Iterative Method-based Subdomain Local Solver with Convergence Tolerance Control in Domain Decomposition Method

Abstract

Domain-decomposition method (DDM) is an efficient parallelization scheme for the fintie element method. To optimize the performance of the DDM subdomain local solver for multi-core CPU architecture, we propose iterative solver-based approaches, which enable a DDM-based ”on-cache” iterative solver. A subdomain local solver using SSOR pre-conditioned CG with Eisenstat’s trick, which is OpenMP parallelized, is not only more efficient in memory usage but also comparable or even faster in computational time than the existing direct solver-based approach on multi-core PC clusters.