Repository navigation
release: train PGO and BOLT on --fast checks too - #976
Conversation
The fast preset runs code default mode never does (lazy members, locality scheduling, the extra checker caches, leaf-file regions), so the release profiles had no counts for it. Add --fast runs of xstate-main and webpack to pgo-train.sh and bolt.sh, and a --fast CLI run to the Windows training. Measurements in notes/perf-pgo-fast-training.md.
Performance change vs mainWorkflow run Β· main Clef gate: Clef estimated 13.1% default and single-threaded wall time, 11.2% peak RSS, 14.6% single-threaded instruction count impact probability. The main and PR tsrs-lintThe main and PR |
Merging this PR will not alter performance1 warning
Comparing Footnotes
|
Codex Review Summary: β CompletedThis comment shows the latest Codex review activity on this pull request.
βΉοΈ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with π while any review is running, comments if it has suggestions, and reacts with π once all reviews finish with no findings. |
There was a problem hiding this comment.
π‘ Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 33879b3810
βΉοΈ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with π.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| - [perf-icount-round4](perf-icount-round4.md) | ||
| - [perf-mimalloc-safe](perf-mimalloc-safe.md) | ||
| - [perf-pgo](perf-pgo.md) | ||
| - [perf-pgo-fast-training](perf-pgo-fast-training.md) |
There was a problem hiding this comment.
Remove the unrequested README edit
The root AGENTS.md explicitly prohibits updating any README.md file unless the user requests it. This notes-index addition modifies notes/README.md without such a request. Revert the index change; the performance note and training-script updates can remain independently.
AGENTS.md reference: AGENTS.md:L6-L6
Useful? React with πΒ / π.
release binaries are PGO trained (plus BOLT on linux) on the conformance suite, fourslash, and
tsrs -p ./tsrs --linton xstate-main and webpack. none of it passes--fast, so everything the fast preset switches on (lazy member tables, lazy d.ts members, locality scheduling + work stealing, the extra checker caches, leaf regions) had zero counts and got laid out as cold. in a--fastprofile of a large private monorepo the lazy member functions are visibly hot, and in the shipped profileget_ready_lazy_member_table_worker,create_lazy_member_tableandget_lazy_declared_memberliterally have all-zero block countsthis adds
tsrs -p . --faston the same two projects topgo-train.shand to the instrumented tsrs runs inbolt.sh, and a--fastrun of the existing cli fixture to the windows training. the end-of-script profile checks now require the new profiles too. default-mode runs stay as they weremeasured on a graviton4 box (m8g.4xlarge), two PGO builds from one instrumented build that differ only in which raw profiles get merged, then BOLT on each with main's vs this branch's
bolt.sh. 15 alternating pairs per workload, CPU time B vs A:--fastprivate monorepo--fastvscode--fastmikro-orm--fastmui-docs--fastsupabase-studio--fastwebpack (trained)22 of the 24
--fastcells have intervals that exclude 0. cycles and wall move the same way, peak RSS moves under 1% everywhere, and output is byte identical between arms everywhere. the win is fetch: frontend stalls -4 to -5%, l1i refills -2 to -7%tradeoff is tiny: two more sub-second runs per training, and default mode wobbles by at most +0.3% CPU because the merged hot set grows. i also tried
--fast --checkers 1, the lint run underTSRS_FAST=1,--fastat 8 cores, and a second pair of projects on top. none of them moved cycles, so they're not in herecaveats: only measured on linux aarch64. the go fourslash suite was left out of both arms (no go on the box, and it only trains the language server). macos/windows get the change unmeasured, and the windows run is a tiny fixture. full numbers, method and repro are in
notes/perf-pgo-fast-training.md