Compile-time generation of regular communications patterns

A major hindrance to writing programs for distributed memory architectures has been the necessity of inserting explicit interprocessor communication. Recent work has shifted much of this detailed work from the programmer onto the compiler. In this paper, we provide explicit formulas describing the communications to be generated for an important class of computations: block-and cyclicly-distributed arrays with linear subscripts. We illustrate these methods with results from a compiler for the Kali language.