Content and context aware networking using semantic tagging

Today’s model of networking primarily concentrates intelligence at the end hosts with the network itself offering a simple“best-effort”,“data agnostic ” communication medium. However, this paradigm has proven to be insufficient to meet todays needs co

ContentandContextAwareNetworkingUsingSemanticTagging

SethuramBalajiKodeswaran,AnupamJoshi

DepartmentofComputerScienceandElectricalEngineering

UniversityofMarylandBaltimoreCounty1000,HilltopCircle,Baltimore,MD21250Email:(kodeswar,joshi)@cs.umbc.eduAbstract

Today’smodelofnetworkingprimarilyconcentratesin-telligenceattheendhostswiththenetworkitselfoffer-ingasimple“best-effort”,“dataagnostic”communicationmedium.However,thisparadigmhasproventobeinsuf -cienttomeettodaysneedsconsideringthediversityofap-plicationsanddevicesthatarenetworked.Tooffervalueaddedservicestotheseendusersandapplications,moreandmoreintelligenceneedstobemigratedawayfromtheedgesandintothenetworkinacontrolledandtractablemanner.Inthispaper,wepresentourapproachofutilizingsemanticdatataggingtoprovidecontentlevelinformationfordatastreams owingthroughanetwork.Apolicybasedmanagementmechanismisutilizedwithinthenetworkfab-ricallowingrouterstoreasonoverthecontentandmakeintelligentdecisionsregardingthehandlingofdatapack-ets.Servicedifferentiation,in-networkcontentadaptation,traf cmonitoringandcontroletc.aresomeofthenewser-vicesthatcannowbeofferedbythenetworkinagenericand exiblemanner.Bydeployingourproposedarchitec-ture,anetworkneednolongerbeviewedasasimpledatatransportmediumbutratherasapolicy-controlledintelli-gentpacket/streamprocessorthatcanofferspecializedhan-dlingbasedonapplicationneeds.

presentourvisionofagenericand exibleframeworkthatenablestheincrementaldeploymentofintelligentservicesintothenetworkwiththeaimofoptimizingtheend-userexperiencefornetworkedapplications.Oneofthekeyen-ablersofthisapproachistoprovidenetworkrouterswithvisibilityintothetypeofdatathattheyarecarrying.Thiscontent-levelinformationcanthenbeusedbytherouterstomakemoreintelligentroutinganddatahandlingdeci-sions.Unlikeotherapproachessuchasactivenetworks,ourframeworkreliesonprovidingmetadataforthestreams,leavingtheactualdecisionmakingtotherouters,whicharecontrolledthroughrulesexpressedaspolicies.Apolicyen-forcerisbuiltintothearchitecturetoenforceruleswhichcaneitherbelocalorsystem-wide.Theserulesarespeci edbythenetworkprovidertocontroltraf c owsinthenet-work.Recentadvancesinnetworkprocessorspointtomoreandmoreprocessingpoweravailableatthenetworkrouters,andinmanycases,allowingforlimitedpacketprocessingatlinespeed.Further,industryinitiativessuchas[1]furthermotivatein-networkprocessingforspecializedhandlingofdatastreams.

Foranycontentlabelingsolutiontobeviableinalargescalenetwork,itmustbeboth exibleandgeneric.Havingaproprietarycontentlabelingschemedoesnotscalewellandforceseveryroutingentitytoknowhowtohandleeachcontentprovider’sindividuallabelingscheme.Forthisrea-son,wehavechosentouseRDF[3]asthemechanismtoconveythisinformation.Asdemonstratedinitsapplicationinthesemanticweb,RDF/RDFSisvery exible,genericandhasseenwidespreadacceptanceasadefactometa-datamarkupforwebcontent.ByutilizingRDF/RDFS(orOWLforschemade nition)asthemechanismtomarkup ows/packets,intermediaryintelligentroutingentitiescanusethismetadatatoreasonovertheirexistingknowledge-basetodeterminehowbesttohandleagiven ow.Inaddi-tion,inferencescanbemadetofurther“generalize”or“spe-cialize”agiven ow.Forexample,arouterthatcanpro-videspecializedhandlingforMPEG-4streamscanchoosetohandleaparticularpacketasthoughitwerepartofamul-

1Introduction

ThecurrentInternetwasoriginallydesignedtoprovide“best-effort”datatransportoverawiredinfrastructurewithendhostsutilizingalayerednetworkstacktoproviderelia-bility,qualityofservice,securityetc.foruserapplications.However,theproliferationofinelasticapplications,coupledwithwidespreadmigrationtowardshybridnetworksutiliz-ingwiredandwirelesslinksandtheplethoraofendhostvariantsrangingfromcellphonestoenterpriseserversne-cessitatesthemigrationofmoreandmoreservicesawayfromtheedgesandintothenetwork.Inthispaper,we

Today’s model of networking primarily concentrates intelligence at the end hosts with the network itself offering a simple“best-effort”,“data agnostic ” communication medium. However, this paradigm has proven to be insufficient to meet todays needs co

timediastream(ie.generalize)orapartofaPFrame(ie.specialize)dependinguponthegranularityofthedescrip-tionprovidedandtheknowledgebaseoftherouter.

ItcanbearguedthatoneofthemostimportantreasonswhythecurrentInternetissowidelyacceptedanddeployedisprimarilyduetoitslayeredarchitecture.Thenetworkstackisimplementedasdifferentlayers,whereeachlayeroffersawellde nedfunctionality,iswellencapsulatedandcanbedevelopedindependentoftheothers.Wellde nedSer-viceAccessPoints(SAPs)arespeci edtoclearlyde nethedataandprimitivesexchangedbetweenadjacentlayersinthestack.Whilethisisthecurrentlyacceptedmodelofanetworkstack,recentyearshaveseenincreasedinterestintheareaofcrosslayeroptimizations.Theideahereistotrytoprovidealayerinthenetworkingstackwithadditionalcontextualinformationsothatamoreinformeddecisioncanbemadeonthedatahandledbythatlayer.Proponentsofthismodelarguethatwithsuchinteractions,bandwidthus-ageoptimization,moreef cientrouting,betterQoSguar-antees,betterpowerutilizationetc.canbeachieved.Ourframeworkprovidesthemeanstoexploitsuchcrosslayertechniquestoallow“hints”providedbyhigherlayerstobeusedlowerdowninthestackandlikewise,applicationsrun-ningontopoftheframeworkcanalsoregistertobenoti edofchangesinthenetworkconditionsthataremeaningfultothem.Alocalpolicyenforcementenginecontrolsthein-teractionsthatarepermissible.Throughthis,systemwidepoliciescanbedistributedthroughoutthenetworkthatareenforcedlocallyandacrossthenetworktocoordinatetheinteractionsanddatatransport.

