mirror of
https://github.com/bilawalsidhu/gods-eye-view.git
synced 2026-11-17 08:42:21 +00:00
Three fixes to how the globe draws and reports loading. **The idle frame rate is the app's rule.** The Cesium frame port capped the loop at `IDLE_FRAME_RATE` whenever nothing held it, so the rule lived inside the renderer. The render governor now sets the rate: the display's own while the loop is held, `IDLE_FRAME_RATE` while not. The frame port gains `onHeldChange(cb)`, which says when its first hold is taken and its last released (animated visual styles included), so the governor hears every hold. The quality setting's frame-rate cap is kept apart from the governor's rate, and the loop draws at the lower of the two, so neither overwrites the other. **Line batching is timed.** A line set waited 30 drawn frames before its unchanged lines settled into a static batch (300 for the long wait), so the batch built twice as fast at 120 Hz as at 60 Hz, and took seconds at the idle rate. It now waits `SETTLE_MS` (500 ms) and `LONG_SETTLE_MS` (5 s), and moved shapes go back into their batches on the same timing. **KEY REQUIRED.** A layer that cannot load without a provider key ended the global status chip LOAD FAILED. The chip now has a KEY REQUIRED outcome, ranked below a real failure: a layer whose stats say `keyRequired` ends KEY REQUIRED, any other failure still ends LOAD FAILED. Vessels report a missing AISStream key as `keyRequired` and name the key in the panel; a later outage clears it, as it is no longer the key that is missing. Checked: the full suite (new tests for `onHeldChange`, the quality cap against the governor's rate, timed settling, and the chip with and without a key), format, boundaries, build, the stub renderer boot and the Cesium conformance run; in the browser, the idle and held frame rates; and the first-run check now asserts that with no keys set the chip never says LOAD FAILED.
157 lines
7.0 KiB
Markdown
157 lines
7.0 KiB
Markdown
# Performance baseline
|
|
|
|
This page records one hardware-rendered Apple M5 comparison captured on 22
|
|
August 2026 in Chrome 150 at 1440 x 900. It is not a minimum hardware
|
|
specification and should not be used to predict performance on untested systems.
|
|
The original capture artifacts are not included here, so this page records
|
|
results rather than defining a runnable benchmark.
|
|
|
|
## Test context
|
|
|
|
The baseline was captured on 22 August 2026 with these conditions:
|
|
|
|
| Setting | Value |
|
|
| --- | --- |
|
|
| Renderer | Apple M5 Metal through the hardware ANGLE path |
|
|
| Browser | Chrome 150 in a fresh isolated profile |
|
|
| Viewport | 1440 x 900 at device pixel ratio 1 |
|
|
| Focus | Page foregrounded for controlled scenes |
|
|
| Scene sample | 5 seconds of scripted motion, then 5 seconds at rest |
|
|
| Startup | Browser cache disabled; three samples |
|
|
|
|
The capture covered three startup samples, 16 cold layer scenarios with 14
|
|
measurements, 23 controlled option and stress scenes, and five
|
|
hardware-rendered overlay scenes.
|
|
|
|
## Startup
|
|
|
|
| Sample | App ready | Initial settle | Load event | Motion / rest | Used JS heap |
|
|
| ---: | ---: | ---: | ---: | ---: | ---: |
|
|
| 1 | 784.980 ms | 2,035.082 ms | 439.5 ms | 60 / 60 FPS | 102.9 MiB |
|
|
| 2 | 604.849 ms | 1,855.836 ms | 442.4 ms | 60 / 60 FPS | 111.6 MiB |
|
|
| 3 | 558.527 ms | 1,809.592 ms | 438.8 ms | 60 / 60 FPS | 105.1 MiB |
|
|
| Median | 604.849 ms | 1,855.836 ms | 439.5 ms | 60 / 60 FPS | 105.1 MiB |
|
|
|
|
These frame rates were measured when the globe was capped at 60 FPS. It now
|
|
draws at the display's refresh rate while anything moves and ticks at most 60
|
|
times a second when parked (the render governor's idle rate,
|
|
`IDLE_FRAME_RATE` in `src/renderGovernor.js`).
|
|
|
|
The initial-settle measurement is the more useful launch reference because it
|
|
includes the first visual and data settling window. All three samples reached
|
|
the display ceiling during both motion and rest.
|
|
|
|
## Cold layer activation
|
|
|
|
Cold activation was measured separately from warm option switching. Live object
|
|
counts are included so that future runs can compare source populations before
|
|
attributing a difference to the client.
|
|
|
|
| Layer | Activation | Source count | Motion / rest | Used JS heap |
|
|
| --- | ---: | ---: | ---: | ---: |
|
|
| CCTV city | 19,608.240 ms | 48 | 60 / 60 FPS | 192.7 MiB |
|
|
| Space Missions (report label: Rocket missions) | 3,581.066 ms | 26 | 60 / 60 FPS | 131.3 MiB |
|
|
| Radio | 3,458.709 ms | 750 | 60 / 60 FPS | 124.8 MiB |
|
|
| Bikeshare | 2,069.498 ms | 633 | 60 / 60 FPS | 157.4 MiB |
|
|
| Datacenters | 817.693 ms | 4,362 | 59.6 / 60 FPS | 328.2 MiB |
|
|
| Flights | 667.671 ms | 247 | 60 / 60 FPS | 118.4 MiB |
|
|
| Submarine cables | 614.727 ms | 2,629 | 60 / 60 FPS | 412.0 MiB |
|
|
| Military Flights | 557.113 ms | 68 | 60 / 60 FPS | 118.4 MiB |
|
|
|
|
CCTV had the largest cold activation cost in this capture. Submarine cables
|
|
used the most heap, followed by datacenters. Completed single-layer samples
|
|
generally reached 60 FPS, so activation time and heap separate these cases more
|
|
clearly than steady-state frame rate.
