高性能計算導論-introduction

高性能計算導論課程,IntroductiontoHighPerformanceComputing,課程內(nèi)容,什么是高性能計算應(yīng)用對高性能計算的需求硬件對高性能計算的支持軟件對高性能計算的支持,參考文獻,黃鎧，徐志偉著，可擴展并行計算技術(shù)，結(jié)構(gòu)與編程，北京：機械工業(yè)出版社，2000陳國良著，并行計算-結(jié)構(gòu)、算法、編程，北京：高等教育出版社，1999都志輝著，高性能計算并行編程技術(shù)-MPI并行程序設(shè)計，北京：清華大學出版社，2001Mpi:/mpiOpenMP:CUDA:,WhatisHPC?,WhatisHPC?Fromwikipedia,High-performancecomputing(HPC)usessupercomputersandcomputerclusterstosolveadvancedcomputationproblems.Today,computersystemsapproachingtheteraflops-regionarecountedasHPC-computers.supercomputersandcomputerclustersparallelcomputerAdvancedcomputationproblemsparallelprocessingalgorithm,ParallelComputer,Aparallelcomputerisasetofprocessorsthatareabletoworkcooperativelytosolveacomputationalproblem.Thisdefinitionisbroadenoughtoincludeparallelsupercomputersthathavehundredsorthousandsofprocessorsnetworksofworkstationsmultiple-processorworkstationsComputerwithmulti-coreprocessorandembeddedsystemsParallelcomputersofferthepotentialtoconcentratecomputationalresources-processors-onimportantcomputationalproblems,Largescaleofconcentratedcomputationalresources,GridCloud,ParallelAlgorithm,Specifymultipleoperationsoneachstep,ApplicationsofHPCDemandsonHPC,應(yīng)用對高性能計算的需求,Areas氣象、環(huán)境、海洋數(shù)值預(yù)報應(yīng)用高能物理研究生命科學應(yīng)用（基因序列搜尋與比對、新藥研發(fā)）計算機輔助工程（工業(yè)生產(chǎn)）石油勘探圖像渲染ReductionofexecutiontimeConcurrencyprovided,Highperformancecomputerarchitecture,硬件對高性能計算的支持,并行計算機模型物理機器模型,物理（并行）機器模型,SIMD計算機：Single-instructionmultiple-dataCPU+GPUMIMD計算機：multiple-instructionmultiple-dataSMPMPPClusterBladeserver,Parallelcomputingmodel,主要用于設(shè)計、分析并行算法,抽象（并行）機器模型（abstractmachinemodel）,PRAM模型（parallelrandom-accessmachine）多處理器之間的通信通過訪問共享存儲器實現(xiàn)BSP模型（bulksynchronousparallel）多處理器之間的通信通過網(wǎng)絡(luò)傳輸實現(xiàn),PRAM模型,Thecommunicationoverheadisignored,BSP模型,Thecommunicationlatencyandsynchronousoverheadareconsidered.,模型中的計算行為,PRAM：ExclusivereadConcurrentreadExclusivewriteConcurrentwrite(unsafe)BSP：Computationsuperstepssynchronization,成本函數(shù),PRAM:TEREWTCREWTCRCWBSP:TBSP=Pi+C+sLs:totalnumberofsuperstepsi.0s-1P:computationtimeC:communicationovertimeL:synchronizationoverhead,programmingmodel,軟件對高性能計算的支持,軟件對高性能計算的支持,編程模型Aprogrammingmodelisacollectionofprogramabstractionsprovidingaprogrammerasimplifiedandtransparentviewofthecomputerhardware/softwaresystem.Parallelprogrammingmodelsarespecificallydesignedformultiprocessors,multicomputers,orvector/SIMDcomputers,編程模型ProgrammingModel,消息傳遞模型（messagepassing）MPI共享存儲模型（sharedmemory）OpenMP數(shù)據(jù)并行模型（dataparallel）CUDA/OpenCLCPU-GPUarchitectureMapReduce編程模型,Programstructure,程序結(jié)構(gòu),SPMD：singleprogrammultipledataMPMD：multipleprogrammultipledataSIMD：singleinstructionmultipledataMaster/Workers,衡量高性能系統(tǒng)性能的評價指標,浮點運算的理論峰值（FLOPS）主頻*總核心數(shù)*每個時鐘周期浮點運算次數(shù)浮點運算的實測峰值（FLOPS）測試組件（Linpack）測試并行系統(tǒng)加速比加速比是衡量“并行收益”的重要指標Amdahl定律適用于固定計算規(guī)模的加速比性能描述Gustafson定律適用于可擴展問題,Amdahl定律,計算負載W(原(未改進)系統(tǒng)度量)串行分量Ws并行分量Wp并行部分的執(zhí)行速度相對于原串行部分的加速比p串行分量比例f，并行分量比例1-f并行計算系統(tǒng)加速比:S=(Ws+Wp)/(Ws+Wp/p)=1/(