Chitra93: a Tool to Analyze System Behavior by Visualizing and Modeling Ensembles of Traces
Files
TR Number
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
A key cause of poor system performance is inherently due to the lack of understanding of system behavior. Performance problems are especially apparent in parallel and distributed programs, for which expected speedup is difficult to achieve. Theoretical models and trace visualization tools are suitable for extracting insights into the behavior of a system. Theoretical models available today work for certain types of systems and require possibly unrealistic assumptions, and hence are not considered here. Existing trace visualization tools have yielded new insights into the behavior of the sequential, parallel, and distributed programs. However, they have two inherent limitations: (1) Each tool visualizes only one execution of a program. (This is dangerous when analyzing concurrent programs, which are prone to non-deterministic behavior.) (2) The applicable domain of a visualization tool will be limited unless the tool incorporates a large variety of methods to visually display data. This is because a single display method may yield new insights into only certain systems. In addition, finding the "right" display that can provide the needed insights is a potentially time-consuming process. This project carries through the previous work-CHITRA92. This project addresses these limitations by providing the following four capabilities to analyze traces: (1) CHITRA93 analyzes a set (or ensemble) of traces to obtain the typical behavior of a system. (2) CHITRA93 incorporates three transforms to simplify ensembles by reducing either the state space size or the interval over which time is defined in the ensemble. (3) CHITRA93 builds compact summary of the dynamic behavior (or model) of a system from an ensemble. (4) CHITRA93, to avoid building models that poorly fit an ensemble, provides a suite of methods to partition ensembles into mutually exclusive, exhaustive, and homogeneous subsets so that each subset displays "similar" behavior. These methods include several visual techniques and statistical methods. Finally, a portion of the project seeks to stabilize and to produce a correct version of CHITRA.