Abstract:

Co-processers with powerful floating-point performance have been developing rapidly in recent years and also draw huge attention in the High Performance Computing (HPC) community. The acceleration of three-dimensional UPML Finite Difference Time Domain (FDTD) method  by using GPU co-processer becomes a hot topic in the numerical simulation of electromagnetic (EM). The paper focuses on the implementation and optimization of 3D UPML-FDTD algorithm on GPU clusters. Using the electric dipole excitation source, the proposed algorithm validates the numerical results of EM simulation and the analytical solution to the EM field, showing that the algorithm has high numerical accuracy. The performance of the parallel FDTD algorithm is tested on both Tesla M2070 and K20m GPU clusters, the results with/without optimization are compared, and the computing performance of GPU is compared with that of CPU. The scalability of the algorithm is shown for up to 80 Tesla K20m GPUs. It is concluded that the optimized FDTD algorithm improves its performance and obtain good parallel efficiency on GPU clusters.

Key words: FDTD;UPML;GPU cluster;MPI