SMP機群上的并行代碼優(yōu)化技術(shù).pdf

1、中國科學(xué)院計算技術(shù)研究所博士學(xué)位論文SMP機群上的并行代碼優(yōu)化技術(shù)姓名：陳莉申請學(xué)位級別：博士專業(yè)：計算機系統(tǒng)結(jié)構(gòu)指導(dǎo)教師：張兆慶2002.6.1▲!竺!迎墮堂叢幽型墮———————一optimizationofParaUelCodesonSMPClustersChenLi(ComputerArctlitecture)DirectedByZhangZhaoqingClustersofSyHunetricMultiprocessorsha

2、vebeen塒despreadinhighperfomancecompming，也earchitecturehavemedr鋤aticfbatureofhy“dmemarymodeIs：sharedmemory、“thinnodesaIlddis仃ibuted—memory鋤ongnodesThehvbridarchitectureputsfoⅣ，ardtllepmblemofpm曠鯽衄ingmodela11dperfo哪anceHyb

3、ridprogrammingmodelmixingMPIandmrcadsbringsoutgreatcrcomplexitytoprogrammers，webelievethatauni6edprogrammingmodelshouldbeadoptedonSMPclustersInthispaper，a11extensiontoOpenMPisdevisedtomatchtheunderlyingarchitecture，aIlda

4、nimplementationwhichsupportsmultiple1eVelparalklismisgivenIn—depmresearchisdor圮onissuesparticulartomishierarchyarchitectureThemaincontributionsofthispaperare：l、BaSedonthenotionof“Dis仃ibutedOpenMP”，deviseanOpenMPextension

5、suitcdtotheclusterarchitecture，severalexeeutionmodelsarediscussed，a目reneml丘ameworkisgiVen，矗mdamemalstrategiesabouthowtodetemlinemultiplelevelparallelismisdiscussed2)Dataandcomputationpartitionisalsoimport鋤tforSMPclusters

6、，aredundantcomputationpanitioningteclHliqueisgiVentooptimizinginternodecommunicationThetecllIliquefocusonp砌lelloopsequences，tlleideaistochooseredundantcompu詛廿onproperlyfbreachpamllelloop，inordertoeliminateinterloopcommun

7、icationswhicharcusuallyintroducedusingordinarypanitioIlingstrategy3)Anewapproachbasedonloopmsionispresentedtoexploitpipelillingparallelism。Thetechniqueexploitspipelining矗omcomplexloopstmctIlres，whichdistinguishesitself舶m

8、缸aditioIlalpipeliningtecmques4)Inaglobalco咖m岫cationoptimization矗齜nework，redundantmessageeliminationa11dmessagecoalescingarerealizedTheIn仰ductionofgloballyshadowedareagreatlyreduces出【協(xié)redistributionsincurredbyprocedllr面ca

9、llsTheinterprocedl砌optimi盈Itionofregularcomm蚰icationsfhrthereliminatesrcdundantmessages5)Thegrainof“innanode”p刪lelismareofimportancetoprogr鋤perfornlallce，aconstmctionalgorimmofparallelregionsisgiVeninterprocedurall弘6)Ane

10、wsynclllonizationeliminationtechniqueispresented，basedonstaticcomputationpartitioIlingT、vokindsOfnotionsareintroducedtodescribetherelationshipbetweenreferences，wheredatadependenceoccursThenotionsarecalIeds仃ongconsistent，