The climb gate scored encode against the negotiated refresh, so a 60 fps game was pinned at the ABR floor #255

Merged

enricobuehler merged 1 commits from worktree-cadence-degraded-floor-lock into main 2026-08-15 21:57:47 +00:00

Conversation 0 Commits 1 Files Changed 3

+142 -5

enricobuehler commented 2026-08-15 17:25:35 +00:00

Owner

Copy Link

Copy Source

A 2026-08-15 field session (2560×1440@120 negotiated, game delivering 53–74 fps) collapsed to the 5000 kbps ABR floor and was then held there for 23 minutes — 94% of the session — by our own climb gate: six bitrate climb refused — encode is behind cadence refusals with loss_ppm=0 throughout.

Root cause. The behind test scored every frame's post-submit work against the negotiated interval (8.33 ms at 120 Hz), but a frame's real budget runs until the next frame that actually exists arrives: a 60 fps source gives every frame twice that. An encoder keeping up with every real frame could be marked behind on most of them, latch behind_score past DEPTH_DEGRADE, and refuse every climb. The client's controller is blameless — its short-ack cap latch plus doubling re-probe ladder behaved exactly as designed against a host that held the flag for 23 minutes.

Fix. The budget is now the observed source-delivery period: an EMA over real frames' arrival spacing (repeats excluded — a keepalive re-encode says nothing about the game's rate; and after a repeat the next real frame lands a tiny delta, so an all-frames EMA would average back to ≈interval and defeat itself), clamped to [interval, 4×interval]. A source at or above the negotiated rate keeps bit-for-bit today's deadline; a hitchy source cannot disarm the detector. This also stops the same mis-scoring from spuriously escalating pipeline depth / pipelined retrieve on below-rate sources.

Observability — the session above sat at the floor with no trace of why:

• every cadence_degraded transition now logs behind_score, escalated, budget/interval/observed period — rate-limited to one line per direction per 5 s with suppressed flips counted (the score can oscillate ±1 around the latch threshold at frame rate);
• the control task's climb-refusal line carries the live behind_score (new shared AtomicU32), so the next field log discriminates "the budget was wrong" from "this encoder genuinely can't keep up" by itself.

Gates. amd64 CI container: cargo clippy --all-targets -p punktfunk-host -- -D warnings clean (non-vacuous — Checking punktfunk-host verified), native::stream tests 25/25 incl. a new budget test pinning the field case; cargo fmt --check clean.

A 2026-08-15 field session (2560×1440@120 negotiated, game delivering 53–74 fps) collapsed to the 5000 kbps ABR floor and was then held there for **23 minutes — 94% of the session** — by our own climb gate: six `bitrate climb refused — encode is behind cadence` refusals with `loss_ppm=0` throughout. **Root cause.** The behind test scored every frame's post-submit work against the negotiated interval (8.33 ms at 120 Hz), but a frame's real budget runs until the next frame that actually exists arrives: a 60 fps source gives every frame twice that. An encoder keeping up with every real frame could be marked behind on most of them, latch `behind_score` past `DEPTH_DEGRADE`, and refuse every climb. The client's controller is blameless — its short-ack cap latch plus doubling re-probe ladder behaved exactly as designed against a host that held the flag for 23 minutes. **Fix.** The budget is now the *observed* source-delivery period: an EMA over real frames' arrival spacing (repeats excluded — a keepalive re-encode says nothing about the game's rate; and after a repeat the next real frame lands a tiny delta, so an all-frames EMA would average back to ≈interval and defeat itself), clamped to `[interval, 4×interval]`. A source at or above the negotiated rate keeps bit-for-bit today's deadline; a hitchy source cannot disarm the detector. This also stops the same mis-scoring from spuriously escalating pipeline depth / pipelined retrieve on below-rate sources. **Observability** — the session above sat at the floor with no trace of why: - every `cadence_degraded` transition now logs `behind_score`, `escalated`, budget/interval/observed period — rate-limited to one line per direction per 5 s with suppressed flips counted (the score can oscillate ±1 around the latch threshold at frame rate); - the control task's climb-refusal line carries the live `behind_score` (new shared `AtomicU32`), so the next field log discriminates "the budget was wrong" from "this encoder genuinely can't keep up" by itself. **Gates.** amd64 CI container: `cargo clippy --all-targets -p punktfunk-host -- -D warnings` clean (non-vacuous — `Checking punktfunk-host` verified), `native::stream` tests 25/25 incl. a new budget test pinning the field case; `cargo fmt --check` clean.

enricobuehler added 1 commit 2026-08-15 17:25:37 +00:00

fix(host): the behind-cadence deadline was the negotiated refresh, so a 60fps game pinned ABR at the floor ... 33d0b77e07

A 2026-08-15 field session (2560x1440@120 negotiated, game delivering
53-74 fps) collapsed to the 5000 kbps ABR floor and was then held there
for 23 minutes - 94% of the session - by our own climb gate: six
'bitrate climb refused - encode is behind cadence' refusals with
loss_ppm=0 throughout. The behind test scored every frame's encode work
against the negotiated interval (8.33 ms at 120 Hz), but a frame's real
budget is the arrival of the next frame that actually exists: a 60 fps
source gives every frame twice that. An encoder keeping up with every
real frame could be marked behind on most of them, latch behind_score
past DEPTH_DEGRADE, and refuse every climb the client asked for.

The budget is now the OBSERVED source-delivery period: an EMA over real
frames' arrival spacing (repeats excluded - a keepalive re-encode says
nothing about the game's rate), clamped to [interval, 4x interval] so a
source at or above the negotiated rate keeps bit-for-bit today's
deadline and a hitchy source cannot disarm the detector. This also
stops the same mis-scoring from spuriously escalating pipeline depth /
pipelined retrieve on below-rate sources.

And the gate becomes observable - the session above sat at the floor
with NO trace of why:
- every cadence_degraded transition logs behind_score, escalated,
  budget/interval/observed-period, rate-limited to one line per 5 s
  with suppressed flips counted (the score can oscillate +-1 around
  the latch threshold at frame rate);
- the control task's climb-refusal line now carries the live
  behind_score (new shared AtomicU32), so a field log can finally
  discriminate 'the budget was wrong' from 'this encoder genuinely
  cannot keep up'.

Gated: clippy --all-targets -D warnings + native::stream tests (25,
incl. a new budget test pinning the field case) in the amd64 CI
container; fmt clean.

enricobuehler merged commit c64a1a6767 into main 2026-08-15 21:57:47 +00:00

enricobuehler deleted branch worktree-cadence-degraded-floor-lock 2026-08-15 21:57:49 +00:00

Sign in to join this conversation.

Reviewers

No Reviewers

Labels

No items

No labels

Milestone

No items

No Milestone

Projects

Clear projects

No projects

Assignees

Clear assignees

FernandoMX FlipperDan HitFrostbite enricobuehler (Enrico Bühler) kszlim luxus nannku neoney sassycrown

No Assignees

1 Participants

Notifications

Subscribe

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: unom/punktfunk#255