidthdistribution,andpossiblyreportresultsfordierentcases. ItisimportanttonotethatQoSroutingspeciccostsdependnotonlyonthenetworktopology, oflinkbandwidths,thatresultsinthemaximumdistinctnumberofpathsbeinggeneratedwhen computingtheqosroutingtable.unfortunately,suchadirectapproachappearstoocomplex topologybydefaultingthepathpre-computationalgorithmtoaminimumhopcountalgorithm. asidentifyingsuchaworstcasebandwidthdistributionseemstorequiretheenumerationofan knowtodependonlinkbandwidthdistribution.specically,weobtainthebestcaseforagiven intermediatesolution,weproposetocomputebothbestandaveragecasesforthosecoststhatwe exponentialnumberofpaths.thismakestheproblemdiculteveninsmalltopologies.asan Onepossibleapproachistoattempttoidentify,foraparticularnetworktopology,anassignment inisolation.however,therearealsodependenciesonexternalfactorssuchasnetworksizeand andreceptionoflsas.thecostoftheseoperationscanbemeasuredbyconsideringeachofthem topology.inparticular,generationandhandlingoflsastypicallyinvolvesaccessingthelinkstate Thisyieldstheminimumpossible(one)numberofpathstoeachdestination.Next,weobtainan randomlinkbandwidthdistribution,andaveragingtheresultingcosts. thatwealsomeasurethecostofoperationsthatarecommontoboth.inparticular,generation estimateoftheaveragecasebyexecutingthepathpre-computationalgorithmforasequenceof database,andthecostofsuchanoperationcandependonthesizeandorganizationofthedatabase. Thisinturnislikelytodependonnetworksize.Weinvestigatetheextentofanysuchdependency byconsideringmultiplenetworksizesandtopologies. ComparisonbetweenQoSroutingenabledandstandardversionsoftheOSPFprotocol,require tiveofthenetworktopologyofatypicalinternetserviceproviderintheus.thisistheisptopology showninfigure3(a),whereallnodescorrespondtorouternodesinterconnectedbypoint-to-point thecorrespondingsetofentries.weconsidertwotypesoftopologies. particular,thismeansarticiallypopulatingthelinkstatetopologydatabaseoftherouterswith stateofourtestrouters5totheonecorrespondingtothenetworktopologyweareassuming.in Therstoneisatopologythathasbeenusedinanumberofpreviousstudies,andisrepresenta- Finally,inordertoperformanyoftheabovemeasurements,weneedtorstinitializetheinternal 5Recallthatweusetwo. 16

17 Test node A Test node A Test node B links.intheisptopology,themaximumpathlengthforthebellman-fordcomputationwassetto 16hops.Thetopologyisdimensionedforuniformtrac,assumingminimumhoprouting,sothat Source linkcapacitiesrangebetween20and70mbits/second.thesecondtypeoftopologyweconsider,is (a)theisptopology (b)thevariablesizemeshtopology Source Transit Networks anarticial,meshliketopologythatisconstructedbyrepeatingabasicbuildingblock.thebasic Figure3:Topologiesusedincostmeasurements Routers smallblackdotscorrespondtorouternodes,whilethewhitecirclesindicatetransitnetworknodes buildingblockwhichconsistsof4routersand5transitnetworksisdepictedinfigure3(b).the intheospftopologydatabase.annnmeshtopologyisconstructedbyrepeatingthebuilding blockalongtwodimensions.figure3(b),alsoillustrates22meshtopology.inthistopology,all Test node B Building Block Varyingthesizeofthemeshtopologyallowsustoobservehowthecomputationcostvarieswith 4.2.1QoSRoutingTableComputation thistopologyrangingfrom11to1010.foratopologyofsizenn,themaximumhoplimit linksareassumedtohaveacapacityof45mbits/second.inourexperimentsweuseinstancesof thatthegureforthenumberofroutersinonearearangesfrom20to350with100beingthe medianand160beingthemean.weshowresultsfornetworksofupto400nodes.thetime