A Methodological Approach to Parallel Simulated Annealing on an SMP System

Abstract

Simulated annealing (SA) is a stochastic optimization technique which guarantees under certain conditions to converge to a global minimum. The major disadvantage of this technique is its very slow convergence: this makes it not suitable for many complex optimization problems. Different parallel versions of the algorithm have been proposed, but none of them addresses recent 2-way symmetric multiprocessor (SMP) machines. In this paper, we present a novel approach to the parallel implementation of SA on an SMP system. In addition, we offer an adaptive method to dynamically change the program execution flow at run time, as to obtain the maximum benefit from these shared memory parallel architectures. Since we only exploit time measures for this purpose, we obtain a problem independent and a general purpose implementation. The effectiveness of the method is demonstrated by extensively analyzing the traveling salesman problem (TSP) as a target case study, on a system under different workload conditions.