资源描述:
《Nano-Threads Library Design,纳米线程库设计 实施与评价》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库。
1、Nano-ThreadsLibraryDesign,ImplementationandEvaluationXavierMartorell,JesúsLabarta,NachoNavarro,EduardAyguadéDAC/UPCReportNo.UPC-DAC-1995-33September19951Nano-ThreadsLibraryDesign,ImplementationandEvaluationXavierMartorell,JesusLabarta,NachoNavarro,EduardAyguadeDepartamentd'Arq
2、uitecturadeComputadors(DAC)UniversitatPolitècnicadeCatalunya(UPC)GranCapitàs/n,CampusNord,MòdulD6,08071,Barcelona,Spain{xavim,jesus,nacho,eduard}@ac.upc.esAbstract:Inthisreportwedescribethedesignandimplementationofauser-levelthreadpackagebasedonthenano-threadsprogrammingmodel,
3、whosegoalistoeffi-cientlymanagetheapplicationparallelismatuser-level.Somemodificationsintroducedtothemodelhaveenhanceditsflexibilityandhaveallowedmoreruntimeparallelism.Itisacurrenttrendincompilertechnologytoapplydataandcontroldependenceanalysestobuildanintermediaterepresentation
4、oftheparallelapplicationintheformofahierarchicaltaskgraph(HTG);inthisgraph,nodesaretherepresentationofpro-gramstructuresandedgesrepresentthedependenciesbetweenthem.Theback-endcompilergeneratesexecutablecodefromtheHTGandtheuser-levelthreadpackageprovidestheservicesinordertocont
5、roltheexecutionoftheHTGnodes.Libraryservicesareexplainedshowingseveraltranslationsfromsequentialtopar-allelcode.Wehavemainlyfocusedourattentionintoloopschedulingandfunctionalparallelism.Inbothcases,thegoalistoobtainanefficientparallelexecutionofthegen-eratednano-threadsbyaccord
6、inglybalancingtheworkassignedtoeachthreadandthethreadmanagementoverhead.Theexperimentsletusdeterminewhichistheappropri-atenumberofoperationsspreadoutamongthethreadstoensurelessthan10%ofover-head.1.IntroductionThenano-threadsprogrammingmodelwasfirstintroducedin[PoBK93]and[Poly93
7、]toprovidehighlyoptimizedlight-weightthreads.Inthispaperwearegoingtodescribethedesignandimplementationofauser-levelthreadpackagebasedonthismodel.Ourenvironmentassumesthatapplications(e.g.,CorFORTRANprograms)areautomaticallydecomposedbyaparallelizingcompiler.Thecompileridentifie
8、sthemaximumparallelismoftheapplicationthroughdataandcontrolde