TheFastestPOPintheWorld课件
日期:2010-12-22 12:48
TheFastestPOPintheWorldTheCrayX1–PremierWeather/ClimateWorkhorseUnderstandingtheX1Weather/ClimateapplicationsoptimizedforX1NLOM,MM5,POP,Hirlam,ROMWeather/ClimateapplicationsbeingoptimizedfortheX1CCSM,HYCOM,IFS,WRF,SomeResultsUnderstandingtheX1MustbewillingtorestructureThecompilerwillneverbegoodenoughtodoitallScalarisverybadShortvectorisgood,butyoucandobetterDONOTBESATISFIEDWITHVECTORIZATIONOFTHEDEPTH/HEIGTHLOOPTheX1inter-connectallowsforscalingtoveryhighsustainGFLOPSPOPOptimizationsImpvmixtandImpvmixuPullingtheinnerloopinsidetheKloopCSDdirectivesontheseconddimensionKloopisinthemiddleAniso_hmixPullingtheinnerloopinsidethequadrantloop(4)PullingtheinnerloopinsidecasestatementsCSDdirectivesontheseconddimensionCo-arraysOnlyusedwhereMPIneededhelpGlobal_summationNine-pointstencilinsolveriterationloopStartedusingLocksinsteadofbarriersUsedataskewingwiththelocksandglobalsummationscalarHowtowriteaGlobal_sumusingCo-arraysAllprocessorscomeintotheroutineEveryonedoeslocalsumand/or,
查看全部