【作者】 战修广； 姜殿恒； 张盛； 陈飙松； 李云鹏；

【Author】 Zhan Xiuguang;Jiang Dianheng;Zhang Sheng;Chen Biaosong;Li Yunpeng;The Chinese Society of Theoretical and Applied Mechanics;Institute of Mechanics,Chinese Academy of Sciences;

【机构】 工业装备结构分析国家重点实验室工程力学系大连理工大学；

【摘要】 本文针对直接法并行求解有限元过程中的难点问题,基于SiPESC平台实现了多重多级子结构有限元MPI+OpenMP混合并行计算框架,讨论了有限元求解过程中内存使用遇到的各类难点及其解决办法。有别于区域分解法使用迭代法求解的并行分析问题,使用直接法求解平衡方程的有限元并行分析面临着巨大的内存压力和并行难度。本文分别通过发展一种新的区间二叉树结构,采用内存池技术,改进多重多级子结构有限元矩阵组装策略,优化内外存通信压力等处理手段,提升了多重多级子结构有限元并行求解的规模和效率,得到了同商用软件一致的计算结果,以及较高的并行效率。更多还原

【Abstract】 Aiming at the difficult problems in parallel finite element solution by direct method, a hybrid parallel computing framework of multi-level substructure finite element with MPI + OpenMP is implemented based on SiPESC platform. Various difficulties encountered in memory usage during finite element solution and their solutions are discussed. Different from the parallel analysis problem solved by iteration method using domain decomposition method, the parallel finite element analysis using direct method to solve equilibrium equation faces huge memory pressure and parallel difficulty. In this paper, by developing a new interval binary tree structure, using memory pool technology, improving the assembly strategy of multi-level substructure finite element matrix, optimizing the internal and external communication pressure and other processing means, the scale and efficiency of multi-level substructure finite element parallel solution are improved, and the results are consistent with those of commercial software.更多还原