Organizing communication of parallel processes during automatic parallelization of loop nests with static control flow for cluster systems using polyhedral model
Abstract:
A method to organize communication of parallel processes is proposed for translation of CLooG output into parallel MPI program preserving optimal data placement. It is assumed that space-time mappings are affine and calculated using polyhedral model methods according to the principle of space and time data locality optimization. The notion of a communication polyhedron is introduced to define a data set participating in any information exchange within two-sided MPI communication scheme. Application of the method is illustrated with parallelization of LU-decomposition algorithm.
(in Russian).
Key words and phrases:automatic parallelization, polyhedral model, data locality, clusters, intercommunications.