有限元结构分析的层级负载均衡并行计算方法

发布时间：2018-07-07 21:24

本文选题：多核机群 + 有限元分析　；参考：《科学通报》2017年13期

【摘要】：由于性价比高、计算能力强,多核机群已经成为当今高性能计算的主流工具.然而,多核机群环境下不同的存储机制和通信延迟特点也为高效并行算法的设计带来了挑战.为充分利用多核机群的硬件资源获取最优性能,本文设计了一种有限元结构分析的层级负载均衡并行计算方法.该方法建立在对计算任务的层次性和粒度性充分挖掘的基础上.为与多核机群的硬件拓扑体系结构相适应,本文将计算任务划分为三个层次:节点间并行、片间并行和核间并行.其中,节点间并行和片间并行采用粗粒度并行计算方法,而核间并行采用细粒度并行计算方法.通过将计算任务映射到多核机群的不同硬件层面执行,该方法不仅有效实现了不同层面的负载均衡,而且大幅度降低了系统的通信开销.此外,它还大幅度减少了子区域的数目,有效提高了界面方程的数值收敛性.为验证算法的有效性,在"天河二号"超级计算机上进行了有限元结构线性静力分析大规模并行计算测试.结果表明:同传统区域分解法相比,层级负载均衡并行计算方法能够获得较高的加速比和并行效率.本文的研究主要集中在线性静力学问题上.对于非线性问题或者动力学问题,由于涉及多个迭代步,因此可以将本文算法封装为一个子函数进行调用.
[Abstract]:Because of its high cost-performance and strong computing power, multi-core cluster has become the mainstream tool of high-performance computing. However, different storage mechanisms and communication delay in multi-core cluster environment also bring challenges to the design of efficient parallel algorithms. In order to make full use of the hardware resources of multi-core cluster to obtain optimal performance, a hierarchical load-balancing parallel computing method for finite element structural analysis is designed in this paper. The method is based on fully mining the hierarchy and granularity of computing tasks. In order to adapt to the hardware topology architecture of multi-core cluster, the computing tasks are divided into three levels: inter-node parallelism, inter-chip parallelism and inter-kernel parallelism. Among them, coarse-grained parallel computing method is used for inter-node and inter-chip parallelism, while fine-grained parallel computing method is used for inter-kernel parallelism. By mapping computing tasks to different hardware levels of multi-core cluster, this method not only realizes load balancing at different levels, but also greatly reduces the communication overhead of the system. In addition, it reduces the number of subregions and improves the numerical convergence of the interface equation. In order to verify the validity of the algorithm, a large scale parallel computation test of finite element structure linear static analysis was carried out on Tianhe 2 supercomputer. The results show that compared with the traditional domain decomposition method, the parallel computing method of hierarchical load balancing can achieve higher speedup and parallel efficiency. The research in this paper mainly focuses on the linear statics problem. For nonlinear or dynamical problems, the algorithm can be encapsulated as a subfunction to be called because it involves multiple iterative steps.
【作者单位】：上海交通大学机械系统与振动国家重点实验室;上海交通大学机械与动力工程学院;上海超级计算中心;
【基金】：国家高技术研究发展计划(2012AA01A307) 国家自然科学基金(11272214,51475287) 国家重点研发计划(2016YFB0201800)资助
【分类号】：TP338.6

【相似文献】