基于超级计算机的矩量法性能分析与优化  

Performance Analysis and Optimization of MoM Based on Super Computer System

作　　者：陈岩[1] 林中朝[1] 张玉[1] CHEN Yan LIN Zhong-chao ZHANG Yu(Key Laboratory of Antennas and Microwave Technology, Xidian University, Xi＇ an 710071, China)

机构地区：[1]西安电子科技大学天线与微波技术重点实验室,西安710071

出　　处：《微波学报》2017年第3期1-5,共5页Journal of Microwaves

基　　金：国家高技术研究发展计划(863计划)(2014AA01A302);教育部新世纪优秀人才支持计划(NCET-13-0949);山东省重点研发计划项目(2015GGX101028);中央高校基本科研业务费(JB160218)

摘　　要：复杂目标的精确电磁特性分析往往需要巨大的存储和极长的计算时间。针对这一问题,结合国内发展迅速的超级计算机系统,研究了具有精确高效仿真能力的高性能电磁算法——高阶矩量法。提出了单元预选法来消除矩阵并行填充过程中的无效计算,加速矩阵填充过程。提出了一种具有更少的通信次数和通信量的新型并行LU分解算法,加速矩阵方程求解过程。数值测试表明提出的矩阵并行填充算法和矩阵方程并行求解算法在超级计算机平台上都能获得较高的并行性能,大幅提高了矩量法的仿真能力。According to the insatiable requirement for storage and CPU time in analysis of electromagnetic characteris- tics of complex objects, a high order MoM with high performance and accuracy based on developing domestic super computer system is studied. The redundant computation in parallel filling step of higher order MoM is analyzed, and an algorithm named ＂element pre-selection＂ is achieved to eliminate the invalid computation thus to accelerate the filling step. The paral- lel implementation of LU decomposition algorithm for matrix equations is discussed in detail, and a novel parallel LU decom- position algorithm that requires less communication overhead compared with the original one is proposed. Numerical tests show that the proposed algorithm achieves a high parallel performance on the super computer platform,which greatly improves the simulation ability of the method of moments.

关 键 词：高阶矩量法 超级计算机 高性能 优化 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构] TP338[自动化与计算机技术—计算机科学与技术]