Detecting Performance Bottlenecks Guided by Resource Usage

Detecting performance bottlenecks is critical to fix software performance issues. A great part of performance bottlenecks are related to resource usages, which can be affected by configurations. To detect configuration-related performance bottlenecks, the existing works either use learning methods to model the relationships between performance and configurations, or use profiling methods to monitor the execution time. The learning methods are time-consuming when analyzing software with large amounts of configurations, while the profiling methods can incur excessive overheads. In this paper, we conduct empirical studies on configurations, performance and resources. We find that 1) 49% performance issues can be improved or fixed by configurations; 2) 71% configurations affect the performance by tuning resource usage in a simple way; and 3) four types of resources contribute the main causes of performance issues. Inspired by these findings, we design PBHunter, a resource-guided instrumentation tool to detect configuration-related performance bottlenecks. PBHunter ranks configurations by resource usage and selects the ones that heavily affect resource usages. Guided by selected configurations, PBHunter applies the code instrumentation technique in resource-related code snippets. The evaluation shows PBHunter can effectively (36/50) expose the culprits of performance issues with minor overheads (5.1% on average).