Profdiff Vizualizér

Daněk, Jiří

Diagnosing performance regressions in dynamic compilers like the GraalVM com- piler is difficult because a single benchmark run can involve thousands of compilation units, each undergoing hundreds of inlining and optimization transformations. While the existing Profdiff command-line tool successfully identifies optimization differences, pre- senting deeply nested compilation trees as formatted text output lacks the interactivity needed to trace how a single inlining decision triggers a chain of dependent optimizations that together cause a non-trivial regression. This thesis addresses this gap by develop- ing a web-based GUI for Profdiff. The developed application allows users to navigate benchmark runs, highlight the hottest compilation units, and visualize interactive tree comparisons for faster regression detection. A Micronaut backend wraps the modified Profdiff logic as a dependency and exposes a REST API, while an Angular frontend pro- vides a comprehensive view of diffed benchmark runs and their metadata. Demonstrated through a case study on the mnemonics Renaissance benchmark, the resulting Profdiff Visualizer reduces the manual effort required to navigate raw text output by providing an interactive interface for diagnosing compilation regressions.

guest :: login Digital Repository
		Search		Submit		Help		About