Therestofthispaperisorganizedasfollows.Section2presentssomeoftherelevantbackground.Section3presentsourproposedapproach.Section4presentsresultsofoursimulationstudieshighlightingbene tsofcontentawarenessinthenetworkfollowedbyourconclusion.

atedassemblylanguagecalledSpannertoencodeacom-pleteprogramintoasingleIPv4orIPv6datagram.TheSwitchWareproject[4]takesasimilarapproachutilizingaspecializedlanguagecalledPLANwhosecapabilitiesarerestrictedtoonlyperforming“safe”operationsonanynode.ANTS[34]utilizesacombinationofmobilecode,demandloadingandcaching.Similartotheothermodels,anANTSnetworkconsistsofnodesrunninganANTSplatform,pack-etsarereplacedwithsmartcapsulescapableofexpressingmainlyforwardingroutinesandmechanismsarebuiltintothemodelforon-demandcodedissemination.Othersim-ilarapproaches[19]havefocusedonensuringquickjust-in-timecompilationofdownloadedcodetooptimizetheirperformancealongwithstaticlanguageveri cation.TheActiveSignalingProtocol(ASP)[13]takesadifferentap-proachtoactivenetworksbygoingawayfromthecapsulemodelandhavingdatapacketscarryareferencetosomeportablecodewhichcanbedownloadedonthe yandrunonanASPExecutionEnvironmentasnecessarytoenableanewactiveapplication.ActiveServices[5]takesanal-ternatemodeltointroductionofintelligenceintothenet-workbyrestrictingtheintelligencemainlytotheapplica-tionlayertherebypreservingtheroutingandforwardingse-manticsoftheInternetarchitecture.Asimilarapplicationlevelprocessingofuser-dataisconsideredin[10]tohandlecongestioncontrolinthenetwork.Thearchitectureallowsforapplicationstospecifyintra-networkprocessingbyus-ingActiveProcessingFunctionIdenti ers(APFI)sothatbandwidthallocatedcanbeintelligentlyreducedinaman-nertailoredtotheapplicationratherthangenerically.Whileactivenetworksareaverypowerfulparadigm,thereareseveraldrawbacksthathavehinderedtheiracceptanceandwidespreaddeployment.Theprimaryconcerniswithsecurityandprotectionagainstactiveroutersrunningmali-ciouscodehiddeninactivepackets.Approachesthatlimittheallowableoperationstoasafesubsetlimitthepowerofspecializedhandlingthatispossible.Anequallyimportantissueisthatnetworkoperatorsprefertohavecompletecon-trolovertheirnetworkandallowinguserde nedcodetobedownloadedandexecutedthatcouldpotentiallymakerout-ingdecisionsfordatastreamsisnotveryamenableforac-ceptance.Manyapproachesutilizedcustomizedplatformsandlanguagesraisinginteroperabilityconcerns.Also,thestateofthetechnologyinnetworkprocessorsatthattimepresentedchallengesforthetruerealizationofthisap-proach.Withourframework,weareworkingtowardsthesamegoalofmakingnetworkssmarterbutdoingsowithacompletelydifferentapproachthanactivenetworks.Ourframeworkallowsforthespeci cationof“whattypeofdata”is owingthroughthenetworkratherthanspecifying“howthatdatashouldbehandled”.Essentially,ourframe-workprovidesthehintstothenetworkfabricthroughcon-tentmetadatabuttheactualdecisionofhowbesttohandle2

2RelatedWork

2.1ActiveNetworks

ActiveNetworksisoneoftherevolutionaryapproachestoinjectintelligenceintothenetwork.Atanextremecase,thiscanbeviewedasanefforttoaugmentdatapacketswithcodefragmentscontainingspecializedprocessinglogicforhandlingthatpacket.“Active”routersexecutethecodecar-riedinthepacketallowinghighlycustomizedhandlingof owsorpackets.Thisarchitecturepermitsmassiveincreaseinthecomputationperformedwithinanetworkallowingforthedeploymentofnewservicesintothenetworkinatotallyseemlessandon-demandmanner.Severalprojectshavetakenthisapproachtoimpartingintelligenceintothenet-work.TheSmartPackets[29]approachutilizesaspecial-izedprogramminglanguagecalledSprocketandanassoci-

Today’s model of networking primarily concentrates intelligence at the end hosts with the network itself offering a simple“best-effort”,“data agnostic ” communication medium. However, this paradigm has proven to be insufficient to meet todays needs co

thattypeofstreamislefttotherouterandthenetworkop-erators.Also,theuseofastandardmechanismtodescribemetadataprovidesaviablesolutiontotheinteroperabilityconcerns.Thepolicymechanismusedinourframeworkal-lowsthenetworkoperatortospecifyhandlersthatcanbetriggeredbasedonreasoningoverthecontentmetadatatoinvokespecializedhandling,wherethehandlersareessen-tiallyvalueaddedservicesofferedbythenetworkproviderthemselvesandthereby,trusted.

