You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Scaling V11 · The page poses and updates only the Souls that show (main thread saturated at 200 Souls) #2618
Tranche:#2504 · Agenda:#1960 · Lane: V (page) · Task: V11 · Ceiling: 4 h (a limit, not a target) · Tier: expected T2; node tools/proof/tier.mjs decides · Depends on: PR 2613 (V9's page harness) merged, or measured on its branch
Main purpose
The page's per-Soul body cost scales with the Souls a viewer can see, not with every Soul in the world.
At 200 Souls the main thread is 99–100 % busy in every arm.
Each Soul's body (the people kit's show, pose and render) is 27–32 % in the Free view.
three.js matrix updates are the largest self-time entries, 15–29 %.
Every view is at 13–24 fps at 200 Souls, and about 10 at 500.
The JS heap reaches 0.8–1.0 GB at 200 and 1.2–1.7 GB at 500, on phone too.
Acceptance
A1. Step 0: per-Soul body cost per frame by part, and how many of the 200 are on screen, or within a distance that matters, in each camera.
A2. Pose and matrix updates run only for Souls that show, or at a lower rate for far ones, under a stated rule.
A Soul that comes into view is posed correctly on its first visible frame. Test: an off-screen Soul pans into view and its pose matches one updated every frame.
Bodies never pop, and none teleport or freeze in view.
A3. Bars declared here before the runs. page at 50, 200 and 500 Souls, Free view and Broadcast, desktop and phone, base against candidate, 2 repeats:
p50 at 200 Souls under 25 ms on desktop Free;
body share at 200 down by at least half;
heap reported.
Touches
The people kit's per-frame body update (name the files in A1, likely apps/web/src/gen-mesh/ and apps/web/src/WorldEntityLayer.ts), their tests and changes/. Renderers observe state and never own a second movement path: this changes only how often an observed pose is applied, never where a Soul is.
Stop rule
The standard one: see #2554's stop rule. A visual change needs an uncut, real-speed clip from the runner (the clip command) at T2 or above. Never meet a bar by hiding Souls a viewer should see, or by lowering what the page shows without a stated rule.
V11 lane: started 15:43 UTC 16 Nov; bars and arms declared before any run
Files (A1 names them). The people kit's per-frame body update is mars/viewer/app/engine-view/humans.ts (HumanViews.render → pose, its pose-rate rule poseDue), reached from take-replay.tsTakePeople.show. Not apps/web/src/gen-mesh/ or WorldEntityLayer.ts: the Souls' bodies are drawn by the people kit. The product change is in humans.ts only, with its test in apps/web/test/body-pose-rate.test.ts.
Premise found in the code (Step 0, before measuring). A pose-rate rule already exists (PERF-EDEN): bodies within nearM are posed every frame, to farM every 2nd, beyond every 4th, off screen every 8th. Three gaps:
A body that comes into view keeps a stale pose until its rate's next turn, up to 3 frames of a 4-frame rate after up to 7 held frames. A2's test fails on main today: palms 0.05 to 0.35 m off on the first visible frame.
A held body's rig still goes through three's matrix walk and bone upload every frame: updateMatrixWorld recomposes every bone of every body, and each skeleton re-uploads its bone texture.
"Off screen" is judged by the sampled place only, so a held pose drawn at its old place could stay in view.
Stated rule (the candidate).
A body is in the view if its 1.3 m sphere is in the frustum at its sampled place or where it is drawn now.
A body that comes into view is posed on that frame, whatever its rate.
A body that is neither posed this frame nor in the view is held. Three's matrix walk and bone upload skip its rig until it is posed again, unless its parent moved. Its drawn pose and its shadow stay exactly what they were.
Rates are unchanged: in view, every frame / 2nd / 4th by distance; off screen, every 8th. No Soul is hidden, and nothing is drawn at a lower detail. Bodies in the view are untouched, so idle life and the other render-time offsets still apply every frame.
