* Introduce `DISCONNECTED` connection quality.
Currently, this state happens when any up stream track does not
send any packets in an analysis window when it is expected to send
packets.
This can be used by participants to know the quality of a potentially
disconnected participant. Previously, it took 20 - 30 seconds for
the stale timeout to kick in and disconnect the limbo participant which
triggered a participant update through which other participants knew
about it.
Previously, `POOR` quality was also overloaded to denote that the
up stream is not sending any packets. With this change, that is a
separate indicator, i. e. `DISCONNECTED`.
* clean up
* Update deps
* spelling
Using time from outside make anachronous samples in expected
distance/bit rate measurement. So, have to let the time be
snap shotted in scorer lock scope.
Need to pass in the correct time. Previously streaming start was
determined by another delta snap shot which as removed for efficiency.
Did not realise that we were passing in zero time for stats.
Also, revert of the change (the part which did not re-pause) from this
PR (https://github.com/livekit/livekit/pull/2037). That change affects
other paths. The edge it was trying to fix is more rare. Need to think
about a way which covers all cases.
* Split RTPStats into receiver and sender.
For receiver, short types are input and need to calculate extended type.
For sender (subscriber), it can operate only in extended type.
This makes the subscriber side a little simpler and should make it more
efficient as it can do simple comparisons in extended type space.
There was also an issue with subscriber using shorter type and
calculating extended type. When subscriber starts after the publisher
has already rolled over in sequence number OR timestamp, when
subsequent publisher side sender reports are used to adjust subscriber
time stamps, they were out of whack. Using extended type on subscriber
does not face that.
* fix test
* extended types from sequencer
* log
Profiling showed updating jitter going through the snapshot maps.
With the reduction of one, there should only be one snapshot
and hopefully that should gain some cycles back.
* Push track quality to poor on a bandwidth constrained pause.
* add tests
* scale distance by divisor
* fix test distance to desired
* wait longer for subscription manager to reconcile
* Prevenet anachronous sample reading.
Not so pretty way of solving this. Please let me know if you have
thoughts.
Passing in time allows testing easier. But, that also leads to
time reversal problems. Example scenario
1. Connection stats worker gets a time and initiates quality
calculation.
2. A layer transition is recorded after that.
3. By the time, scorer is called to calculate score with time from Step
1, there is time reversal and results in anachronous sample.
One option is to use a scorer lock in connection stats module and wrap
all calls to scorer in that lock, but that does not prevent the passed
in time stamps themselves getting out of order. Also, stand alond use
of scorer in some other context will be problematic.
Doing the hybrid thing of taking current time in scorer if passed in
time is zero so that scorer lock domain controls it.
* use zero time everywhere in normal flow
* make APIs with and without time passed in as Paul suggested
* Use receiver report stats for loss/rtt/jitter.
Reversing a bit of https://github.com/livekit/livekit/pull/1664.
That PR did two snapshots (one based on what SFU is sending
and one based on combination of what SFU is sending reconciled with
stats reported from client via RTCP Receiver Report). That PR
reported SFU only view to analytics. But, that view does not have
information about loss seen by client in the downstream.
Also, that does not have RTT/jitter information. The rationale behind
using SFU only view is that SFU should report what it sends irrespective
of client is receiving or not. But, that view did not have proper
loss/RTT/jitter.
So, switch back to reporting SFU + receiver report reconciled view.
The down side is that when receiver reports are not receiver,
packets sent/bytes sent will not be reported to analytics.
An option is to report SFU only view if there are no receiver reports.
But, it becomes complex because of the offset. Receiver report would
acknowledge certain range whereas SFU only view could be different
because of propagation delay. To simplify, just using the reconciled
view to report to analytics. Using the available view will require
a bunch more work to produce accurate data.
(NOTE: all this started due to a bug where RTCP was not restarted on
a track resume which killed receiver reports and we went on this path
to distinguish between publisher stopping vs RTCP receiver report not
happening)
One optimisation to here here concerns the check to see if publisher is sending data.
Using a full DeltaInfo for that is an overkill. Can do a lighter weight
for that later.
* return available streams
* fix test
1. Completely removing RTT and jitter from score calculation.
Need to do more work there.
a. Jitter is slow moving (RFC 3550 formula is designed that way).
But, we still get high values at times. Ideally, that should
penalise the score, but due to jitter buffer, effect may not be
too bad.
b. Need to smooth RTT. It is based on receiver report and if one
sample causes a high number, score could be penalised
(this was being used in down track direction only). One option
is to smooth it like the jitter formula above and try using it.