2.2ContentAdaptation

Contentadaptationisacommonlyusedparadigmtoad-dressissuessuchasef cienttransferofbandwidthinten-sivedataoverabandwidthlimitednetwork,makingthecontentamenabletodeviceswithdiverse/limitedcapabil-ities,personalization,policy-basedadaptations,securityre-latedadaptationsetc.TheOpenPluggableEdgeServices(OPES)[7]isanIETFworkinggrouptaskedwithde n-ingacommonarchitectureforenablingedgebasedcontentadaptation.Theideaistospecifyaplatformprovidingnet-workedservicesattheapplicationlevelforoffloadingori-ginserversandimprovinguserexperience.OPESisgen-erallyapplicableforoperatingonrequestorresponsefordatausingHTTP/SMTP/RTSPwithrulesspeci edinIn-termediaryRuleMarkupLanguage(IRML)[8]todeterminewhatserviceshouldbeusedonwhattypeofcontent.Con-tentadaptationhasalsobeenstudiedextensivelythroughproxybasedsystems.[14]presentsanapplicationlevelcontentadaptationformultimediadataasanef cientso-lutionforhandlingdisseminationofrichcontenttoawidevarietyofdevicesconnectedoverlinkswithdiversechar-acteristics.[16,18]arguesthatwiththediversityofenduserdevices,on-the- yadaptationbytranslationalproxiesattheapplicationlevelisbothanecessaryandacosteffec-tive, exiblesolution.Adaptationofcontenthereisthroughdata-typespeci clossycompressionanddistillation.TheConductor[35]frameworkisasimilarapproachthatcandy-namicallydeploymultipleadapterstooperatealonganap-plication’scommunicationpath.Aplanningalgorithmisbuilt-intodeterminewhatadapterstouseandwhere.

[26]presentsaframeworkforadaptingmultimediawebdocumentstooptimallymatchtheneedsoftherequestingclientdevicesusingamultimodal,multiresolutionrepre-sentationhierarchyforthecontentandaCustomizerthatpicksthebestcontentrepresentationthatprovidesmaxi-mumvaluetotheclient.Inthismodel,thecontentauthorprovidesthetranscodingpoliciesandcontrolstheadapta-tionandthisisdoneinanof inemanner(asopposedtoonrequestinmostproxybasedtranscodingsystems).[31]presentsaContent-awareActiveGateway(CAG)architec-tureallowingforonthe ycontentadaptation.Theyrelyontheabilityto lterspeci ctypesoftraf cbyidentifyingpat-3

ternsintheheaderorpayloadofpackets.Thegeneralideaistorunspecializedprograms(staticallyinstalledordown-loaded)onenterprise/residentialgatewaysthatcanregister lters(portnumber,IPaddressetc)onthepacketroutingfabric.These ltersareactivatedwhenpacketsmatchingthosecriteriaareencounteredandpassedtotheCAGcom-putationlayerforprocessingatclosetolinespeed.[9]isaproxybasedwebcontentadaptationforsupportingbrowsingbymobiledevicesoverwirelesslinksusinguserspeci edpreferences.[32]usesalinklevelredirectionin-frastructurecalledSelNetthattagspacketsatlinklevelwithfunctionidenti erstoenableaproxybasedcontentadapta-tion.[6]proposeaservercentriccontentadaptationframe-workcreatingservicespeci coverlaynetworksthroughtheuseofdynamicproxiesalongthedatapath.[30]pro-poseaclientcentriccontentadaptationusingWebStreamCustomizers(WSC)allowingforsystembasedandcontentbasedcustomization.Inadditiontoresearchprojects,sev-eralcommercialestablishmentshavebeenlaunchedwithcontentadaptationastheirprimarybusinessfocussuchasVoiceAgeNetworks,Volantis,Adamind,LightSurf,Sens-eStream,Mobixelletc.

Oneofthegoalsthatwehaveidenti edforourframeworkistosupportcontentadaptationfordatastreams.Ourap-proachofusingsemantictagsembeddedinthestreamwillallowcontentprovidersto“signal”theiradaptationpoliciesandrequirements.AsdiscussedinSection3,thiscanbebothin-bandorout-of-band.Anintelligentstreamproces-sorcanreasonoverthisinformationandusingitsexistingknowledgebase,theappropriatecontentadaptationservicescanbeapplied.

2.3CrossLayerTechniques

StrictLayeringhasbeentheprimarydesigntechniquefornetworkingstackswhichhasresultedinsimple,scal-ableandinteroperablesolutions.Proponentsofthelay-eredmodelsuchas[21]arguethatcrosslayeroptimiza-tionbreaksthearchitecturalsimplicityoflayeringandistooshortterminvisionandhighlyspeci c(totopology,tech-nology,applicationetc.)leadingtospaghettiimplementa-tionsandwillnotseewidespreaddeployment.Thecounterargumenthasbeenthatrestrictingtheinteractionsbetweenlayerstodataandwellde nedprimitivesimposesaratheruniformhandlingsemanticfordatastreamswhichisnotal-waysthemostef cient.Muchoftheworkoncrosslayeroptimizationshasfocusedonoptimizationsatthenetworkedgesprimarilyforwirelesstechnologies.In[17],inganEligibleRateEstimate,thequal-ityofthevideoisdynamicallyadjustedtooffertheuser,uninterruptedvideoatthebestqualitythatcanbesustained.MobiWeb[25]isaframeworkforsupportingadaptiveap-

Today’s model of networking primarily concentrates intelligence at the end hosts with the network itself offering a simple“best-effort”,“data agnostic ” communication medium. However, this paradigm has proven to be insufficient to meet todays needs co

plicationsthathavereal-timecharacteristicsoverwirelesslinks.Aninter-streampriorityschemeisusedtoaddresstheshortterm uctuationsofwirelessstreamsleavinglong-termadaptationtotheapplication.[33]presentsacrosslayeralgorithmthatadaptsthelinklevelARQtotheend-to-endpacketlossobservedbyTCP.TheideahereistoadapttheMAClayerretransmissionaccordingtoatargetlossrateusedasparametertodescribethedesiredQoSforaTCPconnection.[28]presentanapplicationlevelARQmecha-nismfor802.11MAC.Heretheapplicationdecideswhichpacketsrequireretransmissionandjustthoseareretrans-mitted.[24]presentsatransportlayersolutiontoenhancingnetworkQoSusingapplicationlevelinformation.Hereap-plicationscansplitdataintodifferentstreamswhichcanbedeliveredacrossdiverseroutessothatmorecriticaldataistransmittedoveramorereliableconnectionwhilethelesscriticaldatacanfollowalessreliableroute.[23]presentsanapproachtosupportcongestioncontrolledmulticastreal-timecommunicationusingself-organizedtranscodingtohandlelocalrepairinwirednetworks.Thetranscodingparametersareconstantlyupdatedtore ecttherealtimestateofthelinkbeingusedtoservethegrouptoprovideanetwork-friendlycongestioncontrol.

