文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > An Event-driven Architecture for MPI Libraries
侵权投诉

An Event-driven Architecture for MPI Libraries

Existing MPI libraries couple the progress of message transmission or reception with library invocations by the user application. Such coupling allows for simplicity of implementation, but may increase communication latency and waste CPU resources. This pa

AnEvent-drivenArchitectureforMPILibraries

SupratikMajumderandScottRixner

RiceUniversityHouston,TX77005{supratik,rixner}@rice.eduAbstract

ExistingMPIlibrariescoupletheprogressofmessagetransmissionorreceptionwithlibraryinvocationsbytheuserapplication.Suchcouplingallowsforsimplicityofimplementation,butmayincreasecommunicationlatencyandwasteCPUresources.Thispaperproposestheaddi-tionofanevent-drivencommunicationthreadtomakemes-sagingprogressinthelibraryseparatelyfromtheapplica-tionthread,thusdecouplingcommunicationprogressfromlibraryinvocationsbytheapplication.Theasynchronousevent-threadallowsmessagestobesentandreceivedcon-currentlywithapplicationexecution.Thistechniquedra-maticallyimprovestheresponsivenessofthelibrarytonet-workcommunication.Microbenchmarkresultsshowthatthetimespentwaitingfornon-blockingreceivestocompletecanbesigni cantlyreducedoreveneliminatedentirely.Ap-plicationperformanceasmeasuredbytheNASbenchmarksshowsanaverageof4.5%performanceimprovement,withapeakimprovementof9.2%.

VijayS.PaiPurdueUniversityWestLafayette,IN47907

vpai@purdue.edu

1Introduction

TheMessagePassingInterface(MPI)standardincludesprimitivesforbothblockingandnon-blockingcommunica-tion[10].Thelatteraredesignedtoenableoverlapbetweenanapplication’scommunicationandcomputation,withtheinitiationofamessagesendorreceivefollowedbyinde-pendentcomputation.However,almostallexistingMPIli-braries(andallfreely-availableones)onlymakeprogressonthetransmissionorreceptionofmessageswhentheap-plicationcallsintothelibrary.Thus,ifamessagecouldnotbesenttoorreceivedfromanothernodeatthetimeofthe rstnon-blockingsendorreceivecall,theactualcommu-nicationcouldbearbitrarilydelayedwaitingfortheappli-cationtore-enterthelibrarythroughanMPIcall.Atsuch

1The

MPI1.1progressrulespeci callystates:“AcalltoMPI

WAITthatcompletesasendwilleventuallyreturn

ifamatchingreceivehasbeenstarted,unlessthereceiveissatis edbyanothersend,andevenifnocallisexecutedtocompletethereceive.”[10]

1

第1页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email