GPU集群加速近似逆预条件CG并行求解器.pdfVIP

ISSN 1673—9418 CODEN JKYTA8 JournalofFrontiersofComputerScienceandTechnology E—mail：fcst@Vip．163．eom httpl|| ．ceaj．org 1673-9418／2015／09(09)-1084-09 Tel：+86．10 doi：10．37780．issn．1673—9418．1412060 GPU集群加速近似逆预条件CG并行求解器 赵 莲 ，赵永华H，陈 尧 ．一，赵 慰 1,2 1．中国科学院 计算机网络信息中心，北京 100190 2．中国科学院大学，北京 100190 ApproximateInversePreconditionedCG ParallelSolveronGPU Cluster ZHA0 Lian。，ZHAOYonghuaH，CHEN Yao ，ZHAOWei 1．ComputerNetworkInformationCenter，ChineseAcademyofSciences，Beijing100190，China 2．UniversityofChineseAcademyofSciences，Beijing100190，China +Correspondingauthor：E-mail：yhzhao@sccas．cn ZHAO Lian，ZHAO Yonghua，CHEN Yao，eta1．ApproximateinversepreconditionedCG parallelsolveron GPUcluster．JournalofFrontiersofComputerScienceandTechnology,2015，9(9)：1084-1092． Abstract：ThispapershowsthestudyontheparallelalgorithmofAINV (approximateinverse)andSSOR—AI(sym— metricsuccessiveoverrelaxationapproximateinverse)preconditionersonGPUclustersystems．Withmultilevel k-waygraphpartitioning，thispaperproposestheparallelmethodwhichcanrtansfomr asparsematrixintoblockarrow fomr basedonamethodtoidentifyinterior／boundaryvertexofsubrgaphsandapemr utation．Basedontheblockarrow matrix，withthestrategyofsequentialcomputationapproximateinverseofinnerblockandparallelcomputation bewt eenthedifferentblocks，theparallelalgorithm ofAINV andSSOR-AIisobtained．Basedontheoptimization techniquesofcollaborativecomputingbewteenCPU andGPU，page—lockedhostmemoryandoverlappingtransfers withcomputationondevice，thispapercombinesparallelapproximateinversepreconditionerwithCG (conjugate rgadient)algorithmtoobtainahybridparallelsolverforlinearsystems．Numericalexperimentsindicatethatapplying theabovemethodscanobtainvery goodaccelerationeffectandscalabilitybothAINV parallelimplementat