2016 Volume 2016 Pages 20160006
To improve the total performance of numerical simulation on PC clusters or supercomputers, not onl parallel efficiency, but also performance per computing node is becoming more important than ever before. Here, we propose an new approach to optimize the performance of a finite element code based on domain-decomposition method (DDM) on the current multi-core architecture. We propose LSC (Local Schur Complement) approach, where the Schur complement matrix is explicitly computed. We demonstrate that our approach is not only more efficient in memory usage but also faster than existing approaches, which depend heavily on the forward- and back-substitution stage of a direct solver, with our performance prediction models and also numerical examples on K Computer and Intel architecture.
