Performance Evaluation of 3-D Hybrid Parallel Finite Element Method by MPI/OpenMP

抄録

　We developed a 3-D hybrid parallel finite element method with the process-parallelization by MPI and the thread-parallelization by OpenMP. In this method, the domain decomposition method is used in the process-parallelization and the multicolor ordering is used in the thread-parallelization. In this paper, we analyze the performance of the hybrid-parallelization through the cogging torque analysis of an IPM motor. As a result, it is possible to calculate the magnetic field faster using MPI and OpenMP at the same time. Hybrid parallelized computation at 8 nodes × 4 threads is 8.7 time faster than non-parallelized computation. Parallelization with OpenMP using the multicolor ordering was able to efficiently parallelize the forward-backward substitutions of ICCG. However, number of ICCG iterations increases due to the multicolor ordering in the hybrid parallelization.