|
|
|
|
## Aircraft, detection, and Cockpit
|
|
|
|
| Scene | Motion / rest |
|
|
| --- | ---: |
|
|
| Idle globe | 60 / 60 FPS |
|
|
| Flights, 2D | 60 / 60 FPS |
|
|
| Flights, 3D proximity | 60 / 60 FPS |
|
|
| Flights, all 3D models | 60 / 60 FPS |
|
|
| Military Flights, all 3D models | 60 / 60 FPS |
|
|
| Detection at 25% | 39.3 / 41.1 FPS |
|
|
| Detection at 50% | 37.4 / 39.8 FPS |
|
|
| Detection at 100% | 34.4 / 35.5 FPS |
|
|
| Cockpit | 49.6 / 49.2 FPS |
|
|
|
|
The clean detection scenes processed 8,169 to 8,170 observations. Selected
|
|
labels rose from 14 at 25% density to 28 at 50% and 56 at 100%. The aircraft
|
|
rows came from an earlier loaded, foreground-controlled pass because the clean
|
|
rerun received no live aircraft rows.
|
|
|
|
## Visual styles and combined stress
|
|
|
|
| Scene | Motion / rest |
|
|
| --- | ---: |
|
|
| Normal | 60 / 60 FPS |
|
|
| CRT (report label: Retro) | 60 / 60 FPS |
|
|
| NVG (report label: Surveillance) | 60 / 60 FPS |
|
|
| FLIR (report label: Thermal) | 49 / 60 FPS |
|
|
| Anime | 60 / 59.8 FPS |
|
|
| Noir | 47 / 56.6 FPS |
|
|
| Snow | 42.3 / 45.8 FPS |
|
|
| Combined static | 57.6 / 60 FPS |
|
|
| Combined operational | 39.9 / 43.1 FPS |
|
|
|
|
The combined static scene rendered 11,575 objects, used 872.2 MiB of JavaScript
|
|
heap, and issued 48,665 text draws during motion and 54,106 at rest. The combined
|
|
operational sample contained 3,909 observations and two selected labels, but its
|
|
live aircraft and traffic rows were empty, so it remains a limited stress case.
|
|
|
|
Snow, Noir, dense detection, and text-heavy combined layers are the clearest
|
|
controlled comparison points for later optimization work. Snow has since been
|
|
retired as a visual preset; its row records the measurement as taken.
|
|
|
|
## Keyed live sources
|
|
|
|
NASA FIRMS, AISStream, and TomTom were captured in a separate hardware-rendered
|
|
pass. The page was visible but was not the focused window, so these frame rates
|
|
must not be compared directly with the foreground-controlled scenes above.
|
|
|
|
| Source | Point-in-time population | Activation or coverage | Motion / rest |
|
|
| --- | ---: | --- | ---: |
|
|
| NASA FIRMS | 100,430 detections in 3,557 cells | 30.0 s activation | 32.1 / 55.2 FPS |
|
|
| AISStream | 12,000 vessels | 6.4 s activation | 22.1 / 29.8 FPS |
|
|
| TomTom Traffic | 4,222 road dots | 70% coverage, 2 decoded tiles | 45.0 / 51.7 FPS |
|
|
|
|
These populations change continuously. A future comparison must record the
|
|
live counts again and match the focus conditions.
|
|
|
|
## Controls for a future capture
|
|
|
|
Use the same controls before attributing a difference to the application:
|
|
|
|
1. Record the exact GPU renderer and reject software-rendered or unavailable GPU
|
|
strings.
|
|
2. Use a 1440 x 900 viewport at device pixel ratio 1 and keep the page focused.
|
|
3. Measure cache-disabled startup separately from cold layer activation and warm
|
|
option switching.
|
|
4. Repeat startup three times and compare medians.
|
|
5. Sample each option for 5 seconds in scripted motion and 5 seconds at rest.
|
|
6. Record live object counts before attributing a difference to the client.
|
|
7. Treat a live-source outage as missing coverage, not as evidence of low client
|
|
rendering cost.
|
|
8. Record whether the app is running in Pinokio's embedded view or in an
|
|
external browser. One Pinokio 8.2.0 user reported severe stutter in the
|
|
embedded view that went away when the same local URL was opened in a normal
|
|
browser ([#365](https://github.com/bilawalsidhu/gods-eye-view/issues/365)).
|
|
Do not compare embedded-view and external-browser measurements directly;
|
|
compare the reported GPU renderer and matched scene settings first.
|
|
|
|
## What is not established yet
|
|
|
|
- This report does not establish Windows performance.
|
|
- The report does not record machine memory capacity, so it cannot support a
|
|
minimum-memory recommendation.
|
|
- The report does not cover other GPU renderers or viewport configurations.
|
|
- Military Installations is outside this comparison because it requires close
|
|
camera context.
|
|
- The keyed pass has no controlled rerun suitable for comparison with the
|
|
option scenes.
|
|
|
|
Use this page as a regression baseline for one known hardware and browser
|
|
configuration, not as a compatibility guarantee.
|