Parallelization of Finite Element Analysis Using Broadcast on Workstation Cluster

Abstract

In large-scale numerical simulations based on a rigid-plastic finite element method, most of the processing time is used in solving linear equations called stiffness equations. Therefore, the method of parallel processing is introduced to speed up the process of solving stiffness equations. The conjugate gradient method is adopted for solving the equations and is parallelized on a cluster of workstations. In the present paper, we present a new cluster of workstations which uses broadcasting to transmit data from a master processor to all slave processors, instead of unicasting. Broadcasting can transmit data to all slaves by one transmission regardless of the number of slaves. Therefore, the cluster of workstations with broadcasting can reduce communication time sharply compared with sequential data transmission by conventional unicasting. We achieved more than twelvefold increase in speed for solving the stiffness equations with twenty slave processors.