Formal Framework for Semi-Automatic Parallel Program Generation

摘要

The development of programs for distributed-memory multiprocessors is currently a very laborious and error-prone task. The paper describes a formal framework for the semi-automatic generation of programs for distributed-memory parallel processors. The framework is language independent and allows the handling of static and dynamic data decompositions, like block-, scatter, and combined block/scatter decomposition and is based on a calculus and a set of associated rewrite rules. The applicability of the framework is shown by the automatic translation of Booster programs. Booster is a high-level programming language for the description of parallel algorithms. Given a Booster program and a separate description of the required data-decomposition(s), efficient programs for the individual processors are automatically generated in one of the standard imperative languages (such as FORTRAN or C).