Ourproposedframeworkprovidesthenecessarysupportforimplementingcrosslayeroptimizationsatthelocalnodelevel.Oneoftheprimaryconcernsagainstcrosslayertech-niquesisthatexposingtoomuchinformationbetweenlay-ersmakesthesystemoverlycomplicated.Tothisend,ourframeworktakestheapproachofparameterizingtheinter-actionssimilarto[15].Thegoalistoexposemeaningfulparametersandtheirassociatedcostsateachlayerwhilekeepingtheactualmechanismofachievingthemhiddenwithinthatlayer.Inourframework,thelocalpolicyen-forcercanspecifyparameterstobeusedacrossthedifferentlayersofthestacktomeetapplicationneeds.Theobjec-tiveistokeeptheinteractionstractablewhilestillachievingtheef cienciesofcrosslayertechniques.Anadditionalandnecessarygoalistocoordinateinteractionsacrossanetworksincesometechniquessuchaslinklevelreliabilityneedtobeappliedatasystemleveltoreallyseethebene ts.Ourapproachtocrosslayeroptimizationsgoesonestepfurtherbyallowingforthespeci cationofsystemleveloptimiza-tionsthatcanspanmultiplenodes.Systemlevelinforma-tioncanbeexchangedbetweennodestoinformeachnodeofthedynamicsystemvariationsthatcanthendrivethelo-caloptimizationsthatareapplied.

reservationscheme.Thebasicmodelutilizesapacketclas-si er,scheduler,admissioncontrolandareservationsetupmechanism.ResourceReservationProtocol(RSVP)[12]isthesignalingprotocolusedtoconveya“ ow-spec”and“ lterspec”tointermediaryroutersdescribingthetraf ccharacteristicsandtheresourcereservationsthatneedtobemade.CloselycoupledtoIntServisthepolicyadmis-sioncontrolframework[36]allowinganetworkoperatortospecifypolicybasedadmissioncontrolrulesthatcanbelookedupandenforcedattheroutingelements.TwokeycomponentsarethePolicyDecisionPoint(PDP)andthePolicyEnforcementPoint(PEP).ThePEPreceivesrequestforresource(throughRSVP).ThePEPcontactsthePDPtodeterminewhatpolicytouse.ThePDPusesapolicydata-basetoconveyaparticularpolicydecisiontothePEP.EachPEPalsohasalocalPDPthatcanusedincasethePDPisnotreachable.Throughthismechanism,admissioncontrolpoliciessuchastimeofday,SLAs,prioritization,prepaidtransactions,senderspeci edrestrictionsetc.canbeen-forced.OneofthemaindrawbacksofIntServisthemain-tenanceofsoftstaterepresentingper owresourcereserva-tionsonrouterspreventingitfromscalingforlargenetworksizes.Also,the ltersarebasedonpacketheadersandhavelimitedexpressivity.

DifferentiatedServices(DiffServ)[27]takesadifferentap-proachtoenablingQoSguarantees.Theideahereistomoveawayfromaper- ow, negrainQoSmodeltoamorecoarsegrainedQoSmodeltoallowforscalingtolargenet-works.TheapproachfollowedistousetheIPv4TypeofService(TOS)ortheIPv6Traf cclassbytetoconveyag-gregateQoSrequirements.6bitsofthis eldareusedtospecify64DifferentiatedServicesCodePoints(DSCPs).EachDSCPmapstoaspeci cPerHopBehavior(PHB)ateveryintermediaryrouter.Packetsaremarkedtocorre-spondingDSCPsatthenetworkingressbasedonServiceLevelAgreements.OneofthemainproblemswithDiff-Servisthatsinceitiscoarsegrain,itisreallyusableonlybyISPsandnotbyendusers.Also,markingisstaticanddoesnotensurethatadequateresourcesareactuallyavail-abletohandletheQoSrequirements.ThereisalsolimitedexpressivityasthenumberofDSCPsislimited.

Forourframework,weuseacombinationofapproaches.WeuseaPEP/PDPlikeapproachtoallowforpoliciestobespeci edthatarelocallyenforcedandcanspeci edatasys-temlevelbyanetworkoperator.SimilartotheDiffServap-proach,werelyonpackettaggingtoofferdifferentlevelsofservice.However,unliketheDSCPtags,werelyoncustomtagsthatpointtosemanticmetadataaboutthecontent.Thismetadatacanberetrievedandreasonedovertodeterminewhatpoliciesneedtoenforcedforthatstream.Forexam-ple,apolicyspeci edbyanenterprisethatdropsallmusicMP3streams owingintotheirof cenetworkshouldstillallowanof cialpresentationcontainingMP3audiotoen-4

2.4ServiceDifferentiation

Goingbeyondthe“besteffort”paradigmofthecurrentInternethasbeenthegoalofseveralinitiatives.IntegratedServices(IntServ)[11]proposesanarchitecturefora negrainedQoSsystemthatisbasedonaper- owresource

Today’s model of networking primarily concentrates intelligence at the end hosts with the network itself offering a simple“best-effort”,“data agnostic ” communication medium. However, this paradigm has proven to be insufficient to meet todays needs co

terthenetwork.Approachesbasedonpacket lteringandheadercheckingcannotofferthislevelof exibilitysincetheylackinferencecapabilities.Likewise,applicationsrun-ningontopofourframeworkcanspecifytheirresourceneeds,withthisinputvaryingindetailfrombeinghighlyspeci ctojustcommunicatinghighlevelsemanticinfor-mationlikethe“ owoftypeJPEGwithframeratexandqualityy”.OurframeworkrunningontopofanIntservorDiffServnetworkwillbeabletoconverttheapplicationrequesttothecorrespondingrealizationmechanismsuchassignallinganRSVPreservationormappingthedatapacketstothecorrespondingDSCPasthecasemaybe.

