Characterization of InputOutput Bandwidth

Characterization of InputOutput Bandwidth

ID:39711799

大小:788.90 KB

页数:10页

时间:2019-07-09

Characterization of InputOutput Bandwidth_第1页
Characterization of InputOutput Bandwidth_第2页
Characterization of InputOutput Bandwidth_第3页
Characterization of InputOutput Bandwidth_第4页
Characterization of InputOutput Bandwidth_第5页
资源描述:

《Characterization of InputOutput Bandwidth》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库

1、201342ndInternationalConferenceonParallelProcessingCharacterizationofInput/OutputBandwidthPerformanceModelsinNUMAArchitectureforDataIntensiveApplicationsTanLi,YufeiRen,DantongYu*,ShudongJin,ThomasRobertazziDepartmentofElectricalandComputerEngineering,StonyBr

2、ookUniversity,StonyBrook,NY,11794,USAEmail:tan.li@stonybrook.edu,yufei.ren@stonybrook.edu,shudong.jin@stonybrook.edu,thomas.robertazzi@stonybrook.edu*ComputationalScienceCenter,BrookhavenNationalLaboratory,Upton,NY11973,USAEmail:dtyu@bnl.govTABLEIAbstractDat

3、a-intensiveapplicationsfrequentlyrelyonmul-NUMAFACTOROFDIFFERENTSERVERCONFIGURATIONSticorecomputersystems,inwhichNon-UniformMemoryAccess(NUMA)isadominantarchitecture.TotransferdataintoandoutServertypeNUMAfactorfromthesehigh-performancecomputersbecomesabottle

4、neck,Intel4sockets/4nodes1.5andthusitiscrucialtounderstandtheirI/OperformanceAMD4sockets/8nodes2.7characteristics.However,thecomplexityinNUMAarchitectureAMD8sockets/8nodes2.8presentsanewchallengeinmodelingitsI/Oaccesscost,andthusHPbladesystem32nodes5.5leadto

5、difficultiesinconfiguringproperprocessorandmemoryaffinity.Inthispaper,weshowthatexistingNUMAexperimentalmethodsandmetricsareinappropriateoncontemporaryhigh-endsystems.Wecharacterizeastate-of-the-artNUMAhost,andpropose,tothebestofourknowledge,thefirstmethodologyo

6、ryisnottheonlyresourceaffectedbytheasymmetrictosimulateI/Ooperationsusingmemorysemantics,andmodelnatureofaNUMAarchitecture.PreviousstudiesexposedtheI/Obandwidthperformance.OurmethodologyisthoroughlytestedandvalidatedbymappingmultipleparallelI/Ostreamstoasign

7、ificantI/Operformancevariationinbothlatencyanddifferentsetsofhardwarecomponents(CPU,memory,networkbandwidth.WhenanI/Oaccessisremote,latencyincreasescards,andSSDs)andbymeasuringtheperformanceofeachandmaximumbandwidthusuallydecreasesfordatatransfer.mapping.Thee

8、xperimentalresultsandanalysisrevealthatourForexample,oneresearch[3]benchmarkedTCPperformancemethodologycandramaticallyreducecharacterizationworkload,ontwodifferentNUMAtestbeds.Theirmeasurementsh

当前文档最多预览五页,下载文档查看全文

此文档下载收益归作者所有

当前文档最多预览五页,下载文档查看全文
温馨提示:
1. 部分包含数学公式或PPT动画的文件,查看预览时可能会显示错乱或异常,文件下载后无此问题,请放心下载。
2. 本文档由用户上传,版权归属用户,天天文库负责整理代发布。如果您对本文档版权有争议请及时联系客服。
3. 下载前请仔细阅读文档内容,确认文档内容符合您的需求后进行下载,若出现内容与标题不符可向本站投诉处理。
4. 下载文档时可能由于网络波动等原因无法下载或下载错误,付费完成后未能成功下载的用户请联系客服处理。