networksize.arecentsurvey[24]ofvendorswhohavedeployedospfinrealnetworks,reports forthebellman-fordcomputationwassetton+2. neededforcomputingthestandardspftreeandtheqosroutingtableforthemeshtopologyare isanaspectwhichweinvestigateinsection4.3. QoSpathpre-computationmaybesignicantlylargerthanthatoftheSPFcomputation,andthis averagecase.inbothcases,theprocessingcostofqospathpre-computationisnotsignicantly thecostofde-allocatingthepreviousqosroutingtable,andresultsareshownforthebestand largerthanthatofthespfcomputation.however,oneshouldrememberthatthefrequencyof computationarebasedonarticiallyconguringthelinkstatedatabaseofourtestrouterforthe showninfigure4.theperformanceestimateforboththespfcomputationandtheqospathpre- itcanhelpidentifyitemsthatmaybeworthyoffurtheroptimization,eitherintheimplementation dierentnetworksizesweassume.inthecaseofqosrouting,thepre-computationtimesinclude Itisinterestingtobreakdownthecostofpathpre-computationintoindividualcomponents,as 17

18 SPF QoS Table Best Case QoS Table Average Case computationcost,37%ofthetotaltimeisspentinde-allocatingthepreviousqosroutingtable Another23%isusedforaccessingthelinkstatedatabasetoretrieveinformationaboutnodesand Figure4:Processingtimeforpathcomputation 6000 theirlinks,forsearchingthroughlinkstaterecords,andforfollowingpointersbetweenlinkstate orintheoperationoftheprotocolitself.forthatpurpose,weusetheprolingtoolgprofto conclusionthatcanbedrawnfromtheabovenumbers,isthatgiventhecostofde-allocatingthe spentinsideafunctionanditsdescendants.basedonthis,wendthatoftheoverallpathpre- databaseentries.theremaining40%includesthevariouscomputationsperformedbythealgorithm aswellasmanagementandupdatesofpathstructures,andallocationofnewpathentries.one estimateindividualoperationalcosts.thetoolprovidesuswithinformationaboutthetotaltime Network Size standardgatedroutingtablefordierentnetworksizes.asbefore,themeshtopologyisusedto generatenetworksofvaryingsize.itshouldalsobenotedthatalthoughthegatedroutingtable Figure5illustratesthedierencesinmemoryrequirementsbetweentheQoSroutingtableandthe 4.2.2MemoryRequirementsoftheQoSRoutingTable appeartohavethepotentialtoyieldsignicantperformanceimprovements. previousqosroutingtable,thememorymanagementoptimizationsthatwerepreviouslydiscussed, OSPFprotocolwasactive.Asaresult,thecomparisonbetweenroutingtablesizesismeaningful. Conclusionssimilartotheonesthatcouldbedrawnfromthepreviouscomparisonofprocessing costs,applytomemoryrequirements.specically,whilethememoryrequirementsforqosrouting areclearlyhigher,giventhecostandavailabilityofmemory,thedierenceisagainnotextremely signicant,e.g.,aboutafactorof2fortheaveragecase.thisdierencewastobeexpectedsince bothtablescontainaradixtreeofalldestinations,andtheqosroutingtablerequiresadditional storageforthepathstructures.itisthisadditionalstoragethatcorrespondstothedierences normallycontainsrouteslearnedfromalltheactiveroutingprotocols,inourexperimentsonlythe intheradixtreebasedonthedestination,andsecondasearchofthepathstructureassociatedwith valueandreturningasuitablepath(nexthop).accessingtheroutingtablerequiresrstalookup PathselectionconsistsofaccessingtheQoSroutingtableforagivendestinationandbandwidth showninfigure CostofPathSelection 18 Time (microsec)

