Graphic Engine Resource Management

Modern consumer-grade 3D graphic cards boast a computation/memory resource that can easily rival or even exceed that of standard desktop PCs. Although these cards are mainly designed for 3D gaming applications, their enormous computational power has attrac

GraphicEngineResourceManagement

MikhailBautinAshokDwarakinathTzi-ckerChiueh

ComputerScienceDepartment

StonyBrookUniversity

{mbautin,ashok,chiueh}@cs.sunysb.edu

ABSTRACT

Modernconsumer-grade3Dgraphiccardsboastacomputation/memoryresourcethatcaneasilyrivalorevenexceedthatofstandarddesktopPCs.Althoughthesecardsaremainlydesignedfor3Dgamingapplications,theirenormouscomputationalpowerhasattracteddeveloperstoportanincreasingnumberofscienti ccomputationprogramstothesecards,includingmatrixcomputation,collisiondetection,cryptography,databasesorting,etc.Asmoreandmoreapplicationsrunon3Dgraphiccards,thereisaneedtoallocatethecomputation/memoryresourceonthesecardsamongthesharingapplicationsmorefairlyande ciently.Inthispaper,wedescribethedesign,implementationandevaluationofaGraphicProcessingUnit(GPU)schedulerbasedonDe citRoundRobinschedulingthatsuccessfullyallocatestoeveryprocessanequalshareoftheGPUtimeregardlessoftheirdemand.Thisscheduler,calledGERM,estimatestheexecutiontimeofeachGPUcommandgroupbasedondynamicallycollectedstatistics,andcontrolseachprocess’sGPUcommandproductionratethroughitsCPUschedulingpriority.Measurementsonthe rstGERMprototypeshowthatthisapproachcankeepthemaximalGPUtimeconsumptiondi erenceamongconcurrentGPUprocessesconsistentlybelow5%foravarietyofapplicationmixes.

Keywords:GPUscheduling,ResourceManagement,GPUfairness,DRI,DRMkernelmodule,GERM

1.INTRODUCTION

Modernconsumer-gradegraphiccardsboastacomputation/memoryresourcethatcaneasilyrivalorevenexceedthatofstandarddesktopPCs.Forexample,aGeforce8800GTX-basedcardcontains786MBofmemorywithatotalmemorybandwidthof86.4GB/sec,andaG80GPUthathas128shaderprocessorseachoperatingat

1.35GHz,where16pixelsareprocessedsimultaneouslyineachclockcycle.Althoughthesecardsareoriginallydesignedfor3Dgamingapplications,theirenormouscomputationalpowerhasattractedanincreasingnumberofresearcherstoporttheirapplicationstothem.Non-graphicapplications1thathavebeensuccessfullyportedtoGPUsincludematrixcomputation,video/imageprocessing,collisiondetection,cryptography,databasesorting,etc.AsmoreandmoreGPU-exploitingapplicationsstarttoemerge,aGPU’scomputation/memoryresourcemayneedtobetime-sharedamongmultiplesuchapplicationsthatarerunningconcurrently.Asaresult,GPUneedsanoperatingsystemtomanageitscomputation/memoryresource,justasCPU.Inaddition,theGPU’sresourcemanagershouldcoordinateitsresourceallocationdecisionswiththosemadebythemainOSrunningontheCPU.

ExistingoperatingsystemsupportforGPU2,3allowsmultiplegraphicapplicationstorunonasingleGPUasifeachofthemhasexclusiveaccesstotheGPU’sresource,butdoesnotpreventapplicationsfrommonopolizingtheGPUresource.Typically,agraphicapplicationisbuiltontopofauser-levelgraphiclibrary,e.g.OpenGLorDirect3D,whichconvertshigh-levelcommandsintolow-levelGPUcommandsthroughauser-levelGPU-speci cdriver.AGPUinteractswithitsdriverviaaGPUcommandringbu eronthegraphiccard,towhichtheGPUdriverDMAscommandsfromthehostmemory.AlthoughGPUinterruptsCPUwhenDMAtransactionscomplete,theydonotgenerateinterruptswhenaGPUcommandexitstheGPU’sgraphicpipeline.

ModernGPUsarestatefulinthesensethatagraphicapplication’sGPUcommandsmustbeexecutedwithrespecttoaspeci cGPUstatethatitsetupbeforehand.AGPUstatetypicallyincludestransformationmatrices,view-portspeci cations,lightingparameters,etc.Becausedi erentapplicationsrequiredi erentGPUstates,itisessentialtorestoretheGPUstateofagraphicapplicationbeforeschedulingitsGPUcommands.Restoringagraphicapplication’sstaterequiresissuingasequenceofGPUcommandstoproperlysetupthe

Modern consumer-grade 3D graphic cards boast a computation/memory resource that can easily rival or even exceed that of standard desktop PCs. Although these cards are mainly designed for 3D gaming applications, their enormous computational power has attrac

GPU.ThisstepissimilartocontextswitchingassociatedwithmultiplexingprocessesonaCPU,exceptonekeydi erence:ContextswitchingisperformedbytheOSaspartofCPUschedulingandistypicallytransparenttotheapplications,whereasGPUstateswitchingisdonebytheuser-levelGPUdriver,andisdecoupledfromtheCPUscheduler.

BecausetheCPUschedulerisnotresponsibleforestablishingagraphicapplication’sstatebeforeschedulingitontheCPU,thereisnoguaranteethatitsGPUcommandsareexecutedwithrespecttotheircorrespondingstate.Toaddressthisproblem,theuser-levelGPUdriverofagraphicapplicationmustacquirealockbeforesubmittingcommandstotheGPUtoguaranteethatithasexclusiveaccesstotheGPU.WecallthesetofcommandsanapplicationissuestotheGPUbetweenalockacquisitionandalockreleaseasacommandbatch,whichisdi erentfromacommandgroup,aminimalsequenceofcommandsthatmustbeDMAedintotheGPUatomically.

Asaresultoftheabovedesign,thecomputationandmemoryresourcesontheGPUareallocatedona rstcome rstservebasis.Moreconcretely,theGPUprocessescommandsintheGPUcommandringbu erintheorderinwhichtheyareinsertedwithoutregardstoanynotionoffairnessorpriority.BecauseGPUcommandsmayrequestallocationofphysicalmemoryonthegraphiccard,theseallocationrequestsarealsoservicedona rstcome rstserveorder.

TheprevailingGPUcommandschedulingmodelhasseveralseriousweaknesses.First,itcannotlimittheamountofGPUresourceconsumedbyindividualprocesses,asisdonebyCPUschedulerusingtimequanta,becauseeachgraphicapplicationisallowedtoissueasmanyGPUcommandsasitwantsonceacquiringthelockprotectingthecommandringbu er.Second,ifaprocessacquiresthecommandringbu erlockanddoesnotreleaseitmaliciouslyorbymistake,nootherprocessescanissuecommandstotheGPUandtheGPU’sresourcemayliewastedduringsuchtimes.Finally,theresourceallocationdecisionsoftheGPUschedulerarenotcoordinatedwiththeCPUscheduler’s.Thiscouldhurtagraphicapplication’sperformanceattimesofGPUcommandburstsbecausetheCPUan …… 此处隐藏：38784字，全部文档内容请下载后查看。喜欢就下载吧 ……