|
Improving the Performance of Parallel Hermitian Eigensolvers: The Path
to a Teraflop Code
Kendall S. Stanley
University of California at Berkeley
Hosted by Paul Hovland
10:30AM, Thursday, March 12, 1999
Building 221, A-216
|
| Abstract |
Despite the success of dense linear algebra codes in reducing execution
time on large eigenproblems, there are still opportunities for further performance
improvement. A particularly disturbing fact is that the fastest way to solve small
eigenproblems (1000 by 1000) or less is still a serial code.
I will explain how we achieved 684 Gigaflops in a generalized Hermitian eigensolver within
MP-Quest, a production electronic structure code, while performing no more flops than the
best serial code.
I will also discuss techniques which will make a parallel eigensolver the fastest way to
perform a 1000 by 1000 eigensolution, and techniques that will allow us to further improve
performance on thousands of processors - allowing us to cross the Teraflop barrier for the
first time on an eigencode.
This work, performed jointly with Mark Sears (Sandia National Labs) and Greg Henry (Intel)
earned us second place in the peak performance division of the 1997 Gordon Bell
competition. |
