perf(examples): use mimalloc as global allocator in the bench binary#67
Merged
Conversation
Qodo reviews are paused for this user.Troubleshooting steps vary by plan Learn more → On a Teams plan? Using GitHub Enterprise Server, GitLab Self-Managed, or Bitbucket Data Center? |
3 tasks
Callgrind on the bench binary showed roughly 33 % of cycles inside libc's malloc/free family (`_int_malloc`, `_int_free`, `malloc`, `free`, `malloc_consolidate`, `unlink_chunk`, …). The hot paths are short-lived small allocations: `SmallVec` segments inside `Ranges<RerVersion>` on every range clone, the per-call `FxHashMap<&Requirement, bool>` in `reduce_by`, hashbrown rehashes, `String::clone`s for package names. mimalloc's small-object path is measurably cheaper than glibc's on exactly this shape. Wire `mimalloc::MiMalloc` as `#[global_allocator]` in the bench binary only. Production wheels (`rer-python`) are untouched — switching the global allocator in a PyO3 extension is more nuanced and is left for a separate change. ## Benchmark (188 cases, release, same machine, two consecutive runs) | Stage | Total | Mean | vs rez | |------------------------------------|--------:|-------:|-------:| | Baseline (main) | 43.0 s | 230 ms | 8.8× | | + Shared cache (PR #66) | 34.4 s | 183 ms | 11.1× | | + mimalloc (this branch), run 1 | 29.8 s | 158 ms | 12.8× | | + mimalloc (this branch), run 2 | 29.4 s | 156 ms | 13.0× | Cumulative -32 % from main. Stacks cleanly on top of #66. ## Correctness 188/188 still match rez 1:1 (`cargo test --release -p rer-resolver --test test_rez_benchmark -- --ignored`). Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
doubleailes
added a commit
that referenced
this pull request
May 15, 2026
The phase loop builds a `pending_set: FxHashSet<(usize, usize)>` of scope-pair indices and then iterates it, calling `scopes[x].reduce_by(scopes[y].package_request)` for each. Profile data on rez's 188-case benchmark: 27 M `reduce_by` calls, but **98.5 %** are immediate early-returns because `scopes[y]`'s family doesn't appear in `scopes[x]`'s `fam_requires` set. Doing that check at pair-generation time instead of pair-consumption time avoids tens of millions of function-call round trips, *and* — this is the bigger win in practice — keeps the pending set itself ~100× smaller. The downstream cost reduction is across the board: - the `FxHashSet` doesn't rehash through 27 M inserts - `pending.sort_unstable()` runs over ~400 K entries instead of 27 M - the consume loop pops ~400 K entries instead of 27 M - fewer allocations from the inner `pending_set` growth Implementation: - Add `PackageScope::fam_requires_contains(name)` that delegates to the slice (or returns `false` for conflict / ephemeral scopes). - Replace each `pending_set.insert((x, y))` with a `maybe_push` helper that applies the filter. - Apply the same filter in the cascade path inside the consume loop (when scope `x` narrows, we now only push `(j, x)` for `j` that actually depend on `x`'s family). ## Benchmark (188 cases, release, same machine) | Stage | Total | Mean | vs rez | |----------------------------------|--------:|-------:|--------:| | Baseline (main) | 43.0 s | 230 ms | 8.8× | | + Shared cache (PR #66) | 34.4 s | 183 ms | 11.1× | | + mimalloc (PR #67) | 29.4 s | 156 ms | 13.0× | | **+ this filter, run 1** |**21.9 s**|**117 ms**|**17.5×**| | **+ this filter, run 2** |**22.3 s**|**119 ms**|**17.1×**| -25 % on top of #67 and **-49 % cumulative from main**. The differential test got the same lift — `cargo test … --release` over the 188 cases dropped from ~46 s to ~38 s. ## Correctness 188/188 still match rez 1:1 (`cargo test --release -p rer-resolver --test test_rez_benchmark -- --ignored`). The filter is sound: if `scopes[x].fam_requires_contains(name)` is `false`, then no variant in `scopes[x]` mentions that family, so `reduce_by` would have returned `Unchanged` with no reductions — skipping the call preserves both the result and the reduction list. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
doubleailes
added a commit
that referenced
this pull request
May 15, 2026
