GC nursery pacing (#11645): accepted regressions to recover

[proggeramlug]

The owner accepted these regressions on 2026-09-30 when landing #11645 ("land 11645 now, but let's have an issue for the small regression with as many details as we can right now"). This issue tracks getting them back. Part of #11549.

#11645 (survival-aware nursery pacing) starts the scavenge nursery ladder at 4 MB instead of 16 MB. That wins big on package workloads: dotenv −30% instructions, and validator, date-fns, jwt and uuid −20–30% peak RSS. It costs small or fully-live programs one or more extra minors. The rows below are the ones that got worse.

Numbers (main vs PR)

row	Δ instructions	Δ peak RSS	minors/fulls, main → PR
dotenv_parse	−30.5%	−20.3%	0/7 → 61/0
moment_parse_format	−20.6%	+3.1% (21 reps; +5.4% at 11 reps)	1/33 → 43/0
validator_batch	−3.7%	−26.0%	46/0 → 211/0
date-fns_format_add	−1.0%	−30.0%	17/0 → 71/0
qs_parse_nested	+0.36%	+1.2% (ranges overlap)	11/0 → 14/0
qs_stringify_nested	−0.3%	+0.6%	24/10 → 24/10
jsonwebtoken_decode	−3.6%	−22.0%	2/0 → 17/0
uuid_v4	−0.1%	−21.0%	1/0 → 6/0
alloc loop	+0.57%	−20.0%	67/0 → 203/0
binary-trees n=3	+746% (24.95 M → 211.09 M)	+63.8% (15.5 → 25.3 MB)	0/0 → 1/0
binary-trees n=6	+649% (28.69 M → 214.83 M)	+59.0%	0/0 → 1/0
binary-trees n=10	+553% (33.68 M → 219.81 M)	+54.7%	0/0 → 1/0
binary-trees n=20	−4.0%	−8.9%	1/0 → 2/0
gc_ratchet 01_nursery_churn	−4.5%	−25.5%	1/1 → 4/1
gc_ratchet 02_survivor_promotion	+17.3% (276.8 M → 324.6 M)	−11.6%	1/1 → 2/1
gc_ratchet 12_large_live_set	−5.7%	+2.6% (105.6 → 108.4 MB)	5/1 → 6/1

The rows in bold are the accepted regressions.

Method

• Host: perrymaster (Linux x86_64, Zen 4, 16 cores, shared). Both arms were built with --release (-p perry -p perry-runtime-static -p perry-stdlib-static), compiled with PERRY_NO_AUTO_OPTIMIZE=1, and run with TZ=UTC.
• Main was 7fa094cb4. The PR was 0cce8784f, i.e. perf(gc): survival-aware nursery pacing, 4 MB floor on the existing influx ladder (includes #11612) #11645's five commits (HOLD (RSS trade): perf(regex): stop charging operation scratch to GC pressure; grow lent registers (bounded) #11612 ×2, pacing, root-holders classify, changelog) on that main. perf(gc): survival-aware nursery pacing, 4 MB floor on the existing influx ladder (includes #11612) #11645 was later rebased onto 6a5090751 as 427728113; the numbers were not re-taken on that base.
• THP was off (MIMALLOC_ALLOW_THP=0, i.e. 4 KiB accounting). The host has THP in madvise mode, and THP-on RSS changes from run to run by several %; see perf(gc): survival-aware nursery pacing, 4 MB floor on the existing influx ladder (includes #11612) #11645's body.
• Instructions: perf stat -e instructions:u. Package and loop rows are per iteration, from a two-N differential (median of 3 at n1, median of 11 at n2). binary-trees and gc_ratchet rows are totals for a fixed size.
• Peak RSS: /usr/bin/time %M on the binary, median of 11, arms interleaved. moment, 12 and 02 were re-run at 21 reps.
• Minor and full counts come from PERRY_GC_DIAG=1.
• Everything ran under the shared bench lock /tmp/perry-bench-lock.d.
• Raw data on perrymaster:
  • /root/claude-gc-nursery-pacing-bench/m6.txt: instructions and RSS, all rows.
  • /root/claude-gc-nursery-pacing-bench/m7.txt: the 21-rep RSS re-run of moment, 12 and 02.
  • Binaries: /root/claude-gc-nursery-pacing-bench/main2/ and /root/claude-gc-nursery-pacing-bench/pr2/.
  • Scripts: /root/claude-gc-nursery-pacing-{meas2,meas3,compile}.sh, tabulated by bench/tab.py and bench/tab3.py.
• Fixtures: benchmarks/packages/* (dotenv 18.0.1, qs 6.16.0, moment 2.31.0, …) and benchmarks/gc_ratchet/probes/{01,02,12}. binary-trees, alloc and retain are the .ts files in the bench dir.

