Fr0zkaandClaude Opus 5.5
2ef0ca72b2
Policy 3 movement traces (12 s move + 5 s dwell, 3 runs each). Unreal 30 m/s: near LOD0-2 ready while moving 31.6 -> 44.2% (LOD0 15.4 -> 34.7%), world tick p99/max 23.1/28.3 -> 11.3/13.6 ms, dwell completion 3.55 -> 3.22 s; far holes 2.5 -> 4.3%. Walking unchanged (~99%). Stale work is small (1.9% of moving LOD0 tiles finish out of range); the remaining limit is sustained generation throughput (8 task slots, p95 worker 105 ms). Near:far stays 4:1. World unchanged (pinned hashes pass). Fast tier PASS; Linux UE builds. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
32 lines
3.8 KiB
Markdown
32 lines
3.8 KiB
Markdown
# Linux108 movement results
|
||
|
||
Baseline is `ef22b03`. Policy 3; three runs per speed and platform; 12 s moving plus 5 s dwell. Percentages and dwell times are the median with the three-run range. UE tick is `commandlet_world_tick_frame.world_tick_ms`; headless records sampled queue-update work, not frame ticks. Tick/update values are medians of the per-run p99 and max. Columns show baseline → current.
|
||
|
||
| Speed / platform | Near LOD0–2 | Walkable ring | Far holes | Tick or update p99 / max (ms) | Dwell complete (s) |
|
||
|---|---:|---:|---:|---:|---:|
|
||
| 1.4 m/s / headless | 89.40 [89.23–89.43] → 89.57 [89.43–89.85] | 97.92 → 97.92 | 13.72 [13.59–14.11] → 13.46 [12.98–13.69] | 4.37 / 4.37 → 4.44 / 4.44 | 0.00 → 0.00 |
|
||
| 1.4 m/s / UE | 99.27 [99.21–99.28] → 99.31 [99.28–99.43] | 100 → 100 | 0.62 [0.56–0.85] → 1.26 [1.18–1.36] | 11.48 / 26.39 → 6.25 / 14.28 | 0.07 [0.04–0.18] → 0.02 [0.02–0.24] |
|
||
| 30 m/s / headless | 80.64 [80.60–81.10] → 82.61 [82.41–82.77] | 95.14 → 96.53 [96.30–96.53] | 6.55 [6.35–6.64] → 14.42 [14.14–14.66] | 4.22 / 4.22 → 4.35 / 4.35 | 0.25 → 0.25 |
|
||
| 30 m/s / UE | 31.55 [29.13–31.83] → 44.24 [43.85–45.70] | 100 → 100 | 2.52 [2.14–3.15] → 4.25 [4.03–4.66] | 23.11 / 28.29 → 11.29 / 13.60 | 3.55 [3.44–3.61] → 3.22 [3.18–3.29] |
|
||
|
||
UE per-LOD readiness, baseline → current (median [range]):
|
||
|
||
| Speed | LOD0 | LOD1 | LOD2 |
|
||
|---|---:|---:|---:|
|
||
| 1.4 m/s | 99.02 [99.00–99.06] → 99.11 [99.06–99.14] | 99.79 [99.77–99.82] → 99.84 [99.82–99.84] | 98.90 [98.83–99.04] → 98.92 [98.90–99.38] |
|
||
| 30 m/s | 15.42 [13.16–15.49] → 34.68 [34.50–35.94] | 25.62 [23.29–26.44] → 42.06 [41.44–43.91] | 59.07 [57.33–61.11] → 59.60 [59.20–60.88] |
|
||
|
||
## What the traces show
|
||
|
||
At 30 m/s, the scheduler changes desired center 45 times per run, about once per 8 m tile crossing. 88.4% of moving LOD0 admissions are ahead of the heading; median lead is 13.2 m (p95 27.6 m). Of 4,835 generated LOD0 tiles admitted during movement, 94 (1.94%) finish outside the desired range. 29 (0.60%) are cancelled before worker completion. Stale work is not the main readiness loss.
|
||
|
||
LOD0 readiness starts at 83–85% in the first two seconds and falls to 14–16% in the final two. About 512–516 LOD0 tiles remain desired; the trace shows a mean of about 328 missing. The worker task cap is 8, with p95 worker time 105 ms and p95 request-to-apply time 155 ms. This points to sustained generation/backlog throughput as the remaining fast-travel limit. The 2 ms apply budget is rarely hit (4–7 frames per run); tile-apply max is under 2.5 ms. `UWorld::Tick` max is 14.7 ms; the custom streaming-work tick max is 13.7 ms. The commandlet loop includes a deliberate ~16 ms sleep and is excluded from these tick values.
|
||
|
||
Fast near readiness improved by 12.7 points, but remains 44.2% in UE versus 82.6% headless. Near-first service also costs far coverage: 30 m/s far holes rose from 2.52% to 4.25% in UE and 6.55% to 14.42% headless. A headless 2:1 near/far trial measured 80.1% near and 14.8% far holes, so the final scheduler retains 4:1. Walk/fast hysteresis remains 75/35 and 1200/800 cm/s.
|
||
|
||
## Checks
|
||
|
||
The Unreal build succeeded with `-NoUBA -MaxParallelActions=4` before timing. Unreal timings used `taskset` only; headless used `taskset` plus the 6 GiB `prlimit`. The final headless build stamp is `f3ec37a389bb6c984bac0fa360b96e49b6b53fbda92c6461c1246b690e21afaf`, matching the fast-tier manifest. Fast tier passed: 274 checks, 80 runs; pinned L100 hashes and W1/W8/W12 comparisons passed. `git diff --check` is clean. No passage-support or mesher geometry semantics changed.
|
||
|
||
Artifacts: `Saved/Linux/l108/movement/{baseline,after}/summary.json`, `after/tile-timelines.csv`, Unreal traces, and `Saved/Linux/l108/verify-fast/fast-summary.json`.
|