Communication Architecture Tuners A Methodology for the Design of High-Performance Communic

KanishkaLahiriAnandRaghunathanGaneshLakshminarayanaSujitDey

Dept.ofElectricalandComputerEngg.,UniversityofCalifornia,SanDiego,CA

C&CResearchLabs,NECUSA,Princeton,NJ

Abstract

Inthispaper,wepresentageneralmethodologyforthedesignofcustomsystem-on-chipcommunicationarchitectures.Ourtechniqueisbasedontheadditionofalayerofcircuitry,calledtheCommunica-tionArchitectureTuner(CAT),aroundanyexistingcommunicationarchitecturetopology.Theaddedlayerenhancestheabilityofthesystemtoadapttochangingcommunicationneedsofitsconstituentcomponents.Forexample,morecriticaldatamaybehandleddiffer-ently,leadingtolowercommunicationlatencies.TheCATmonitorstheinternalstateandcommunicationtransactionsofeachcomponent,and“predicts”therelativeimportanceofeachcommunicationtrans-actionintermsofitspotentialimpactondifferentsystem-levelper-formancemetrics.Itthenconfigurestheprotocolparametersoftheunderlyingcommunicationarchitecture(e.g.,priorities,DMAmodes,etc.)tobestsuitthesystem’schangingcommunicationneeds.

WeillustrateissuesandtradeoffsinvolvedinthedesignofCAT-basedcommunicationarchitectures,andpresentalgorithmstoauto-matethekeysteps.Experimentalresultsindicatethatperformancemetrics(e.g.numberofmisseddeadlines,averageprocessingtime)forsystemswithCAT-basedcommunicationarchitecturesaresignifi-cantly(sometimes,overanorderofmagnitude)betterthanthosewithconventionalcommunicationarchitectures.1Introduction

TheevolutionoftheSystem-on-Chip(SOC)paradigminelec-tronicsystemdesignhasthepotentialtoofferthedesignerseveralbenefits,includingimprovementsinsystemcost,size,performance,powerdissipation,anddesignturn-around-time.Theabilitytoreal-izethispotentialdependsonhowwellthedesignerexploitsthecus-tomizabilityofferedbythesystem-on-chipapproach.Whileonedi-mensionofthiscustomizabilityismanifestedinthediversityandcon-figureabilityofthecomponentsthatareusedtocomposethesystem(e.g.,processoranddomain-specificcores,peripherals,etc.),another,equallyimportantaspect,isthecustomizabilityofthesystemcom-municationarchitecture.Inordertosupporttheincreasingdiversityandvolumeofon-chipcommunicationrequirements,whilemeetingstringentperformanceconstraintsandpowerbudgets,communicationarchitecturesneedtobecustomizedtothetargetsystemorapplicationdomaininwhichtheyareused.

1.1PaperOverviewandContributions

Inthispaper,wedemonstratetheneedforthedesignofflexiblecommunicationarchitecturesbyanalyzingexamplesystemsandsce-nariosinwhichnostaticcustomizationoftheprotocolscancom-pletelysatisfythesystem’stime-varyingcommunicationrequire-ments.Thepresentedtechniquecanbeusedtooptimizeanyun-derlyingcommunicationarchitecturetopologybyenhancingitwithCommunicationArchitectureTuners(CATs)whichmakeitcapableofadaptingtothechangingcommunicationneedsofthecomponentsconnectedtoit.InthispaperweillustratetheissuesandtradeoffsinvolvedinthedesignofCAT-basedcommunicationarchitectures,

Example1:ConsidertheexamplesystemshowninFigure1thatrepresentspartoftheTCP/IPcommunicationsprotocolusedinanet-workinterfacecard.Thesystemperformschecksum-basedencod-ing(foroutgoingpackets)anderrordetection(forincomingpackets),andinterfaceswiththeEthernetcontrollerperipheral(whichimple-mentsthephysicalandlinklayernetworkprotocols).SincepacketsintheTCPprotocoldonotcontainanynotionofqualityofservice(QoS)[23],wehaveenhancedthepacketdatastructuretocontainafieldintheheaderthatindicatesadeadlineforthepackettobepro-cessed.Weassumethattheobjectiveduringtheimplementationofthesystemistominimizethenumberofpacketswithmisseddead-lines.