But, for now, disabling that also.
2. When receiving lesser number of packets (for example DTX), reduce the
weight of packet loss with a quadratic relationship to packet loss
ratio. Previously using a square root and it was potentially
weighting it too high. For example, if only 5 packets were received
due to DTX instead of 50, we were still giving 30% weight
(sqrt(0.1)). Now, it gets 1% weight. So, if one of those 5 packets
were lost (20% packet loss ratio), it still does not get much weight
as the number of packets is low.,
3. Slightly slower decrease in score (in EWMA)
4. When using RED, increase packet loss weight thresholds to be able to
take more loss before penalizing score.
data.
Without the check, it was getting tripped by publisher not publishing
any data. Both conditions returned nil, but in one case, the receiver
report should have been received, but no movement in number of packets.
* Run quality scorer when there are no streams.
In the down stream direction, receiver report is used for scoring.
If there are no receiver reports, it should go to `dry` state and report
poor quality.
Update scorer on dry condition only when update score has not happened
for longer than some multiple of update interval. Cannot update on every
interval when there are no streams as receiver report might be just
missed. Waiting for longer to ensure that report is definitely not
received.
* update last stats time
This fixes the case of screen share forwarding. We should probably also
look at proper AddTrack. The problem was that
- AddTrack used two layers for screen share from JS sample app
- Track was published with rid = f. Given that and the track info,
consistent layer mapping set the layer as 1.
- `getBufferLocked` always uses the highest layer for SVC
- Between the two, when down track was requesting PLI, there was
no buffer at the requested layer and hence no PLI went out.
A few other notes
- Tried locking SVC to layer 0 (instead of layer 2), but that resulted
in PLI layer lock spamming. It did not happen in v1.3.0 of the server
though. Not sure what causes that. Need to investigate later.
But, that does not happen when using layer 2 buffer as SVC buffer.
- When using layer 2 for SVC, the PLI throttle config will be using that
of layer 2. Is that okay?
- `buffer` structure should maintain more stats about spatial layers for
SVC case so that layer stats can be reported to analytics/scoring etc.
- In general, `buffer` may need some more hooks to make it SVC aware so
that it can handle various spetial layer aware/specific bits.
There are cases where the layer bit rate configuration is such that
the expected bitrate difference is very high. For example,
setting up layer 2 (f) layer for 1.7 Mbps and layer 1 (h) for 180 kbps.
With bitrate based quality, a layer drop results in going to `POOR`
quality rating. With layer based, it will drop one level only.
Also, cleaning up the distance to desired calculation a bit.
With push model (i. e. connection quality evaluation triggered
by reception of RTCP receiver report), it is possible that a report
is received quickly after a track is started (especially with video).
Those should not trigger a quality evaluation.
Set `lastStatsAt` in `Start` routine and ensure that start has been
called and enough time has passed since last stats time to avoid
small windows.
* Expected vs actual Layer based connection quality.
With VBR streams (like screen share), bit rate is not a good indicator
of whether desired layer (spatial/temporal) is achieved due to high
variance.
Using expected vs actual layer (i. e. distance to desired) can capture
any short fall and include it in quality scoring.
This PR uses distance to desired, i. e. how many steps it would take to
go from actual spatial/temporal -> desired spatial/temporal and that
distance is propotionally used (currently it is just linear) to decrease
score.
* wire up layer transitions for screen share tracks
* Use EWMA (Exponentially Weighted Moving Average) for score updates.
Makes code simpler, but makes it harder to test as the inflection points
are not exact.
Score falls a bit slower to be conservative on dropping quality too
quickly. Still fall factor is higher (i. e. newer scores get more
weight) than rise factor (i. e. newer scores get lower weight).
Slower rise factor to introduce hysteresis on things climibing back too
quickly.
In the extreme case, asympttotic conditions could cause unexpected
results. For example, having 4% loss of video continously will never
drop quality to `POOR`. It will get close to 60, but it will always
stay above 60 forever and hence quality will never drop to POOR.
Maybe, need some sort of variable thresholding to deal with that. But,
that is an extreme case and may not happen in real life.
* remove unused stuff
* Push/pull for connection stats/quality scoring.
Was not happy with pure pull method missing a window because
of RTCP RR timing is slightly off for audio and using a much
larger window of data in the next update.
That also resulted in RTP stats getting some bits of code.
As that is per-packet processing, was not a good idea.
Switching to push-pull method.
For up track, it is pull, i. e. connection stats worker will pull stats.
