Hybrid CPU-GPU Implementation of Fast Multipole Method for Large-Scale Electromagnetic Scattering Problems

摘要

In this paper, the parallelization of the single level Fast Multipole Method (FMM) is investigated on a Central Processing Unit (CPU) - Graphical Processing Unit (GPU) high-performance computing platform for solving large-scale electromagnetic scattering problems. By using both CPU and GPU hardware resources, the proposed framework enables accommodating problem sizes with millions of unknowns. The performance is evaluated on a 13-node cluster that populates NVIDIA Tesla M2090 GPUs. The potential performance achievements including speedup and scalability for large-scale problems using FMM are shown and analyzed.