Per-row mechanisms

• binary-trees n=3/6/10. The program builds a ~5 MB tree that stays fully live and exits after 7–9 MB of allocation. Main never collects it. The PR runs one minor at the 4 MB floor, and that minor promotes the whole tree in place: about +186 M instructions. Earlier profiling put this minor's cost at ~120 M of trace, which is the collector's general per-object visit/classify cost (visit_gc_layout_slot_descriptors, scan_object_fields, classify_arena, visit_slot_with_weak_fact), plus ~33 M of non-trace fixed setup. Main pays the same per-object cost from n≥20 on, where n=20 is −4.0% here. The extra RSS comes from the minor's worklist and moved_headers doubling from a zero estimate. A never-reallocating header list was tried on wip/11549-header-list-experiment: it cut about 3 MB, but traded RSS elsewhere under THP (binary-trees 10→40 +7%, gc_ratchet 01 +3.6%), so it was not landed.
• gc_ratchet 02 (+17.3%). The PR runs 2 minors where main runs 1. Minor 2 re-copies minor 1's surviving cohort, and the final gc() full then sweeps a young generation still full of garbage. The old perf(gc): survival-aware nursery pacing, 4 MB floor on the existing influx ladder (includes #11612) #11645 table showed +4.3%, but that compared main without perf(gc): the first collection's barrier-arming walk skips wholly-nursery blocks #11668 against the PR with it. With perf(gc): the first collection's barrier-arming walk skips wholly-nursery blocks #11668 on main (d7df6e7), this is the second minor's true cost.
• gc_ratchet 12 (+2.6% RSS). One extra early minor (6 vs 5) promotes part of the large live set earlier than main does.
• alloc loop (+0.57%). 203 minors against 67. Each minor has a fixed cost of ~100 k instructions (perf(gc): cut a copying minor's fixed cost (intern young log, skip empty array-tail tables, young-only prunes) #11634 line of work). The 136 extra minors × ~106 k account for the whole delta. To break even, the fixed cost would have to fall to ~35 k per minor.
• qs_parse_nested (+0.36%). 14 minors against 11. See the qs prunes below.

Unexplained: moment's RSS shift

On 10ece9958 (the previous #11645 measurement), moment's THP-off RSS was main 40.8 MB against PR 35.3 MB (−13.5%). On 7fa094cb4, main is 39.6 MB (range 37.7–45.2) and the PR is 40.8 MB (range 40.0–41.2). Main barely moved; the PR's median rose ~5.5 MB. Minor and full counts are unchanged (1/33 against 43/0), and the −20.6% instruction win still holds.

Nothing has been bisected. The candidates are the GC- or rooting-touching commits in 10ece9958..7fa094cb4:

