erik/acdream
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

docs(perf): commit physical Slice A baselines

Browse source 
Capture the authoritative capped, uncapped, and dense-Arwic routes on the physical AMD display path. Preserve the portal allocation, exception-storm, and repeated-Caul memory failures as MP1c's before-state while unblocking the prepared-asset cutover.
This commit is contained in:
Erik 2026-07-24 14:41:50 +02:00
parent 49f3a48ee6
commit 2421d09f5d
6 changed files with 335 additions and 19 deletions
Download patch file Download diff file Expand all files Collapse all files

89
docs/research/2026-07-24-slice-a-physical-local-baselines.md Normal file
View file

@ -0,0 +1,89 @@
# Slice A physical-local baselines
**Date:** 2026-07-24
**Authority:** performance-gate baseline
**Source and measured binary:** `49f3a48ee63b82d00731b0f91497930925279c86`
The corrected Slice A tooling was exercised on the authoritative physical AMD
display path defined by
[`2026-07-24-performance-reference-configuration.md`](2026-07-24-performance-reference-configuration.md):
RX 9070 XT driver `32.0.31021.5001`, 2560 x 1440 at 240 Hz, with the client
rendering 1280 x 720 Release gameplay without developer UI. The tracked source
tree was clean at every launch, source and binary commits matched exactly, all
route actions completed, and all three clients disconnected gracefully.
The full sanitized metrics and every portal window are in
[`2026-07-24-slice-a-physical-local-baselines.json`](2026-07-24-slice-a-physical-local-baselines.json).
Raw CSV, EventPipe, screenshots, logs, and process reports remain local
untracked diagnostic artifacts.
## Headline measurements
| Capture | Frames | CPU p50 / p95 / p99 | GPU p50 / p95 / p99 | Frame-thread allocation p50 / max |
|---|---:|---:|---:|---:|
| 240 Hz capped nine-stop | 107,081 | 4.287 / 6.972 / 9.404 ms | 1.702 / 3.261 / 3.847 ms | 28.1 KiB / 203.5 MiB |
| Uncapped nine-stop | 232,974 | 1.520 / 5.694 / 7.825 ms | 0.897 / 3.239 / 3.406 ms | 27.3 KiB / 202.9 MiB |
| Uncapped dense Arwic | 15,785 | 3.672 / 6.561 / 14.121 ms | 2.476 / 3.284 / 3.752 ms | 40.7 KiB / 28.5 MiB |
| Dense Arwic, stable 5 s | 1,044 | 4.239 / 6.139 / 6.760 ms | 2.712 / 3.235 / 3.263 ms | 37.5 KiB / 13.0 MiB |
The physical path confirms that ordinary steady rendering is fast. The stable
dense-Arwic window contained 15,530 world entities and 111 animated entities
while remaining below 6.8 ms CPU p99 and 3.3 ms GPU p99. The 32 ms RDP envelope
was therefore a remote-presentation artifact rather than local renderer cost.
Portal materialization remains the dominant transient. Sawato allocated up to
203.5 MiB in one capped frame and 202.9 MiB in one uncapped frame. The
uncapped route allocated 11.25 GB on the frame thread in 501 seconds, while the
process-wide runtime counter averaged 50.0 MB/s and recorded 58 Gen-2
collections. `ArgumentOutOfRangeException` was also emitted 4,659 times during
that route. This is the invalid typed-Setup probe storm already in Slice C's
scope, not an accepted steady-state mechanism.
## Repeated-portal memory result
Both nine-stop sessions completed behaviorally but deliberately failed the
existing memory plateau assertion:
- The capped final Caul checkpoint was 685.6 MiB higher in working set and
699.8 MiB higher in private memory than the first Caul checkpoint.
- The uncapped final Caul checkpoint was 295.1 MiB higher in working set.
- In the uncapped comparison, world entities returned from 21,031 to 21,025,
animated entities from 18 to 13, and owned composite textures from 163 to
120. Managed live bytes changed only 271.95 MB to 274.69 MB, while committed
managed memory rose from 438.01 MB to 867.14 MB.
- The capped comparison likewise ended with fewer mesh records, atlas arrays,
and composite textures than the first Caul checkpoint, but committed managed
memory rose from 449.30 MB to 1,235.05 MB.
The evidence does not point to retained world ownership. It points to very high
allocation churn, LOH fragmentation, and GC heap segment reservation after
portal decode. The baseline is intentionally recorded as a failure instead of
loosening the assertion; prepared-content Slice C must remove the decode/probe
traffic and then repeat this exact gate.
## Contention and ownership interpretation
The physical capped route recorded 7,292 runtime contentions. The committed
stack-attributed RDP capture remains the attribution oracle because the same
code and route ranked `DatDatabaseWrapper.TryGet` first (75.3%), followed by
`ObjectMeshManager.ProcessQueue` (16.3%) and
`DecodedTextureCache.GetOrCreate` (4.2%). Slice C removes demand from those
paths before any lock redesign is considered.
Canonical ownership counters and graceful teardown passed on every run. No
pending mesh uploads, landblock retirements, live-entity teardowns, or
composite warmups remained at the final checkpoints.
## Gate state
Slice A's physical-local exit criterion is complete: a committed 240 Hz capped
nine-stop baseline, uncapped nine-stop baseline, pinned uncapped dense-Arwic
baseline, and per-portal frame/allocation windows now exist on the reference
display path. This unblocks Slice C.
It does **not** mark the measured memory behavior healthy. The plateau failures,
~200 MiB portal frames, ~50 MB/s process allocation, exception storm, and
heap-reservation growth are the before-state that Slice C must materially
improve without changing pixels or gameplay.