The phase loop builds a `pending_set: FxHashSet<(usize, usize)>` of scope-pair indices and then iterates it, calling `scopes[x].reduce_by(scopes[y].package_request)` for each. Profile data on rez's 188-case benchmark: 27 M `reduce_by` calls, but **98.5 %** are immediate early-returns because `scopes[y]`'s family doesn't appear in `scopes[x]`'s `fam_requires` set. Doing that check at pair-generation time instead of pair-consumption time avoids tens of millions of function-call round trips, *and* — this is the bigger win in practice — keeps the pending set itself ~100× smaller. The downstream cost reduction is across the board: - the `FxHashSet` doesn't rehash through 27 M inserts - `pending.sort_unstable()` runs over ~400 K entries instead of 27 M - the consume loop pops ~400 K entries instead of 27 M - fewer allocations from the inner `pending_set` growth Implementation: - Add `PackageScope::fam_requires_contains(name)` that delegates to the slice (or returns `false` for conflict / ephemeral scopes). - Replace each `pending_set.insert((x, y))` with a `maybe_push` helper that applies the filter. - Apply the same filter in the cascade path inside the consume loop (when scope `x` narrows, we now only push `(j, x)` for `j` that actually depend on `x`'s family). ## Benchmark (188 cases, release, same machine) | Stage | Total | Mean | vs rez | |----------------------------------|--------:|-------:|--------:| | Baseline (main) | 43.0 s | 230 ms | 8.8× | | + Shared cache (PR #66) | 34.4 s | 183 ms | 11.1× | | + mimalloc (PR #67) | 29.4 s | 156 ms | 13.0× | | **+ this filter, run 1** |**21.9 s**|**117 ms**|**17.5×**| | **+ this filter, run 2** |**22.3 s**|**119 ms**|**17.1×**| -25 % on top of #67 and **-49 % cumulative from main**. The differential test got the same lift — `cargo test … --release` over the 188 cases dropped from ~46 s to ~38 s. ## Correctness 188/188 still match rez 1:1 (`cargo test --release -p rer-resolver --test test_rez_benchmark -- --ignored`). The filter is sound: if `scopes[x].fam_requires_contains(name)` is `false`, then no variant in `scopes[x]` mentions that family, so `reduce_by` would have returned `Unchanged` with no reductions — skipping the call preserves both the result and the reduction list. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This was referenced May 15, 2026
doubleailes
added a commit
that referenced
this pull request
May 15, 2026
`PackageVariantSlice::extractable` was a `HashSet::is_subset` call that iterated `common_fams` and tested membership in `extracted_fams` on every probe. Fresh callgrind on the rez 188-case benchmark (after #66/#67/#68/#70) put `HashSet::is_subset` at **11.8 %** of inclusive cycles — nearly half of `PackageVariantSlice::extract`'s total 25.3 %. Every `extract()` call hits this guard; the vast majority return early with "nothing left to extract." The set operation is unnecessary. `extracted_fams` is only ever populated by inserting an element of `common_fams.difference( extracted_fams)` in `extract`, and `copy_with_entries` resets it to empty. So `extracted_fams ⊆ common_fams` always holds, and under that invariant the `is_subset` check is equivalent to a length compare: common_fams.is_subset(extracted_fams) ⟺ common_fams ⊆ extracted_fams ⟺ (since extracted ⊆ common) common_fams == extracted_fams ⟺ |common| == |extracted| Replace the body with `common_fams().len() > extracted_fams.len()`. ## Benchmark (188 cases, release, same machine, two runs) | Stage | Total | Mean | vs rez | |--------------------------------------|--------:|-------:|-------:| | Baseline (main, post-#70) | 18.6 s | 99 ms | 20.5× | | + this change, run 1 | 13.1 s | 70 ms | 29.1× | | + this change, run 2 | 13.9 s | 74 ms | 27.4× | **-30 % on top of #70**, **-69 % cumulative from main** (43.0 s → 13.1 s, 8.8× rez → 29.1× rez). Differential test got the same lift: 188/188 still match rez 1:1, in 20.77 s (down from 27.67 s). Predicted gain was 5–10 %. Like #68, hidden downstream costs (the `is_subset` iterator setup/teardown, the hash lookups it performed, and the now-unnecessary `common_fams_cache` first-time computation on slices that never need extraction) made the actual gain larger. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
doubleailes
added a commit
that referenced
this pull request
May 15, 2026
