initially.
Out-of-order reception is detected and dropped before this change. By
moving the adjustment to that point, potentially more reports can be
processed.
My previous PR was trying to fix a premature transition from CONGESTED -> NONE state.
But, that introduced a bug which accelerated the transition from EARLY_WARNING -> CONGESTED state due to holding the `queuingRegion` state variable.
Fix by not holding `queuingRegion` in state and take result of
processing runs.
NOTE: This can still use some work to make the state machine cleaner.
Initially, I was passing around a bunch of variables which was uncouth.
Then moved things to state which made code easier to read and also log
different things, but it causes problems like the ones above. Will take
another look and think more about making it more robust.
* Do not seed if stream is already writable.
It is possible that in migration case, when the forwarder state is
fetched from migrating out node and used to seed downtrack, it has
already started due to the time it takes to get the state. Seeding in
that state will reset things and cause large sequence number gaps
potentially.
* do not take lock
* Debug FF TWCC
* - TURN off TWCC for Firefox. Seems to fail with VP9 send, i.e. there are
no TWCC feedback packets when sending VP9.
- Relax thresholds for congestion as staging data is showing
oscillations.
- Clean up some logging.
* debug log a few more signal messages
* revert config
* revert config
* clean up
Seeing some negative loss count (seems to be when interacting with
Firefox, but don't have definitive proof it does not happen with
others). Debug logging a bit to understand what could cause it.
Also, consolidating some common code to process packet feedback.
* move unrolled mime type check for broader use
* Use in IsSvcCodec and make MimeType exported
* test
* tidy branches
* tidy
---------
Co-authored-by: Paul Wells <paulwe@gmail.com>
* Remove alloc in packet forwarding path.
Unlikely logger creation was doing allocs. Replace it with a function
like in rtpstats_receiver.go so that allocations do not happen
unnecessarily.
* variable rename
* one more place
* Better naming/logging for send side bwe
* Check BWE congestion state before doing optimal allocation.
It is possible that BWE declares congestion, but the estimated bandwidth
may still be enough to accommodate all tracks. So, stream allocator
would still not in DEFICIENT state. On a new track allocation, it will
get optimal allocation although BWE is in congested state.
Take BWE congestion state into consideration before doing any track
allocation.
* get congestion state from BWE rather than caching it in stream allocator
* get states from update
* log contributing groups only when they are small in number
* get oldest group always
* notify only when estimate is dropping
* require stronger CTR down trend
* maintain min/max group
* relax thresholds to declare congestion a bit
* Thottle the publisher data channel sending when subscriber is slow
Avoid the publisher overwhelm the sfu data channel buffer when
the subscriber has lower receive bitrates. It will drop message
if the subscriber is considered too slow to block the entire room.
* Enable nack in mediaengine and disable it in transceiver as need
pion doesn't support per transciver codec configuration, so the nack of this session will be disabled
forever once it is first disabled by a transceiver.
https://github.com/pion/webrtc/pull/2972
There are cases where the probe result has enough headroom to up
allocate all deficient tracks. Mainly happens after a loss scenario
where the estimate is actually still high.
After boosting once, there was a check for the track to hit the desired
layer before boosting again. But, that is not really necessary. Can
boost target and forwarder should resolve to the latest target.
Removing that check in the forwarder.
Also, adding a gratuitous boost check in stream allocator periodic ping
when deficient to look for opportunities to boost.
* Don't wait rtp packet to fire track
Create track from sdp instead of first rtp packet,
it is consistent with the browser behavior and
will accelerate the track publication.
* fix test
Raja Subramanian