资源描述:
《The_‘Uniform_Heterogeneous_Multi-threaded’_Processor_Architecture.pdf》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、CommunicatingProcessArchitectures2001103AlanChalmers,MajidMirmehdiandHenkMuller(Eds.)IOSPress,2001The`UniformHeterogeneousMulti-threaded'ProcessorArchitectureDanielTownerDavidMayDepartmentofComputerScience,UniversityofBristol,UKAbstract.Multi-threadedprocessorar
2、chitecturesarecapableofconcurrentlyexecut-ingmultiplethreadsusingasharedexecutionresource.Twooftheiradvantagesaretheirabilitytohidelatencywithinathread,andtheirhighexecutionefficiency.Un-fortunately,singlethreadperformanceisoftenpoor.Inthispaperwepresentasimplemo
3、delofamulti-threadedprocessor,andshowhowanoccam-likelanguagemaybecompiledintofinegrainedthreadssuitableforexecutingonthisprocessor.Thesefinegrainedthreadsallowallbutthemostserialprogramstobecompiledintomultiplethreads.Thus,poorsinglethreadperformanceisavoidedbyens
4、uringthatsufficientthreadsarealwaysavailable,evenattheinstructionlevel.Wecallthistechnique`uni-formheterogeneousmulti-threading'(UHM).AcompilerimplementingUHMhasbeenbuilt,alongwithacycleaccuratesimulatorofaUHMprocessor.Wedemon-stratethattheprocessoriscapableofgoo
5、dperformance,whilstbeingsimpletodesignandbuild.1IntroductionMulti-threadedprocessorshavebeenthesubjectofresearchformanydecades[1].OneofthefirstwastheDenelcorHEP[2],andrecentlyinteresthascentredaroundtheSimultane-ousMulti-threading(SMT)processor[3,4,5].Multi-threa
6、dedprocessorshavetwomajoradvantagesoverothertypesofprocessor:1)theyoffertheabilitytohidelatencywithinathread(e.g.,memoryorexecutionlatency)and2)theyachievehighefficiency.Apossiblethirdadvantageisthattheyarepotentiallylesscomplexthanalternativehighperformanceproce
7、ssors.Unfortunately,theseadvantagesareoftenmarredbypoorsinglethreadperfor-mance.Methodsofimprovingsinglethreadperformancehavebeenwidelyresearched.Thenor-malapproachtoimprovingsinglethreadperformancehasbeentomaketheprocessormuchmorecomplex.Forexample,byaddingfeat
8、uressuchasdynamicinstructionissue,out-of-orderexecution,andregisterrenamingtoextractinstructionlevelparallelism(ILP)theperformanceofasinglethreadcanbegreatlyimproved.