Callgrind on rez's 188-case benchmark (post-#71/#72) showed `SmallVec::extend` + `Drop` at ~4 % of cycles, almost entirely from `VersionRange::clone`. Every `Requirement::clone()` (in `extracted_request.clone()`, the per-pair `package_request.clone()` in `reduce_by`, the `req.clone()` and `package_request.clone()` in `Reduction`, etc.) deep-copies the inner `Ranges`'s `SmallVec` of `(Bound, Bound)` segments. After the rest of the perf stack (#66/#67/#68/#70/#71/#72), this is the largest non-amortised allocation cost left. Switch the inner from `Ranges<RerVersion>` to `Rc<Ranges<RerVersion>>`. `Rc<T>::clone` is a refcount bump; `Rc<T>::Hash`/`Eq` defer to the inner `T`, so the derived semantics on `VersionRange` are unchanged. Methods that build a new range (`intersection`, `union`, `complement`, `from_versions`, `span`, `split`, ...) still produce a fresh `Ranges` internally and wrap it with `Rc::new` — the win is on the read / clone path, not the construction path. `as_ranges()` still returns `&Ranges` (via `Rc::deref`). `into_ranges` now uses `Rc::unwrap_or_clone` — falls back to a clone if the `Rc` is shared, but is the consume-the-`VersionRange` API and rare in practice. ## Benchmark (188 cases, release, same machine, two runs) | Stage | Total | Mean | vs rez | |------------------------------------|--------:|-------:|-------:| | Baseline (post-#71/#72), median | ~12.7 s | 68 ms | ~30× | | + this change, run 1 | 11.2 s | 60 ms | 34.1× | | + this change, run 2 | 11.3 s | 60 ms | 33.7× | **~11 % on top of #72**, **~74 % cumulative from main** (43.0 s → 11.2 s, 8.8× rez → 34.1× rez). Differential test (`cargo test … --ignored`): 17.73 s, **188/188 still match rez 1:1**. Predicted 3–5 %. The slightly bigger gain reflects that `VersionRange::clone` cascades into a lot more than just the `SmallVec::extend` it was attributed to in the callgrind exclusive view — it also drove allocator-side work and the matching `Drop`s. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
3 tasks
doubleailes
added a commit
that referenced
this pull request
May 15, 2026
Callgrind on rez's 188-case benchmark (post-#71/#72) showed `SmallVec::extend` + `Drop` at ~4 % of cycles, almost entirely from `VersionRange::clone`. Every `Requirement::clone()` (in `extracted_request.clone()`, the per-pair `package_request.clone()` in `reduce_by`, the `req.clone()` and `package_request.clone()` in `Reduction`, etc.) deep-copies the inner `Ranges`'s `SmallVec` of `(Bound, Bound)` segments. After the rest of the perf stack (#66/#67/#68/#70/#71/#72), this is the largest non-amortised allocation cost left. Switch the inner from `Ranges<RerVersion>` to `Rc<Ranges<RerVersion>>`. `Rc<T>::clone` is a refcount bump; `Rc<T>::Hash`/`Eq` defer to the inner `T`, so the derived semantics on `VersionRange` are unchanged. Methods that build a new range (`intersection`, `union`, `complement`, `from_versions`, `span`, `split`, ...) still produce a fresh `Ranges` internally and wrap it with `Rc::new` — the win is on the read / clone path, not the construction path. `as_ranges()` still returns `&Ranges` (via `Rc::deref`). `into_ranges` now uses `Rc::unwrap_or_clone` — falls back to a clone if the `Rc` is shared, but is the consume-the-`VersionRange` API and rare in practice. ## Benchmark (188 cases, release, same machine, two runs) | Stage | Total | Mean | vs rez | |------------------------------------|--------:|-------:|-------:| | Baseline (post-#71/#72), median | ~12.7 s | 68 ms | ~30× | | + this change, run 1 | 11.2 s | 60 ms | 34.1× | | + this change, run 2 | 11.3 s | 60 ms | 33.7× | **~11 % on top of #72**, **~74 % cumulative from main** (43.0 s → 11.2 s, 8.8× rez → 34.1× rez). Differential test (`cargo test … --ignored`): 17.73 s, **188/188 still match rez 1:1**. Predicted 3–5 %. The slightly bigger gain reflects that `VersionRange::clone` cascades into a lot more than just the `SmallVec::extend` it was attributed to in the callgrind exclusive view — it also drove allocator-side work and the matching `Drop`s. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
3 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Stacks on top of #66. Callgrind on the bench binary showed roughly 33 % of cycles inside libc's malloc/free family (`_int_malloc`, `_int_free`, `malloc`, `free`, `malloc_consolidate`, `unlink_chunk`, …). The hot allocations are short-lived small ones: `SmallVec` segments inside `Ranges` on every range clone, the per-call `FxHashMap<&Requirement, bool>` in `reduce_by`, hashbrown rehashes, `String::clone`s for package names. mimalloc's small-object path is measurably cheaper than glibc's on exactly that shape.
This wires `mimalloc::MiMalloc` as `#[global_allocator]` in the bench binary only. Production wheels (`rer-python`) are untouched — switching the allocator in a PyO3 extension has Python-allocator interactions worth treating as a separate change.
Benchmark (188 cases, release, same machine)
Additive -13 % on top of #66; -32 % cumulative from main.
Correctness
Base
This PR targets `shared-variant-cache` (`#66`). Once #66 merges, GitHub will retarget this to `main` automatically; rebase will be a no-op.
🤖 Generated with Claude Code