3.1Node

Level

2.5SemanticMarkup

ResourceDescriptionFramework(RDF)[3]isaframe-workforrepresentingmetadataregardingwebcontentusingXMLastheencodingmechanism.RDFprovidesastan-dardframeworkforinterchanginginformationacrossappli-cations.RDFisbasedontheideaofidentifyingobjectsusingUniformResourceIdenti ers(URI)anddescribingresourcesintermsofsimplepropertiesandpropertyvalues.Astatementisatriplespecifyingasubject(usingURI),apredicate(representingaproperty)andanobject(repre-sentingthevalue).Thisresultsinstructuredgraphswithnodesforsubjectsandobjectswitharcsrepresentingthepredicates.RDFSchema[2]isthemechanismforspecify-ingvocabulariesthatcanbeusedtoformRDFstatements.Thevocabularygivestheactualmeaningtothestatement.Throughthis,objectsandtheirpropertiesandwhattheyre-fertocanbespeci ed.Morepowerfulschemade nitioncanbeprovidedusingOWL.

Providingmarkupforcontentisprevalentinmanyotherar-easalbeitnotalwaysusingastandardmachineprocessablemeanssuchasRDF.SessionDescriptionProtocol(SDP)-Anytimeisaspeci cationforaudio-videocontentmarkup.ID3tagsappliedtoMP3 lescanconveymetadatasuchastitlesandcomposers.ManyoftheseapproachesuseunstructuredXMLwhichisdif culttouseforconveyingknowledge.WehavechosentouseRDFforourframe-workmainlytoprovideastandardandsimplemechanismthroughwhichdatastreamscanbedescribed.Themainob-jectiveistokeepourframework exibletoallowfortheintroductionofnewconceptsasneededratherthanenforcearigidschematobefollowedbyallparticipants.

Figure1.CoCoNetNodeFramework

Atthenodelevel,thearchitectureweproposeintroducesourNodeFrameworkasanadditionallayercalledtheCo-CoNetlayerbetweentheapplicationandthetransportlayer.Thislayerisresponsibleforinterceptingsocketcallsmadebyapplicationstothetransportlayer.TheAPIisenhancedtoallowtheapplicationtoprovidesemanticlevelinforma-tionformessagestransmittedoverthisinterface.ALocalPolicyDecisionPoint(LPDP)isusedtodeterminewhatpoliciestoenforcebasedonthecontent.Inourframework,eachPolicyEnforcementPoint(PEP)isateverylayerinthenetworkingstackwhile[36]treatsthePEPatanodelevel.PlacementofthePEPateverylevelofthestackal-lowsustoimplementcoordinatedcrosslayerinteractionsinitiatedandcontrolledbyourframework.ThePEPex-posestheinterlayeroptimizationpointsthatanyparticularlayersupports.TheframeworkutilizesthepoliciesstoredintheLPDPtodrivethesettingstobeappliedtoeachofthePEPsinthestack.Essentially,weareproposingtoex-poseanetworkstackasacollectionofswitchesanddialsandallowanexternalpolicytodeterminetheexactsettingsofeachofthesedials(basedoncontentandcontext).Wewanttoexposefunctionality,notnecessarilythemechanismofhowitisachieved(thisfallsunderintra-layeroptimiza-tion).Forexample,aMACcanadvertisetwodifferentdataratesandtheirassociatedpacketerrorprobabilitieswith-outexposingtheFECschemeusedtoachievetheserates.Thepoliciescanbespeci edasproductionrules(if(condi-tion)then(action))orevent-condition-actionrules(oneventif(condition)then(action)).Inessence,theNodeFrame-workprovidesarich,extensibleoptionforrealizingpolicycontrolledcross-layerinteractionswithinanode’snetworkstack.Byparameterizingthepossiblesetofinteractions5

3ProposedApproach

Thissectionpresentsanoverviewofourproposedframework.Webreakitdownintotwocomponents;atanodelevelandatasystemlevelthatspansthenetwork.

Today’s model of networking primarily concentrates intelligence at the end hosts with the network itself offering a simple“best-effort”,“data agnostic ” communication medium. However, this paradigm has proven to be insufficient to meet todays needs co

thatarepermissible,thecrosslayerinteractionsarekepttractablewithoutmakingtheimplementationoverlyconvo-luted[21].

Alayer2LSP.

ADiffServawarenetwork. AnIntServawarenetwork.

ACoCoNetRouterFrameworkwillperformthenecessarymappingbasedonpolicy,contentandcontext.Forinstance,supposeapacketarrivesatarouterindicatingthatitre-quiresreliabletransfersemantics.Thedataplanechosentothenexthop,inthiscase,couldbeoveraTCPconnec-tion.Likewise,adatapacketindicatingthatitissensitiveinformation(telnetloginsforexample)butcurrentlynoten-cryptedcanberoutedtothenexthopoveranIPSECtun-nelordroppedifnoneisavailable(ifthatisthepolicy).WhereaCoCoNetRouterFrameworkrunsisveryimple-mentationdependent.Forexample,incaseofawirelessadhocnetwork,everyhostisarouterandhencecanpoten-tiallyruna(albeitsimpli ed)CoCoNetRouterFramework.Likewise,inanenterprisesetting,thehostmachineswithintheenterprisewilllikelyrunonlytheCoconetNodeFrame-workwithonlytheexteriorgatewayroutersrunningtheCo-conetRouterFramework.AnetworkserviceproviderwillmostlikelyhaveonlyedgeroutersruntheCoconetRouterFrameworkleavingthecoreoptimizedforfastdata owhandling.TheroleoftheGlobalPolicyDistributionPoint(GPDP)istodisseminateanynetworkwidepoliciesthatneedtobeenforced.Thiscanincludeitemssuchaspref-erentialtreatmentthatneedstobegiventocontentorigi-natingfromaparticulardomain,preferentialtreatmentforaparticulartypeofcontent,anycontentbasedadaptationtechniquesthatneedtobeemployedinthenetworketc.ItisenvisionedthattheGPDPiscontrolledbytheISPtosetforthglobalruleswhiletheLPDPhostedatanenterpriselocationispossiblysharedbetweentheISPandtheenter-prise.ThiscanfurtherbeextendedtosaythattheLDPDisunderlocalusercontrol(basedonuserpoliciesandpref-erences)andcanadditionally,hostuserpreferences.Inor-dertopropagatecontentlevelinformationforpacketsand ows,weproposetotakeoneofthefollowingapproaches. ThemetadatacanbedirectlyencodedintotheIPop-tions eldofanIPpacket(sizeisanissue).Werefertothisas“in-bandtagging”.

