...
首页> 外文期刊>PASJ: Publications of the Astronomical Society of Japan >Implementation and performance of FDPS: a framework for developing parallel particle simulation codes
【24h】

Implementation and performance of FDPS: a framework for developing parallel particle simulation codes

机译:FDPS的实现和性能:开发并行粒子模拟代码的框架

获取原文
           

摘要

Wepresentthebasicidea,implementation,measuredperformance,andperformancemodelofFDPS(FrameworkforDevelopingParticleSimulators).FDPSisanapplication-developmentframeworkwhichhelpsresearcherstodevelopsimulationprogramsusingparticlemethodsforlarge-scaledistributed-memoryparallelsupercomputers.Aparticle-basedsimulationprogramfordistributed-memoryparallelcomputersneedstoperformdomaindecomposition,exchangeofparticleswhicharenotinthedomainofeachcomputingnode,andgatheringoftheparticleinformationinothernodeswhicharenecessaryforinteractioncalculation.Also,evenifdistributed-memoryparallelcomputersarenotused,inordertoreducetheamountofcomputation,algorithmssuchastheBarnesa€“HuttreealgorithmortheFastMultipoleMethodshouldbeusedinthecaseoflong-rangeinteractions.Forshort-rangeinteractions,somemethodstolimitthecalculationtoneighborparticlesarerequired.FDPSprovidesallofthesefunctionswhicharenecessaryforefficientparallelexecutionofparticle-basedsimulationsasa€?templates,a€?whichareindependentoftheactualdatastructureofparticlesandthefunctionalformoftheparticlea€“particleinteraction.ByusingFDPS,researcherscanwritetheirprogramswiththeamountofworknecessarytowriteasimple,sequentialandunoptimizedprogramofO(N2)calculationcost,andyettheprogram,oncecompiledwithFDPS,willrunefficientlyonlarge-scaleparallelsupercomputers.AsimplegravitationalN-bodyprogramcanbewritteninaround120lines.Wereporttheactualperformanceoftheseprogramsandtheperformancemodel.Theweakscalingperformanceisverygood,andalmostlinearspeed-upwasobtainedforuptothefullsystemoftheKcomputer.Theminimumcalculationtimepertimestepisintherangeof30a€‰ms(N=107)to300a€‰ms(N=109).Thesearecurrentlylimitedbythetimeforthecalculationofthedomaindecompositionandcommunicationnecessaryfortheinteractioncalculation.Wediscusshowwecanovercomethesebottlenecks.
机译:Wepresentthebasicidea,实施,measuredperformance,andperformancemodelofFDPS(FrameworkforDevelopingParticleSimulators).FDPSisanapplication-developmentframeworkwhichhelpsresearcherstodevelopsimulationprogramsusingparticlemethodsforlarge-scaledistributed-memoryparallelsupercomputers.Aparticle-basedsimulationprogramfordistributed-memoryparallelcomputersneedstoperformdomaindecomposition,exchangeofparticleswhicharenotinthedomainofeachcomputingnode,andgatheringoftheparticleinformationinothernodeswhicharenecessaryforinteractioncalculation.Also,evenifdistributed-memoryparallelcomputersarenotused,inordertoreducetheamountofcomputation,algorithmssuchastheBarnesa€“HuttreealgorithmortheFastMultipoleMethodshouldbeusedinthecaseoflong-rangeinteractions.Forshort-rangeinteractions,somemethodstolimitthecalculationtoneighborparticlesarerequired。 FDPS提供了所有这些功能,这些功能可以有效地并行执行基于粒子的模拟,例如模板,而该模板独立于实际数据tructureofparticlesandthefunctionalformoftheparticlea€“particleinteraction.ByusingFDPS,researcherscanwritetheirprogramswiththeamountofworknecessarytowriteasimple,sequentialandunoptimizedprogramofO(N2)calculationcost,andyettheprogram,oncecompiledwithFDPS,willrunefficientlyonlarge-scaleparallelsupercomputers.AsimplegravitationalN-bodyprogramcanbewritteninaround120lines.Wereporttheactualperformanceoftheseprogramsandtheperformancemodel.Theweakscalingperformanceisverygood,andalmostlinearspeed-upwasobtainedforuptothefullsystemoftheKcomputer.Theminimumcalculationtimepertimestepisintherangeof30a€‰MS(N = 107)to300a€‰毫秒(N = 109)。目前,这些限制了域分解的计算时间和进行交互计算所需的通信时间。我们讨论表明可以克服许多瓶颈。

著录项

相似文献

  • 外文文献
  • 中文文献
  • 专利
获取原文

客服邮箱:kefu@zhangqiaokeyan.com

京公网安备:11010802029741号 ICP备案号:京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有
  • 客服微信

  • 服务号