To/FromnetworkETHERDRIVERPKT_QUEUEIP_CHECKCHECKSUMCommn.requestsCATCHECKSUMMIPSR3000CATIP_CHECKCATCHECKSUMCATData&controlsignalsPriority=f1(size,deadline)PacketDMAsize=f2(size,deadline)header…(size,...deadline)ProtocolParametersBUSINTERFACEOptimizedArbiterMemoryTocommunicationarchitecture(a)Pkt.iPkt.i+1arrivesarrivesCAT-basedarch.(b)

Pkt.i+1Pkt.ideadlinedeadlinePkt.jPkt.j+1Pkt.jarrivesarrivesdeadlinePkt.j+1deadlineAllpktsmeetdeadlines(a)SpecificationProtocolParameters:PriorityDMA_modeDMA_size…...SharedMemoryTo/fromapplicationlayerFigure3:CAT-basedarchitecturefortheTCPexample

MIPSR3000IP_CHECKCHECKSUMBusinterfaceii+1ijj+1(b)ImplementationArbiterMemoryFigure1:TCPsystemfromanetworkinterfacecard:(a)Specificationand(b)Implementationutilizingaconventionalbus-basedcommuni-cationarchitecture

Figure1(a)showsthebehavioroftheTCPsystemasasetofconcurrentcommunicatingprocesses.Figure1(b)showsthesystemarchitectureusedtoimplementtheTCPsystem.The

queueprocessesaremappedtoem-ether

beddedsoftwarerunningonaMIPSR3000processor,whiletheip

Figure4:ExecutionoftheoptimizedTCPsystemwithaCAT-basedarchitecture

Weattemptedtoeliminatetheproblemmentionedabovebyusing

checkchecksum)adifferentpriorityassignment(ether

forthebusprotocol.Theexecutionofthesystemunderthenewpri-orityassignmentisdepictedinthesecondwaveformofFigure2.Asaresultofthenewpriorityassignment,packetsiandi1meettheirdeadlines,butpacketjmissesit,duetocontentionwitharequestfromether

ether_driverip_checkchecksumcheckether

driverprocessforpacketi,thearbiterreceivestwoconflictingbusaccessrequests:processip

driverrequestsbusaccesstoprocesspacket

i1(sincepacketi1hasalreadyarrivedfromthenetwork).Basedonthepriorityassignment,thearbitergivesbusaccesstoprocessip

checkandchecksumhavecompletedprocessingpacketi.Thisleadstopacketi1missingitsdeadline.Packetsjandj1domeettheirdeadlines.

Case1:checksum>ip_check>ether_driverCase2:ether_driver>ip_check>checksumPkt.iPkt.i+1arrivesarrivesPkt.i+1Pkt.ideadlinedeadlinePkt.jPkt.j+1Pkt.jarrivesarrivesdeadlinePkt.j+1deadlinei+1missesdeadlinejmissesdeadlineii+1jj+1ii+1ether_driveriip_checkjj+1checksumjj+1Figure2:ExecutionoftheTCPsystemforvariousbuspriorityas-signments

Theexampleshowsasituationwheretherelativeimportanceofthecommunicationtransactionsgeneratedbythevarioussystemcomponentsvariesdependingonthedeadlinesofthepacketstheyareprocessing.Itdemonstratesthatconventionalcommunicationarchi-tectures(i)providetoolimitedadegreeofcustomizabilityforsystemswithstringentperformancerequirements,and(ii)aretypicallynotca-pableofsensingandadaptingtothevaryingcommunicationneedsofthesystemandthevaryingnatureofthedatabeingcommunicated.Example2:ACAT-basedcommunicationarchitecturefortheTCPsystemisshowninFigure3(a).CATsareaddedtoeachbusmas-terandthebuscontrollogic(arbiterandcomponentbusinterfaces)isenhancedtofacilitatetheiroperation.Amoredetailedviewofacom-ponentwithaCATisshowninFigure3(b).ThecomponentnotifiestheCATwhenitgeneratescommunicationrequests.TheCATalsoobservesselecteddetailsaboutthedatabeingcommunicatedandthecomponent’sinternalstate.

