DOE ACTS COLLECTION Workshop Robust and High Performance Tools for Scientific Computing

September 4-7, 2002

Computational Problem:

Abstract

I am participating in a SciDAC project, the Terascale Supernova Initiative (TSI), headed by Tony Mezzacappa (ORNL), http://www.phy.ornl.gov/tsi/.

The goals of the project are to model supernova physics on terascale computers. My own role in the project is to help develop scalable linear solvers for the large linear systems that arise in these simulations. We are interested in the parallel linear solvers available in PETSc, Hypre and Aztec. We may need to combine solver components from different packages to construct preconditioners that are effective for our problems.

Performance analysis tools like TAU could be valuable in trying to scale up the application codes to terascale performance.

I have considerable experience with the parallel numerical algorithms used in the libraries in the ACTS toolkit, and in my work at NCSA, have used a number of performance analysis tools, for applications running on NCSA platforms, like the SGZI Origins and Itanium based Linux clusters. Tony Mezzacappa, our lead PI, feels it would be helpful for our project if were better connected to the DoE ACTS Collection of Tools, and in this context, I am interested in attending the September Workshop.

Publications:
1. R. Eigenmann, G. Gaertner, F. Saied, and M. Straka. "SPEC HPG Benchmarks: Performance Evaluation with Large-Scale Science and Engineering Applications", submitted.

2. Saied, F. , and G. Mahinthakumar, 1998. Efficient Parallel Multigrid Based Solvers for Large Scale Groundwater Flow Simulations,, Computers Math. Applic., Vol 35, No. 7, p.45-54, 1998.

3. M. J. Holst and F. Saied. "Multigrid and Domain Decomposition Methods for Electrostatics Problems". in Domain Decomposition Methods in Scientific and Engineering Computing. Contemporary Mathematics 180, pp 231--238. Editors: D. E. Keyes and J. Xu, American Mathematical Society, Providence, 1995.

putational results are hindered by the non-robustness of current non-linear solvers. An efficient methodology, with efficient and robust non-linear solvers, would allow to conduct these computational testing on non-linear convex problems.