摘要
【目的】充分利用国产高性能平台计算资源,增强电磁场数值仿真算法的工程应用能力。【方法】结合平台硬件异构特性,提出针对电磁场积分方程方法中矩量法(MoM)、多层自适应交叉近似算法(MLACA)及多层快速多极子算法(MLFMA)的MPI-OpenMP-HIP三级混合并行架构及高效率实现方法。【结果】通过架构设计、任务分配、通信控制和负载均衡等方式,实现电磁场仿真算法在分布式节点、多核、类GPU国产加速卡配置的计算集群上高效率并行计算,应用于器件参数提取、目标散射截面评估等工业电磁仿真。【结论】实验结果表明三级混合并行极大地提升了电磁仿真算法的计算效率和应用能力,为基于国产高性能计算平台的新一代电磁仿真软件研制提供了实践范例。
[Objective]To fully utilize the computing resources of domestic high-performance platforms and enhance the engineering application capabilities of electromagnetic field numerical simulation algorithms,[Methods]this paper proposes a three-level hybrid parallel architecture of MPI-OpenMP-HIP and a high-efficiency implementation scheme for the method of moments(MoM),a multi-level adaptive cross approximation algorithm(MLACA),and a multi-level fast multipole algorithm(MLFMA)in the electromagnetic field integral equation method by considering the heterogeneous hardware characteristics of the platform.[Results]Through architecture design,task allocation,communication control,and load balancing,the high-efficiency parallel computing of electromagnetic field simulation algorithms have been achieved on a computing cluster configured with distributed nodes,multi-cores,and domestic GPU-like accelerator cards,which can be applied to industrial electromagnetic simulations such as device parameter extraction and target scattering cross-section evaluation.[Conclusions]Experimental results show that the three-level hybrid parallel architecture significantly improves the computing efficiency and application capabilities of electromagnetic simulation algorithms,providing a practical example for the development of new generation electromagnetic simulation software for domestic high-performance computing platforms.
作者
何十全
王旭
张钰
HE Shiquan;WANG Xu;ZHANG Yu(University of Electronic Science and Technology of China,Chengdu,Sichuan 611731,China;Zhejiang University,Hangzhou,Zhejiang 310027,China)
基金
光合基金A类(202302013877)。
关键词
电磁积分方程
三级混合并行
国产异构平台
electromagnetic integral equations
three-level hybrid parallel
domestic heterogeneous platforms
