Dynamics consumer memory growth exceeds the existing flatness limit before next release #1623

Activity

1. github-actions

   added

   needs-triageFiled without an area: label; triage and label itFiled without an area: label; triage and label it

   on Oct 4, 2026
2. github-actions commented on Oct 4, 2026

   

   github-actionsBot

   on Oct 4, 2026 – with GitHub Actions

   More actions

   Thanks for filing. Labelled needs-triage because it has no area: label yet. The scheme is area:<subsystem> (runtime-vm, jit, concurrency, observer, memory, trace-tape, packages, http, gfx, embed, docs, ci, gates, lint-tooling, consumer, aot, stdlib) plus a kind (kind:*, or the stock bug/enhancement) — add the ones that fit if you know them, and a maintainer will confirm and drop needs-triage. See CONTRIBUTING.md, "Reporting Bugs".
3. InauguralPhysicist

   added

   area:memorySubsystem: memorySubsystem: memory

   bugSomething isn't workingSomething isn't working

   and removed

   needs-triageFiled without an area: label; triage and label itFiled without an area: label; triage and label it

   on Oct 6, 2026
4. InauguralPhysicist commented on Oct 6, 2026

   

   InauguralPhysicist

   on Oct 6, 2026

   CollaboratorAuthor

   More actions

   Consumer-wave re-run on a candidate containing #1629: memory gate PASSES

   Candidate: v0.44.0 = 01bbf7ea843e04aa779199a105cdc0802271c241. This is also current main: git rev-list --count v0.44.0..origin/main = 0. dynamics pins the same tag: dynamics#49 merged as f20a0a5 and sets .devcontainer/Dockerfile EIGS_REF=v0.44.0.

   Containment:

   $ git merge-base --is-ancestor 45a71e1d 01bbf7ea && echo CONTAINS
   CONTAINS
   

   (45a71e1d is the merge commit of #1629, "gc: retire sole-pin candidates between full cycle scans".)

   The original wave measured 96bbbb54, a stale pre-rebase head of the #1616 branch that doesn't contain #1629. It was the wrong candidate.

   Method (mirrors runtime's bisect measurements)

   • Fresh clone at 01bbf7ea, make gfx CC=gcc. That's the same build as the dynamics devcontainer.
   • dynamics at main f20a0a5, clean tree. tests/test_bif_mem.sh was not modified: sha256 f789ca14…, GROWTH_KB=25600, CEILING_KB=204800, VCAP=1500000.
   • Same setup as CI and runtime's runs: DISPLAY unset, so the gate uses xvfb-run -a, and every process pinned to 4 CPUs with taskset -c 0-3. A GitHub runner has 4 vCPUs. Every leg is a cold process: GNU time %M peak RSS under ulimit -v 1500000.
   • A: 8 full runs of the unchanged gate. Each run measures static and toggle at 30/100/300 frames, and growth is peak(300) − peak(30). Each run also includes both planted faults.
   • B: 6 interleaved cold pairs (30 vs 300 frames) per cell, for static/toggle × default/forced-VM (EIGS_JIT_OFF=1). These use runtime's unchanged measure.sh / run_pairs.sh method, with the order rotated each round.
   • Control: runtime's 96bbbb54 build ran through the same unchanged gate, interleaved at rounds 2, 5 and 8. This shows the instrument still catches the regression on this box today.

   A. Unchanged gate, 8 runs on v0.44.0 (KB; peaks at 30 / 100 / 300 frames)

   run	static peaks	static growth	toggle peaks	toggle growth	exit	planted faults caught
   1	128,676 / 132,048 / 132,308	3,632	128,492 / 130,944 / 132,596	4,104	0	both
   2	129,428 / 131,160 / 132,692	3,264	127,968 / 130,696 / 131,748	3,780	0	both
   3	129,216 / 132,424 / 132,992	3,776	127,872 / 130,028 / 131,256	3,384	0	both
   4	129,144 / 131,132 / 131,656	2,512	128,628 / 130,332 / 132,316	3,688	0	both
   5	129,028 / 130,924 / 132,416	3,388	128,584 / 130,644 / 132,872	4,288	0	both
   6	129,584 / 131,868 / 131,788	2,204	128,552 / 130,512 / 131,944	3,392	0	both
   7	129,072 / 131,852 / 132,544	3,472	128,656 / 130,788 / 132,444	3,788	0	both
   8	128,884 / 131,792 / 132,100	3,216	128,804 / 131,572 / 131,852	3,048	0	both
   min / median / max		2,204 / 3,326 / 3,776		3,048 / 3,734 / 4,288	8/8 PASS

