Analysis of the Residual Optimization Potential and Runtime Analysis of the GW code on SNI Supercomputers

Summary

The GW code has been designed in 1987 for the computation of multi-aquifer groundwater flow problems on vectorcomputers as the Cyber 205 and the Cray X—MP. For a description of the code see i.g. Pelka, Peters (1986) and Peters (1988). Vectorcomputers have undergone further development since the early stages of the code implementation. In cooperation between the Aachen University of Technology (RWTII) and Siemens Nixdorf (SNI) the GW code was analyzed for residual optimization potential with respect to SNI supercomputers. Based on the existing code, vector compiler directives and recoding such as inlining were used to improve the performance. Most emphasis was put into the optimization of the solvers for the linear systems of equations resulting from the finite element formulation. The global matrix is random sparse, symmetric and positive definite. A direct skyline solver as well as iterative preconditioned conjugated gradient (PCG) solvers were tested. Especially the performance of the diagonally scaled PCG could be optimized substantially by recoding. The paper shows runtime results from two application data sets on SNI supercomputers and an IBM 3090 VF as al mainframe supercomputer with vector facility.