rustc-perf is a project aimed at continuously collecting and visualizing the performance of rustc. During Google Summer of Code (GSoC), I worked on improving rustc-perf in various ways.
I have worked on the following four major topics.
The first two topics were the primary goals outlined in my project proposal. However, I also worked on the third and fourth areas, which were outside the scope of my initial objectives.
As a result of my contributions, the load time for the query trace visualizer has been so much improved, bringing its performance on par with compile-time visualization on certain pages, and contributors can now check benchmark changes without needing to run the benchmark suite website locally.
More details about these topics are provided below.
Previously, Perfetto was used to visualize query traces. It took several minutes to load traces in rustc-perf, and a less-than-ideal experience due to the transition to a separate website. To address this, we aimed to embed the query trace visualization tool directly into the rustc-perf project.
Perfetto did not provide an easy way to embed it. So I have tried other visualizers such as Catapult, Speedscope, Firefox Profiler. However, I found a better way, to host a slightly modified version of the Perfetto UI under RLO, after discussing with the author of Perfetto. (PR)
Before reaching this decision, we explored various visualization tools. Here are some of the options we considered:
trace2html Python script to dynamically generate pages. However, since the project was no longer being maintained and would increase our Python dependencies, we finally decided not to adopt it.During discussions on the embedding process, we discovered that we were sending data to Perfetto uncompressed, sometimes over 800MB at a time. To resolve this, I implemented Brotli compression, which reduces the size of the traces by 90%. Therefore, the load times has been substantially improved. (PR)
Both compile-time and runtime benchmarks are collected in rustc-perf. The runtime benchmarks are relatively new and were not as well visualized as the compile-time benchmarks. I added graphs for runtime performance on both the details and dashboard pages, achieving the same level of visualization as compile-time performance.
I implemented a CLI tool that makes it easy to compare benchmarks locally, which had been a long-requested feature. With this command, contributors can now check benchmark changes without needing to run the website.
Additionally, this command has been ported to rustc’s bootstrap (the Rust compiler build system) as well, so that Rustc contributors can now test the performance of the compiler in a much easier way.
I also focused on several other improvements, spending a lot of time refactoring the codebase. Consequently, the status page, which hadn’t been working locally for quite some time, is now functional, and the module structure is much easier to understand.
As I contributed to rustc-perf, I started to develop an interest in rustc itself, so I’ve submitted a few patches to the compiler.
Working on an open-source project for the first time in my life has been an incredibly valuable experience. Everything—from communicating in English to discussions on GitHub and Zulip—was new and exciting for me.
I would like to express my deepest gratitude to my mentor @Kobzol, who always supported me throughout this journey, from discussing the proposal to helping with this final report, all while serving as an Organization Admin for the Rust Foundation. Thank you so much.
I am also grateful to the Rust Foundation, to Google for providing the GSoC opportunity, and to everyone who helped me throughout the project.