Events & News
Computer Science Colloquium
| Category | Lecture |
| Date | Thursday, March 13, 2008 |
| Time | 11:00 am |
| Location | GS 906 |
| Details | Light refreshments served in the 9th floor atrium at 10:45 AM. |
| Speaker | David Lowenthal |
| Affiliation | The University of Georgia |
Power Management and Scalability Prediction in High-Performance Computing
The number of processors in high-end parallel architectures continues
to increase rapidly. At the same time, these architectures are
becoming more difficult to program effectively, which often leads to
poor parallel program scalability. Combined, these two factors are
causing a power management problem at the nation's supercomputer
centers.
In this talk we investigate two different approaches to the power
management problem---dynamic voltage scaling (DVS) and scalability
prediction. We first determine near-optimal energy savings available
using DVS through a linear-programming approach. We use the results
to argue that while DVS-based energy savings is important, a larger
problem---one that DVS alone cannot address---is that many parallel
programs do not scale well. To this end, we introduce a novel
technique, based on statistical regression, to accurately predict
program scalability. Our approach relies only on training runs
performed on a subset of the processors that are originally requested.
This work allows programmers or system administrators to make better
choices about how many processors should be allocated to a particular
application.
Biography
David Lowenthal is a professor of Computer Science at the University
of Georgia. He received his Ph.D. in the Computer Science department
at the University of Arizona in 1996. His research centers on
parallel and distributed computing, operating systems, and networks.
Most of his current and past research projects involve addressing
fundamental parallel computing problems, such as data distribution,
scalability prediction, and energy reduction, through a system
software perspective.