Evaluating Trace Aggregation Through Entropy Measures for Optimal Performance Visualization of Large Distributed Systems
Abstract
Large-scale distributed high-performance applications are involving an ever-increasing number of threads to explore the extreme concurrency of today's systems. The performance analysis through visualization techniques usually su ers severe semantic limitations due, from one side, to the size of parallel applications, from another side, to the challenges to visualize large-scale traces. Most of performance visualization tools rely therefore on data aggregation in order to be able to scale. Even if this technique is frequently used, to the best of our knowledge, there has not been any real attempt to evaluate the quality of aggregated data for visualization. This paper presents an approach which lls this gap. We propose to build optimized macroscopic visualizations using measures inherited from information theory, and in particular the Kullback-Leibler divergence. These measures are used to estimate the complexity reduced and the information lost during any given data aggregation. We rst illustrate the applicability of our approach by exploiting these two measures in the analysis of work stealing traces using squari ed treemaps. We then report the e ective scalability of our approach by visualizing known anomalies in a synthetic trace le with the behavior of one million processes, with encouraging results.
Origin : Files produced by the author(s)
Loading...