Optimizing Parallel Program Execution by Self-Organizing Maps

This paper presents self-organizing feature maps as an efficient tool generating solutions of the mapping problem. Parallel program and parallel computer are modeled by graphs. Based on the Kohonen learning rule some adaptations are necessary to apply self-organizing maps. Special metrics reflecting properties of the parallel program respectively the parallel computer allows to map arbitrary parallel programs onto most of the common parallel architectures (two-dimensional lattice, threedimensional torus, hypercube, etc.). Simulations and applications to reference problems show that self-organizing maps are more efficient than other optimizing methods (e.g. Simulated Annealing) applied to this NP-hard problem. self-organizing maps, optimization, parallel processing, load balancing, mapping 1 INTRODUCTIO