An implementation of direct linear equation solver using a many-core CPU for mechanical dynamic analysis

This research proposes an effective implementation of linear equation solver for an implicit integration on a many-core CPU. Although this implementation is applied to a flexible body simulation in mechanical dynamics, it could be also utilized in a wide range of other fields. BFS-based nested dissection and its numerical factorization enables adaptive control of setting operational range as well as positive parallelization compared with traditional DFS-based nested dissection. It brings better parallel efficiency when various sized separators are divided into blocks under a certain size. This study presents an experiment to identify an optimal maximum block size. Sparse matrices from mechanical dynamics software are numerically factorized, and the time results show that CACHE memory mode is appropriate for a better performance than FLAT mode. And it is recommended to split the operational region in accordance with MCDRAM size in this experiment. Our research shows fairly similar performance to DSS included in MKL and speeds up the time approximately 8 - 14 times in comparison with CHOLMOD, a part of SuiteSparse.