• d7df6e7 perf(gc): the first collection's barrier-arming walk skips wholly-nursery blocks #11668 (first-minor arming walk skips wholly-nursery blocks). It was in both PR measurements, but main only gained it now.
• d2dcafe fix(codegen): root the function across Func.prototype.x = <call> (#11635) #11639: roots the function across Func.prototype.x = <call>. moment does this pervasively at startup, and its peak is set in the first ~0.2 s by the startup cohort's copy.
• 329bc71 fix(gc): pinned objects are marked, traced and scanned as roots (use-after-free in cross-thread promises) #11664: pinned objects are marked, traced and scanned as roots.
• d4ea6be perf: function objects point to a static JsFunctionInfo; delete the closure-body registry #11654: function objects point to a static JsFunctionInfo; the closure-body registry is deleted.
• 0abf224 shapes: link-time static ShapeIds by content — runtime adoption + driver pass; REGISTERED_TYPED_SHAPES deleted (charter step 4) #11653 and e8ae819 perf(runtime): store facts are shape identity; a store is one shape compare (charter step 3) #11646: static ShapeIds and store facts.
• 767a791 arrays: exact allocation size, a hole-filled tail, and a one-word element read guard (S0, S1) #11659: exact array allocation size.

The first step is to bisect the PR arm over this range with MIMALLOC_ALLOW_THP=0, looking at moment's peak RSS.

Expected help from #11676 (per-object trace cost)

#11676 cuts the per-object trace ~35%. With #11645 stacked on it earlier, binary-trees came out at +474/+412/+351% (n=3/6/10) instead of +746/+649/+553%, and 02 at −2.8% instead of +17.3%. So once #11676 lands, 02 should drop off this list, and binary-trees should roughly halve.

Candidate next steps

1. Prune the fixed per-minor work that qs's minors spend time in. Shares of qs minor time: closure box-capture prune 16%, layout-owner prune/sort 12%, remembered-set rebuild 12%. Making these incremental or skipping them when nothing changed helps qs, the alloc loop and every high-minor-count row.

2. Cut the first minor's fixed setup (the ~33 M of non-trace work in binary-trees' single minor).

3. A survival-aware first-minor policy. At the first minor, survival alone does not separate binary-trees' fully-live tree from a package's startup cohort, which is 10–31% alive (0.4–1.3 MB). Measured alternatives:

   • Powering on at the 16 MB base fixes binary-trees and 02 but costs dotenv +22.6% and moment +25.1% RSS.
   • An 8 MB start is a size-specific fit that costs 02 +20.6%.

   A policy would need a signal beyond first-minor survival, e.g. live bytes relative to allocated bytes, or deferring the promotion decision.

[proggeramlug]

Re-measured after #11676 merged. Main 034b1ea38 against #11645 at 1412ce8f2, with the same method as the issue body: THP off, the bench lock, and median of 11. Raw data is in /root/claude-gc-nursery-pacing-bench/m8.txt on perrymaster.

row	Δ instr (was)	Δ instr now	Δ RSS (was)	Δ RSS now	minors/fulls, main → PR
binary-trees n=3	+746%	+497.7% (24.96 M → 149.18 M)	+63.8%	+60.2%	0/0 → 1/0
binary-trees n=6	+649%	+432.9%	+59.0%	+54.6%	0/0 → 1/0
binary-trees n=10	+553%	+368.7%	+54.7%	+49.1%	0/0 → 1/0
gc_ratchet 02	+17.3%	+19.8% (257.56 M → 308.65 M)	−11.6%	−12.0%	1/1 → 2/1
gc_ratchet 12	−5.7%	−5.1%	+2.6%	+2.5%	5/1 → 6/1
moment	−20.6%	−20.4%	+3.1%	+13.8% (37.4 → 42.5 MB)	0/33 → 43/0 (main had 1/33)
qs parse_nested	+0.36%	+0.75%	+1.2%	−2.4%	11/0 → 14/0
alloc loop	+0.57%	+0.57%	−20.0%	−19.3%	67/0 → 203/0