Inthisexample,theCATobservesthepacketsizeanddeadlinefieldsfromtheheaderofthepacketcurrentlybeingprocessedbythecomponentandperformsthefollowingfunctions:(i)itgroupscommunicationeventsbasedonthesizeanddeadlineofthepacketcurrentlybeingprocessed,and(ii)foreventsfromeachgroup,itde-terminesanappropriateassignmentofvaluestothevariousprotocolparameters.Therationalebehindusingthedeadlineisthatpacketswithcloserdeadlinesneedtobegivenhigherimportance.Incaseswhenallthepacketsinthesystemhaveroughlyequaldeadlines,itisadvantageoustofavorthecompletionofpacketswhicharesmaller,sincetheyhaveabetterchanceofmeetingthedeadline.Forexample,thepriorityiscomputedusingtheformulastdtawheres,tdandtarepresentthepacketsize,deadline,andarrivaltime,respectively.ThepacketsequenceofExample1wasusedtodriveexecutionoftheoptimizedsystemofFigure3.UnderaCAT-basedarchitecture,thesystemmeetsthedeadlinesforallthepackets.Whenpacketi1(whichhasatightdeadline)arrives,theCATassignstothecommuni-cationrequestsgeneratedbyether

checkandchecksum,whicharestillpro-cessingpacketi.Thisleadstopacketi1meetingitsdeadline.When

packetj1arrives,however,thecommunicationrequestsgenerated

checkandbyether

checksumtoprocesspacketjtocompletioninordertomeetitstightdeadline.

e2e3.Thisidentifiesmostofthecriticalevents,and

veryfewnon-criticalones.

CATtokensCOMPONENTComm.t1t1requestst2+t3cntillustrationofCAT-optimizedcommunicationar-controllers/arbiterschitecture.

forthevariouschannelsinthecommunicationar-generatedAsimplesystemwithaCAT-basedcommunicationdetailedCATviewusingofouracomponenttechniqueswithwasashownCATisinshownFigurein3(a).architectureFigureA8.moreTheationconsistschitecturecircuitsprotocolthatofagenerate“partitiondetector”circuit,andparametergener-parametersvaluesduringforthesystemvariousexecution.

communicationar-Partitioncommunicationdetector:temthefollowingexecution.transactionsAcommunicationinformationThepartitiongeneratedpartitionisasubsetofthegenerateddetectorbybycircuitthecomponentthecomponent:

monitorsandduringanalyzessys-Tracerexecutingtokensgeneratethesespecificgeneratedtokensoperations.bythepurelyforThecomponentthepurposecomponenttoindicateoftheisenhancedthatitCAT.

toisTheeratedcommunicationbythecomponent.

transactioninitiationrequeststhataregen-Anydataotherapplication-specificpropertieswhichbeingindicategenerateditsrelativebytheimportance).

component(ofe.g.,thefieldscommunicationinthedataTheendpartitiondetectorusesthisinformationbelongofautomaticallytoasequenceapartition.ofcomputeInconsecutivetheSectionstartand4.3,communicationtoidentifyendwepresentconditionsgeneraltransactionsthestartandforeachtechniquesthatpartition.toParametercommunicationgenerationsizes,protocolcircuits:parametersThese(e.g.,circuitspriorities,computeDMAvaluesblockfortectoretc.)basedonthepartitionIDgeneratedbythepartitionde-byarbitersthecircuit,systemanddesigner.otherTheapplication-specificdatapropertiesspecifiedinAutomaticachangeandincontrollersthecharacteristicsinthevaluescommunicationoftheseparametersaresenttotheofthecommunicationarchitecture,architecture.resultingpresentedintechniquesSection4.2.

todesigntheparametergenerationcircuitsarelustratedThefunctioningformsusingsymbolicofaCAT-basedwaveformscommunicationarchitectureisil-thirdwaveformrepresentrepresentstracertokensthegeneratedinFigurecommunicationbythe9.Thefirsttwowave-transactionscomponentgenerated

whiletheAnalyzesystem,PartitioncreateComm.CAGcommunicationEvaluateAssignparameterAnalysisGraphinstancesPartitions/clustervaluesto12clustersstatistics3clusters4Inputs:sloPartitionedht.ci/mappedsystem,ImprovedRe-Analyzesystem,wmotComm.Arch.topology,performance?Re-computemmoorecp.Inputtraces,performancemetricstsPerformancemetrics5ywhecSnraSynthesizeCATstorealizeoptimizedprotocols6OptimizedCAT-basedsystemcommunicationFigurearchitecturearchitectures

