Optimization of the Parallel Finite Element Method for the Earth Simulator

抄録

The feasibility of the GeoFEM as a platform for the parallel finite element method on the earth simulator was investigated. Since the earth simulator consists of 640 SMP nodes, each of which has eight vector processors, there are three levels of hierarchical parallelization methods: inter-node, intra-node, and vectorization. GeoFEM has extremely high inter-node parallel efficiency. However, the application of GeoFEM in an environment involving over 1,000 processors has not yet been examined. Furthermore, the hierarchical architecture of the Earth Simulator requires optimization for intra-node parallelization and vectorization for better practical performance. Various ordering methods have been used to accomplish intra-node parallelization and vectorization, and we eventually achieved a performance of 10 TeraFLOPS for a 6.4-GDOF problem.