Directed Research Project Details
PerfExplorer: Parallel Performance Analysis Using Data Mining Techniques
| Author: | Kevin Huck |
|---|---|
| Date: | December 15, 2004 |
| Time: | 10:30 |
| Location: | 220 Deschutes |
| Committee: | Allen Malony (Chair) Sarah Douglas Steve Fickas |
Abstract
A growing number of todays parallel computers have well over one thousand processors. Performance analysis on very large systems is a challenge for scientists working with currently available analysis applications. One solution to this problem is to perform cluster analysis on the data and analyze the representative behavior of the clusters. To aid in this analysis, we have built a performance analysis management console called PerfExplorer. PerfExplorer provides an interface for selecting parallel performance trials, setting analysis parameters, and performing cluster analysis on the data. PerfExplorer also provides an interface for browsing the results of the analysis. Using PerfExplorer, we have performed analysis tests on three benchmark application and one real world application, and reproduced previous ad-hoc analysis results. We will discuss the current functionality of PerfExplorer, and discuss the future directions for research in this area.