大规模结构网格数据的相关性统计建模轻量化方法杨阳1武昱1汪云海2曹轶1,31（北京应用物理与计算数学研究所北京100094）2（山东大学计算机科学与技术学院山东青岛266237）3（中物院高性能数值模拟软件中心北京100088）（yang1993@mail.ustc.edu.cn）CorrelationStatisticalModelingReductionMethodforLarge-ScaleStructuralGridDataYangYang1,WuYu1,WangYunhai2,andCaoYi1,31（InstituteofAppliedPhysicsandComputationalMathematics,Beijing100094）2（SchoolofComputerScienceandTechnology,ShandongUniversity,Qingdao,Shandong266237）3（CAEPSoftwareCenterforHighPerformanceNumericalSimulation,Beijing100088）AbstractDatavisualanalysisisessentialforlarge-scalenumericalsimulations.Thestoragebottleneckofhigh-performancecomputersmakesitchallengingtoanalyzeandvisualizedatawithoriginalhigh-resolution.Themethodbasedonstatisticalmodelingcansignificantlyreducethedatastoragecost,withthereconstructionuncertaintybeinghigh.Therefore,weproposealarge-scaledatareductionmethodforefficientanalysisandvisualizinglarge-scalemulti-blockvolumedatageneratedbymassivelyparallelscientificsimulations.Thetechnicalcoreofthismethodistoguidethestatisticalmodelingofadjacentdatablocksthroughthestatisticalrepresentationofcorrelationbetweendatablocks.Bydoingso,ourmethodefficientlypreservesthestatisticaldatapropertieswithoutmergingdatablocksstoredindifferentparallelcomputingnodesandrepartitioningthemaccordingtothehomogeneityrequirementsofthevisualization.Comparedwithexsitingmethods,theoriginaldatacanbereconstructedmoreaccuratelybycouplingnumericaldistributioninformation,spatialdistributioninformation,andcorrelationinformation,furtherreducingthevisualuncertainty.Theexperimentaltestsusefivesetsofscientificdatawiththelargestscaleofonebilliongrids.Thequantitativeanalysisresultsshowthatourmethodimprovesthedatareconstructionaccuracybyuptotwoordersofmagnitudeatthesamedatacompressionratiocomparedwiththecurrentstate-of-the-artmethods.Keywordsdatareduction；massivelyparallelscientificsimulation；large-scalemulti-blockvolumedata；correlationstatistica...
