Fr0zkaandClaude Opus 5.5
5e091d254c
Opt-in (L77 build + --linux77-profile; compiled out by default, 0.02% handler cost). FI/TN cost account closes against direct grid-fill timing within 0.15% (LOD0-2 bins within 1.4%). Passage carving = 66% of LOD0-2 CPU (52-54% whole view), passage support 10-14%, base packed + analytic field ~3%. Hotspots: value noise 17%, lower_bound in VisitNearestPassageSpatialNode 14%, ApplyResolvedStratePassageCarves 6.5%. LOD0/1 complete LAST (4.1/4.6 s) while LOD2 is ready at 2.2 s. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
82 lines
7.1 KiB
Markdown
82 lines
7.1 KiB
Markdown
# Linux77 — FI/TN whole-view cost
|
||
|
||
## Run and method
|
||
|
||
Final profile runs were FI/TN r12 and r13, source stamp `6532cc3d9ca7137b1b62a3ae1c7bc54ebffc8fa3d58a9f12740c11d2048f4a0d`. They used eight workers, LOD0–8, `hash=off`, `force_generate=off`, and no `--stream-hash`; whole-tile skips remained active (481 FI, 578 TN). The route and phase tags are enabled only in the L77 build plus `--linux77-profile`. A separate default build with the profiler compiled out emitted zero samples.
|
||
|
||
The sampler used a per-thread CPU timer on each worker and the main thread. It recorded one PC per timer delivery and expanded delayed signals by `1 + si_overrun`. Across four profiles it recorded 32,235 events / 128,923 weighted ticks, with nine timers per process, zero ring loss, and 129.06 s of profiled thread CPU. Handler CPU was 23.9 ms total (0.018%); per-run cost was 0.017–0.020%. The handler figure excludes kernel signal-delivery cost. The full per-call tag/counter overhead could not be isolated from run-to-run timing variation.
|
||
|
||
## Cold readiness
|
||
|
||
Times are from the two final profiled runs; values show mean (range). LOD readiness means every desired tile at that LOD completed.
|
||
|
||
| Ready point | FI | TN |
|
||
|---|---:|---:|
|
||
| First tile | 0.377 s (0.376–0.377) | 0.399 s (0.388–0.409) |
|
||
| 27-tile neighborhood | 0.744 s (0.714–0.774) | 0.563 s (0.559–0.566) |
|
||
| LOD0 | 4.139 s (4.118–4.160) | 4.046 s (3.921–4.172) |
|
||
| LOD1 | 4.618 s (4.593–4.642) | 4.233 s (4.106–4.361) |
|
||
| LOD2 | 2.226 s (2.209–2.242) | 2.040 s (1.983–2.096) |
|
||
| LOD0–2 all ready | 4.618 s (4.593–4.642) | 4.234 s (4.106–4.361) |
|
||
| Whole view ready | 4.617 s (4.593–4.641) | 4.233 s (4.105–4.360) |
|
||
|
||
LOD1 is last, so LOD0–2 completion and whole-view readiness coincide in these runs. The ~1 s target is met for the first tile and neighborhood, but not for complete LOD0–2. The compiled-out controls measured 4.54–5.55 s FI and 4.70–4.74 s TN for whole-view readiness; they emitted no sampler records. Their timing scatter is too large to infer a small instrumentation delta. This headless run measures readiness, not frame-time stutter.
|
||
|
||
## Exclusive cost account
|
||
|
||
Worker CPU is the sum of tile-worker time, averaged over r12/r13. Each cell is seconds (% of that scope); rounded values may not sum exactly. The phase split assigns field and composition work wherever it runs. MC, skirt, and refine columns contain the remaining non-field work in those caller stages, so composition inside a mesher call is not counted twice. `Other` is unclassified grid work plus the residual tile work.
|
||
|
||
| Strate / scope | Worker CPU | Base packed / analytic / other | Graph | Bands | Passages | Support | Caps + edits | Dispatch | MC | Skirts | Refine | Other |
|
||
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
|
||
| FI LOD0–2 | 23.02 s | 0.45 (1.9%) / 0.29 (1.3%) / 0.00 | 0.03 (0.1%) | 0.52 (2.3%) | 15.17 (65.9%) | 3.26 (14.2%) | 0.06 (0.3%) | 0.84 (3.7%) | 0.36 (1.6%) | 0.02 (0.1%) | 1.73 (7.5%) | 0.30 (1.3%) |
|
||
| FI whole view | 33.87 s | 0.45 (1.3%) / 1.29 (3.8%) / 0.08 (0.2%) | 0.04 (0.1%) | 0.66 (2.0%) | 18.16 (53.6%) | 3.50 (10.3%) | 0.09 (0.3%) | 2.27 (6.7%) | 0.47 (1.4%) | 0.05 (0.2%) | 6.35 (18.8%) | 0.47 (1.4%) |
|
||
| TN LOD0–2 | 21.50 s | 0.31 (1.4%) / 0.41 (1.9%) / 0.00 | 0.02 (0.1%) | 0.57 (2.6%) | 14.07 (65.4%) | 2.90 (13.5%) | 0.05 (0.2%) | 0.78 (3.6%) | 0.26 (1.2%) | 0.04 (0.2%) | 1.85 (8.6%) | 0.25 (1.2%) |
|
||
| TN whole view | 30.63 s | 0.31 (1.0%) / 1.39 (4.5%) / 0.04 (0.1%) | 0.04 (0.1%) | 0.70 (2.3%) | 15.91 (52.0%) | 3.06 (10.0%) | 0.06 (0.2%) | 2.10 (6.8%) | 0.38 (1.2%) | 0.07 (0.2%) | 6.21 (20.3%) | 0.38 (1.2%) |
|
||
|
||
“Base other” is native plus any other base route. Native appears at LOD4–8 in the whole view; legacy-surface and shared-source routes recorded no calls here. Analytic-batch also recorded no calls.
