Tracking down software changes responsible for performance loss

Continuous software change may inadvertently introduce a drop in performance at runtime. The longer the performance loss remains undiscovered, the harder it is to address. Current profilers do not efficiently support performance comparison across multiple software versions. As a consequence, identifying the cause of a slow execution caused by a software change is often carried out in an ad-hoc fashion. We propose multidimensional profiling as a way to repeatedly profile a software execution by varying some variables of the execution context. Having explicit execution variation points is key to understanding precisely how a performance aspect evolves along with the version history of the software. We present the key ingredients to make multidimensional profiling effective, and sketch the design of Rizel, an implementation in the Pharo programming language.