Abstract

Scalable parallel numerical libraries and automatically parallelizing compilers seem to be contrary approaches to the same goal: the user--friendly generation of efficient parallel numerical programs for shared--memory and distributed--memory multiprocessors. We propose a framework that integrates the library approach and parallelizing compiler technology. It is based on fast and powerful pattern recognition in sequential source programs and considerate local algorithm replacement. By a simplified prototype implementation, we demonstrate the functionality of this approach for a massively parallel shared--memory target machine, the SB-PRAM. We further propose constructive guidelines to adapt the method to distributed--memory multiprocessors by integrating an automatic array distribution engine, by synthetic performance prediction, and by data--distribution--independent design of library routine specification.

Citations