Performance Analysis of Parallel FDTD Algorithm on IBM BlueGene Supercomputer Series
The advances in high performance computing nowadays have increased the need of the computational resources to solve the large-scale problems. The large-scale electromagnetic and optical problems with the size of the order of 400 wavelengths in every dimension for a problem with arbitrary complex geometry structure can be solved. One of the most widespread numerical methods for Maxwell equations solving is Finite difference Time Domain (FDTD) method. FDTD is based upon Yee lattice and Maxwell equations in integral form and differ-entional relations from which finite difference approximation is obtained. In this paper, IBM BlueGene/P and BlueGene/Q performance is compared to calculate the large electromagnetic problem using parallel Finite difference Time Domain method. Introduced early EMWSolver3D was implemented on Lomonosov Moscow State University IBM BlueGene/P supercomputer. Using 1D data decomposition for the highly elongated spatial domains together with the asynchronous transfer operations between nodes of supercomputer it is allows to solve problems up to 1010 computational cells on more than a thousand nodes. Parallel implementation of FDTD algorithm based on hybrid MPI/OpenMP approach is used in this paper. The scalability of the algorithm on both BlueGene/P and BlueGene/Q systems that runs in different parallel modes is considered. Parallel processing performance comparison of both systems shows that the IBM BlueGene/Q supercomputer gives near linear speedup and the great efficiency on FDTD algorithm.
A.P.Smirnov A.N.Semenov A.V.Pozdneev
Lomonosov Moscow State University, Moscow, Russia IBM East Europe/Asia Ltd., Moscow, Russia
国际会议
Progress in Electromagnetics Research Symposium 2014(2014年电磁学研究新进展学术研讨会)
广州
英文
253-256
2014-08-01（万方平台首次上网日期，不代表论文的发表时间）