Research On Parallel Algorithm Of Electromagnetic Simulation Based On Heterogeneous Computing

Finite difference time domain (FDTD) method is one of the important methods of solving maxwell's equations in electromagnetics, which has been widely used since.Finite difference time domain method is used to directly convert the Maxwell differential equation with time variable into difference equation by using central difference method in Yee's grid space. In order to ensure the stability of the solution, to suppress the numerical dispersion, the space step size and the time step of the FDTD should not be too large. Especially in solving electrically large problems, the FDTD method is very time consuming and memory consuming.FDTD algorithm has the natural parallelism, the use of parallel computing method can reduce the computing time and improve computing speed. In recent years, with the rapid development of hardware technology, the actual floating point operation performance of the graphics processor (GPU) is far more than the same period of the CPU, and low cost. With the advent of Compute Unified Device Architecture (CUDA) model, using graphics processors for parallel program development becomes very simple and efficient.This article using the Quadro 4000 and Tesla M2050 GPU, CUDA as a development tool, has completed three dimensional FDTD calculation of high performance.The main work is as follows:(1)On the platform, the first time the application TH-1A supercomputer system extended to the electromagnetic field calculations. The study of Hunan University Supercomputer Center supercomputing system as the platform, with its high peaks and LINPACK performance characteristics of heterogeneous together in parallel,achieved a three-dimensional FDTD parallel algorithm.(2) On the algorithm optimization, the use of page locked memory to store the field to be used on the host side, reduce the transmission of data consumption; The field quantity using a low cost cache L1 cache, improve the hit rate of data access. After optimization of the three-dimensional FDTD parallel algorithm, compared to not to optimize the algorithm to accelerate the ratio increased by about 4 times.(3)On the basis of given parallel algorithm with the corresponding serial algorithm GPU-based FDTD comparative analysis, from the three aspects described that application of GPU technology, three-dimensional FDTD parallel algorithm accelerated significantly.(4)In electromagnetic applications, the three-dimensional FDTD parallel algorithm is extended to the practical application of electromagnetic calculation. This paper makes a simulation of a band stop filter based on the three-dimensional CPML-FDTD parallel algorithm. The experimental results show that the speedup is 55 times faster than the serial algorithm, and the computation error is within the range of 10-4.