Detection and Visualization of Performance Variations to Guide Identification of Application Bottlenecks

摘要

The identification of performance bottlenecks in parallel applications is a challenging task. Without some form of performance measurement tool, this task lacks any guidance and purely relies on trial-and-error. At the same time, data sets from parallel performance measurements are often large and overwhelming. We provide an effective solution to automatically identify and highlight several types of performance critical sections in an application run. Our approach first identifies time dominant functions of an application that are subsequently used to analyze runtime imbalances throughout the application run. We then present the resulting runtime variations in an intuitive visualization that guides the analyst to performance hot spots. We demonstrate the effectiveness of our approach in a case study with three applications, detecting performance problems and identifying their root-causes in all cases.