基于超级计算机的并行FDTD关键技术与应用

发布时间：2018-01-13 13:08

本文关键词：基于超级计算机的并行FDTD关键技术与应用　出处：《西安电子科技大学》2016年博士论文　论文类型：学位论文

【摘要】：电磁场与电磁波在生产生活中发挥着越来越重要的作用。数值计算具有精确高效、灵活方便等显著优势,因此成为电磁特性分析与设计的现代化手段,也日益发挥着越来越重要的作用。时域有限差分(FDTD)法作为电磁场三大经典数值方法之一,具有自己独特的优势。对于电大尺寸及具有复杂(细小)结构的电磁目标仿真,巨大的内存资源和计算时间消耗成为了该算法的限制因素。随着当今计算机软硬件的飞速发展,大型计算机集群为大规模复杂电磁问题的解决提供了可能。利用计算能力更强的平台来求解大规模电磁场问题成为计算电磁学领域中的热点和挑战之一。作为大型计算机集群中新兴的纯国产计算机而言,由于其编译器等环境都是针对国产CPU的架构自主开发的,与通用计算机平台的环境存在着较大的差异,且在纯国产超级计算平台中对电磁应用领域进行相关的研究工作具有更高的安全性,因此在国产计算机中对自主开发的并行FDTD算法程序进行移植和测试显得尤为重要。作为提升求解效率的重要手段,采用新型的硬件资源国产众核处理器,可进一步加速求解速度,缩短数值计算的时间。基于上述原因,本文基于国内超级计算机硬件平台,对并行FDTD算法的性能优化及其大规模应用进行了深入研究。本文的主要工作可以概括为：(1).基于FDTD方法的运算特性,采用笛卡尔(Cartesian)拓扑结构来描述整个计算空间子区域间的逻辑关系,实现了基于MPI的并行FDTD方法,并建立影响并行FDTD并行效率性能的数学模型。通过理论分析,提出可获得较高并行性能的虚拟拓扑选取理论准则。同时给出一种采用并行策略实现基三角面片模型来提取FDTD立方体模型的方法,使得核心程序与模型无关,相比于串行建模法,极大的提高了建模速度。当问题规模较大时,效率可提升60%以上：(2).基于上述理论准则,在国家超级计算天津中心和国家超级计算深圳中心两个通用平台上对并行FDTD虚拟拓扑优化进行了大量测试验证工作。通过对天津中心测试结果的分析,得出跨节点通信少的虚拟拓扑可获得较好的并行性能。而通过对深圳中心测试结果的分析,得出在整个通信任务中,占主要通信任务的通信量小时,算法可获得较好的并行性能。综合两个通用平台上的结果以及数学模型得出的理论准则,给出通用平台上并行FDTD最优化虚拟拓扑的一般性准则,即：a、基本准则是选择拓扑使得总通信量最小：b、通信总量相当的情况下,跨节点通信少的拓扑结构性能一般较好；c、跨节点通信量相当时,选择通信负载较为均衡的拓扑。基于该准则,计算时可获得较高的并行性能；(3).考虑到纯国产CPU的计算机与通用计算机平台的区别,研究了纯国产平台上并行FDTD方法获得高效率的虚拟拓扑准则。受限于国产CPU的内存,整个测试采用分段测试的方式进行。通过对测试结果的分析,得出了与通用平台上基本一致的最优化虚拟拓扑选取准则；(4).基于国产众核平台的特点,对并行FDTD方法进行了算法上的移植。并针对国产众核平台上独有的编程方式,对并行FDTD方法的并行策略进行了优化。通过对多个算例模型的测试,得出了在国产众核平台上算法的加速性能；(5).在不同的测试平台上,对FDTD的并行性能进行了测试。在纯国产CPU计算机平台上突破了10万核,以128 CPU核为基准,算法在10240 CPU核时的并行效率仍可达67%以上。在通用平台上突破了30万核,以6万CPU核为基准,算法在30万CPU核时的并行效率可达80%以上。在国产众核平台上突破了400万核,最大计算Yee网格量约为1万亿；值得指出,根据2015年5月的查新报告显示,这是目前国际上最大并行规模的国产电磁仿真应用。(6).在不同的超级计算机平台上就几种典型的电磁应用进行了仿真分析,包括天线的辐射特性分析、载机平台的散射特性分析、机载天线的受扰方向图分析以及雷电袭击飞机时,飞机周围的近场分析等；(7).基于GiD软件进行了二次开发,形成了一套自主可控的并行FDTD软件。通过该软件可实现从建模、参数设置、媒质赋值、MESH剖分、计算和后处理显示等一体化过程。
[Abstract]:Electromagnetic field and electromagnetic wave in production and life plays an increasingly important role. The numerical calculation is accurate and efficient, flexible and convenient and other significant advantages, therefore become a modern means for the analysis and design of electromagnetic characteristics, plays a more and more important role. The finite difference time domain (FDTD) method is one of the three classical numerical electromagnetic field methods, has its own unique advantages. For large and complex (small) electromagnetic simulation structure, memory resources and huge computing time consumption has become the limitation of the algorithm factors. With the rapid development of computer hardware and software, provides the possibility of a large cluster of computers for solving large-scale complex electromagnetic problems. By calculation a stronger platform to solve the problem of electromagnetic field calculation becomes one of the hot and challenges in the field of electromagnetics. As a large-scale computer cluster In the emerging pure domestic computer, because the compiler environment is based on CPU architecture is developed, there is a big difference with the general computer platform environment, and in the pure domestic super computing platform in the field of electromagnetic application of associated work has higher security, so the parallel FDTD algorithm program for independent development in the domestic computer transplant and test is very important. As an important means to improve the solving efficiency, the hardware resources of domestic multi-core processor model, can further accelerate the solving speed, shorten the calculation time. Based on the above reasons, the domestic super computer hardware platform based on the performance of the parallel optimization FDTD algorithm and its application are studied. The main work of this paper can be summarized as follows: (1). The operation characteristics based on FDTD method, The Cartesian (Cartesian) topology to describe the logical relations of the calculation space between subdomains, realizes the parallel FDTD method based on