Arms. These run on souls-timed via timed-run.ymlpage (V9's harness), cores world=1, 2 repeats each, dispatched together.
Both measurement branches add the same page timers (measurement only, never for main): per frame, bodies on screen (near / mid / far), posed and held, and the ms of show (sample, setFrame, pose), the bodies' matrix walk and skeleton updates.
agent/2618-measure-base = PR 2613's head + the timers.
agent/2618-measure-cand = that + this PR's commit.
arm
shape
sizes
base E / cand E
event (uneven)
50, 200 (V9: event at 500 restarts the world)
base U / cand U
uniform (even)
50, 200, 500
Bars. Scored once, as declared; nothing is re-scored.
B1. Desktop Free at 200 Souls: the frame p50 (the median of the two repeats' p50s) is under 25 ms in the candidate, in both shapes.
B2. Body share at 200 Souls is down by at least half in the candidate, in every tier × camera × shape cell (8 cells), each the mean of 2 repeats.
Body share is (show + the bodies' matrix walk + skeleton updates) ÷ the window's wall time, from the page timers.
V9's reading is also reported: show inclusive in the CPU profile.
B3. The JS heap is reported for every arm, base and candidate.
A1. The base arms answer Step 0: body ms per frame by part, and how many of the 200 are on screen and within nearM/farM in each camera.
A2. The unit test proves it: an off-screen Soul pans into view at every phase of both rates, and its bone matrices and palms on the first visible frame equal those of a body posed every frame. A held body uploads nothing, and no body in view is held.
Clip (T2): clip --kit crowd-pan, base main against candidate (the PR head). 120 Souls gather at the bandstand, and a Free camera turns so the crowd pans into view and out, slowly and then fast. Bodies must not pop, teleport or freeze in view.
V11 checkpoint, 16:14 UTC: handed back while the page runs finish
Done.
Draft PR 2619 (agent/2618-bodies-that-show, head d21fe5b, T2) carries the change, its test and the crowd-pan clips. The base main and candidate clips are linked in the PR, and A2 is met with evidence.
CI: ci-ok is red only because the PR is a draft (the CI window is closed); everything else passes or skips.
Running (dispatched together at 15:59 UTC, as declared above): page E base / cand, page U base / cand.
Exact next step. When all four have finished:
Download each run's artifact with gh run download <id> -R soulsimulation/souls -D .local/v11/<id>.
From a checkout of agent/2618-measure-base, run node tools/test-harness/page-souls-v11-score.mjs <baseE> <candE> <baseU> <candU>, each argument being .local/v11/<id>/timed-run-page-<id>. It prints B1, B2, V9's show share, every arm with its heap (B3), A1's Step 0 tables and the candidate's held counts.
Post that output here unchanged and tick A1 and A3 in the PR (met or unmet:).
Write proof.json, upload it, then mark the PR ready and re-add ci:run.
B1: unmet. Desktop Free at 200 Souls is at 66.7 ms (event) and 58.3 ms (uniform). The bar was 25 ms.
B2: unmet.
In 7 of 8 cells the candidate's body share is 0.75 to 0.93 of base.
In event desktop Broadcast it is 1.22. There, the director's shots in the candidate arm had 70 and 53 bodies within nearM (40 m), against 20 and 2.6 in base, and the unchanged rule poses those every frame. Shot choice drives that cell, not the change.
B3: met. The heap is reported for every arm.
A1: met. The A1 table below has the numbers.
At 200 Souls, pose() is most of show(): 6.6 to 21.8 ms of 12.3 to 28.1 ms per frame.
The bodies' matrix walk is 2.6 to 4.2 ms per frame, and skeleton updates 1.5 to 3.0 ms.
In the Free view the crowd is on screen, 40 to 90 m away: 143 of the 185 drawn bodies in event desktop.
The change (A2).
The candidate holds 29 to 151 bodies per frame at 200.
