Projects like the PTP, CDT, Photran, and TPTP are drawing the High Performance Computing community to Eclipse. As part of the HPC software cycle, application performance must be monitored and analyzed to find bottlenecks and other performance issues. Unfortunately, profiling application performance on large-scale parallel systems generates massive amounts of performance data. For such large-scale data, current analysis and visualization tools typically either show only summary information insufficient for practical performance analysis, or bombard the user with all of the performance details with no way to pinpoint useful patterns. HPCVision attempts to solve this dilemma using statistical clustering algorithms to efficiently locate patterns and identify anomalies in parallel performance profiles. This tool can help automate the application tuning cycle and increase the productivity of the human analyst.
This demo will show how to use the HPCVision tool to find performance anomalies and pinpoint bottlenecks in a parallel application. We will also discuss Eclipse implementation issues and how to extend the framework for custom functionality (data types, analysis, and display). This session targets Eclipse tool developers and people interested in performance tools and the HPC environment.
Adam Bordelon is a Computer Science graduate student at Rice University. His research focuses on parallel performance analysis using scalable data mining techniques to extract useful patterns from large-scale performance data. He is also working on integrating Rice's HPCToolkit performance profiling suite with Eclipse and the Parallel Tools Platform.