Understanding software performance regressions using differential flame graphs

Abstract

Flame graphs are gaining rapidly in popularity in industry to visualize performance profiles collected by stack-trace based profilers. In some cases, for example, during performance regression detection, profiles of different software versions have to be compared. Doing this manually using two or more flame graphs or textual profiles is tedious and error-prone. In this `Early Research Achievements'-track paper, we present our preliminary results on using differential flame graphs instead. Differential flame graphs visualize the differences between two performance profiles. In addition, we discuss which research fields we expect to benefit from using differential flame graphs. We have implemented our approach in an open source prototype called FLAMEGRAPHDIFF, which is available on GitHub. FLAMEGRAPHDIFF makes it easy to generate interactive differential flame graphs from two existing performance profiles. These graphs facilitate easy tracing of elements in the different graphs to ease the understanding of the (d)evolution of the performance of an application.