Summary
International Symposium on Antennas and Propagation
2010
Session Number:2WB2
Session:
Number:2WB2-2
Robustness and Stability of a Massively Parallel Out-of-Core Solver for Solving Pure MoM Problems with Million Level Unknowns
Yu Zhang, Xun-Wang Zhao, Sio-Weng Ting, Hang Su,
pp.-
Publication Date:2010/11/23
Online ISSN:2188-5079
Summary:
With the rapid development of computer technology, especially with the advent of multicore technology in recent years, parallel computation is playing a more and more important role in computational electromagnetics (CEM). Parallel technology has been used in various CEM methods, such as the method of moments (MoM) [1-3], fast multipole method (FMM) [4] and finitedifference time-domain method (FDTD) [5]. Because of its numerical accuracy, MoM has been a very popular method in radiation and scattering analysis. As is well known, MoM needs to deal with very large, full-density matrices for solving complex problems, and hence it incurs huge memory requirement and computational complexity. Our previous works [1-3] provided a solution to overcome these drawbacks rooted in MoM by taking advantage of the higher order basis functions (HOBs) and the technological advancements in high-performance computing (HPC) hardware, in particular, multi-core CPUs. In detail, we solved a pure MoM problem with million level unknowns by using a parallel out-of-core solver combined with the higher order basis functions [6]. The project with million level unknowns was simulated by using 512 cores and only 1 TB RAM for approximately 16 TB RAM (double precision) problems. In this paper, the robustness and stability of the parallel out-of-core solver is verified by simulating an aircraft and a formation of aircraft. The solver can deal with general structures composed of metallic and dielectric materials. The ability to solve large, complex problems in reasonable time is the greatest strength of the solver. Unlike fast algorithms (e.g., fast multipole method and adaptive integral method), the solver does not lose any accuracy of MoM for reducing memory requirement.