Solving Large Sparse Linear Systems to End-to-end Accelerator Structure Simulations.

Lee, L. Q. ; Ge, L. ; Kowalski, M. ; Li, Z. ; Ng, C. K. ; Schussman, G. ; Wolf, M. ; Ko, K.

Technical Information Center Oak Ridge Tennessee

关键词: Accelerators;    Simulation;    Linear systems;    Convergence;    Performance;   

RP-ID  :  DE2004826910

学科分类：工程和技术（综合）

美国|英语

来源: National Technical Reports Library

【 摘 要 】

This paper presents a case study of solving very large sparse linear systems in end-to-end accelerator structure simulations. Both direct solvers and iterative solvers are investigated. A parallel multilevel preconditioner based on hierarchical finite element basis functions is considered and has been implemented to accelerate the convergence of iterative solvers. A linear system with matrix size 93,147,736 and with 3,964,961,944 non-zeros from 3D electromagnetic finite element discretization has been solved in less than 8 minutes with 1024 CPUs on the NERSC IBM SP. The resource utilization as well as the application performance for these solvers is discussed.