Beyond Software Performance Visualization

Abstract

Performance visualization tools of the last decade have yielded new insights into the new behavior of sequential, parallel, and distributed programs. However, they have three inherent limitations: (1) They only display what happened in one execution of a program. (This is dangerous when analyzing concurrent applications, which are prone to non-deterministic behavior.) (2) A human uses one or more bandwidth-limited senses with a visualization tool. (This limits the scalability of a visualization tool.) (3) The relationship of "interesting" program events are often separated in time by other events; thus discerning time dependent behavior often hinges on finding the "right" visualization - a possibly time-comsuming activity. CHITRA93 complements visualization systems, while alleviating these limitations. CHITRA93 analyzes a set (or ensemble) of traces by combining the visualization of a few traces with a statistical analysis of the entire ensemble (overcoming (1)); and reduces the ensemble to empirical models that capture the time dependent relationships of "interesting" program events through application, programming language, and computer architecture independent analysis tools (addressing (2) and (3)). CHITRA93 incorporates: transforms, such as aggregation, that simplify the ensemble and reduce the state space size of the models generated; a user interface that allows some transforms to be selected by editing the visualization with a mouse; homogeneity tests that allow partitioning of an ensemble; an efficient semi-Markov model generation algorithm whose computation time is linear in the sum of the lengths of the traces comprising the ensemble; and a CHAID-based model that can fathom non-Markovian relationships among transitions in the traces. The use of CHITRA93 is demonstrated by partitioning ten parallel database traces with nearly 8,000 states into two homogeneous subsets, each modeled by a 20 state irreducible, periodic (non-Markovian) stochastic process.