The bodies' matrix and skeleton ms drop by 30 to 40 % in the Free view, and by 60 to 87 % in uniform Broadcast.
Bodies in the view cost what they did, so this change cannot halve body share where the crowd is on screen.
500 Souls. Neither uniform arm has page data at 500. The world on one core went black and made its planned restart (exit 75) before the page arms, as V9 found at 500 in the event shape. That is a server limit, not the page.
Noise. Arms differ run to run in crowd and Broadcast shots. For example, event desktop Free had 146 and 143 bodies on screen in base, and 99 and 127 in the candidate.
B1: desktop Free at 200 Souls, frame p50 (median of repeats; bar: candidate under 25 ms)
shape
base p50 (r1 / r2)
cand p50 (r1 / r2)
cand median
met
event
66.8 / 66.8
66.7 / 66.7
66.7
no
uniform
66.6 / 66.6
50.1 / 66.6
58.3
no
B1 unmet.
B2: body share at 200 Souls (mean of repeats; bar: candidate at most half of base in every cell)
shape
tier
camera
base share
cand share
ratio
base body ms/frame
cand body ms/frame
met
event
desktop
broadcast
0.251
0.305
1.22
24.59
33.31
no
event
desktop
free
0.409
0.380
0.93
34.84
30.45
no
event
phone
broadcast
0.239
0.217
0.91
17.08
17.62
no
event
phone
free
0.412
0.316
0.77
25.40
18.37
no
uniform
desktop
broadcast
0.260
0.211
0.81
22.98
18.55
no
uniform
desktop
free
0.431
0.376
0.87
27.13
22.45
no
uniform
phone
broadcast
0.280
0.211
0.75
17.68
12.75
no
uniform
phone
free
0.400
0.359
0.90
22.68
20.44
no
B2 unmet.
V9's reading: show inclusive share of the CPU profile (first repeat, largest size)
shape
Souls
tier
camera
base
cand
event
200
desktop
broadcast
0.234
0.288
event
200
desktop
free
0.296
0.321
event
200
phone
broadcast
0.191
0.183
event
200
phone
free
0.265
0.294
Every arm: frame p50 / p95, fps, heap (B3), body ms per frame (mean of repeats)
shape
Souls
tier
camera
arm
p50 (r1 / r2)
p95 (r1 / r2)
fps
heap MB
body ms
body share
long tasks ms
event
50
desktop
broadcast
base
33.4 / 50
83.3 / 83.4
20.9
605.3
7.22
0.153
31169
event
50
desktop
broadcast
cand
33.4 / 33.4
83.4 / 83.3
23.1
710.3
6.81
0.155
26221
event
50
desktop
free
base
16.7 / 16.7
50 / 49.9
41.8
597.8
7.46
0.310
1573
event
50
desktop
free
cand
16.7 / 16.7
50 / 49.9
41.7
655.6
7.03
0.292
1711
event
50
phone
broadcast
base
33.3 / 33.4
66.7 / 83.4
26.1
576.5
6.16
0.159
17987
event
50
phone
broadcast
cand
33.3 / 33.3
66.7 / 66.7
28.9
842.2
5.14
0.148
11434
event
50
phone
free
base
16.7 / 16.7
33.4 / 33.4
50.2
666.6
5.63
0.282
563
event
50
phone
free
cand
16.7 / 16.7
33.4 / 33.4
49.1
695.6
4.92
0.241
667
event
200
desktop
broadcast
