Runtime Optimizations for a Java DSM Implementation ABSTRACT

Jackal is a fine-grained distributed shared memory implementation of the Java programming language. Jackal implements Java’s memory model and allows multithreaded Java programs to run unmodified on distributed-memory systems. This paper focuses on Jackal

RuntimeOptimizationsforaJavaDSMImplementation

R.Veldema

R.F.H.Hofman

R.A.F.Bhoedjang

H.E.Bal

DepartmentofComputerScience

VrijeUniversiteit

Amsterdam,TheNetherlands

rveldema,rutger,bal@cs.vu.nl

DepartmentofComputerScience

CornellUniversityIthaca,NY,USAraoul@cs.cornell.edu

objectorasectionofaJavaarray.Incontrastwithpage-basedDSMs,Jackalusessoftwareaccesscheckstodetermineifaregionispresentinlocalmemoryandup-to-date.Ifanaccesscheckdetectsthataregionisabsentorout-of-date,itinvokesJackal’sruntimesystemwhichimplementsamultiple-writercachecoherenceprotocolthatresolvesreadandwritemisses.Aregionismanagedbyitshomenode,whichistheproces-sorthatcreatedtheassociatedobject.Jackaldoesnotuseasingle-writerprotocol,becausethatwouldrequirethecom-pilertoinformtheruntimesystemwhenaread/writeoperationhas nished;thatwouldincreasecodesizeandprotocolover-head,andposecomplicationsforthecompilerin(re)movingaccesschecks.

JackalconformstotheJavamemorymodel,whichallowscachingofobjectsin(thread)localmemoryandlogicallyre-quirescomplete ushingoflocalmemoryuponeachentryandexitofasynchronizedblock.Inoursystem,mainmemoryequatestoanobject’shomenode,andlocalmemorytotherequestingmachine’smemory.Flushingregionsandsubse-quentlyrequestingthemagainmaycausealargeoverheadun-deranaiveimplementation(especiallyusingtheclasslibrarieswhichperformmanyunnecessarysynchronizations[1]).Toreducethisoverhead,weinvestigatepossibilitiesofferedbytheJavamemorymodeltocacheregionsacrossasynchroniza-tionoperation.Thisispossibleforregionsthatareread-sharedandregionsthatareaccessedbyasinglemachine.

JackalusesanoptimizingJavacompilertogenerateaccesschecks.Intheoptimizationpassesofthecompiler,accesschecksmayberemoved,liftedorcombined.Forexample,arrayaccessesmaybecombinedandliftedfromaloopthat(partially)traversesthearray,oraccessesmaybeaggregatedwhenthecompilerdeterminesthatanobjectisusedtogetherwithitsreferencedsubobjects.Thecompileropimizationsaredescribedindetailin[24].

Thecontributionsofthispaperareasfollows:

WedescribevariousRTSoptimizationstoreducethenum-berofregion ushes.

WemeasuretheimpactoftheRTSoptimizationsforsev-eralJavaapplicationsandcomparethemtotheimpactofcompileroptimizations.

Thepaperisstructuredasfollows.Section2treatsJava’smemorymodel.Section3describesJackalanditsimplemen-tation.Section4summarizesJackal’scompileroptimizationsanddescribesournewRTSoptimizations.Section3andanex-tendedversionofSubsection4.1appearedearlierin[24],butwerepeattheseintroductorysectionsheretomakethispaper

ABSTRACT

Jackalisa ne-graineddistributedsharedmemoryimplemen-tationoftheJavaprogramminglanguage.JackalimplementsJava’smemorymodelandallowsmultithreadedJavaprogramstorununmodi edondistributed-memorysystems.

ThispaperfocusesonJackal’sruntimesystem,whichim-plementsamultiple-writer,home-basedconsistencyprotocol.ProtocolactionsaretriggeredbysoftwareaccesschecksthatJackal’scompilerinsertsbeforeobjectandarrayreferences.WedescribeoptimizationsforJackal’sruntimesystem,whichmainlyconsistofdiscoveringopportunitiestodispensewith ushingofcacheddata.Wegiveperformanceresultsfordif-ferentruntimeoptimizations,andcomparetheirimpactwiththeimpactofonecompileroptimization.We ndthatourrun-timeoptimizationsarenecessaryforgoodJackalperformance,butonlyinconjunctionwiththeJackalcompileroptimizationsdescribedin[24].Asayardstick,wecomparetheperformanceofJavaapplicationsrunonJackalwiththeperformanceofequivalentapplicationsthatuseafastimplementationofJava’sRemoteMethodInvocation(RMI)insteadofsharedmemory.

1.INTRODUCTION

Jackalisacompiler-supported, ne-graineddistributedsharedmemory(DSM)systemforJava.Thesystemcanrununmodi- ed,multithreadedJavaprogramsonaclusterofworkstations.Together,Jackal’scompilerandruntimesystem(RTS)hidethedistributednatureofthecluster:Jackalprogramsusethreadsandsharedvariablesinsteadofmessage-passingabstractionslikeRemoteMethodInvocation[19].ThispaperfocusesontheimplementationoftheRTSanditsoptimizations,whichmainlyconsistofdiscoveringopportunitiestodispensewith ushingofcacheddatatomainmemory.

Jackalresembles ne-grainedDSMsystemslikeShasta[22]andSirocco[13]inthatitusesasmallunitofcoherencethatismanagedentirelybysoftware.InJackal,theunitofcoherenceiscalledaregion.EachregioncontainseitheracompleteJava

Permissiontomakedigitalorhardcopiesofallorpartofthisworkforpersonalorclassroomuseisgrantedwithoutfeeprovidedthatcopiesarenotmadeordistributedforpro torcommercialadvantageandthatcopiesbearthisnoticeandthefullcitationonthe rstpage.Tocopyotherwise,torepublish,topostonserversortoredistributetolists,requirespriorspeci cpermissionand/orafee.

Copyright2000ACM0-89791-88-6/97/05..$5.00

Jackal is a fine-grained distributed shared memory implementation of the Java programming language. Jackal implements Java’s memory model and allows multithreaded Java programs to run unmodified on distributed-memory systems. This paper focuses on Jackal

self-contained.Section5studiestheimpactoftheRTSopti-mizationsonJackal’sperformanceonaMyrinet-basedclustercomputer.Section6discussesrelatedwork.Finally,Section7concludes.

2.JAVA’SMEMORYMODEL

Webrie ysummarizeJava’smemorymodel;foradetaileddescriptionwerefertothelanguagespeci cation[10]andPugh’scritiqueofthememorymodel[21].

Java’smemorymodelspeci esthateachthreadhasawork-ingmemory,whichcanbeconsideredathread-privatecache.Theentireprogramhasamainmemorywhichisusedforcom-municationbetweenthreads.Thedatamodi edbyathreadis ushedtomainmemoryuponencounteringasynchronizationpoint.(Inthisrespect,themodelresemblesreleaseconsis-tency[8,16].)SynchronizationpointsinJavacorrespondtotheentryan …… 此处隐藏：38919字，全部文档内容请下载后查看。喜欢就下载吧 ……