Abstract: Aiming at the problem that existing communication optimization algorithms can't make MPI automatic parallelizing compilers to generate messag}passing programs of ideal speedups, this paper proposed a communication optimination algorithm using reordering transformation and loop distribution. According to the interprocedural side effect sets and reordering transformation rule based on mpi_wait/mpi_irecv movements, this algorithm makes orderly use of rcordering transformation and loop distribution, and then expands communication-computation overlap windows of the point to-point non-blocking communication as safely as possible, so that MPI automatic parallelizing compilers can generate message-passing codes of overlapping communication with more computation. Experimental results show that this algorithm can hide more point-to-point non-blocking communication overheads than others and improve their specdups significantly.

Key words: Reordering transformation, Loop distribution, Communication optimization, Parallelizing compilation, Point-to-point non-blocking communication, Communication-computation overlap