github selkies-project/selkies ab9e23a

pre-release14 minutes ago

fix(rtc): Leave a full path's loss out of the FlexFEC density

The repair density followed the smoothed loss of every feedback window,
including the windows a standing queue overflowed in. When a path
narrowed under the stream, its first windows lost a third of their
packets to the full queue, the density went to three repairs a group,
and those repairs went into the same full queue: congestion control cut
the video, and the repairs kept the path full in its place. The density
now reads only the windows no queue stood in, as the choice to repair
every group already did; loss on a path with room still raises it.

Measured with the congestion step (WebRTC, an 8 Mbit/s target on a
10 Mbit/s path narrowed to 4 Mbit/s for 5 s, the picture's age read off
the screen), four runs each, the same binary with and without:

                  before   after

median picture age 815 ms 408 ms
worst picture age 2.19 s 1.48 s
longest freeze 1.50 s 0.90 s
path queue full 3.4 s 2.2 s

Turning FlexFEC off altogether gave 320 ms, so the repairs a congested
path was sent were most of what it held. The steady phases are
unchanged, and a path narrowed to 1.5 Mbit/s is about the same either
way (its stream drops to a few frames a second for other reasons).

For #620.

Don't miss a new selkies release

NewReleases is sending notifications on new releases.