Fr0zkaandClaude Opus 5.5
8d62ba57ed
FI/TN LOD0-2 W12 worker CPU 13.9/12.2 s: passages 40/35%, far-shell refinement (HasDownwardHit/ HasUpwardHit ~1.7-1.9 s) in 'other' 27/30%, support 14/13%, dispatch 8%, base field 6%, bands 6%. Workers 28%/24% idle during LOD0-1. Near-ready W12: SW 0.77, FI 1.51, TN 1.38, VS 0.96 s. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
6.7 KiB
Linux103 — current near-ready cost account
Source stamp: c5f49a45f51856b857a1ee8877b7d777208e67daf64281caefc7ef84e8006410.
L77 cost rows are medians of three W12 profiles per stratum. Phase CPU is estimated by
sample share, calibrated to direct tile-worker CPU. Timer weights include 1 + si_overrun.
W12 cost
| LOD0–2 worker cost | FI | TN | Measurement |
|---|---|---|---|
| Direct tile-worker CPU | 13.919 s | 12.155 s | Direct worker timers |
| Passage phase | 5.538 s (39.8%) | 4.209 s (34.6%) | L77 sample allocation |
| Other phase | 3.749 s (26.9%) | 3.681 s (30.4%) | L77 sample allocation; mostly far-shell refinement |
| Support | 1.864 s (13.7%) | 1.623 s (13.0%) | L77 sample allocation |
| Dispatch lookup | 1.107 s (8.0%) | 1.043 s (8.6%) | L77 sample allocation |
| Base field | 0.831 s (6.1%) | 0.808 s (6.5%) | L77 sample allocation |
| Composition bands | 0.773 s (5.7%) | 0.781 s (6.3%) | L77 sample allocation |
| Graph, caps, edits | 0.084 s | 0.053 s | L77 sample allocation |
Phase rows partition the worker total. Direct stage timers overlap that account: MC total/grid
fill was 9.838 / 9.263 s FI and 7.827 / 7.340 s TN; skirt mesh was 0.656 / 0.661 s.
Grid-fill sample estimates closed against direct mc_grid_ms at 100.0% FI and 99.8% TN.
Whole near-worker sample closure was 99.2% FI and 99.1% TN.
Passage share is 39.8% FI / 34.6% TN. L99's earlier W8 census reported 34.78% / 28.81%; that comparison crosses source stamps and worker counts. Current near exact-fallback calls (901,055 / 607,012) make off-grid exact work a plausible contributor, but its causal share was not isolated.
Top route × phase allocations (direct-worker-equivalent seconds):
| Stratum | Largest route × phase costs |
|---|---|
| FI | packed-scalar passages 4.533; analytic-scalar passages 0.936; packed support 1.651; packed bands 0.641; dispatch 1.095; unknown/other 3.745 |
| TN | packed-scalar passages 3.119; analytic-scalar passages 1.089; packed support 1.394; packed bands 0.671; dispatch 1.035; unknown/other 3.673 |
Passage sweep work: FI used 3,371,576 base and 347,335 detail samples (90.7% / 9.3%);
TN used 3,145,983 and 332,547 (90.4% / 9.6%). Symbolized samples attribute about
0.247 / 0.143 s FI and 0.261 / 0.169 s TN to base/detail preparation, 1.010 / 0.739 s
to sweep lookups, and 0.822 / 0.585 s to exact-route visits plus miss-bound tests. Those
subcosts are sampled estimates, not separate direct timers.
Near LOD0–2 sweep disposition was FI 13,903,288 misses = 901,055 exact fallbacks + 13,002,233 skips; TN 9,086,664 = 607,012 + 8,479,652. The exact-fallback totals include 255,529 FI and
250,311 TN skirt queries. Near far-shell stratum refinement made 410,980 FI and 447,978 TN
off-grid queries. Symbolized HasDownwardHit / HasUpwardHit samples attribute about 1.73 s
FI and 1.85 s TN to those hit checks.