19 SPF QoS Table Best Case QoS Table Average Case alongwhichaveragingcantakeplace.foragiventopologyanddistributionoflinkbandwidth, Figure5:Comparisonofmemoryrequirements averagingcanbedonebasedonthedestinationnodeandtheamountofrequestedbandwidth.the destinationaectsnotonlythecostofthelookupintheradixtree,butalsothepotentialdepthof thesearchinthepathstructureassociatedwiththedestination,asthenumberofentriesinthepath thedestinationuntilapathcapableofsatisfyingtherequestedbandwidthisfound.thecostof 0 structureislikelytodierfromdestinationtodestination.thesearchinthepathstructureisalso totheminimumhopcount.asfarastheaveragecaseisconcerned,therearetwodimensions bestcaseisobtainedbyforcingallpathstructurestoonlycontainoneentry,theonecorresponding theseoperationsisshowninfigure6forboththebestandaveragecases.asmentionedearlier,the Network Size aectedbytherequestedbandwidthaslargevalueswilltypicallyrequiresteppingthroughmore networklinks.thisresultsinthemaximumsearchtimethroughthepathstructure,sothatweare entriesinthepathstructure.inourmeasurements,weaveragebasedonlyondestinations,andthe ineectmeasuringanaverageworstcase. isavoidedbyforcingallrequeststobeforanamountofbandwidthlargerthanthecapacityofthe averageiscomputedacrossallpossibledestinationsinthenetwork.averagingbasedonbandwidth pathsandfewlinkdisjointpathsofhigherhopcount.thisagainminimizesanydierencebetween averageandbestcases.finally,thelimitedprecisionoftheprolingtoolmightalsohaveaected theresulttosomeextent.themainmessagefromfigure6isthatwhilethecostofpathselection causeisthearticialnatureofthemeshtopology,whichhasalargenumberofminimumhopcount boththeaverageandthebestcase,thedierencesbetweenthetwoshouldbeminimal.another rstisthatthelookupintheradixtreeisthedominantcomponent,andsinceitisidenticalfor fornetworksofsizeover300nodes.thereareanumberofreasonsfortheobservedbehavior.the Figure6showssimilarperformancefortheaverageandbestcasesfornetworksof50nodesand costsofqosroutingfromothercostsoftheroutingsystemsuchasrsvpandinterfacecosts,we mayslightlyincreasewithnetworksize,thissensitivityissmallandsoistheactualcostofselecting apath,atleastinthecontextofapre-computedqosroutingtable. simulatedthearrivalofarequestbydirectlyinvokingthepathselectionfunction. andpassedtogatedthroughthersvp-gatedinterface.sinceinthisstudywewanttoisolatethe Weshouldalsonotethatinarealoperationalenvironment,requestswillbeoriginatedbyRSVP 19 Size (bytes)

20 Time (microseconds) Figure6:Costofpathselection Best Case 9800 Average Case Network Size Average Case QoSroutingversionsofOSPF.TheyconsistofthecostofgeneratingandreceivingLSAs,that Thelastsetofparameters,whosecostswewanttoestimate,arecommontothestandardand Figure7:Costofaccessingthelinkstatedatabase isreceived,theinformationabouttheremoterouterneedstobeenteredinthelinkstatedatabase, 4.2.4LinkStateAdvertisementsGenerationandReception 3800 andthereforeadatabaseaccessneedstobeperformed. router,constructingthelinkstateadvertisementpacket,andoodingitoverallinterfaces.we areincurredinbothcases.generatinganlsainvolvesscanningalltheactiveinterfacesofthe assumedinourmeasurementthateachlsawassentinitsownpacket.conversely,whenanlsa ThetimerequiredforgeneratingandreceivingLSAswasmeasuredusingtwosimilarmachines Network Size machineswereconguredsoastoformanospfadjacency,andthenexchangelsasbetweenthem. connectedtoeachother,andrunninggatedwithourqosroutingenabledversionofospf.the TheOSPFclampdownmechanismbasedonMinLSIntervalandMinLSArrivalwasdisabled,in ordertogetaccuratemeasurementsofthelsagenerationandreceptioncosts. crosecondsinourtestsystem.thisishowevernotamarginalcostwhencomparedtosomeof Accordingtoourmeasurements,eithergeneratingorreceivinganLSAtakesabout200mi- Time (microsec)