大规模并行高阶矩量法的容错算法研究

发布时间：2018-01-26 07:29

  本文关键词： 超级计算机 并行矩量法 容错算法 现场保护 可靠性　出处：《电子与信息学报》2017年09期 　论文类型：期刊论文

【摘要】：基于超级计算机的大规模并行电磁计算对于解决实际工程中的复杂电磁难题具有重要意义,但超级计算机中由节点故障导致的进程崩溃事件的概率远远高于普通计算机。该文针对传统电磁计算难以有效应对进程崩溃的现状,提出一种高效的、适用于大规模并行高阶矩量法的容错算法。在现有并行高阶矩量法的基础上,基于"硬盘缓存"和"直接内存读取"设计高效率、高可靠性的现场保护算法,同时设计了高效的断点恢复算法。算法的有效性主要在于"固定的现场保护点"这一特点,它使得算法在有故障的情况下仍然可以正常有序地进行;而原算法每次碰到故障,则只能从头计算。数值仿真实验验证了容错算法在应对进程崩溃事件时的有效性,大幅提高了大规模并行高阶矩量法的可靠性。
[Abstract]:Large-scale parallel electromagnetic computation based on supercomputer is of great significance for solving complex electromagnetic problems in practical engineering. But the probability of process crash caused by node failure in supercomputer is much higher than that of common computer. This paper proposes an efficient method to solve the problem that traditional electromagnetic computation is difficult to deal with process crash effectively. Fault tolerant algorithm for large-scale parallel high order moment method. Based on the existing parallel high order moment method, a high efficiency and high reliability field protection algorithm is designed based on "hard disk cache" and "direct memory reading". At the same time, an efficient breakpoint recovery algorithm is designed. The effectiveness of the algorithm is mainly due to the feature of "fixed field protection point", which enables the algorithm to proceed normally and orderly in the event of failure. The original algorithm can only be calculated from scratch every time the fault occurs. Numerical simulation results show that the fault-tolerant algorithm is effective in dealing with the process crash event and greatly improves the reliability of large-scale parallel high-order moment method.
【作者单位】： 西安电子科技大学天线与微波技术重点实验室;
【基金】：国家自然科学基金(61301069) 教育部新世纪优秀人才支持计划(NCET-13-0949) 中央高校基本科研业务费(JB160218) 国家863计划项目(2012AA01A308)~~
【分类号】：TP338
【正文快照】： 作为电磁特性仿真中最精确的数值方法,矩量Research Funds for the Central Universities(JB160218),TheNational 863 Program of China(2012AA01A308)法(Method of Moments,Mo M)可以有效处理各种复杂电磁问题[1]。矩量法需要建立并求解矩阵方程:AX=B,其中A代表阻抗矩阵,它是，

本文编号：1465059