小程序中英文外文文献翻译.docx

小程序中英文外文文献翻译.docx

《小程序中英文外文文献翻译.docx》由会员分享，可在线阅读，更多相关《小程序中英文外文文献翻译.docx（20页珍藏版）》请在冰豆网上搜索。

本科毕业设计（论文）

中英文对照翻译

（此文档为word格式,下载后您可任意修改编辑!

）

外文文献翻译原文及译文

标题：

ENHANCINGAPPLICATIONPERFORMANCEUSINGMINI-APPS:

COMPARISONOFHYBRIDPARALLELPROGRAMMINGPARADIGMS

作者：

GaryLawsonMichaelPoteatMashaSosonkinaRobertBaurle

期刊：

ComputerScience

年份：

2016原文

ENHANCINGAPPLICATIONPERFORMANCEUSINGMINI-APPS:

COMPARISONOFHYBRIDPARALLELPROGRAMMINGPARADIGMS

GaryLawsonMichaelPoteatMashaSosonkinaRobertBaurle

ABSTRACT

Inmanyfields,real-worldapplicationsforHighPerformanceComputinghavealreadybeendeveloped.Fortheseapplicationstostayup-to-date,newparallelstrategiesmustbeexploredtoyieldthebestperformance;however,restructuringormodifyingareal-worldapplicationmaybedauntingdependingonthesizeofthecode.Inthiscase,amini-appmaybeemployedtoquicklyexploresuchoptionswithoutmodifyingtheentirecode.Inthiswork,severalmini-appshavebeencreatedtoenhanceareal-worldapplicationperformance,namelytheVULCANcodeforcomplexflowanalysisdevelopedattheNASALangleyResearchCenter.Thesemini-appsexplorehybridparallelprogrammingparadigmswithMessagePassingInterface（MPI）fordistributedmemoryaccessandeitherSharedMPI（SMPI）orOpenMPforsharedmemoryaccesses.PerformancetestingshowsthatMPI+SMPIyieldsthebestexecutionperformance,whilerequiringthelargestnumberofcodechanges.Amaximumspeedupof

23wasmeasuredforMPI+SMPI,butonly10wasmeasuredforMPI+OpenMP.Keywords:

Mini-apps,Performance,VULCAN,Shared

Memory,MPI,OpenMP1INTRODUCTION

Inmanyfields,real-worldapplicationshavealreadybeendeveloped.Forestablishedapplicationstostayup-to-date,newparallelstrategiesmustbeexploredtodeterminewhichmayyieldthebestperformance,especiallywithadvancesincomputinghardware.However,restructuringormodifyingareal-worldapplicationincursincreasedcostdependingonthesizeofthecodeandchangestobemade.Amini-appmaybecreatedtoquicklyexploresuchoptionswithoutmodifyingtheentirecode.Mini-appsreducetheoverheadofapplyingnewstrategies,thusvariousstrategiesmaybeimplementedandcompared.Thisworkpresentstheauthorsexperienceswhenfollowingthisstrategyforareal-worldapplicationdevelopedbyNASA.

VULCAN（ViscousUpwindAlgorithmforComplexFlowAnalysis）isaturbulent,noequilibrium,finite-ratechemicalkinetics,Navier-Stokesflowsolverforstructured,cell-centered,multiblockgridsthatismaintainedanddistributedbytheHypersonicAirBreathingPropulsionBranchoftheNASALangleyResearchCenter（NASA2016）.Themini-appdevelopedinthisworkusestheHouseholderReflectorkernelforsolvingsystemsoflinearequations.Thiskernelisusedoftenbydifferentworkloads,andisagoodcandidatetodecidewhatstrategytypetoapply

toVULCAN.VULCANisbuiltonasingle-layerofMPIandthecodehasbeenoptimizedtoobtainperfectvectorization,thereforetwo-levelsofparallelismarecurrentlyused.Thisworkinvestigatestwoflavorsofshared-memoryparallelism,OpenMPandSharedMPI,whichwillprovidethethird-levelofparallelismfortheapplication.Athird-levelofparallelismincreasesperformance,whichdecreasesthetime-to-solution.

MPIhasextendedthestandardtoMPIversion3.0,whichincludestheSharedMemory（SHM）model（MikhailB.（Intel）2015,MessagePassingInterfaceForum2012）,knowninthisworkasSharedMPI（SMPI）.ThisextensionallowsMPItocreatememorywindowsthataresharedbetweenMPItasksonthesamephysicalnode.Inthisway,MPItasksareequivalenttothreads,exceptSharedMPIismoredifficultforaprogrammertoimplement.OpenMPisthemostcommonshared-memorylibraryusedtodatebecauseofitsease-of-use（OpenMP2016）.Inmostcases,onlyafewOpenMPpragmasarerequiredtoparallelizealoop;however,OpenMPissubjecttoincreasedoverhead,whichmaydecreaseperformanceifnotproperlytuned.

Asearlyastheyear2000,theauthorsin（CappelloandEtiemble2000）foundthatlatencysensitivecodesseemtobenefitfrompureMPIimplementationswhereasbandwidthsensitivecodesbenefitfromhybridMPI+OpenMP.Also,theauthorsfoundthatfasterprocessorswillbenefithybridMPI+OpenMPcodesifdatamovementisnotanoverwhelming

bottleneck（CappelloandEtiemble2000）.Sincethistime,hybridMPI+OpenMPimplementationshaveimproved,butnotwithoutdifficulties.In（Drosi-nosandKoziris2004,ChorleyandWalker2010）,itwasfoundthatOpenMPincursmanyperformancereductions,including:

overhead（fork/join,atomics,etc）,falsesharing,imbalancedmessagepassing,andasensitivitytoprocessormapping.However,OpenMPoverheadmaybehiddenwhenusingmorethreads.In（Rabenseifner,Hager,andJost2009）,theauthorsfoundthatsimplyusingOpenMPcouldincurper-formancepenaltiesbecausethecompileravo