Large Scale Elasto-Plastic Analysis Using Domain Decomposition Method Optimized for Multi-core CPU Architecture

摘要

To solve a large scale elasto-plastic dynamics analysis of a complicated structure, such as a seismic analysis of a nuclear power plant and a skyscraper, a new implementation strategy for a parallel finite element code, suitable on a parallel supercomputer with modern multi-core / many core scalar CPUs, has been required. In this work, we propose a new design and programming style to optimize the performance of a parallel finite element code based on the domain-decomposition method (DDM) on multi-core CPUs, considering their cache hierarchy. Instead of a traditional, memory access-intensive approach, DS (Direct solver-based matrix Storage), two new matrix storage-free approaches, DSF (Direct solver-based matrix Storage-Free) and ISF (Iterative solver-based matrix Storage-Free), are proposed. Our new DSF/ISF-based DDM solver is not only more efficient in memory usage but also comparable in computational time against existing DS-based DDM solvers on multi-core CPU architectures.