|
||
|
||
### Route census and halo work
|
||
|
||
[route-census.csv](route-census.csv) has exact route calls and points, plus expanded CPU time, for every LOD × caller-stage × route and the LOD0–2 / whole-view aggregates. Grid-fill points for LOD0–2 were:
|
||
|
||
| Strate | Packed scalar | Packed X8 calls / points | Analytic scalar | Total grid points |
|
||
|---|---:|---:|---:|---:|
|
||
| FI | 10.519 M | 1.044 M / 6.319 M | 0.989 M | 17.827 M |
|
||
| TN | 6.639 M | 1.398 M / 8.423 M | 1.548 M | 16.610 M |
|
||
|
||
Route-census time is exclusive by outer route and stage; composition time remains inside its packed/analytic route. The cost table is a separate, exclusive phase-first view of those same samples. `dispatch_lookup` and `unknown` rows have zero density-call counts by design.
|
||
|
||
Halo counts are exact; halo CPU is its sampled share of grid fill and overlaps the cost categories above. FI LOD0–2: 1.534 M of 17.827 M points (8.6%), about 1.92 s CPU (8.4%). FI whole view: 2.524 M of 24.439 M (10.3%), about 2.65 s (7.8%). TN LOD0–2: 1.485 M of 16.610 M (8.9%), about 1.89 s (8.8%). TN whole view: 2.382 M of 22.604 M (10.5%), about 2.47 s (8.1%). The run recorded halo points, not a set of coordinates, so exact cross-tile duplicate-coordinate counts are unavailable.
|
||
|
||
## Closure and sampled hotspots
|
||
|
||
The grid estimate uses each LOD’s weighted `grid_fill` sample share of direct worker time; the comparison is direct `mc_grid_ms` timing:
|
||
|
||
| Strate / scope | Sample estimate | Direct grid fill | Closure |
|
||
|---|---:|---:|---:|
|
||
| FI LOD0–2 | 19.051 s | 19.058 s | 99.96% |
|
||
| FI whole view | 23.931 s | 23.903 s | 100.11% |
|
||
| TN LOD0–2 | 17.365 s | 17.377 s | 99.93% |
|
||
| TN whole view | 20.818 s | 20.848 s | 99.86% |
|
||
|
||
Every LOD0–2 bin closed within 1.4%. Sparse far bins were noisier (FI LOD7 reached +18%; TN LOD6 reached −17%), while the whole-view totals still closed within 0.15%. Those per-LOD tails are sample variance, not missing work in the target scopes.
|
||
|
||
Top symbolized PC samples below are expanded-tick shares, grouped by `addr2line` leaf symbol; the `lower_bound` caller resolves into passage spatial-node visitation:
|
||
|
||
| Sampled function | FI LOD0–2 / whole | TN LOD0–2 / whole |
|
||
|---|---:|---:|
|
||
| `VF_FieldValueNoise` | 16.6% / 12.8% | 17.0% / 12.9% |
|
||
| `std::lower_bound` in `VisitNearestPassageSpatialNode` | 14.2% / 11.4% | 11.5% / 9.2% |
|
||
| `ApplyResolvedStratePassageCarves` | 6.5% / 5.6% | 6.9% / 5.9% |
|
||
| `VF_FieldLerp` | 6.3% / 4.8% | 6.5% / 4.8% |
|
||
|
||
## Biggest lever
|
||
|
||
Passage carving is the largest phase: 65–66% of LOD0–2 CPU and 52–54% of whole-view CPU. A same-world 2× speedup in this phase would save about 7.0–7.6 worker CPU seconds for LOD0–2 and 8.0–9.1 seconds for the whole view: roughly one third of early CPU and one quarter of whole-view CPU. Removing the phase entirely is the measured hard ceiling (52–54% whole view); it is not a realistic target.
|
||
|
||
## Failed or excluded runs
|
||
|
||
The first `ITIMER_PROF` attempts (`fi-profile-r1`, `tn-profile-r1`) recorded 1,822 and 1,701 samples, all on the main thread (`lod=255`), and are excluded. All final r12/r13 profiles and compiled-out controls exited successfully; none timed out.
|
||
|
||
No `--stream-hash` run was made, as required to keep skip behavior. The source diff adds only opt-in profiling tags, counters, and the sampler; generator and meshing formulas were not changed. Hash equality to `90378a5` was therefore not directly measured.
|