资源描述:
《Hard Data on Soft Errors A Large-Scale Assessment ofReal-World Error Rates in GPGPU软错误地硬数据:大规模评估 GPGPU地真实世界错误率.pdf》由会员上传分享,免费在线阅读,更多相关内容在教育资源-天天文库。
1、HardDataonSoftErrors:ALarge-ScaleAssessmentofReal-WorldErrorRatesinGPGPUImranS.Haque1VijayS.Pande212StanfordUniversity,DepartmentsofComputerScienceandChemistryMay28,2018Abstractaccessibleonlyondedicatedsupercomputers,withpeakthroughputimprovementsofoveranorderofGraphicsprocessingunits(GP
2、Us)aregainingmagnituderelativetoconventionalCPUs.Thisex-widespreaduseincomputationalchemistryandtremelyhighperformancemakesGPUsattractiveotherscientificsimulationcontextsbecauseoftheirincomputationalchemistry,asmanyapplicationshugeperformanceadvantagesrelativetoconven-areboundbythelimitso
3、favailablecomputationaltionalCPUs.However,thereliabilityofGPUsinpower(e.g.,thesimulationtime,simulatedsystemerror-intolerantapplicationsislargelyunproven.Insize,andforcefielddetailinmoleculardynamicsareparticular,alackoferrorcheckingandcorrect-allfundamentallyconstrainedbyavailablecompu-i
4、ng(ECC)capabilityinthememorysubsystemsoftation).Indeed,GPUshavebeenappliedtosev-graphicscardshasbeencitedasahindrancetotheeralimportantproblemsincomputationalchem-acceptanceofGPUsashigh-performancecoproces-istryincludingmoleculardynamics[1,2],Poisson-sors,buttheimpactofthisdesignhasnotbe
5、enpre-Boltzmannelectrostatics[3],DFTandMP2quantumviouslyquantified.chemistrymodels[46],andmolecularcomparisonInthisarticlewepresentMemtestG80,oursoft-[7].Scientificproblemsoutsidechemistry,includ-wareforassessingmemoryerrorratesonNVIDIAingbiologicalsequencealignment[8]andmachineG80andGT200
6、-architecture-basedgraphicscards.learning[9]havealsoshownsignificantspeedupsFurthermore,wepresenttheresultsofalarge-fromreimplementationforGPUexecution.scaleassessmentofGPUerrorrate,conductedWhileGPGPU(general-purposecomputationonbyrunningMemtestG80onover20,000hostsonGPUs)isattractivefrom
7、theperspectiveofthrough-theFolding@homedistributedcomputingnetwork.put,itsoriginintherelatively-error-tolerantareaofOurcontrolexperimentsonconsumer-gradeandconsumergraphicshasraisedconcernsaboutrelia-arXiv:0910.0505v1[cs.AR]3Oct2009dedicated-GPGPUhardwareinacontrolledenvi