Heuristics for Mapping Parallel Computations to Parallel Architectures

Abstract

This paper studies the problem of allocating the interacting task modules, of a parallel program, to the het,erogeneous processors in a parallel architecture. The goal is to provide a load balanced allocation which minimizes the completion time of the program. The problem of finding an optimal allocation is known to be an NP-hard problem, even if the processors are homogeneous, and thus necessitates development of heuristic schemes. This paper presents three heuristic algorithms for task assignment, based on simulated annealing, tabu search, and stochastic probe approaches respectively. We present an experimental analysis of these three heuristics and compare their performance. Experiments reveal that our new stochastic probe approach always yields the best solutions while requiring significantly less CPU time.