Fr0zkaandClaude Opus 5.5
5c519f5206
After the normal visible-apply count is reached and <50% of the 2 ms budget is spent, apply one horizon result (off during startup boost; voxel.StreamingFarApplySupplement=0 disables). 30 m/s UE, 5 pairs: far holes 4.38% -> 2.70%, far finish-to-apply p50 1.105 -> 0.114 s, worst tick 14.74 ms (<15.43), near coverage +0.3 pt, arrival whole-view -313 ms. Geometry unchanged (24/24 hashes), fast tier green. Adds trace-only RMC proxy publication timing. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
27 lines
2.6 KiB
Markdown
27 lines
2.6 KiB
Markdown
# Linux 121 — finished far tiles
|
||
|
||
One change: after four normal visible applies, allow one far apply if less than half of the 2 ms budget is used. Disable the lane during startup boost. Geometry and worker limits stay unchanged.
|
||
|
||
Five interleaved UE pairs at 30 m/s (8 slots, ready depth 8) reduced median far coverage holes from 4.38% to 2.70%; near settled coverage rose from 45.12% to 45.43%. Far worker-finish-to-apply wait fell from 1.105 s / 3.25 s p50/p95 to 0.114 s / 2.40 s. The lane applied 672 extra far tiles across the five runs.
|
||
|
||
The worst tick was 14.74 ms, below L120's 15.43 ms ceiling. The worst per-run p99 was 11.00 ms versus 11.58 ms for baseline. Five 140 cm/s arrival pairs improved paired median LOD0–2 settlement by 20 ms and whole-view/far-clear by 313 ms. Both arms had zero far coverage holes at 20 s.
|
||
|
||
## Stage timings
|
||
|
||
Five paired UE traces per arm. Result wait is worker finish to `ApplyTileResult` (p50/p95 ms); generation, apply, and proxy are p95 ms. Proxy lag is reconstructed from callback and tile-apply timestamps because profile-only proxy-lag fields are zero when profiling is off.
|
||
|
||
| Speed / LOD | Far generation p95 | Result wait baseline → lane | Apply baseline → lane | Proxy baseline → lane |
|
||
|---|---:|---:|---:|---:|
|
||
| 1.4 m/s near | — | 12 / 43 → 9 / 37 | 0.362 → 0.340 | 0.411 → 0.406 |
|
||
| 1.4 m/s far | 5.6 → 6.3 | 209 / 229 → 142 / 209 | 0.314 → 0.252 | 0.221 → 0.196 |
|
||
| 30 m/s near | — | 12 / 125 → 12 / 120 | 0.206 → 0.197 | 0.250 → 0.237 |
|
||
| 30 m/s far | 10.0 → 10.3 | 1,105 / 3,250 → 114 / 2,397 | 0.351 → 0.324 | 0.335 → 0.312 |
|
||
|
||
The four-apply count limit was hit on 44% of baseline moving frames; the 2 ms time limit was hit on under 1%. Stale and obsolete results were discarded before apply and were under 1% of tile events.
|
||
|
||
Admission had idle-worker/queued-result windows while far coverage was missing: 504/2,135 baseline and 446/2,109 lane samples. The 256-result cap did not cause those windows: backlog p95/max was 101/206, with zero cap declines in them. Core-region compilation was pending in 64%/62% of those windows.
|
||
|
||
RMC reported `Updated` after proxy publication, but many replacement components were still hidden by the coverage handoff. NullRHI runs cannot measure the first presented pixel; the trace's far coverage is explicitly not visual enclosure.
|
||
|
||
UE build passed with `-NoUBA -MaxParallelActions=4`; fast tier passed (269 gates, 5 diagnostics). Clang/GCC W1/W8/W12 matched L104 across four strata: 24/24 exact tile and whole-view hashes. Headless has no UE apply budget, so the lane has no headless effect. Logs are under `Saved/Linux/l121`; timing used CPUs 0–11 with no build running.
|