Slope-Based Sequencing Yardstick for Analyzing Unsatisfactory Performance of Multithreaded Programs

As users are increasingly concerned about energy efficiency, they are also increasingly intolerant of performance anomalies of programs that may cause significant energy waste. Bug localization is a bottleneck in the development of multithreaded programs. Although both static and dynamic performance bug localization techniques have been proposed, they cannot handle performance anomalies with unforeseen patterns, and cannot work well if the concept of performance anomaly is fuzzy or evolves over time for the same program. We propose a novel model-based approach to performance bug localization. The approach is based on curve fitting and trend estimation over program executions with performance data. We describe our trend estimation model and illustrate it with the result of a case study on locating three real-world performance bugs in MySQL.