IPpacketsusetheIPoptions eldtocarryaspecial32bitidenti er.Thisidenti erisusedtoindicateawellknown(global)e-fulfornon- owbasedone-timepacketssuchastelnetloginorHTTPURLrequest.Werefertothisas“out-of-bandtagging”.

IPpacketsusetheIPoptions eldtocarryaspecialkey.Thiskeyislookedupinadirectoryservicetoidentifythemetadatadescribingthatpacketor ow.Thisisalso“out-of-bandtagging”.Usingastructured6

3.2Network

Level

Figure2.OverlayNetwork

Atthenetworklevel,weenvisionthattherewillbeanoverlaynetworkcomprisedofroutersthatruntheCoCoNetRouterFramework.ClientmachinesrunningourNodeFrameworkcommunicateoverthisoverlay.Theoverlaycomprisesoftwocomponents;

AcontrolplanecomponentthatinvolvesinteractionsbetweentheCoCoNetRouterLayersattheroutingel-ements.

Adataplanecomponentthroughwhichthedatapack-etsare owing.

OvertheCoCoNetRoutercontrolplane,routerscanex-changetraditionalmanagementinformationsuchaslinkstates,bufferlengthsetc.Inaddition,informationsuchascontenttypescurrentlybeinghandled,adaptationscurrentlyavailablecanbeadvertised.Anadditionalkeypieceofin-formationexchangedisthelocalpoliciesthatarecurrentlybeingappliedtoadatastreamthatisbeingrouted.LocalPEPsettingsforagivenstreamor owhaveglobalimpli-cations.Forexample,unlesseveryhopisreliable,adatapacketcannotbereliablyroutedthroughanetwork.Thedataplanecanbeimplementedaseither: AUDPconnectionbetweentworouters. ATCPconnectionbetweentworouters. AnIP-in-IPtunnelbetweentworouters.

Today’s model of networking primarily concentrates intelligence at the end hosts with the network itself offering a simple“best-effort”,“data agnostic ” communication medium. However, this paradigm has proven to be insufficient to meet todays needs co

Peer-to-Peeroverlaytoallowforkeylookupswillbeneededforthisoption.Inthisapproach,beforeaclientstartsa owthroughanetwork,itregistersitscontentmetadatawiththe rsthoprouterandgeneratesakey.ThiskeyiscarriedintheIPpacketsandisavailabletoanyintermediaryrouter.Atanypointalongthedata ow,thiskeycanbeusedbyintermediaryrouterstofetchthemetadatathroughanoutofbandmechanism.Theinformationconveyedinthemetadataisreallyleftuptotheapplication.Forexample,anMP3streammayhavethefollowingdescriptionwhichcanbeusedbytheexam-pleenterprisedescribedintheearliersectiontodeterminewhethertodropornot(inthiscase,thestreamisallowedtopasssinceitispartofof cialbusiness).

<?xmlversion="1.0"?><rdf:RDFxmlns:rdf=

"/1999/

02/22-rdf-syntax-ns#"

xmlns:mmschema=

"/mms#"><rdf:Descriptionrdf:about=

"/SalesReport.mp3">

<rdf:typerdf:resource=

"/mms#audio"/><mms:LengthInMin>5</mms:LengthInMin><mms:LengthInMB>4</<mms:LenghtInMB><mms:technicalType>

/mms#MP3</mms:technicalType><mms:semanticType>

/mms#Lecture</mms:semanticType></rdf:Description>

Asimilarframeworkisproposedin[22]specifyingameta-dataarchitectureforapolicybasedcontentdeliverynet-work.Here,theauthorsenvisionusingametadatagate-waycapableofextractingnetworkdeliveryrelatedmeta-datafromcontentmetadata.ThisdeliveryrelatedmetadataisusedtoimplementnetworkservicessuchasQoSandcon-tentcachingcontrol.Ourapproachisdifferentmainlyinthatweencodethecontentmetadataontothestreamsthem-selves(in-bandorout-of-band).Theintermediaryrouters(andnotjusttheedgeasin[22])cantheninferbasedonlocalknowledgebaseandcontext,whatpoliciesareapplica-bleandtakelocalaction.Thisallowsforon-demandcon-tentadaptationbasedontransientconditionsasopposedtoend-to-endadaptations.Furthermore,providingcontentin-formationsothataroutercandifferentiatebetween,forex-ample,videostreamingfromasurveillancecameraanda

7

streamingmovieallowsthenetworktomakesmartdeci-sionsonroutingdatastreamsacrosslinkswithdifferentre-liabilitycharacteristics.Also,forourarchitecture,weareusingRDFwhichprovidesagenericmechanismtoconveymetadatawhichcanbereasonedover.

4PreliminaryResults

Toinvestigatesomeofthepotentialbene tsthatcon-tentawarenesswithinthenetworkcanbring,wesimulatedacrosslayerapproachwheretheroutinglayerproactivelymaintainsalternateroutesthatcanbeusedimmediatelyuponfailureoftheprimaryroute.Weconsideredawire-lessnetworkrunningtheDSR[20]routingprotocoloverwhicha2000frameMPEG-4videoisstreamed.DSRisareactiveprotocolie.routesarediscoveredwhenneededbyanapplicationordatapacket.Dataisthentransmittedthroughsourceroutingandwhentheroutebreaks,thenodeupstreamfromthebreaksignalsarouteerrortothesourceandcanoptionallytrytosalvagethecurrentpacketthroughre-routing.Routesarediscoveredtoadestinationonde-mandbythesourcebybroadcastingarouterequestmes-sage.Onekeybene twithDSRisthefactthatasourcecan,inonerequest,potentiallyreceivemultiplerouteresponsesindicatingdifferentpathsthatcanbeusedforroutingdatafromsourcetodestination.DSRthenreliesonsourcerout-ingtotransferdata.Inourwork,weexploittheinher-entmultipathcapabilityofDSRbymaintainingalternateroutes.WeassumethattheapplicationmarkspacketsasIntracodedFrames(IFrames),Predictive-codedFrames(PFrames)orBidirectionally-interpolatedFrames(BFrames).The“marking”inoursimulationissimple;theapplicationinterfacetothetransportlayerhasbeenextendedtocarryalongwitheachsegmentofdata,agloballyknownidenti errepresentingtheframetype.OurextensiontoDSRutilizesthe“lessimportantdata”(BFramesinthiscase)astracerstocheckfortheviabilityofanalternateroute.Therationalhereisthatevenifwelosethisframe,theuserisnotgoingtoseeabigdropinthequalityofthereceivedsignalsincethesearethemostcompressedframes.WeusedNS2foroursimulationsandconsideredanareaof1000mx1000mwithvaryingnumberofnodesrangingfrom25to100withspeedsbetween0and20m/susingthe802.11MAC.ThetwoapproachescomparedwereregularDSRandourcon-tentawareenhancement.FortheregularDSR,theonlymodi cationmadewastoturnoff owstateanddisabledrouterequestsfromintermediarynodesforthepurposeofsalvaging.Withourenhancement,thesourcenodecontin-uestousetheprimarypathforbulkofthedatatransfer.Periodically,thesourcewillpickaBFrame(theapplica-tionspeci edmarkingisusedtodetectattheroutinglayer,whichframesareBFrames)toberoutedalongalternateroutesthatwereobtainedthroughtheinitial(andsnooped)

