A Simulated Annealing Algorithm for GPU Clusters

Simulated Annealing (SA) is a powerful global optimization technique that is frequently used for solving many practical problems from various scientific and technical fields. In this article we present a novel approach to parallelization of SA and propose an algorithm optimized for execution in GPU clusters. Our technique exploits the basic characteristics of such environments by using hierarchical problem decomposition. The proposed algorithm performs especially well for complex problems with large number of variables. We compare our approach with traditional parallel Simulated Annealing, both in terms of speed and result accuracy.