10:MethodologyfordesigningCAT-basedcommunicationbytitionthecomponent.ThefourthAlltorcommunicationdetectorcircuittransactionswhichchangeswaveformshowsthestateofthepar-thatinresponsetothefirstthree.TheFSMisinstateS3areclassifiedoccurasbelongingwhenthetopartitionpartitiondetec-CP1.cuit.fifthtoHerewaveformitassignsashowsprioritytheleveloutputof4oftotheprioritygenerationcir-delaypartitionpartitionassociatedCPincreaseinpriorityallleadstransactionstoadecreasethatbelong1.ThisintheCPaswithshowntheincommunicationthelastwaveformtransactionsofFigurethat9.

belongtotionTheoverall1,analysisarchitecturesmethodologyorderisappliedistoshowninforFiguredesigning10.CAT-basedInstep1,communica-performanceIninourtowork,derivewetheinformationthepartitioned/mappedandstatisticsusedsysteminthedescriptionlatersteps.inulation,[24],whichmanner.whilebeingisusecomparabletheperformanceanalysistechniquepresentedmuchinaccuracytocompletesystemsim-Graphtem’s,(CAG)TheoutputwhichofmoreefficienttoemployinaniterativeisthisanalysisisaCommunicationAnalysisrepresentexecutionformedclustersundertheagivenhighlyinputcompacttraces.representationTheverticesinofthethegraphsys-edgescomputationsinbythethegraphvariousofcomputationsandabstractcommunicationsper-representcomponentsduringthesystemexecution.Thedetailedminevarioussystemandperformanceexecutioncommunications.thedependenciestracestatistics.

andThecanbeCAGeasilyisbetweenconstructedthevariousanalyzedtofromdeter-anumberInstep2,wegroupthecommunicationverticesintheCAGintoathatofpartitionssuchthateachpartitioncontainsasettionshavesimilarcommunicationrequirements,whiledifferentofeventsparti-eachrepresentdistinctrequirements.Notethatintheextremecase,beandassignedcommunicationtoadistinctvertexpartition.inthecommunicationHowever,thisanalysiscancausegraphtheareacan3delayoverheadincurredintheCATtobecomeprohibitive.Steponevaluatesueswhich,colsforeachstepvarious4determinesstatisticsanforassignmenteachcommunicationofprotocolpartition,parameterbasedval-theforthesystempartition.communicationTheoutputofsteparchitecture.4isasetStepofcandidate5re-evaluatesproto-Ifnoasystemperformanceperformanceimprovementfortheoptimizedresults,stepsprotocols1to5derivedinstep4.synthesisfurtherthatwereofperformancedeterminedhardwarein(CATs)improvementstep4.toimplementisobtained.theoptimizedSteparerepeated6dealsuntilprotocolswith4AlgorithmsfortheDesignofCATs

detail.InthissectionwedescribetheshadedtionparametereventWeinstances,presenttechniquesstepsofFigure10inmorediscusshowtoobtainpartitionsofthecommunica-communicationvalueseventandhowtoselectanoptimalsetofprotocolinstancestosynthesizeintopartitions.

CAThardwareforclassifying4.1ProfilingandpartitioningcommunicationeventinstanceseratedInstep2ofFigure10,weperformananalysisoftheCAGinstanceinsimilardelaysstep1ontomeasurethesystemtheperformance.impactofindividualgen-Instancescommunicationwhichhaveapartition.impactparameterTheonperformancethesystemimpactperformanceofaninstancearegroupedismeasuredintothesamebymancefollowingwhencalledexamplethesensitivitycommunicationthatcapturesillustratesourdelaypartitioningofthethechangeprocedure.

instanceinsystemachanges.perfor-Thesystemexecutiontime(clockcycles)systemcomponentsc1Start=16Finish=32c2Start=33Finish=44z1Start=45Finish*=50=10,O=t(z1)+t(z2)s(c1)=20s(c2)=10s(c3)=10s(c4)=0c3Start=4Finish=9Start=33Finish=34Start=10Finish=14c4z2Start=35Finish*=40c1c2c3c4CP1CP2CP3Figure11:SensitivitycalculationandpartitioninginstancesintheCAG

Figure11showsasectionofaCAGgeneratedfromarepresen-tativeexecutionofanexamplesystem.Shadedverticesc1throughc4representinstancesofcommunicationevents.Verticesz1andz2representthefinaloutputsofthesystem.Theobjectivefunctiontobe

tz2,wheretvisthefinishtimeminimizedisthequantitytz1

ofavertexvintheCAG.

