Simulated annealing task to processor mapping for domain decomposition methods on distributed parallel computers

Abstract

The parallel solution of large sets of equations derived from finite element or finite difference methods often involves the use of domain decomposition methods. This paper is concerned with the related problem of mapping subdomain partitioning to a processor topology in such a way that the communication overheads during solution are minimized. In this manner, processors far away from each other in the processor network are not required to communicate directly. This is achieved by solving an optimization problem using the simulated annealing method. The techniques described here can be applied more generally to distribute tasks (as opposed to subdomains) to processors. Copyright  1999 John Wiley & Sons, Ltd.