Scientific Computing Seminar

Date:
Thursday, April 28, 2005
Time:
11:00am-12:00pm
Location:
70-191
Seminar Speaker:
Christof Voemel
Computer Science Division
UC Berkeley
http://www.cs.berkeley.edu/~voemel/
Title:
New Eigenvalue and SVD Algorithms for LAPACK and ScaLAPACK
Abstract:
Since 2005, two groups at the University of California at Berkeley and atthe University of Tennessee at Knoxville are funded by NSF to improve the widely used LAPACK and ScaLAPACK libraries. Key improvements address novel algorithms, ease of use, performance, accuracy, reliability, and user support.

After a brief introduction of the general project, the main part of this talk will be about our own work in this framework. We speak about the ongoing development of an eigensolver and an SVD code based on Dhillon's and Parlett's algorithm of Multiple Relatively Robust Representations (MRRR).

Our talk will address the following aspects:
1. Algorithmic issues and how to overcome them
2. Performance analysis of the eigenvalue and the SVD code, a comparison with other LAPACK codes
3. Bottlenecks and performance tuning
4. Possibilities for parallelization

Sponsor of Seminar:
Osni Marques
Scientific Computing

Contact Esmond G. Ng EGNg@lbl.gov