Skip to content

Chromium stress run records

Recorded evidence; these results apply to the named runs and artifacts.

Run verdicts

Run Verdict Evidence boundary
Nov 1 short suites, all three OSes/both backends Pass Twelve reconnects, eight resizes, bounded pressure/soak
Linux eight-hour, four configurations Fail X11 Alloy cgroup growth 486.3 MiB exceeded 256 MiB
Linux ten-minute diagnostic Pass Does not explain old eight-hour growth
Instrumented Linux eight-hour rerun Pending Six-hour observations are progress, not acceptance
Nov 2 macOS Chrome, requested 900 s Fail 127 throughput windows below 70% baseline
Nov 2 macOS Alloy Fail Browser-ready timeout; soak never ran

Short-suite limits

macOS/Windows sampled 122–124 seconds and did not exhaust six-viewer encoder pools. Linux sampled 94–95 seconds and returned explicit codec-unavailable responses at the fourth delayed viewer. It ran beside endurance fixtures: not idle-GPU throughput. Linux median server RSS growth was zero, child RSS −1.00–3.17 MiB, descriptors 0–1. Across reported runs, RSS/child growth was 0–0.68 / −0.20–24.62 MiB, descriptors 0–3. Sixteen Rust video tests and both macOS native suites passed.

Eight-hour failure and follow-up

All four configurations reached 960 samples; last sample 28,770 s. X11 Alloy peaked at 4.47 GiB, with server RSS +2.93 MiB and no median GPU/descriptor growth. All cleanup records exited 0 without cgroup leaks. The failure prevented its final decode and GPU-reclamation assertion.

New sampling adds memory.stat/events, smaps_rollup, and component storage. shmem belongs within file, not an additive charge. Initial diagnostics found ~870 MiB file cache and ~875 MiB downloaded components per browser. The ten-minute X11 run had 120 samples, total-charge changes −0.6/−2.1 MiB, 16 KiB file growth, decoded video, and clean teardown.

At six hours, X11 Chrome grew ~1,432 MiB (file ~1,387; kernel 44; anonymous −0.6). All streamed ~14.6 FPS with no recorded OOM. At 18,210 s all profiles reported a 2,730 MiB OptGuideOnDeviceModel after earlier cgroup peaks. This is correlation, not proof of the old failure's cause. Limits/security updates remain unchanged. Final decode and cleanup are attempted even on limit failure.

Longer macOS follow-up

Chrome: 904 s/171 samples, baseline 49.98 FPS, minimum 5.80, longest gap 1.47 s; RSS −0.14 MiB, children +42.41 MiB, descriptors unchanged; final 60 frames changed. Alloy timed out before ready. Both fixtures cleaned up; later desktop-load observations do not establish either root cause.

Acceptance rules

Baseline ≥20 FPS; healthy viewers ≥70% baseline and ≥10 FPS, gaps <2 s. Growth limits: server 128 MiB, summed children 256 MiB, descriptors/handles 32. Summed RSS is not GPU accounting. Actual exhaustion requires an observed error. Resource/decoder failures remain failures; three units preserve those limits.

Scheduling correction

Shared renditions run first; private renditions run least-recently-encoded first with a 16.667 ms budget. Deferred work retains references/keyframe requests. Windows resize checks use actual native DPI. Samplers preserve process identity and fail missing live-process metrics rather than counting zero.

Run isolated pressure tests separately from endurance on the same GPU. Windows needs an interactive principal and psutil; all drivers need FFmpeg. These are local-capability streaming checks, not remote login or native input acceptance.

Recorded measurements

Results — 2026-11-01

Platform/backend Final suite Baseline fps Viewer-pressure fps (up to six)
macOS Chrome Pass 50.0 48.5
macOS Alloy Pass 49.9 51.3
Windows Chrome Pass 48.9 50.2
Windows Alloy Pass 48.6 51.3
Linux Wayland Chrome Pass 38.2 38.7
Linux Wayland Alloy Pass 35.0 38.9
Linux X11 Chrome Pass 35.6 39.0
Linux X11 Alloy Pass 34.2 39.4

Running the suites

Check Assertion
Baseline Hardware video at least 20 fps; 60 sampled frames decode into changing images
Mixed codecs and slow viewer H.264 rendition at most 360 pixels high, throttled/delayed peer decodes, healthy peer stays responsive
Stalled viewer Viewer stops reading entirely, healthy peer continues, original connection resumes and decodes
Bounded encoder pressure Up to six delayed viewers get video or an explicit error; healthy peer retains throughput; a fresh rendition decodes after release
Churn Twelve fresh connections alternate codec requests, decode fresh frames, then disconnect
Resize Eight alternating CSS sizes replace encoder generations while respecting measured capture DPI; final frames decode
Bounded soak Document identity and video remain live; server RSS, descendant RSS and file descriptors/handles are sampled; final frames decode
Cleanup No surviving owned descendants; normal server exit on Unix, explicitly forced process-tree shutdown on Windows

Commands and examples

Running the suites

uv run tools/test_chromium_platform_stress.py \
  --platform macos --server /path/to/gft --host /path/to/host.app/Contents/MacOS/probe \
  --shell /bin/sh --artifacts .cache/platform-stress/macos

uv run tools/test_chromium_platform_stress.py \
  --platform linux --ozone wayland \
  --server /path/to/gft --host /path/to/graffitty_chromium \
  --shell /bin/sh --artifacts .cache/platform-stress/wayland

uv run tools/test_chromium_platform_stress.py \
  --platform windows --ssh windows-test-host \
  --server C:/test/gft.exe --host C:/test/cef/graffitty_chromium.exe \
  --shell 'C:/Program Files/Git/bin/bash.exe' --python C:/test/python.exe \
  --remote-dir C:/test/stress --principal-task ExistingInteractiveTestTask \
  --artifacts .cache/platform-stress/windows

Evidence files

  • .cache/mac-alloy-resize/stress-scheduler-final
  • .cache/chromium-acceptance/linux-endurance-8h
  • .cache/chromium-acceptance/linux-memory-6h-snapshot
  • .cache/linux-memory-followup
  • .cache/chromium-acceptance/linux-memory-*
  • .cache/chromium-platform-stress
  • .cache/chromium-acceptance/linux-portable-current-stress-{wayland,x11}
  • .cache/chromium-acceptance/mac-stress-15m-per-backend