Whole-view off-grid caller counts, per run:
| Caller | FI | TN |
|---|---|---|
| MC midpoint unique points | 29,101 | 29,864 |
| Skirt surface / extruded / clipped vertices | 21,380 / 192,868 / 26,365 | 21,082 / 191,186 / 22,643 |
| Skirt refinement face / coarse centroid | 188,382 / 210,056 | 184,353 / 209,077 |
| Far-shell surface / projection-normal queries | 539,321 / 192 | 539,321 / 192 |
| Far-shell stratum refinement | 1,131,895 | 1,145,781 |
Across all four strata, all-view sweep misses totaled 63,036,259: 1,908,562 exact fallbacks and 61,127,697 skips. Skip reasons were feature bound 59,646,691; tile free 147,272; vertical bound 160,272; negative reach 1,158,700; positive reach 14,762. Per-caller miss totals equal exact fallbacks plus skips; unaccounted misses are zero. All 12 W12 profiles had zero lost samples, and their parsed sample counts and expanded-tick sums matched.
Critical path and readiness
In the W12 timeline traces, FI LOD0–1 settled at 1.605 s; TN at 1.218 s. The far-parent build then took 171 ms FI / 114 ms TN. The LOD2 batch started as the far parent completed and drained in 56 / 79 ms. Shape compile was 311 / 200 ms in the traces. Worker-pool busy time was 70.7% FI / 75.0% TN, with 28.3% / 24.3% explicitly idle. The long interval is LOD0–1 tile work; the LOD2 tail is short.
L80 protocol medians, near LOD0–2 / whole view seconds, three interleaved runs per cell:
| Stratum | W8 | W12 |
|---|---|---|
| SW | 0.764 / 1.914 | 0.772 / 1.786 |
| FI | 1.612 / 2.593 | 1.508 / 2.404 |
| TN | 1.510 / 2.432 | 1.377 / 2.199 |
| VS | 1.107 / 2.061 | 0.958 / 1.822 |
The W12 whole-view pipeline reported shape / far-shape times of 0.199 / 0.114 s FI and 0.215 / 0.114 s TN.
Next lever
No single value-preserving code change was clear from the timers, so L103 ships the following ranked candidates. Savings are inferred CPU scenarios; elapsed-time gains and hash preservation need a changed build to measure.
- Reduce passage sweep/carve work. It is 5.54 s FI / 4.21 s TN; a 10% reduction would save about 0.55 / 0.42 worker CPU seconds.
- Reduce far-shell hit checks. The symbolized estimate is 1.73 / 1.85 s; a 25% reduction would save about 0.43 / 0.46 worker CPU seconds. The million-plus whole-view refinement queries make this the best bounded target to investigate.
- Reduce support traversal. It is 1.86 / 1.62 s; a 25% reduction would save about 0.47 / 0.41 worker CPU seconds.
Hashes and checks
All 24 whole-view runs matched across Clang/GCC W1/W8/W12. The 20 peer comparisons had identical tile counts and zero density, mesh, payload, or topology mismatches. All 24 hashes also equal the corresponding L101 runs.
| Stratum | Tiles | Clang W1 whole hash: density / mesh / payload / topology |
|---|---|---|
| SW | 2,850 | 8723d8c61348bf31 / fe1c3e410373a343 / bf111bbc45e2f017 / c9c562b100b51f7f |
| FI | 2,249 | 6bb633f756e2b8be / fe930e83b0dc6bf3 / 5298602285b6892c / e306f87ae432a4fc |
| TN | 2,233 | 612b2d11fc6bd946 / c026a40c6127d649 / 0f9036914e28fe46 / 94b36e5f73694eff |
| VS | 2,329 | 99c93f4c7155ffb3 / 488d421c6438f529 / 249b5afc12be4540 / 9c5e22f6dc1e44b8 |
Fast tier: PASS, 274 checks (269 pass, five retained diagnostics), 274.4 s. No production
Source/ files changed, so no Unreal build or arrival trace was required.
Failed or excluded runs
- The first fast-tier launch exited before running checks because its temporary binary symlink was wrong. The corrected run passed.
- The first readiness launch had no output directory and started no headless process; all 32 subsequent warmup/measured runs completed.
- The W8 census and an earlier policy-3 census are retained separately and excluded from the final W12 account. All final W12 census, timeline, readiness, and hash headless runs passed.
Headless runs used taskset -c 0-11 prlimit --as=6442450944 and a 1,740-second timeout.
Artifacts are under Saved/Linux/l103/.