MPI, and to establish the mathematical model of FDTD parallel parallel efficiency performance. Through theoretical analysis, proposed can obtain high performance parallel virtual topology selection criterion. Then a method is adopted to realize base triangle model method to extract FDTD cube model of parallel strategy, the core program and model independent, compared to the serial modeling method, which greatly improves the modeling speed. When the scale of the problem is large, can improve the efficiency of 60% above: (2). The theory based on the criterion in the National Supercomputing Center in Tianjin and the National Supercomputing the center of Shenzhen two general platform for parallel FDTD virtual topology optimization of a large number of testing work. Through the analysis of the test center of Tianjin, the Cross communication virtual topology few nodes can obtain good parallel performance. Through the analysis of the results obtained in the test center of Shenzhen, the task of communication, communication accounts for the main communication task hours, algorithm can obtain good parallel performance. Two general results based on the mathematical model and the theory of criterion. Parallel optimization FDTD virtual topology general criteria, namely: a general platform is given, the basic criterion is chosen so that the total amount of the minimum communication topology: B communication, total equivalent, cross topology node communication performance less good; C, cross section by letter amounts, the communication topology is load equilibrium. Based on this criterion, the calculation can obtain high parallel performance; (3). Considering the difference between pure domestic CPU computer and general computer platform, the pure domestic platform parallel FDT Virtual topology criterion D method to obtain high efficiency. Due to the domestic CPU memory, the test adopts subsection test method. Through the analysis of the test results, the selection criterion of optimal virtual topology is consistent with the general platform; (4). The characteristics of domestic nuclear platform based on parallel FDTD method the algorithm of the transplant. And according to the unique domestic many core platform programming, the parallel strategy of parallel FDTD method was optimized. Through several examples of model test, the acceleration performance of the algorithm in domestic multi-core platform; (5). In different test platform, the the performance of parallel FDTD are tested. In the pure domestic CPU computer platform exceeded 100 thousand to 128 nuclear, nuclear CPU as a benchmark, the CPU core algorithm in 10240 parallel efficiency is still above 67%. On a common platform to break the 300 thousand nuclear, to 60 thousand CPU as a benchmark, the CPU core algorithm in 300 thousand parallel efficiency can reach more than 80%. In the domestic nuclear platform exceeded 4 million nuclear, the maximum calculated Yee grid is about 1 trillion; it should be pointed out that according to the new report found in May 2015 showed that it is made of electromagnetic simulation application of the largest international parallel scale. (6). The simulation analysis was carried out on the application of several typical electromagnetic supercomputers in different platforms, including the analysis of the radiation characteristics of the antenna and scattering characteristics of aircraft platform, airborne antenna interference pattern analysis and lightning hit the aircraft, the aircraft around the near field analysis; (7). Based on the GiD software for the two time development, has formed a set of independent controllable parallel FDTD software. The software can be achieved through the modeling, parameter setting, medium assignment, MESH triangulation, calculation and postprocessing display integration process.

【学位授予单位】：西安电子科技大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP338;O441.4

【参考文献】