Linpack测试指南v1

Linpack测试指南-V1.0Author:DongjianWuEmail:[e-mail]说明，此文档基于IntelHPC系列软件编写。1.确认安装了IntelC++Compiler\IntelFortranCompiler，IntelMKL，IntelMPI本文档系统和软件版本OSRHEL5u3IntelCompiler11.1.072IntelMPI3.2.2.0062.编译Linpacka.配置好环境变量source/data1/intel/Compiler/11.1/072/bin/iccvars.shintel64source/data1/intel/Compiler/11.1/072/bin/ifortvars.shintel64source/data1/intel/impi/3.2.2.006/bin64/mpivars.shsource/data1/intel/Compiler/11.1/072/mkl/tools/environment/mklvarsem64t.shb.编译linpackcd/data1/intel/Compiler/11.1/072/mkl/benchmarks/mp_linpackmakearch=em64t编译完成后cd/data1/intel/Compiler/11.1/072/mkl/benchmarks/mp_linpack/bin/em64t在此可以看到HPL.dat和xhpl2个文件，其中HPL.dat为配置文件，xhpl为可执行程序。3.配置HPL.dat把它们拷到共享目录。cpHPL.datxhpl/lustre/test/linpackviHPL.datHPLinpackbenchmarkinputfileInnovativeComputingLaboratory,UniversityofTennesseeHPL.outoutputfilename(ifany)6deviceout(6=stdout,7=stderr,file)1#ofproblemssizes(N)50760Ns2#ofNBs192224NBs0PMAPprocessmapping(0=Row-,1=Column-major)2#ofprocessgrids(PxQ)32Ps46Qs16.0threshold1#ofpanelfact012PFACTs(0=left,1=Crout,2=Right)2#ofrecursivestoppingcriterium42NBMINs(=1)1#ofpanelsinrecursion2NDIVs1#ofrecursivepanelfact.102RFACTs(0=left,1=Crout,2=Right)1#ofbroadcast0BCASTs(0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM)1#oflookaheaddepth0DEPTHs(=0)2SWAP(0=bin-exch,1=long,2=mix)256swappingthreshold1L1in(0=transposed,1=no-transposed)form1Uin(0=transposed,1=no-transposed)form0Equilibration(0=no,1=yes)8memoryalignmentindouble(0)主要注意N，NB，P和Q的值N是矩阵的大小，其公式为：N*N*8=总内存*80%，在intelNehalem以上的cpu,N*N*8=总内存*90%测试的效果会更好些。NB是求解矩阵分块大小，一般为192,224P和Q要满足：P*Q=进程数；PQ这2个关系4.运行linpackmpdbootmpiexec–machinefilehosts–n12./xhpl5.看输出结果，并算出效率。一台双路服务器配置了2颗Xeonx5650@2.66GHz的理论峰值为2.66（主频）*4(每个时钟执行4次浮点运算)*2（2颗cpu）*6(每颗cpu有6个核)=127.68GFlops实测的一组输出结果：================================================================================T/VNNBPQTimeGflops--------------------------------------------------------------------------------WR00C2L25076022426720.911.210e+02--------------------------------------------------------------------------------||Ax-b||_oo/(eps*(||A||_oo*||x||_oo+||b||_oo)*N)=0.0024948......PASSED============================================================================Finished8testswiththefollowingresults:8testscompletedandpassedresidualchecks,0testscompletedandfailedresidualchecks,0testsskippedbecauseofillegalinputvalues.----------------------------------------------------------------------------EndofTests.可以看出输出结果为121Gflops效率=121/127.68=94.7