Skip to content

40b5093

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 18 Nov 03:31

fix(rtc): Measure a round trip longer than the gap between sender reports

A receiver report's LSR names the last sender report the peer received. Only the newest one
sent was remembered, and one goes out every 0.5 to 1.5 s, so once a round trip passed that gap,
as it does behind a standing queue, no report matched it and the round trip stayed at what it
was before the queue stood: 33 ms behind a 1.3 s queue. A key frame's time on its way
(key_on_its_way) was timed from that, so a NACK the queue's loss sent asked for another key
frame while the first was still crossing; the resync and frame-lost waits read it too.

The last eight sender reports are kept and a receiver report is matched against any of them.

On the WebRTC congestion step (tests/perf/congestion_step.spec.ts, a 250 KB bottleneck queue:
1.33 s at 1.5 Mbit/s), interleaved pairs, the time from the narrowing until the picture holds
under 150 ms behind for a second:

20 -> 1.5 Mbit/s held 12 s, 12 pairs newest only last eight
median 7.6 s 6.7 s
slowest 11.1 s 8.5 s
(repairs at the whole steered rate in both arms of 8 pairs, at half in the newer arm of 4;
nothing is retransmitted there with the newest report alone)

10 -> 4 Mbit/s held 5 s, 4 pairs newest only last eight
median 3.6 s 3.5 s
median picture age 310 ms 339 ms

At 10 -> 4 the round trip measured makes a peer's repeated NACKs count a frame lost only after
it, and the retransmissions they draw ran up to 4.2 Mbit/s on the 4 Mbit/s link while repairs
could take the whole steered rate; under the half they take now (the commit before) the two
match.

For #620.