Automatic Parallelization for Distributed Memory Multiprocessors
Anne Dierstein, Roman Hayer, Thomas Rauber · 1994
This paper describes a framework for a parallelizing compiler for distributed memory multiprocessor machines (DMMs). The framework provides a compiler and a runtime support library which allows to use the DMMs with a sequential language. The compiler computes a data distribution for the arrays of the source program and parallelizes the inner loops of the program. The data distribution is computed by a branch-and-bound algorithm that uses a performance estimator to evaluate the relative efficiency of different data decomposition schemes for any given program. The performance estimation takes place at compile time and uses several parameters of the used DMM like the startup time and the byte transfer time. The paper also describes a prototype implementation of the framework on an Intel iPSC/860 for the language Pascal and discusses some experimental evaluations.