Repository navigation
Perf: batch same-kind cell publication under one allocation lock #15
Description
Activity
Implemented and pushed as Chris with no coauthors:
65441fe— general same-kindHeap.createBatch, one-lock publication, exact accounting/parallel/OOM tests, and README usage3183d08— prefix-return OOM ordering so consumers commit successful cells before retry/recovery179ffe2— inline the single-cell helpers to preserve direct-path codegen- zig-js consumer
84d2f9f4— batch up to 16 fixed-shape object allocations per shared quick-loop checkpoint
Validation/evidence:
- zig-gc full unit suite and full TSan suite passed locally; the OOM tests caught and prevented publication of uninitialized output slots during development.
- zig-js focused semantic/GC tests and focused TSan tests passed, including a real four-Thread assertion that at least 2,048 cells used the batch API.
- zig-js full validation passed 761/761 tests with zero leaks plus 7/7 comparison-harness tests.
- Exact-parent direct object-churn A/B: 195.410 ms -> 192.974 ms median (-1.25%, candidate won 5/7 pairs), checksum exact.
- Exact-parent shared 8-lane A/B: 902.841 ms -> 393.507 ms median (-56.42%, candidate won 7/7 pairs), checksum exact.
- The validated 1,540-sample publication at zig-js
3d790a60reduced shared 8-lane object churn from 18,472.129 ms to 12,368.464 ms (-33.0%) and improved scaling from 0.21x to 0.31x amid high published dispersion.
Remote CI was missing from zig-gc/main, so #16 restores it in
a659d66. Current two-job unit/20x-TSan run: https://github.com/zig-utils/zig-gc/actions/runs/29364477008All #15 API, correctness, accounting, and consumer-performance acceptance items are satisfied; I will close this after the restored remote gate is green.
Remote dependency gate is green after restoring it to main and fixing explicit Linux libc linkage: https://github.com/zig-utils/zig-gc/actions/runs/29364720539
- unit: success
- 20-iteration TSan: success
Together with the previously posted exact-parent consumer A/B and full zig-js validation, all acceptance criteria are satisfied.
Parent consumer issue: zig-utils/zig-js#97
Post-quickening profiling of the exact zig-js shared object_churn row (ReleaseFast, 8 lanes) finds Heap.create at 14,907 top-of-stack samples in a five-second sample, versus GcCellBacking.allocFn at 1,936 and headerForPayload at 1,371. Heap.create serializes every parallel allocation through the single alloc_lock even when a caller can safely initialize several same-kind cells before its next safepoint.
Scope:
Acceptance: