基于CPU-GPU异构机群的FDTD并行算法加速研究  

Accelerating Parallel FDTD on CPU-GPU Heterogeneous Cluster System

作　　者：邵宗有[1,2] 王昭顺[1] 刘新春 

机构地区：[1]北京科技大学信息工程学院,北京100083 [2]曙光公司,北京100193 [3]无锡城市云计算中心有限公司,无锡214315

出　　处：《系统仿真学报》2013年第2期235-240,共6页Journal of System Simulation

基　　金：国家高技术研究发展计划(863)(2011AA040502);核高基重大专项(2012ZX01028001-003)

摘　　要：时域有限差分法(FDTD)求解电磁学中麦克斯韦方程组是科学与工程计算中一个非常重要的算法。通过对FDTD求解麦克斯韦旋度方程的直接时间域的分析,给出其基于多个GPU组成异构机群系统上的并行加速算法,用OpenCL、CUDA和MPI编程模型实现了并行程序。在目前的主流NVIDIA和ATI的GPU平台上,加速的并行FDTD程序相对CPU串行程序和8个CPU核的MPI并行程序,分别获得了超过8倍和1.5倍的加速,并在多个GPU卡上获得了接近线性加速的扩展性能。Finite-Difference Time-Domain （FDTD） for computational electrodynamics modeling techniques is an important algorithm in scientific and engineer computing applications. Parallel FDTD algorithms of time-dependent Maxwell＇s equations were investigated, and accelerated algorithms on a CPU-GPU heterogeneous cluster system were proposed. The parallel FDTD program was implemented in a hybrid model ofOpenCL, CUDA and MPI. In state of the art GPU processors from both NVIDIA and ATI, the accelerated FDTD achieves speedup of 8 times and 1.5 times over a serial program on one CPU core and parallel program on 8 CPU cores, respectively. The parallel hybrid program also achieves an approximate linear speedup with multiple GPUs.

关 键 词：电磁场仿真 时域有限差分法 GPU 机群 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]