Study On Parallelization Of Time-and Freguency-Domain MoM Based On Graphics Processing Units

The conventional frequency-domain Method of Moments(MoM) has been widely used to analyze electromagnetic scattering and radiation problems, due to its good accuracy and its Green’s function satisfying radiation boundary conditions directly, while the time-domain MoM has great advantages in analyzing transient electromagnetic scattering and radiation problems. However, both of them suffer from large computational burdens, so they have encountered computational bottleneck when used to solve electromagnetic scattering and radiation problems including electrically large objects. The predecessors in computational electromagnetics have leveraged two main ways to deal with it, namely, fast algorithms and parallel computation. While fast algorithms have been developed to reduce memory requirement and computational complexities, parallel computation has been introduced to decrease computational time and improve computational efficiency.For the above reason, a novel parallel computation technique proposed in recent years, called as Graphics Processing Units (GPU) parallel technology, is leveraged to accelerate the time-and frequency-domain MoM computation. Based on studying NVIDIA GPU hardware architecture and Compute Unified Device Architecture (CUDA) programming model in detail, the way to use streaming in CUDA is specially discussed. With the streaming, the latency of communication between host and device can be hidden. Taking the analysis of frequency-domain electromagnetic scattering from metal objects as an example, GPU is applied to accelerate two time-consuming procedures in frequency-domain MoM, i.e., impendence matrix fill and GMRES iterative solution of matrix equation. The general way and asynchronous way based on streaming are compared to speed up impendence matrix fill and matrix equation solution. The former transfers the whole impendence matrix on to the video memory of GPU. By contrast, the latter carries out the impendence matrix fill and matrix-vector multiplication block-by-block in order to overcome the disadvantage of limited video memory. Several numerical examples are given to show the accuracy and efficiency of the proposed algorithm, and furthermore the effect of the number of streaming on the acceleration performance is studied. According to the frequency-domain GPU accelerated MoM, parallelization of time-domain MoM based on GPU is developed. The procedure of application of Marching-On-in-Time (MOT) to time-domain electric field integral equation (TDIE) is derived step-by-step. In TDIE-MOT algorithm, as the increasing of time iteration step, the two main time-consuming stages include multiple impendence matrix fills and matrix-vector multiplications. Therefore, an asynchronous parallel way based on streaming is utilized to speed up the impendence matrix fill and matrix-vector multiplication. Some numerical examples are given to demonstrate good accuracy and performance of the proposed algorithm.