• binary-trees shrank as predicted. The instruction cost fell by about a third, to roughly the +474/+412/+351% measured when perf(gc): survival-aware nursery pacing, 4 MB floor on the existing influx ladder (includes #11612) #11645 was stacked on perf(gc): cut the copying minor's per-live-object trace cost (straight-line object scan, hoisted per-object facts, exact memos) #11676. The RSS cost is unchanged, which fits: it comes from the minor's tracking lists, not from the trace.
• 02 did not shrink. perf(gc): cut the copying minor's per-live-object trace cost (straight-line object scan, hoisted per-object facts, exact memos) #11676 made main's own collections cheaper too (277 M → 258 M), so the second minor's relative cost stayed about the same. The earlier stacked "−2.8%" reading for 02 came from a different comparison and does not hold on main.
• moment RSS got worse: +3.1% → +13.8%. Main tightened to 37.2–38.5 MB, now with 0 minors where it previously ran 1. The PR went from 40.8 MB to 42.5 MB (range 42.2–42.9). This is now the largest unexplained item here. The bisect suggested in the body should cover both arms over 10ece9958..034b1ea38, including perf(gc): cut the copying minor's per-live-object trace cost (straight-line object scan, hoisted per-object facts, exact memos) #11676.
• qs parse went from +0.36% to +0.75%. That is still small, but it doubled.

Correctness on the rebased head is clean:

• cargo test -p perry-runtime passes.
• The seeded PROTECT_FROMSPACE sweeps pass on dotenv, moment, qs and binary-trees: 200/200 on the previous base and 50/50 on the final head.
• The gc/array/regex gap A/B matches main.

#11645 is now ready at 356e189a3, on main 5fbc2c3b0. The numbers above were not re-taken after that last rebase: main's P4 shape-only GC flip is in between and could move them.

[proggeramlug]

Owner decision 2026-09-30: #11645 lands with moment peak RSS at +13.8% (37.4 → 42.5 MB, main 034b1ea vs PR, THP off, median of 11). That is above the +3.1% accepted earlier. The change comes from main itself: main now runs 0 minors on moment, where it used to run 1. Recovering it is tracked here, together with gc_ratchet 02 at +19.8% instructions. Next step: bisect main between 7fa094c and 034b1ea for moment's minor-count change.

[proggeramlug]

Additional nursery/headroom tradeoff on a cyclic graph workload, measured after #11645 landed:

Measured 2026-10-01 on macOS arm64, Perry main d40ed1a47019bcae547819ebb5756bcd96ba1057 (v0.5.1655), with matching compiler and runtime archives. This snapshot already includes #11645. These are diagnostic probes; no compiler/runtime fix was applied and the original comparison remains unchanged.

Arm	Median CPU seconds	Median peak RSS MiB
Original/default policy	10.64	131.63
Original, PERRY_GC_SCAVENGE_NURSERY_MB=8	11.42	99.42
Original, PERRY_GC_SCAVENGE_NURSERY_MB=4	11.90	78.52
Typed child-array local/default	4.90	131.55
Typed local + nursery base 4 MiB	6.03	79.58

Three sequential repeats per arm; CPU = child user + system time from wait4, RSS = peak resident memory. Every run matched the original full stdout and exit status, including parent identity and closure/cache checks. Default, nursery-8, nursery-4 and typed-local arms rotated order across three rounds; the combined arm ran afterward. Same constants, node count and 1 GiB limit.

Reducing the base alone costs approximately 12% CPU while reducing RSS approximately 40%. Separate GC diagnostic captures report 981 copying minors at base 4 versus 274 at default, and exit arena capacity 36 MiB versus 72 MiB. Diagnostic timing/counts are not substituted for uninstrumented CPU/RSS. The combined arm is 43.3% lower CPU and 39.5% lower RSS versus fresh default; its CPU win comes from a separate compiler-call-lowering probe.

The explicit base override is not equivalent to #11645's 4 MB starting influx-ladder floor: the base can be dynamically scaled and is not the total/effective nursery size. This is residual workload evidence after the adaptation landed, not a pre-#11645 comparison or a recommendation to lower defaults globally. It does not establish which additional policy change would recover the CPU cost.

Compiler-side opportunity and complete unchanged reproducer: #11743.

Evidence in the project-comparison workspace: demo/results/stress/perry-profile/summary.md, profile-manifest.json, cpu-summary.json, experiment-summary.json, experiments.json, gc-diagnostics.stderr, nursery-4-gc-diagnostics.stderr, census.jsonl, vmmap-*.txt, and the original/typed-receiver makeTree disassemblies. These paths are local artifacts, not publicly hosted links.