Parallel computing and domain decomposition

Domain decomposition techniques appear a natural way to make good use of parallel computers. In particular, these techniques divide a computation into a local part, which may be done without any interprocessor communication, and a part that involves communication between neighboring and distant processors. This paper discusses some of the issues in designing and implementing a parallel domain decomposition algorithm. A framework for evaluating the cost of parallelism is introduced and applied to answering questions such as which and how many processors should solve global problems and what impact load balancing has on the choice of domain decomposition algorithm. The sources of performance bottlenecks are discussed. This analysis suggests that domain decomposition techniques will be effective on high-performance parallel processors and on networks of workstations.