Re: [rmcat] Hackathon: RMCAT (and CCFB) in Jitsi (fwd)

[Sergio Garcia Murillo <[e-mail]>]

On 01/11/2019 21:32, Jonathan Lennox wrote:
>
>
>> On Oct 22, 2019, at 10:37 AM,Sergio Garcia Murillo 
>> <[e-mail] 
>> <mailto:[e-mail]>> wrote:
>>
>> Hi Jonathan,
>>
>> My C++ interface is slightly different, although I have it 
>> implemented for google's transport wide feedback message:
>> https://github.com/medooze/media-server/blob/master/include/SendSideBandwidthEstimation.h
>>
>> Main difference is that I have split your packet arrival processing 
>> in two calls, one for notifying when one rtp packet is sent and one 
>> to notify when a feedback packet is received with a map of the 
>> received stats. If you prefer to have an individual call per 
>> feedbacked rtp packet, I think it would be important to notify when 
>> the last packet of the feedback has been received as for example I do 
>> the final bwe calculations there.
>
> I had a single call in hopes of using the same algorithm interface to 
> implement REMB (receiver-side) feedback with the abs-send-time 
> extension; in that case you only have information at the time of 
> packet receipt, not packet transmission, and each piece of “feedback” 
> information comes in separately.


I tried that, but mixing REMB and transport wide cc feedback on same API 
didn't worked for me.


>
> I agree that some ability to tell the algorithm which feedback came in 
> at once would be useful.
>
>> I miss also to be able to control the congestion control part of the 
>> SFU, in my implementation which is based on BBR, I rely heavily on 
>> continuously sending RTX as probing bitrate as described on Varum's 
>> FRACTAL algorithm.
>
> How tightly coupled do you think probing needs to be with the 
> algorithm?  I’ve generally considered probing to be a separable piece 
> — the prober’s responsibility is to consume the output of the bwe 
> algorithm, and to generate traffic when appropriate (i.e. when the 
> overall application has some larger discrete quantum of bandwidth that 
> it wants to be able to start sending).

Most probably is a requirement given my design choices. As I said 
before, my implementation is a delay based algorithm inspired on BBR and 
I always send RTX as probing bitrate when not sending the upper layer 
available on the SFU. So when I detect a increase on the delay when the 
bloat bitrate is hit, I update the bwe estimation, but set the probing 
bitrate to a proportionate amount lower than the bloat bitrate to 
decrease the delay back to best case scenario


[...]


>> Last topic is troubleshooting, I have created a tool for viewing bwe 
>> stats dumps from my SFU:
>> https://github.com/medooze/bwe-stats-viewer
>>
>> If we can agree on a common format for the logs, we could enhance it 
>> and use it for troubleshooting and comparing behaviors.
>
> That does look useful.  Do you have documentation for the log format 
> you’re using?


I have updated the readme with that file 
(https://github.com/medooze/bwe-stats-viewer/blob/master/README.md). 
Basically it is a csv file with a per packet stats (so should be quite 
easy to add with your API

Best regards

Sergio