Draton GC Benchmarks - 2026-03-18

This report publishes the next public GC benchmark snapshot for commit f90dcdc.

This pass was intentionally narrower than the earlier March 18 hasher work:

• fuse young survivor promotion into the scan loop instead of building a separate promotion list first
• stop pushing pointer-free objects onto the major mark stack
• keep active major-cycle allocations and protect() semantics safe while that mark-stack optimization is in place

The result is mixed, and this report keeps that visible instead of smoothing it over.

Published artifacts

• GitHub Actions scorecard: docs/benchmarks/gc-scorecard-gha-2026-03-18-run-23230366650.json
• GitHub Actions OCaml compare: docs/benchmarks/gc-ocaml-compare-gha-2026-03-18-run-23230366650.json
• Previous public report: docs/benchmarks/gc-results-2026-03-18-e8f4d9e.md

GitHub Actions run

• workflow: gc-perf
• run id: 23230366650
• repository commit: f90dcdc
• result: success

GitHub Actions scorecard summary

Current run:

Scenario	elapsed_ns
young-burst	3525346
promotion-chain	15976933
barrier-churn	7115994
old-reuse-churn	27672566
large-object-burst	4179019

Delta against the previous public run 23228923625:

Scenario	Previous elapsed_ns	Current elapsed_ns	Change
young-burst	3347438	3525346	5.31% slower
promotion-chain	18399359	15976933	13.17% faster
barrier-churn	7276970	7115994	2.21% faster
old-reuse-churn	31130749	27672566	11.11% faster
large-object-burst	4062662	4179019	2.86% slower

Interpretation:

• the promotion-heavy path improved again on the same runner
• old-generation reuse also improved materially
• young-only and large-object-only scenarios regressed slightly in this pass
• the net effect is mixed rather than a clean across-the-board win

Toolchain-facing workloads in the same run:

Workload	Status	build_elapsed_ns	run_elapsed_ns	Extra
gc_stress_linked_list	ok	35756516727	1800659	exit code 50
selfhost_bootstrap	blocked	80490507928	-	LLVM ERROR: unknown special variable

GitHub Actions OCaml comparison

Current run summary:

• status: ok
• scorecard_result: mixed
• draton_wins: 0
• scenario_count: 3
• geometric_mean_ocaml_over_draton_median_speed_ratio: 0.26372264511307886

Per-scenario medians:

Scenario	Draton median ns	OCaml median ns	ocaml_over_draton_median_speed_ratio	Result
young-burst	647943	357866	0.5523109285847675	Draton slower
promotion-chain	1706554	130891	0.07669900864549262	Draton much slower
barrier-churn	882131	381946	0.43298104249822306	Draton slower

Delta against the previous public run 23228923625:

Metric	Previous	Current	Change
young-burst ratio	0.5400071740737759	0.5523109285847675	2.28% better
promotion-chain ratio	0.08526001286120798	0.07669900864549262	10.04% worse
barrier-churn ratio	0.47182918553087844	0.43298104249822306	8.24% worse
geometric mean ratio	0.2790250801150224	0.26372264511307886	5.48% worse

Interpretation:

• the shared OCaml comparison is still honest and still unfavorable overall
• the scorecard-side promotion improvement did not translate into a better relative OCaml result on this run
• the current pass is therefore a targeted runtime improvement, not a proof that Draton is now closing the whole gap

Local release spot checks

Direct release-binary spot checks on the same commit during development were:

Scenario	Direct run elapsed_ns
young-burst	1702670
promotion-chain	2570247
barrier-churn	2489458

These local numbers were useful to validate direction while iterating, but the GitHub Actions data above remains the authoritative same-runner comparison.

Current honest status

• Draton improved some hot runtime paths again, especially promotion and old-gen reuse.
• Draton still does not beat OCaml on the shared microbench suite.
• selfhost_bootstrap is still blocked in the scorecard for the same LLVM reason as before.