PCA2 Abstract

Title: Structural Analysis Problems with Direct Solvers (PCA2)

Authors: F.X. Roux

Email: roux@onera.fr

Date: July 1993

Abstract:

Industrial structural analysis packages generally make use of direct solvers because of their robustness. The most popular solvers for symmetric matrices are the skyline Choleski factorization and the frontal method. The most computational intensive parts of these codes lie in the computation of the elemental matrices that can be very expensive in nonlinear cases, the assembly of the global stiffness matrix and its factorization. The solution of the system takes minor part in the global cost. This paper reports the study of the implementation of a typical code with the above structure on a massively parallel MIMD system, the iPSC-860 with 128 processors of ONERA, with message-passing programming environment. For each step, an parallelization strategy has been developed and adequated data structures have been designed. The data transfers involved in the reorganization of the data between each step has been studied too. The tests performed have shown that the stiffest parallelization problem is with the factorization step. Further developments have to be made in order to design algorithms better suited for the parallel factorization of sparse matrices on distributed memory parallel computers.

This report is available through