Diagnosing Performance Variations by Comparing Multi-Level Execution Traces

摘要

Tracing allows the analysis of task interactions with each other and with the operating system. Locating performance problems in a trace is not trivial because of their large size. Furthermore, deep knowledge of all components of the observed system is required to decide whether observed behavior is normal. We introduce TraceCompare, a framework that automatically identifies differences between groups of executions of the same task at the user space and kernel levels. Many performance problems manifest themselves as variations that are easily identified by our framework. Our comparison algorithm takes into account all threads that affect the completion time of analyzed executions. Differences are correlated with application code to facilitate the correction of identified problems. Performance characteristics of task executions are represented by a new data structure called enhanced calling context tree (ECCT). We demonstrate the efficiency of our approach by presenting four case studies in which TraceCompare was used to uncover serious performance problems in enterprise and open source applications, without any prior knowledge of their codebase. We also show that the overhead of our tracing solution is between 0.2 and 9 percent depending on the type of application.