Tomeasurethesensitivityofthesystemperformancetocommu-nicationinstancec1,theexistingdelayofc1isperturbedbyavalue∆,andatraversalofthetransitivefanoutofc1intheCAGisusedtore-computethestartandfinishtimesoftheaffectedvertices.Theupdatedfinishtimesoftheverticesareusedtocalculatethechangeinthesystemperformancemetric.Inthisexample,perturbingthedelayofc1by10unitsdelaysthefinishofbothz1andz2by10unitseach,whileperturbingthedelayofc2delaysz1alone.

Inthemannerdescribedabove,wecalculateasensitivitysciforeachinstanceci,whichmeasuresthechangeinthevalueoftheobjec-tivefunctionOafterperturbingthedelayofciby∆.Next,wegroupcommunicationeventinstancesthathavesimilarsensitivityvaluestoformapartition.Inthisexample,c1isassignedtopartitionCP1,c2andc3areassignedtoCP2,andc3isassignedtoCP3.4.2ModifyingProtocolParameters

Inthissectionwedescribesteps3and4oftheoverallflow,i.e.,howtoexamineeachpartitionandthenassignoptimizedprotocolpa-rametervaluestothem.Whileourdiscussionisconfinedtodetermin-ingtheprioritythatshouldbeassignedtoeachpartition,itcouldbeextendedtoincludeotherprotocolparameterssuchaswhetherburstmodeshouldbesupportedornot,andifsowhatthecorrectDMAsizeshouldbe.

Thesensitivityofapartitionindicatestheimpactitseventshaveontheperformanceofthesystem.Howeverassigningprioritiesbasedonthesensitivityalonemaynotleadtothebestassignment.Thisisbecausesensitivitydoesnotcapturetheindirecteffectsofacommuni-cationevent(orsetofevents)onthedelaysofotherconcurrentcom-municationevents.Weaccountforthisbyderivingametricthatpe-nalizespartitionswhicharelikelytonegativelyimpacttheexecutionofcommunicationeventsinotherpartitions.Inordertoobtainthisinformation,weanalyzetheCAGandevaluate,foreachpairofpar-titionsCPiCPj,theamountoftimeforwhichcommunicationeventsthatbelongtoCPiaredelayedduetoeventsfromCPj.Table1showsexampledataforasystemwiththreepartitions.Column2givesthesensitivityofeachpartition.Columns3,4and5givesthetotaltime(wij)thatinstancesinpartitionCP1,CP2andCP3waitforinstancesineachoftheotherpartitions.Forexample,instancesinCP1induceatotalwaitingtimeof100cyclesforinstancesofCP2tofinish.Col-umn6givesthesumofcolumns3,4and5toindicatethetotalwaitingtime(Wi)eventsinpartitionCPihaveintroducedinotherpartitions.ToassignaprioritytopartitionCPi,weuseaheuristicmetricPiwhichiscalculated(usingthenotationofTable1),as

nsjw

ij

Pisi∑

j1

t1C2C1=3Formula123t1t21=41=1t312CP1CP1CP1t1t2t3CP1CP1CP1t1t2t3=3CP1CP1CP1t3CP12=0time(cycles)100908070605040302010090.977.2763.63c&&cnt2=3c&&cnt2<3Partition1=“yes”S0t1t1S3t1S1xt1c&&cnt1=4S2Formula1Formula2Formula3c&&cnt1<4Inputs:c=communicationinstance;t1=tracerinstance;cnt1=counterfor1;cnt2=counterfor2FSM:S0=initialstate;S3=acceptingstate(a)Alternativepredictorstrategies

Figure12:Designofa

(b)Performanceofpredictorsforpar-titionCP1

predictorfunctionforpartitionCP1of

componentC1

(c)FSMimplementationofFormula1

informationscribedsuchastheweightsoneachrequestanddeadlines,asde-protocol.

