Trasgo 2.0: Code generation for parallel distributed- and shared-memory hierarchical systems

T asgo 2.0: Code gene a ion o pa allel dis ibu ed- and sha ed-memo y hie a chical sys ems Ana Mo e on-Fe nandez, A u o Gonzalez-Esc ibano, and Diego R. Llanos Depa amen o de In o m´a ica, Uni e sidad de Valladolid [email p o ec ed],{a u o,diego}@in o .u a.es 1 Ex ended Abs ac Cu en mul icompu e s a e ypically buil as in e connec ed clus e s o sha ed- memo y mul ico e compu e s. A common p og amming app oach o hese clus- e s is o simply use a message-passing pa adigm, launching as many p ocesses as co es a ailable. Ne e heless, o be e exploi he scalabili y o hese clus- e s and highly-pa allel mul ico e sys ems, i is needed o e icien ly use hei dis ibu ed- and sha ed-memo y hie a chies. This implies o combine di e en p og amming pa adigms and ools a di e en le els o he p og am design. P og amming in his kind o en i onmen is challenging. Many success ul pa allel p og amming models and ools ha e been p oposed o speci ic en i on- men s. Howe e , he applica ion p og amme s ill aces many impo an deci- sions no ela ed wi h he pa allel algo i hms, bu wi h implemen a ion issues ha a e key o ob aining e icien p og ams. Fo example, decisions abou pa - i ion and locali y s. synch oniza ion/communica ion cos s; g ain selec ion and iling; p ope pa alleliza ion s a egies o each g ain le el; o mapping, layou , and scheduling de ails. Mo eo e , many o hese decisions may change o di e - en machine de ails o s uc u e, o e en wi h da a sizes. This pape p esen s an au oma ic code gene a ion sys em o mixed dis i- bu ed- and sha ed-memo y pa allel mul icompu e s. We p esen an ex ension o he T asgo p og amming model. This ex ended model suppo s a wide ange o pa allel s uc u es and applica ions whe e coo dina ion is exp essed a an abs ac le el. T anspa en modula objec s a e in oked o guide he pa i ion and mapping o bo h da a and p ocesses, ac oss he whole sys em. We p esen a echnique ha , o a ine exp essions, compu e exac agg ega ed communica- ions a he dis ibu ed le el. I uses in e sec ion o emo e and local oo p in s in e ms o he mapping policies selec ed. Mo eo e , T asgo 2.0 in eg a es poly- hed al analysis ools o ob ain op imiza ions inside each sha ed-memo y pa allel node a he sha ed le el. This app oach allows o au oma ically gene a e mul- ile el pa allel p og ams ha adap hei communica ion and synch oniza ion s uc u es o he a ge machine. Ou expe imen al esul s o bo h, sha ed- and dis ibu ed-memo y en i onmen s, show how his app oach can au oma ically p oduce e icien codes when compa ed wi h manually-op imized codes using MPI o OpenMP models.