Efficient Data-Parallel Computations of Element-by-Element Finite Element Method by Ordering

Abstract

This study aims to develop a partitioning algorithm for data parallel computation. The greedy partitioning algorithm for the message passing developed by Farhat was modified for the data parallel computation of an element-by-element finite element method. The influence of the numbering of nodes and elements onto the parallel efficiency is examined through the parallel computations using CRAY-T 3 D. It was shown that the proposed partitioning algorithm considering the ordering of nodes and elements for data parallel computation can save time of solving problems and will improve the efficiency of parallelism. In the case of CRAY-T 3 D with 128 PEs, the maximum number of communications of the cyclic ordering version in the flow inside a subway station model with about 1 410 000 elements is 89% smaller than the original one. The rate of CPU usage of the cyclic ordering version is 92%.