Performance of NIKE3D with PVSOLVE on vector and parallel computers

The cost of large implicit finite element calculations is dominated by the linear equation solver. In this work, several versions of the direct linear equation solver PVSOLVE were implemented in the general purpose, nonlinear finite element code NIKE3D. Timing studies were performed on Sun and IBM workstations, CRAY Y/MP and C90 supercomputers, and the distributed memory MEIKO CS-2 parallel computer. Cost breakdowns and efficiency curves are presented for a benchmark problem with several mesh densities. The distributed memory implementation is shown to be most efficient for large problems, where MEIKO performance is comparable to that of a single CRAY processor.