Self-Profile Meeting Notes
This document contains meeting notes from the self-profile working group.
2019-03-05: Meeting
Written by: @wesleywiser
Purpose of meeting: Initial meeting to get organized and decide on some next steps.
Agenda:
- Current status update
- Where to track work?
- What does an Minimum Viable Product (MVP) look like?
- Next work items?
Current Status
After a few rounds of experimentation, we have a -Z self-profile option available in the nightly compliler.
The profiler measures a few kinds of things such as when queries start and stop, when the incremental cache is hit and when a cache miss occurs, and when queries are blocked by a parallel query (when the compiler is built in parallel mode).
The profiler currently writes this data in a rather verbose JSON format to a file in the working directory.
This is very slow.
Team updates:
- @michaelwoerister is drafting a tracking issue for our MVP.
- @wesleywiser is currently working on replacing some custom LLVM codegen timing data with the profiler.
Where to track work?
We decided to track just the MVP in a tracking issue on GitHub. The overall roadmap is a bit nebulous at the moment and we felt it would be most productive to focus on getting something usuable on perf.rust-lang.org ASAP.
What does an MVP look like?
Our first MVP will be targeted at getting profiling data available on perf.rust-lang.org.
As a rough outline, perf.rust-lang.org will allow the user to click on a crate from the comparison page and they will be shown a comparison of query performance from both runs. This will be displayed like the NLL Dashboard page but with individual queries on the left instead of crate names.
There’s a number of steps to getting to this point which will be fleshed out in the tracking issue.
Next work items?
In order to get to the MVP, dumping the raw events to disk during compilation needs to get a lot faster. Currently, we’re using JSON for this because it’s easy to parse without sharing a lot of code between the compiler and out-of-tree crates. We’re going to switch this to a compact binary format and create a crate that will form the stable interface that out-of-tree crates can consume. More detail is available in the tracking issue.