Tracking performance across software revisions

Abstract

Repository-based revision control systems such as CVS, RCS, Subversion, and GIT, are extremely useful tools that enable software developers to concurrently modify source code, manage conflicting changes, and commit updates as new revisions. Such systems facilitate collaboration with and concurrent contribution to shared source code by large developer bases. In this work, we investigate a framework for "performance-aware" repository and revision control for Java programs. Our system automatically tracks behavioral differences across revisions to provide developers with feedback as to how their change impacts performance of the application. It does so as part of the repository commit process by profiling the performance of the program or component, and performing automatic analyses that identify differences in the dynamic behavior or performance between two code revisions. In this paper, we present our system that is based upon and extends prior work on calling context tree (CCT) profiling and performance differencing. Our framework couples the use of precise CCT information annotated with performance metrics and call-site information, with a simple tree comparison technique and novel heuristics that together target the cause of performance differences between code revisions without knowledge of program semantics. We evaluate the efficacy of the framework using a number of open source Java applications and present a case study in which we use the framework to distinguish two revisions of the popular FindBugs application.