DOMAIN DECOMPOSITION STRATEGIES FOR SOLVING THE MAXWELL EQUATIONS ON DISTRIBUTED PARALLEL ARCHITECTURES

摘要

Domain decomposition strategies for solving hyperbolic systems of partial differential equations on distributed-memory parallel computing platforms are investigated. The logically-rectangular computational domain is divided either one, two, or three dimensionally into a series of computational blocks, and each block is assigned to a single processor. Theoretical predictions using standard parallel performance models indicate that higher-dimensional decompositions provide superior parallel program performance in terms of scalability. The theory is tested using a finite-volume time-domain (FVTD) Maxwell equations solver to compute the electromagnetic fields inside a rectangular waveguide using various grid sizes and processor numbers on three different parallel architectures- the Intel Paragon, the IBM SP2, and the Cray T3D. The specific performance of the FVTD algorithm on the three machines is investigated, the relation between processor connection topology and message passing performance of a typical grid-based hyperbolic equation solver are identified, and the results are used to augment the classical parallel performance model. Although clear performance trends emerge in terms of the dimensionality of the decomposition, results indicate that higher-dimensional decompositions do not always provide superior parallel performance. [Vol. 12, No. 3 (1997), pp 4-15]