Using asynchronous and bulk communications to construct an optimizing compiler for distributed-memory machines with consideration given to communications costs

The very nature of distributed-memory parallel architectures demands the serious consideration of interprocessor communications, by which non-local memory access can be implemented. In this paper, we propose a method for optimization that utilizes asynchronous and bulk communications. We constructed an HPF^- compiler, a subset of HPF, and evaluated it on the CM5, a true distributed-memory machine using three non-trivial benchmark programs. With optimized code, there was considerable improvement in communications over non-optimized code and we had much better results than when optimizing by means of CM Fortran.