Kaleido/docs/performance-2026-10-27.md
Elijah 0c62d7414e
All checks were successful
Validate and publish Kaleido / android (push) Successful in 2m24s
Release Kaleido 0.3.0 with categories, alphabet jumps and audio preferences
2026-10-27 20:33:20 -07:00

37 lines
3.3 KiB
Markdown

# Navigation performance comparison - 2026-10-27
The onn shows measurable navigation hitches in all three releases. This comparison does not establish a broad regression from the 0.2.1 bugfixes: optimized horizontal browse p95 was 32 ms in both 0.1.11 and 0.2.1, and 34 ms in 0.3.0. Some first-entry transitions are slower in the newer releases, while warm results overlap. No individual change or author can be identified as the cause from these measurements.
## Method
- Same onn 4K Plus Streaming at 192.168.50.195:5555, Android 14, armeabi-v7a; existing 1920x1080 display override and density 320 retained. Launcher, accessibility services and system settings were unchanged.
- Original signed 0.1.11/code 12 and 0.2.1/code 14 APKs were extracted from their tags and checksum-verified. The final signed 0.3.0/code 15 APK was tested. In-place installs retained the encrypted session and app preferences. Theme music and boost were Off.
- Each fresh install began in ART status=verify. Ran the horizontal browse harness, installed the bundled baseline profile, forced speed-profile compilation, confirmed ART status=speed-profile, then ran three cycles of all 14 Home/Movies transitions and another horizontal browse check. Thermal status was 0 throughout the version-start snapshots; corresponding reported sensor temperatures were essentially unchanged.
- Commands: scripts/repro_runtime.py browse and scripts/profile_transitions.py <label> --cycles 3. No competing device/UI tests ran. Raw framestats, compilation states and thermal snapshots remain in ignored evidence/private/.
## Results
| Release | Fresh-install browse median / p95 | Profiled browse median / p95 | Missed intervals across 42 transitions |
| --- | --- | --- | --- |
| 0.1.11 | 27 / 48 ms | 14 / 32 ms | 47 |
| 0.2.1 | 30 / 61 ms | 16 / 32 ms | 50 |
| 0.3.0 | 30 / 53 ms | 16 / 34 ms | 54 |
Fresh-install browse exceeded the 40 ms harness threshold in every release. All profiled browse runs passed, with stable horizontal Y geometry and zero gfxinfo Missed Vsync in that specific horizontal sample. That does not mean every frame met the 16.7 ms budget or every navigation transition was smooth.
Missed intervals below are inferred display stalls from adjacent framestats, not a count of dropped video frames. Each array is cycles 1, 2 and 3; the first cycle is the first pass after launch.
| Transition | 0.1.11 | 0.2.1 | 0.3.0 |
| --- | --- | --- | --- |
| home: hero -> row | 1 / 0 / 0 | 2 / 0 / 0 | 1 / 0 / 0 |
| home: hero <- row | 0 / 0 / 0 | 1 / 1 / 0 | 1 / 2 / 1 |
| tabs: home -> movies | 7 / 6 / 7 | 10 / 6 / 5 | 8 / 6 / 5 |
| tabs: home <- movies | 6 / 4 / 3 | 6 / 3 / 3 | 4 / 6 / 4 |
| movies: row -> row | 2 / 1 / 0 | 2 / 1 / 1 | 1 / 2 / 1 |
| movies: row <- row | 2 / 2 / 2 | 2 / 2 / 2 | 2 / 2 / 2 |
First Home hero-to-row p95 was 28.4 / 42.0 / 43.0 ms respectively; later cycles were 26.5-27.6 / 25.9 / 23.9-27.5 ms. First-pass work is a visible source of occasional stutter. Repeated tab switches also missed display intervals in every version.
This is a small on-device comparison, not a statistical attribution study. It cannot reconstruct the original compilation state when the user noticed the issue. Fresh-install optimization has a large effect here, but it is not proven to explain that earlier observation. The final 0.3.0 release was reinstalled, baseline-profile compiled and returned to Home afterward.