base
66.5 / 99.9
150 / 166.6
13.1
945.7
24.59
0.251
48889
event
200
desktop
broadcast
cand
83.4 / 83.4
183.4 / 150
11.7
951.1
33.31
0.305
52905
event
200
desktop
free
base
66.8 / 66.8
116.8 / 133.4
12.6
958.9
34.84
0.409
55592
event
200
desktop
free
cand
66.7 / 66.7
116.7 / 133.3
13.6
1055.8
30.45
0.380
49831
event
200
phone
broadcast
base
50 / 66.7
133.4 / 133.3
15.6
976.6
17.08
0.239
43341
event
200
phone
broadcast
cand
66.7 / 66.6
150 / 150
13.9
1019.5
17.62
0.217
45294
event
200
phone
free
base
50.1 / 50.1
83.4 / 100
16.5
879.0
25.40
0.412
43371
event
200
phone
free
cand
50 / 50
133.3 / 100
18.0
974.2
18.37
0.316
32664
uniform
50
desktop
broadcast
base
16.8 / 16.7
50 / 50
38.5
605.7
5.25
0.202
4005
uniform
50
desktop
broadcast
cand
16.7 / 16.7
50.1 / 50.1
37.7
546.6
3.83
0.144
11273
uniform
50
desktop
free
base
16.7 / 16.7
33.4 / 49.9
49.2
614.8
4.99
0.245
1707
uniform
50
desktop
free
cand
16.7 / 16.7
33.3 / 33.3
54.3
741.8
3.65
0.198
456
uniform
50
phone
broadcast
base
16.7 / 33.3
49.9 / 66.6
39.1
666.1
4.88
0.189
7213
uniform
50
phone
broadcast
cand
16.7 / 33.3
33.4 / 50
41.2
692.2
3.19
0.131
4273
uniform
50
phone
free
base
16.7 / 16.7
33.3 / 33.2
55.9
657.3
4.40
0.245
417
uniform
50
phone
free
cand
16.7 / 16.7
16.8 / 16.8
57.5
669.0
3.01
0.173
357
uniform
200
desktop
broadcast
base
83.3 / 83.3
133.3 / 133.4
12.6
1061.0
22.98
0.260
51339
uniform
200
desktop
broadcast
cand
66.7 / 66.7
133.3 / 166.7
12.8
1119.9
18.55
0.211
50070
uniform
200
desktop
free
base
66.6 / 66.6
83.4 / 100
16.1
1145.3
27.13
0.431
48230
uniform
200
desktop
free
cand
50.1 / 66.6
100 / 99.9
17.2
966.8
22.45
0.376
45142
uniform
200
phone
broadcast
base
66.6 / 50
133.3 / 83.4
16.5
1049.3
17.68
0.280
39997
uniform
200
phone
broadcast
cand
50 / 50.1
100 / 116.7
17.5
951.6
12.75
0.211
34153
uniform
200
phone
free
base
50 / 50
100 / 83.3
17.9
961.4
22.68
0.400
29441
uniform
200
phone
free
cand
50 / 50
100 / 83.4
18.0
925.9
20.44
0.359
27366
uniform
500
desktop
broadcast
base
no data
uniform
500
desktop
broadcast
cand
no data
uniform
500
desktop
free
base
no data
uniform
500
desktop
free
cand
no data
uniform
500
phone
broadcast
base
no data
uniform
500
phone
broadcast
cand
no data
uniform
500
phone
free
base
no data
uniform
500
phone
free
cand
no data
A1 (Step 0): the base arms' body work per frame, by part, and the Souls on screen (mean of repeats)
shape
Souls
tier
camera
bodies
on screen
near / mid / far
posed
show ms
sample
setFrame
pose
matrix ms
skeleton ms (calls)
event
50
desktop
broadcast
50.0
17.2
10.5 / 5.4 / 1.4
19.1
5.92
0.57
0.30
4.67
0.81
0.49 (50.0)
