Optimal Processor Mapping for Massively Parallel Finite Element Computations

by Sang Y. Synn, Georgia Inst of Technology, Atlanta, United States,
Robert E. Fulton, Georgia Inst of Technology, Atlanta, United States,



Document Type: Proceeding Paper

Part of: Computing in Civil and Building Engineering

Abstract:

The exploitation of massively parallel computer for finite element analysis requires efficient usage of processors as well fast speed. For example, the efficiency of a parallel matrix decomposition program depends on processor utilization and matrix bandwidth size. Most parallel algorithms proposed in the literature for finite element analysis have been targeted only toward high speedup. In this paper, a theoretical formulation for two massively parallel machines(BBN/ Butterfly, KSR1) is developed as a basis for estimating the optimal number of processors. Using that formulation, the algorithm for the efficient usage of processor in massively parallel computations is discussed for two classes of finite element computation: global solution and substructuring approach.

Subject Headings: Finite element method | Computer analysis | Computer models | Parallel computing | Algorithms | Substructures | Matrix (mathematics)

Services: Buy this book/Buy this article

 

Return to search