Advanced methods of performance data processing and analysis

The analysis of program performance on massively parallel (MPP) systems is a non-trivial task which is increasingly performed using visualization tools. Conventional processing, analysis, and display methods typically do not adequately support the evaluation of program performance on large-scale parallel systems. These methods can be limited by the sheer volume of performance data and often lack any rigorous approach to analyze it. The objective of the approach outlined is to develop a more powerful method to analyze performance data visually as well as statistically which is scalable and extensible and which represents some physical and/or logical structure of the parallel system state. The approach focuses on applying image processing and analysis techniques to MPP performance analysis. This approach has been prototyped and tested for proof-of-concept. The AVS scientific visualization tool is used to analyze and display images of performance data from program executions on actual machines. Results of these experiments are presented