event
50
desktop
free
50.0
35.7
0.0 / 35.0 / 0.6
20.8
6.26
0.55
0.27
5.08
0.78
0.41 (50.0)
event
50
phone
broadcast
50.0
5.6
4.5 / 0.7 / 0.5
12.3
4.63
0.54
0.28
3.45
1.08
0.45 (50.0)
event
50
phone
free
50.0
10.9
0.0 / 9.4 / 1.6
11.9
4.16
0.52
0.26
3.04
1.03
0.43 (50.0)
event
200
desktop
broadcast
124.6
66.9
11.3 / 39.9 / 15.7
42.7
19.84
1.67
4.90
12.29
2.83
1.91 (124.6)
event
200
desktop
free
184.8
144.6
0.0 / 142.6 / 2.0
77.4
28.11
1.63
3.72
21.80
4.12
2.62 (184.8)
event
200
phone
broadcast
123.9
14.5
5.3 / 7.7 / 1.5
23.8
12.93
1.65
3.74
6.59
2.65
1.50 (123.9)
event
200
phone
free
181.9
74.7
0.0 / 68.0 / 6.7
49.5
18.71
1.57
3.17
13.10
4.14
2.55 (181.9)
uniform
50
desktop
broadcast
50.0
12.4
3.6 / 1.2 / 7.7
10.8
3.89
0.56
0.26
2.70
0.91
0.45 (50.0)
uniform
50
desktop
free
50.0
8.5
0.0 / 7.1 / 1.4
9.1
3.66
0.56
0.26
2.45
0.91
0.42 (50.0)
uniform
50
phone
broadcast
50.0
6.7
0.4 / 0.2 / 6.1
7.5
3.28
0.53
0.25
2.16
1.14
0.45 (50.0)
uniform
50
phone
free
50.0
1.6
0.0 / 0.1 / 1.5
6.5
2.85
0.52
0.23
1.75
1.10
0.45 (50.0)
uniform
200
desktop
broadcast
174.8
63.1
7.2 / 4.3 / 51.5
36.5
16.58
1.63
4.38
9.60
3.74
2.66 (174.8)
uniform
200
desktop
free
190.2
80.4
0.0 / 79.8 / 0.5
53.9
19.91
1.65
2.53
14.86
4.20
3.01 (190.2)
uniform
200
phone
broadcast
145.2
31.8
2.7 / 0.1 / 29.0
24.4
12.34
1.59
3.25
6.64
3.34
2.00 (145.2)
uniform
200
phone
free
177.7
60.7
0.0 / 49.3 / 11.5
42.3
16.76
1.55
2.95
11.39
3.60
2.32 (177.7)
The candidate arms: held and posed per frame
shape
Souls
tier
camera
on screen
posed
held
matrix ms
skeleton ms (calls)
event
50
desktop
broadcast
19.9
18.3
25.1
0.52
0.26 (24.9)
event
50
desktop
free
30.3
18.9
16.5
0.58
0.24 (33.5)
event
50
phone
broadcast
13.5
11.9
30.6
0.47
0.17 (19.4)
event
50
phone
free
22.7
11.2
23.0
0.58
0.22 (27.0)
event
200
desktop
broadcast
75.5
74.8
63.4
2.09
1.20 (84.8)
event
200
desktop
free
113.0
61.0
28.6
2.87
1.58 (117.2)
event
200
phone
broadcast
31.1
30.7
101.2
1.13
0.58 (46.6)
event
200
phone
free
57.5
37.8
70.1
1.74
0.92 (67.8)
uniform
50
desktop
broadcast
12.4
8.5
32.9
0.38
0.17 (17.1)
uniform
50
desktop
free
9.6
9.3
35.3
0.26
0.10 (14.7)
uniform
50
phone
broadcast
2.7
7.0
41.4
0.23
0.06 (8.7)
uniform
50
phone
free
2.2
6.8
41.8
0.24
0.06 (8.2)
uniform
200
desktop
broadcast
34.1
34.7
120.4
1.44
0.71 (51.5)
uniform
200
desktop
free
84.0
50.7
61.0
2.21
1.22 (92.7)
uniform
200
phone
broadcast
2.2
22.8
151.0
0.68
0.26 (24.0)
uniform
200
phone
free
77.6
47.4
84.5
2.03
1.19 (89.8)
Host: renderer, cores and load per arm are in each results.json (gl, cores, rows[].host) and timed-run.json (cpuSplit).