For down track, there is a new notification about receiver report
reception. Using this to check for time to run stats. And adding a bit
of tolerance for processing window (currently set so that as long as it
is > 95% of usual processing interval). This allows two things
- for video, RTCP RR are more frequent, but we will still not process
till enough time has passed
- for audio, RTCP RR could be once in 5 seconds or so. Can process when
it is available rather than miss a window and use a much larger window
later.
* uber atomic
* Connectino quality misc changes
1. Call scorer.Update() with nil stat when no data available so that
scorer can synthesise window with proper window time.
2. Substract out loss in interval to account for packets not sent at
all.
3. Fix `packetsNotFound` variable in `getIntervalStats`. I remember this
working at some point. Not sure if I fat fingered in another PR and
deleted the increment line.
4. Logging a bit more when no packets expected. Those can get noisy
especially when track is muted. But, seeing some unexplained
instances of no packets leading to quality drop. So, temporary logging
to get a bit more information.
* correct spelling
* Limit packet score minimum to 0.0
* Make connection quality not too optimistic.
With score normalization, the quality indicator showed good
under conditions which should have normally showed some badness.
So, a few things in this PR
- Do not normalize scores
- Pick the weakest link as the representative score (moving away from
averaging)
- For down track direction, when reporting delta stats, take the number
of packets sent actually. If there are holes in the feed (upstream
packet loss), down tracks should not be penalised for that loss.
State of things in connection quality feature
- Audio uses rtcscore-go (with a change to accommodate RED codec). This
follows the E-model.
- Camera uses rtcscore-go. No change here. NOTE: THe rtscore here is
purely based on bits per pixel per frame (bpf). This has the following
existing issues (no change, these were already there)
o Does not take packet loss, jitter, rtt into account
o Expected frame rate is not available. So, measured frame rate is
used as expected frame rate also. If expected frame rate were available,
the score could be reduced for lower frame rates.
- Screen share tracks: No change. This uses the very old simple loss
based thresholding for scoring. As the bit rate varies a lot based on
content and rtcscore video algorithm used for camera relies on
bits per pixel per frame, this could produce a very low value
(large width/height encoded in a small number of bits because of static content)
and hence a low score. So, the old loss based thresholding is used.
* clean up
* update rtcscore pointer
* fix tests
* log lines reformat
* WIP commit
* WIP commit
* update mute of receiver
* WIP commit
* WIP commit
* start adding tests
* take min score if quality matches
* start adding bytes based scoring
* clean up
* more clean up
* Use Fuse
* log quality drop
* clean up debug log
* - Use number of windows for wait to make things simpler
- track no layer expected case
- always update transition
- always call updateScore
* WIP commit
* Connection quality changes
- Fix Firefox showing poor quality
o The issue was that we were using max available layer and
calculating quality. The rationale being that even if
server sends dynacast messages, client may not implement
dynacast and still stream all layers. But, with Firefox
(maybe a Firefox bug), it sends some small amount of
data on layer 2 even when that layer is disabled.
Guessing it is probing (or actually we might be using
some small value for high layers as Firefox cannot turn off
layers). That higher layer gets used in quality calculation.
As the bit rate on that layer is extremely low, it yields low
score.
Fixed by considering the max expected layer. That is of most
interest. Yes, clients may ignore dynacast and stream all layers,
but, max expected is the one of interest. So, look for
quality in the max expected layer and not max available layer.
- Lots of clean up around connection quality stuff
o Use a dynamic scaling thing to ensure that we do not get bitten
by absolute values. Calculate best possible scenario score and
map that to maximum MOS score. This will ensure that different
codecs, different settings do not mess up the scoring. For example,
a client might use 1 Mbps for 720p, but a different client could
use 2 Mbps for 720p. As an SFU/infrastructure middlebox, we do
not have control over quality at those rates. We can only ensure
that streaming happens smoothly at those rates. So, in that
example, for client 1, 1 Mbps will map to MOS 5.0 and for client 2,
2 Mbps will map to MOS 5.0. Any impairments after that will
reflect in the score.
o Penalise for missing target layer by one level for one layer missed.
o Move tests to connection quality directory. The participant test
was not super useful.
* Add missed file
* Remove debug code
* use more constants and initialise normalisation factor
* rtcscore pointer
* Use media payload size in scoring.
Subtract out header bytes when calculating score.
This does not seem to affect the score (under perfect conditions),
but, using header bytes will inflate the bit rate and
will affect scoring.
* Add header bytes to ToProto
* protocol pointer
* fix test
With a small window, the quality is volatile even on small disturbances.
For example losing 2 audio packets in a 2 second window could
drop the quality metric.
Raja Subramanian