   Limit is 25,600 KB, so the worst run is under 17% of it. The highest peak in any leg was 132,992 KB, against the 204,800 KB ceiling. Structure counters held in every leg (built=1, series=1, points=4000, markers=4).

   Control: stale 96bbbb54, same gate, same box, same session:

   run	static growth	toggle growth	exit
   ctrl2	46,692	34,580	1 (FAIL)
   ctrl5	46,592	35,704	1 (FAIL)
   ctrl8	46,440	35,644	1 (FAIL)

   This matches the wave's 45,952 / 35,200 and runtime's 6 of 6 cold-pair failures.

   B. Cold pairs, v0.44.0: peak RSS at 30 → 300 frames = growth (KB)

   round	static, default	static, forced VM	toggle, default	toggle, forced VM
   1	128,800 → 131,520 = 2,720	129,160 → 132,196 = 3,036	127,852 → 132,128 = 4,276	128,080 → 131,728 = 3,648
   2	129,232 → 131,716 = 2,484	129,656 → 131,212 = 1,556	127,904 → 132,624 = 4,720	128,880 → 131,908 = 3,028
   3	128,900 → 132,084 = 3,184	128,876 → 132,308 = 3,432	128,292 → 132,864 = 4,572	128,208 → 131,644 = 3,436
   4	128,780 → 132,316 = 3,536	129,824 → 132,484 = 2,660	129,040 → 132,064 = 3,024	127,172 → 131,804 = 4,632
   5	128,668 → 132,248 = 3,580	129,104 → 131,904 = 2,800	128,316 → 131,844 = 3,528	128,332 → 131,844 = 3,512
   6	129,012 → 132,312 = 3,300	129,000 → 132,140 = 3,140	128,632 → 131,980 = 3,348	127,580 → 132,084 = 4,504
   min / med / max	2,484 / 3,242 / 3,580	1,556 / 2,918 / 3,432	3,024 / 3,902 / 4,720	3,028 / 3,580 / 4,632

   0 of 24 pairs went over 25,600 KB. All 48 processes exited 0 with empty stderr, BIF-UI-OK, and the exact work counters.

   Full dynamics suite on v0.44.0 (CI order, same setup)

   test_lint, test_smoke, test_lab, test_orbit_hist, test_bifurcation, test_orbit_oracle, test_orbit_mouse (16 checks, all three planted faults caught) and test_bif_mem: 8/8 PASS.

   Independent check: dynamics main CI on the v0.44.0 pin (run 37298446691, push, f20a0a5) is green. Its gate logged static growth 1,288 KB and toggle growth 2,016 KB.

   Environment note (not a #1623 issue)

   On this 8-CPU box, running with all 8 CPUs visible makes any build abort with eigenscript: out of memory (requested 80 bytes) under the gate's ulimit -v 1500000. That happens on v0.44.0 too, and the gate then reports rc=134 FAIL. With 7 or fewer CPUs it runs. The likely cause is per-thread address-space reservation in the software renderer. Runtime's runs and this one were pinned to 4 CPUs, the same as a GitHub runner. This is a dynamics gate portability caveat and doesn't affect this verdict.

   Verdict

   PASS. On a candidate containing #1629, the unchanged dynamics memory gate passes 8 of 8 runs, middle endpoint and toggle arm included. All 24 cold pairs (static and toggle, default and forced-VM) are at least 5× under the 25,600 KB flatness limit. The stale 96bbbb54 still fails on the same box. The full dynamics suite passes. Closing as fixed by #1629 (45a71e1d, shipped in v0.44.0).
5. InauguralPhysicist

   closed this as completedon Oct 6, 2026