Activity
genekogan commentedon Nov 16, 2026
V11 lane: started 15:43 UTC 16 Nov; bars and arms declared before any run
Files (A1 names them). The people kit's per-frame body update is
mars/viewer/app/engine-view/humans.ts(HumanViews.render→pose, its pose-rate ruleposeDue), reached fromtake-replay.tsTakePeople.show. Notapps/web/src/gen-mesh/orWorldEntityLayer.ts: the Souls' bodies are drawn by the people kit. The product change is inhumans.tsonly, with its test inapps/web/test/body-pose-rate.test.ts.Premise found in the code (Step 0, before measuring). A pose-rate rule already exists (PERF-EDEN): bodies within
nearMare posed every frame, tofarMevery 2nd, beyond every 4th, off screen every 8th. Three gaps:updateMatrixWorldrecomposes every bone of every body, and each skeleton re-uploads its bone texture.Stated rule (the candidate).
Arms. These run on souls-timed via
timed-run.ymlpage(V9's harness), coresworld=1, 2 repeats each, dispatched together.Both measurement branches add the same page timers (measurement only, never for main): per frame, bodies on screen (near / mid / far), posed and held, and the ms of
show(sample, setFrame, pose), the bodies' matrix walk and skeleton updates.agent/2618-measure-base= PR 2613's head + the timers.agent/2618-measure-cand= that + this PR's commit.Bars. Scored once, as declared; nothing is re-scored.
showinclusive in the CPU profile.nearM/farMin each camera.clip --kit crowd-pan, basemainagainst candidate (the PR head). 120 Souls gather at the bandstand, and a Free camera turns so the crowd pans into view and out, slowly and then fast. Bodies must not pop, teleport or freeze in view.genekogan commentedon Nov 16, 2026
V11 checkpoint, 16:14 UTC: handed back while the page runs finish
Done.
agent/2618-bodies-that-show, head d21fe5b, T2) carries the change, its test and the crowd-pan clips. The basemainand candidate clips are linked in the PR, and A2 is met with evidence.ci-okis red only because the PR is a draft (the CI window is closed); everything else passes or skips.Running (dispatched together at 15:59 UTC, as declared above): page E base / cand, page U base / cand.
Exact next step. When all four have finished:
gh run download <id> -R soulsimulation/souls -D .local/v11/<id>.agent/2618-measure-base, runnode tools/test-harness/page-souls-v11-score.mjs <baseE> <candE> <baseU> <candU>, each argument being.local/v11/<id>/timed-run-page-<id>. It prints B1, B2, V9'sshowshare, every arm with its heap (B3), A1's Step 0 tables and the candidate's held counts.unmet:).ci:run.genekogan commentedon Nov 16, 2026
V11 scored as declared (comment 6921574122); nothing is re-scored
The four page runs, dispatched together at 15:59 UTC, all succeeded:
Reading
nearM(40 m), against 20 and 2.6 in base, and the unchanged rule poses those every frame. Shot choice drives that cell, not the change.pose()is most ofshow(): 6.6 to 21.8 ms of 12.3 to 28.1 ms per frame.B1: desktop Free at 200 Souls, frame p50 (median of repeats; bar: candidate under 25 ms)
B1 unmet.
B2: body share at 200 Souls (mean of repeats; bar: candidate at most half of base in every cell)
B2 unmet.
V9's reading:
showinclusive share of the CPU profile (first repeat, largest size)Every arm: frame p50 / p95, fps, heap (B3), body ms per frame (mean of repeats)
A1 (Step 0): the base arms' body work per frame, by part, and the Souls on screen (mean of repeats)
The candidate arms: held and posed per frame
Host: renderer, cores and load per arm are in each results.json (
gl,cores, rows[].host) and timed-run.json (cpuSplit).