Today’s model of networking primarily concentrates intelligence at the end hosts with the network itself offering a simple“best-effort”,“data agnostic ” communication medium. However, this paradigm has proven to be insufficient to meet todays needs co

routediscoveryprocess.Theaimhereistousealessim-portantframetovalidatealternateroutessothatwhentheprimaryroutefails,acachedalternateroutecanbeusedthathasahighprobabilityofbeingvalid.Fig.[3,4,5,6]

show

Figure5.PFrame

Losses

Figure3.HFrame

Losses

Figure6.BFrameLosses

testalternateroutesandremovethemfromcacheifbro-keninaproactivemanner.Fig.[7]showstheimpactofourenhancementasthetransmissionrangeof802.11isin-creasedkeepingthenumberofnodes xedat100andspeedat15m/s.Atlowranges,ourenhancementbehavessimilartoregularDSRmainlyduetothelackofalternateroutes.Astherangeincreases,ourenhancementstartstoshowim-provementsovertheregularscheme.Fig.[8]showstheim-pactofspeedofthemobilenodes.Atlowspeeds,bothschemesarecomparablesinceroutebreakagesarenottoofrequent.Withincreasingspeeds,routesbreakmorefre-quentlywithneighborsgoingoutofrangewhereouren-hancementshowsmarkedimprovementoverregularDSR.

Figure4.IFrameLosses

theimprovementsinthelossesobservedforH,I,PandBFrameswhenregularDSRiscomparedagainstouren-hancedversion.Inthisexperiment,wemodi edthenumberofnodesinthenetworkwhilekeepingthespeedofnodes xedat15m/sand802.11rangeto250m.Foreachnodecount,thesimulationwasrun5timeswithdifferentrandommobilityscenariosandtheresultsaveraged.Inallcases,ourenhancedversionout-performsregularDSRwithverylittleoverhead(salvagingifthealternateroutechosenforvalidationisbroken).Whenaroutebreakoccursinregu-larDSR,acachedalternaterouteischosen.Inmanycases,thisroutealsoendsupbeingbrokenbutthesourcedoesnotknowthisuntilitreceivesaroutebreakerror.Ifavalu-ableframesuchasaIFramehadbeentransmittedoverthisbrokenroute,ingcontentinformation,ourenhancementisabletoidentifylowimpactcandidateframestouse

to

8

5Conclusion

Inthispaper,wepresentanarchitecturetoenableintel-ligentprocessingofdatastreamswithinthenetwork.Ourarchitecturereliesonsemantictagsprovidingcontentmeta-datathatcanbereasonedoverontheroutingelementstoenablespecializedhandlingfor ows.Ouraimistofacili-tateapplicationsanduserstocommunicatetheirpreferences

Today’s model of networking primarily concentrates intelligence at the end hosts with the network itself offering a simple“best-effort”,“data agnostic ” communication medium. However, this paradigm has proven to be insufficient to meet todays needs co

Figure7.FrameLossesvs.

Range

Figure8.FrameLossesvs.Speed

tothenetworkandletthenetworktrytoprovidetheneededlevelofservice.Policiesareusedtodriveadaptationsbothlocallyatanodeandcoordinatedacrossanetwork.Con-trolledcrosslayerinteractionsareusedtoprovidemoreef- cientprocessingwhilestillretainingtractability.Wehavealsopresentedtheresultsofoursimulationstudiesshowingthatcontentawarenessattheroutinglayeroffersbene tsovertraditionalcontentagnosticapproachestodatadistrib-utionoverawirelessnetwork.

References

[1]Application-OrientedNetworking(AON)

.

[2]RDFVocabularyDescriptionLanguage1.0:(RDFSchema)

/TR/rdf-schema/.[3]ResourceDescriptionFramework(RDF)

/RDF/.

[4]D.S.Alexander,W.A.Arbaugh,M.Hicks,P.Kakkar,

A.Keromytis,J.T.Moore,C.A.Gunter,tles,andJ.M.Smith.TheSwitchWareactivenetworkarchitecture.IEEENetworkMagazine,12(3):29–36,1998.SpecialissueonActiveandControllableNetworks.

9

[5]E.Amir,S.McCanne,andR.H.Katz.Anactiveser-viceframeworkanditsapplicationtoreal-timemultimediatranscoding.InSIGCOMM,pages178–189,1998.

[6]S.Ardon,P.Gunningberg,ndFeldt,M.P.Y.Ismailov,

andA.Seneviratne.March:adistributedcontentadaptationarchitecture.InternationalJournalofCommunicationSys-tems,SpecialIssue:WirelessAccesstotheGlobalInternet:MobileRadioNetworksandSatelliteSystems.,16(1),2003.[7]A.Barbir,R.Penno,R.Chen,M.Hofmann,andH.Orman.

RFC3835:AnArchitectureforOpenPluggableEdgeSer-vices(OPES),August2004.

[8]A.BeckandM.Hofmann.IRML:ARuleSpeci cation

LanguageforIntermediaryServices:draft-beck-opes-irml-00.txt,Febuary2001.

