Thursday, July 13

Multivectors in Computational Science

4:00 PM-6:00 PM
Caribbean 1

Computational science applications involving large, sparse matrices often employ standard numerical algorithms using single n-dimensional vectors. However, a reformulation using k of these vectors, grouped as columns of an n x k matrix where n >> k, enjoys substantially better memory efficiency on modern computer architectures with deep memory hierarchies and limited memory bandwidth. The speakers in this minisymposium will discuss issues in the implementation of these so-called multivectors and several areas of scientific computing that can benefit from their use. They will analyze the role of multivectors in large-scale applications including iterative methods for sparse linear systems and eigenproblems, and discuss automatic differentiation.

Organizers: H. Martin Bücker
Aachen University of Technology, Germany
Paul D. Hovland
Argonne National Laboratory, USA
4:00-4:25 Improving the Performance of Sparse Matrix Vector Multiplication
W. D. Gropp and Dinesh Kaushik, Argonne National Laboratory, USA; D. E. Keyes, Old Dominion University, USA; and B. F. Smith, Argonne National Laboratory, USA
4:30-4:55 Solution of Linear Systems Using Multivectors
Elizabeth Jessup, University of Colorado, Boulder, USA
5:00-5:25 Multivectors and Automatic Differentiation
Paul Hovland, Organizer
5:30-5:55 Multivectors and PETSc
Kristopher Buschelman

©2000, Society for Industrial and Applied Mathematics
Designed by Donaghy's Web Consulting
Created 4/12/00; Updated 7/9/00