On random sampling for parallel simulated annealing

We investigate the parallelism of simulated annealing for graph partitioning. A random sampling technique, in which disjoint partitions of the graph are distributed across the processors so that moves can be proposed and evaluated asynchronously in distinct processors, is proposed. Synchronizations among processors are not necessary until the equilibrium is reached at each processor. Such an approach may thus achieve a speedup that is approximately linear in the number of processors. Furthermore, since no interacting moves can occur in our strategy, it is obvious that our scheme is free of errors in cost evaluation