Soft Computing Approach to Performance Analysis of Parallel and Distributed Programs

This paper describes a novel approach to performance analysis for parallel and distributed systems that is based on soft computing. We introduce the concept of performance score representing the performance of code regions that is based on fuzzy logic. We propose techniques for fuzzy-based performance classification. A novel high-level query language is designed to support the search for performance problems by using linguistic expressions. We describe a fuzzybased bottleneck search, a performance similarity measure for code regions and experiment factors, and performance similarity analysis. Our approach focuses on the support of making soft decisions on evaluation, classification, search and analysis of the performance of parallel and distributed programs.