inSection4,toprovideforamoreflexiblecommunicationTable2:PerformanceofsystemsusingCAT-basedarchitecturesExampleMetricInputStaticCATsbasedImproveSystemTraceProtocolarchitecture-mentTCP/IPmissed20100-deadlinespacketsSYSmissed5734131724.3deadlinestrans.ATMmissed16940162.5deadlinespacketsBRDGavg.no.of10,000304.722.11.2cyclescyclesTable3:ImmunityofCAT-basedarchitecturestovariationininputsInputstotheInputTraceStaticCATsbasedPerformanceSYSInformationProtocolarchitectureimprovementexampleTrace18483181611.98transactionsTrace25734131724.3transactionsTrace31070316388.37transactionsusingForeachsystem,column4reportsperformanceresultsobtainedsultsareinreportedgeneratedastaticcommunicationincolumnbysimulatingprotocol,whilecolumn5reportsre-6.TheresultsaCAT-basedindicatearchitecture.thatsignificantSpeed-upsoverperformancenumberaprotocolcanbeobtainedbyusingaCAT-basedarchitecturebenefitsSYS,a24ofusingfixedparametervalues.InthecaseofTCP,theXmissedperformancedeadlinesimprovementwasreduced(reductiontozero,whileintheinthenumbercaseofofmisseddeadlines)wasobserved.

dependsThedesignformingourtheontheofvariousselectionanefficientstepsofCAT-basedcommunicationarchitectureofathegoodalgorithmrepresentativeoftracewhenper-arealgorithmsdisplaynotattempttogeneratecommunicationFigurearchitectures10.However,thattraces.improvedspecifictoperformancetheexactinputtracesusedtodesignthem,butmanceInordertoanalyzetheoverinputawiderangeofcommunicationarchitectures,improvementstracesensitivityoftheperfor-example,weperformedobtainedthroughCAT-basedcommunicationcommunicationwesimulatedthesystemanadditionalwithCAT-basedexperiment.FortheSYSwidelyperiments.varyingcharacteristics.architecturesforthreedifferentinputandtracesconventionalthathadtoasimulateTherun-timeparametersunpredictability.oftheTableinput3presentstheresultsofourex-Intracesallthewerecases,chosenatrandomandCAT-basedcommunicationsignificantcommunicationthesystemwithimprovementarchitecturedemonstratedaconsistentofarchitecture.overThisthedemonstratessystembasedthatontheaconventionalperformancetheCAT-basedinputstimuli.architecturesarenotoverlysensitivetovariationsin6Conclusions

cuitry,Thispapershowedhowthethejudiciousadditionofalayerofanytem’sexistingcalledcapabilitycommunicationtheCommunicationofadaptingarchitectureArchitectureTuner(CAT),aroundcir-tothechangingtopologycommunicationcanenhanceaneeds

sys-ofinvolveditsconstituentcomponents.WeillustratedissuesandtradeoffsandexperimentpresentedinthealgorithmsdesignoftoCAT-basedcommunicationarchitectures,missedbaseddeadlines,weconcludeaveragethatautomateprocessingperformancetheprocess.Fromconductedtime)metricsfor(e.g.numberofancommunicationarchitecturesaresignificantlysystems(sometimes,withCAT-overcationorderarchitectures.ofmagnitude)betterthanthosewithconventionalcommuni-References

[1]D.beddedD.Gajski,SystemsF..Vahid,PrenticeS.Hall,Narayan1994.

andJ.Gong,SpecificationandDesignofEm-[2]R.trollers,”Ernst,J.IEEEHenkel,Designand&T.TestBenner,Magazine“Hardware-software,pp.–75,Dec.cosynthesis1993.

formicrocon-[3]T.aware/SoftwarecommunicatingB.Ismail,M.Abid,Codesignsystemand,pp.,”M.17–24,inProc.Jerraya,“COSMOS:Acodesignappraochfor

1994.

IEEEInternationalWorkshoponHard-[4]A.theKalavadeandE.Lee,“Agloballycritical/locallynationalconstrainedWorkshophardwareonHardware/SoftwaresowftwarepartitioningCodesignproblemphase,pp.42–48,,”driveninProc.algorithm1994.

