Skip to content

QUIC settings for provisioned contribution paths: hold the rate through random loss, relax reorder thresholds, cap the probe backoff #5953

Closed

Description

@t0ms

Problem

A broadcast contribution feed is a constant-rate MPEG-TS stream on a provisioned path. It cannot slow down, and it has to survive outages, random loss and reordering. We measured a moq import ts → moq-relay → moq export ts --delay 2s lane against SRT carrying the same feed through the same emulated path: one host, netem, 100 ms RTT, a 10 Mb/s stream on a 20 Mb/s link, and SRT's latency matched to the lane's measured latency. (Method and data.) SRT loses less programme on every impairment except light reorder. Each gap traces to a QUIC default that suits a shared path, not a provisioned one. Items 1 and 2 were measured on the noq backend, before the move to moq-quic; item 3 on current main.

  1. Random loss is read as congestion. At 5 % random loss for 40 s the lane lost 30–34 s of programme under BBRv3 and 37 s under CUBIC. On the older quinn builds, whose BBRv1 ignores loss, it lost 0–0.6 s; SRT lost none. A loss-responsive sender at 5 % and 100 ms is held far below 10 Mb/s. max_outgoing_bytes_per_second is a ceiling, not a floor.
  2. Reordered packets are declared lost. Under 20 % reorder the relay declared 1,619 packets lost and the subscriber acknowledged every one of them afterwards. The controller still cut its window to a fourteenth, and the lane lost 37–38 s against SRT's 0–0.2 s. With the packet threshold at 1000 and the time threshold at 2 RTT, it lost 2.1–3.0 s; relaxing either alone did nothing. On the quinn backend the same relaxation did not help (27–33 s lost): the larger window overran the 20 Mb/s bottleneck, which then dropped real packets. moq-quic is quinn's fork, so the thresholds may only pay alongside a sender that holds its rate (item 1). That pairing is not measured. moq-quic's TransportConfig has both thresholds, but moq-tokio doesn't expose them.
  3. Recovery after an outage waits on the probe-timeout backoff. On main (ad87ad3), a 5 s outage cost 8.9 s (median of three). The relay's probes back off to 2.1 s apart, so in two of three replicates its next send came 3.6–3.8 s after the link returned, although a packet from the subscriber reaches it after 0.3 s. Capping the backoff at one doubling cut the loss to 4.9–6.1 s, against SRT's 5.0–5.4 s on the same rig. Never declaring persistent congestion changed nothing. RFC 9002 §6.2.1 says the backoff MUST double, so this would be opt-in.
  4. A dead peer is noticed only when the idle timeout expires. A relay or publisher restart cost 14.2 s, against SRT's 5.4–6.3 s. With --quic-idle-timeout 3s and --quic-keep-alive 1s it cost 4.4–5.1 s, but that timer also sets the longest outage a route rides through. A restarted relay that could stateless-reset its predecessor's connections (#5861, #5910) would remove its share without shortening the timer.

Proposal

Opt-in settings for provisioned paths, each with a test:

  • expose packet_threshold, time_threshold and persistent_congestion_threshold through moq-tokio's QUIC config;
  • add a cap on the probe-timeout backoff exponent to moq-quic;
  • add a congestion choice that holds a configured rate through random loss (paced at it, still capped);
  • possibly a named profile that sets these together, with the idle timeout and keep-alive.

Would you take these? And if so, as separate knobs or as one profile? A default-friendly alternative to the second item would be raising the thresholds adaptively once losses prove spurious (the way RACK widens its reorder window). Re-arming the probe timer when the peer is heard from again could stand in for the third. I'm happy to send pull requests in whichever shape you prefer.

(Written by Claude Opus 5.5)

Activity

  1. added
    questBeing tracked/planned in a quest. See `quest/`
    on Oct 19, 2026
  2. kixelated commented on Oct 19, 2026

    @kixelated
    Collaborator

    Decision: take the outage-recovery part only, as a default rather than a knob. The sender re-arms its probe timer when it hears from the peer again, which stays within RFC 9002 (no exponent cap).

    Declined: a rate-holding congestion controller, exposing the reorder thresholds, and a provisioned-path profile. Random loss is not a quality target here, and the reorder evidence came from random netem reorder; the thresholds can be revisited with a bufferbloat or real-path measurement. Item 4 is covered by #5861.

    Tracked by quest/m2/quic-outage-probe.md.

    Planned in #5969. Anyone who wants to donate tokens can run /quest-start quest/m2/quic-outage-probe to pick it up.

    (Written by Claude Opus 5.5)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    questBeing tracked/planned in a quest. See `quest/`

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions