Solving Electromagnetic Scattering Problems With Tens of Billions of Unknowns Using GPU Accelerated Massively Parallel MLFMA

• Published in: IEEE transactions on antennas and propagation Vol. 70; no. 7; pp. 5672 - 5682
• Main Authors: He, Wei-Jia, Yang, Zeng, Huang, Xiao-Wei, Wang, Wu, Yang, Ming-Lin, Sheng, Xin-Qing
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.07.2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Central processing units; Computational efficiency; Computational modeling; Compute unified device architecture (CUDA); Computer architecture; Computing time; CPUs; Electromagnetic scattering; extremely large-scale problems; Graphics processing units; Instruction sets; Memory management; Message passing; message-passing interface (MPI) parallelization; multilevel fast multipole algorithm (MLFMA); Multipoles; Octrees; OpenMP; Optimization; Random access memory; scattering problems
• ISSN: 0018-926X; 1558-2221
• DOI: 10.1109/TAP.2022.3161520

In this article, a massively parallel approach of the multilevel fast multipole algorithm (PMLFMA) on graphics processing unit (GPU) heterogeneous platform, noted as GPU-PMLFMA, is presented for solving extremely large electromagnetic scattering problems involving tens of billions of unknowns, In this approach, the flexible and efficient ternary partitioning scheme is employed at first to partition the MLFMA octree among message-passing interface (MPI) processes. Then, the computationally intensive parts of the PMLFMA on each MPI process, matrix filling, aggregation and disaggregation, and so on are accelerated by using the GPU. Different parallelization strategies in coincidence with the ternary parallel MLFMA approach are designed for GPU to ensure high computational throughput. Special memory usage strategy is designed to improve computational efficiency and benefit data reusing. The CPU/GPU asynchronous computing pattern is designed with the OpenMP and compute unified device architecture (CUDA), respectively, for accelerating the CPU and GPU execution parts and computation time overlapped. GPU architecture-based optimization strategies are implemented to further improve the computational efficiency. Numerical results demonstrate that the proposed GPU-PMLFMA can achieve over three times speedup, compared with the eight-threaded conventional PMLFMA. Solutions of scattering by electrically large and complicated objects with about 24 000 wavelengths and over 41.8 billion unknowns are presented.