IEEEInter-for

[5]P.cosynthesisH.Chou,R.systemB.Ortega,,”inProc.andG.Int.B.Symp.Borriello,System“TheLevelCHINOOKSynthesishardware/software

,pp.22–27,1995.[6]B.telecommunicationLin,“Asystemdesignmethodolgyforsoftware/hardwarecodevelopmentof

pp.672–677,1996.

networkapplications,”inProc.DesignAutomationConf.,[7]B.cosynthesisP.Dave,G.Lakshminarayana,andN.K.708,1997.

ofembeddedsystems,”inProc.DesignJha,“COSYN:Automationhardware-software

Conf.,pp.703–[8]P.partitioningKnudsenandSynthesis,pp.in111–116,hardware/softwareJ.Madsen,“IntegratingDec.1998.

codesigncommunication,”inProc.Int.protocolSymp.selectionSystemLevelwith

[9]T.,”inYenProc.andInt.W.Wolf,Conf.“CommunicationComputer-AidedDesignsynthesis,pp.for288–294,distributedNov.embedded1995.

systems

[10]J.tionDaveau,pp.150–155,byanallocationT.B.Ismail,andA.A.Jerraya,“Synthesisofsystem-levelcommunica-Sept.1995.

basedapproach,”inProc.Int.Symp.SystemLevelSynthesis,[11]M.,”inGasteierACMTrans.andM.DesignGlesner,Automation“Bus-basedElectroniccommunicationSystems,synthesispp.1–11,onJan.system1999.level

[12]R.dedB.1998.

systemsOrtegaand,”inG.Proc.Borriello,Int.Conf.“CommunicationComputer-AidedsynthesisDesignfor,pp.distributed437–444,embed-Nov.[13]“SonicsIntegrationArchitecture,SonicsInc.(http://www.sonicsinc.com/).”.

[14]“VSIv1.1.0Alliance(http://www.vsi.org/library/specs.htm).”.

on-chipbusDWG.”Onchipbusattributesspecification”version

[15]G.logic,”BorrielloinProc.andInt.R.Conf.H.Katz,Computer“SynthesisDesignand,Nov.optimization1987.

ofinterfacetransducer

[16]J.Conf.S.SunComputer-AidedandR.W.Brodersen,Design,“Designpp.478–481,ofsystemNov.interface1992.

modules,”inProc.Int.

[17]P.chronousGutberlet1994.

dataandpaths,”W.Rosenstiel,inProc.Int.“SpecificationSymp.SystemofinterfaceLevelSynthesiscomponents,pp.134–139,forsyn-[18]S.faceNarayananandD.D.Gajski,“Interfacingincompatible1995.

processgeneration,”inProc.DesignAutomationConf.protocols,pp.468–473,usinginter-June[19]P.forH.Chou,R.B.Nov.embedded1995.

systems,”Ortega,inandProc.G.Int.B.Borriello,Conf.Computer-Aided“Interfaceco-synthesisDesign,pp.techniques

280–287,[20]J.communicationOberg,A.Kumar,1996.

protocols,”andA.inHemani,Proc.Int.“Grammar-basedSymp.SystemLevelhardwareSynthesissynthesis,pp.of14–19,data

[21]R.sisPasserone,Conf.of,interfacesJ.A.pp.8–13,JunebetweenRowson,1998.

incompatibleandA.Sangiovanni-Vincentelli,protocols,”inProc.“AutomaticDesignAutomationsynthe-[22]J.inSmithProc.DesignandG.DeAutomationMicheli,Conf.“Automated,pp.14–19,compositionJune1998.

ofhardwarecomponents,”

[23]A.19.

S.Tanenbaum,ComputerNetworks.EnglewoodCliffs,N.J.:PrenticeHall,

[24]K.system-on-chipLahiri,A.Raghunathan,Design,pp.566–572,communicationandNov.1999.

architectures,”S.Dey,“FastperformanceinProc.Int.Conf.analysisComputer-Aidedofbus-based

[25]G.A.F.Seber,C.J.Wild,Non-linearReagression.Wiley,NewYork,19.

[26]K.multi-channelLahiri,A.Raghunathan,communicationandarchitectures,”S.Dey,“Performanceinvlsi,pp.analysis530–537,ofjanuarysystems2000.

with