Communication vs. computation-experiences within a finite element analysis application on transputers

Abstract

Many parallel numerical algorithms that speed up Finite Element computations have been developed, particularly f o r local memory multiprocessors. The overall performance on a multitransputer machine is governed by the eficiency of the communication amongst the processors and the calculation carried out by each processor. Hence in the distributed calculation of the element matrices and the subsequent assembling of the global matrix there are inherent problems associated with the optimisation of the relative times f o r computation and communication. This paper examines the relative merits of two alternative strategies f o r the calculation of the element matrices and the assembly of the global matrix. Eight noded and twenty noded brick elements have been used f o r the solution of steady state potential and plane strain problems in a comparative study.