Improving matrix-vector product performance and multi-level preconditioning for the parallel PCG package
In this study we consider parallel solution of sparse linear systems arising from discretized PDE`s. As part of our continuing work on our parallel PCG Solver package, we have made improvements in two areas. The first is improving the performance of the matrix-vector product. Here on regular finite-difference grids, we are able to use the cache memory more efficiently for smaller domains or where there are multiple degrees of freedom. The second problem of interest in the present work is the construction of preconditioners in the context of the parallel PCG solver we are developing. Here the problem is partitioned over a set of processors subdomains and the matrix-vector product for PCG is carried out in parallel for overlapping grid subblocks. For problems of scaled speedup, the actual rate of convergence of the unpreconditioned system deteriorates as the mesh is refined. Multigrid and subdomain strategies provide a logical approach to resolving the problem. We consider the parallel trade-offs between communication and computation and provide a complexity analysis of a representative algorithm. Some preliminary calculations using the parallel package and comparisons with other preconditioners are provided together with parallel performance results.
- Research Organization:
- Front Range Scientific Computations, Inc., Lakewood, CO (United States)
- OSTI ID:
- 433345
- Report Number(s):
- CONF-9604167-Vol.1; ON: DE96015306; CNN: Grant DABT63-92-C-0024; TRN: 97:000720-0018
- Resource Relation:
- Conference: Copper Mountain conference on iterative methods, Copper Mountain, CO (United States), 9-13 Apr 1996; Other Information: PBD: [1996]; Related Information: Is Part Of Copper Mountain conference on iterative methods: Proceedings: Volume 1; PB: 422 p.
- Country of Publication:
- United States
- Language:
- English
Similar Records
ML 3.0 smoothed aggregation user's guide.
ML 3.1 smoothed aggregation user's guide.