Periodic ~26 ms hitch every 10 s: vk_ingestTick / sweep_total does one 12 ms batch #534

New issue

Closed

opened 2026-10-08 04:50:10 +00:00 by sky_line · 0 comments

sky_line commented

2026-10-08 04:50:10 +00:00

Owner

Observed (Baltimare, busy region, 2026-10-08, VKHitch log): a ~26 ms frame exactly every 10 s (04:46:51, 47:01, 47:11, 47:21...). The rest of the time frames are a steady ~14.2 ms (70 fps).

Each of those hitch lines shows the same breakdown: vk_ingestTick=13 ms, sweep_total=12 ms, display_total=22 ms. The steady-state average for sweep_total is about 1 ms.

Likely cause: a timer-driven pass of the registration/ingest sweep (selfRegisterSweep or something periodic beside it) does a whole batch in a single frame instead of amortising it.

Done when:

the work is spread over frames, or capped per frame by a time budget;
there is no 10 s periodic spike in the VKHitch log in the same scene;
steady-state cost doesn't regress.

**Observed** (Baltimare, busy region, 2026-10-08, VKHitch log): a ~26 ms frame exactly every 10 s (04:46:51, 47:01, 47:11, 47:21...). The rest of the time frames are a steady ~14.2 ms (70 fps). Each of those hitch lines shows the same breakdown: `vk_ingestTick=13 ms`, `sweep_total=12 ms`, `display_total=22 ms`. The steady-state average for `sweep_total` is about 1 ms. **Likely cause:** a timer-driven pass of the registration/ingest sweep (selfRegisterSweep or something periodic beside it) does a whole batch in a single frame instead of amortising it. **Done when:** - the work is spread over frames, or capped per frame by a time budget; - there is no 10 s periodic spike in the VKHitch log in the same scene; - steady-state cost doesn't regress.

sky_line added the

labels

2026-10-08 04:50:10 +00:00

sky_line added this to the Slipstream project

2026-10-08 04:50:10 +00:00

sky_line added