Progress and Problems

This page is designed as a place where group members can communicate with each other.


Problems with Domain Decomposition

by Adam Huang, April 26

To improve the parallelism of Howard's original code, which uses a global GMRES solver with an incomplete LU preconditioner, I have been working on the Domain Decomposition approach to the solver. I applied the partitioning routines from the METIS 2.0 package. This is a DD library developed by George Karypis of the CS Department of the University of Minnesota. The algorithms implemented by METIS are based on the multilevel graph partitioning scheme. I took the lazy approach here and just applied the routines to the assembled global matrix instead of to the mesh.

The problem is how to solve this system. Different approaches have been explored:

Overall the Domain Decomposition approach I'm taking has not shown any speedup over the original global GMRES solver when running in a sequential mode, because of the slow convergence of the new system. Of course we get quite a lot of parallelism. Larger problems need to be tested to see if there is some benefit.


Tentative Schedule

by Matt Knepley, June 26

Serial Improvement

I hope to implement the improvements listed below by July 15 (this should include testing).

Code Management

I hope to implement the improvements listed below by July 15 (this should include testing).

Diagnostics and Testing

I hope to implement the improvements listed below by August 1.


Project Home

| AEM Home | Institute of Technology |

| Academics | Research | People | Information | Contact AEM |
| Major Projects | by Specialty | Other Programs |

Last updated October 16, 2000