Solving the mapping problem with a genetic algorithm on the MasPar-1

Abstract

Good mapping algorithms can significantly reduce the total execution time of a program. However, the mapping problem is NP-complete. Consequently, heuristic methods should be used. Massively parallel systems allow the implementation of genetic algorithms running on large populations. In this paper, an algorithm based on a neighbourhood model is presented. The program has been implemented on 4096-processor MasPar-1 multicomputer. Experimental results for three genetic operators are presented and compared. The influence of initialisation strategies and selection techniques is also considered. A new initialization strategy based on grouping of adjacent tasks into approximately equal clusters is proposed.<>