[9]H.Bharadvaj,A.Joshi,andS.Auephanwiriyakul.AnAc-tiveTranscodingProxytoSupportMobileWebAccess.In17thIEEESymposiumonReliableDistributedSystems,Oc-tober1998.

[10]S.Bhattacharjee,K.L.Calvert,andE.W.Zegura.Onactive

networkingandcongestion.TechnicalReportGIT-CC-96-02.

[11]R.Braden,D.Clark,andS.Shenker.RFC1633:Inte-gratedServicesintheInternetArchitecture:anOverview,June1994.

[12]R.Braden,Ed.,L.Zhang,S.Berson,S.Herzog,and

S.Jamin.RFC2205:ResourceReSerVationProtocol(RSVP)—version1functionalspeci cation,Sept.1997.[13]B.Branden,B.Lindell,S.Berson,andT.Faber.TheASP

EE:AnActiveNetworkExecutionEnvironment.InDARPAActiveNetworksConferenceandExposition(DANCE2002),pages238–254,June2002.

[14]S.Chandra,C.S.Ellis,andA.Vahdat.Application-LevelDifferentiatedMultimediaWebServicesUsingQual-ityAwareTranscoding.InIEEEJournalonSelectedAreasinCommunications,volume18,December2000.

[15]L.-U.Choi,M.T.Ivrlac,E.Steinbach,andJ.A.Nossek.

Bottom-upapproachtocross-layerdesignforvideotrans-missionoverwirelesschannels.In62ndIEEEVehicularTechnologyConference(tobepublished),September2005.[16]A.Fox,S.D.Gribble,Y.Chawathe,andE.A.Brewer.

AdaptingtoNetworkandClientVariationUsingActiveProxies:LessonsandPerspectives.InIEEEPersonalCom-munications,September1998.

[17]M.Gerla,L.-J.Chen,T.Sun,andG.Yang.Ubiquitousvideo

streaming:Asystemperspective.InAdvancesinPervasiveComputingandNetworking,2004.

[18]S.D.Gribble,M.Welsh,J.R.vonBehren,E.A.Brewer,

D.E.Culler,N.Borisov,S.E.Czerwinski,R.Gummadi,J.R.Hill,A.D.Joseph,R.H.Katz,Z.M.Mao,S.Ross,puterNetworks,35(4):473–497,2001.

[19]J.Hartman,U.Manber,L.Peterson,andT.Proebsting.Liq-uidsoftware:Anewparadigmfornetworkedsystems.tech-nicalreport96-11.Technicalreport,UniversityofArizona,1996.

[20]D.B.Johnson,D.A.Maltz,andY.-C.Hu.TheDy-namicSourceRoutingProtocolforMobileAdHocNet-works(DSR)draft-ietf-manet-dsr-10.txt,July2004.

Today’s model of networking primarily concentrates intelligence at the end hosts with the network itself offering a simple“best-effort”,“data agnostic ” communication medium. However, this paradigm has proven to be insufficient to meet todays needs co

[21]V.KawadiaandP.Kumar.ACautionaryPerspectiveon

Cross-layerDesign,Febuary2005.

[22]M.KawarasakiandR.S.Atarashi.Policybasedcontent

deliverymanagementusingmetadata.InSAINTWorkshops,pages272–279,2004.

[23]I.KOUVELAS,V.HARDMAN,andJ.CROWCROFT.

Networkadaptivecontinuous-mediaapplicationsthroughselforganisedtranscoding,1998.

[24]V.K.MadisettiandA.D.Argyriou.TransportLayerQoS

ManagementforWirelessMultimediaServices.

[25]M.MargaritidisandG.C.Polyzos.WirelessNetwork

SupportforAdaptiveReal-TimeApplications.InAd-vancedSimulationTechnologiesConference(ASTC’99),April1999.

[26]R.Mohan,J.Smith,andC.-S.Li.AdaptingmultimediaIn-ternetcontentforuniversalaccess.InIEEETransactionsonMultimedia,volume1,March1999.

[27]K.Nichols,S.Blake,F.Baker,andD.Black.RFC2474:

De nitionoftheDifferentiatedServicesField(DSField)intheIPv4andIPv6headers,Dec.1998.

[28]P.BucciolandG.DaviniandE.MasalaandE.Filippi

andJ.C.DeMartin.Application-levelPerceptualARQforH.264VideoStreamingover802.11WirelessLAN’s.InTheSeventhInternationalSymposiumonWIRELESSPERSONALMULTIMEDIACOMMUNICATIONS(WPMC2004),September2004.

[29]A.Schwartz,A.W.Jackson,W.T.Strayer,W.Zhou,

D.Rockwell,andC.Partridge.Smartpacketsforactivenet-works.InOpenArch’99,1999.

[30]J.SteinbergandJ.Pasquale.Awebmiddlewarearchitecture

fordynamiccustomizationofcontentforwirelessclients.InWWW’02:Proceedingsofthe11thinternationalconferenceonWorldWideWeb,pages639–650,NewYork,NY,USA,2002.ACMPress.

[31]S.Subramanian,P.Wang,R.Durairaj,J.Rasimas,

F.Travostino,vian,andD.Hoang.PracticalAc-tiveNetworkServiceswithinContent-AwareGateways.InDARPAActiveNetworksConferenceandExposition(DANCE2002),pages344–355,June2002.

[32]C.TschudinandR.Gold.Selnet:Atranslatingunderlay

network-tr2003-020.Technicalreport,UppsalaUniversity,2001.

[33]F.Vacirca,A.D.Vendictis,andA.Baiocchi.Investigating

interactionsbetweenarqmechanismsandtcpoverwirelesslinks.InEuropeanWireless(EW’04),Febuary2004.

[34]D.Wetherall,J.Guttag,andD.Tennenhouse.Ants:A

toolkitforbuildinganddynamicallydeployingnetworkpro-tocols,1998.

[35]M.Yarvis,P.L.Reiher,andG.J.Popek.Conductor:A

frameworkfordistributedadaptation.InWorkshoponHotTopicsinOperatingSystems,pages44–,1999.

[36]R.Yavatkar,D.Pendarakis,andR.Guerin.RFC2753:A

FrameworkforPolicy-basedAdmissionControl,Jan.2000.

10

本文来源：https://www.bwwdw.com/article/xnjm.html