[R-C] SCTP as an RTCWEB option (Re: Modular Congestion Control in rtcweb)

[Stefan Holmer]

On Tue, Apr 10, 2012 at 1:43 PM, Stefan Holmer <stefan at webrtc.org> wrote:

>
>
> On Sun, Apr 8, 2012 at 7:51 PM, Michael Welzl <michawe at ifi.uio.no> wrote:
>
>> Hi,
>>
>> I changed the subject as we're really talking about SCTP, not DCCP here
>> now. More below:
>>
>> [snip]
>>
>>
>>  Yes, this means an optional different ACKing behavior, with some
>>>> congestion control logic on the receiver side, but only for the case of
>>>> (partially) unreliable transfers. We're blowing up SCTP more and more
>>>> here... still, this strikes me as a less blown up than having SCTP +
>>>> DCCP, and possibly (note, only possibly) more efficient than having
>>>> SCTP-for-data + some_new_cc-over-RTP/UDP
>>>>
>>>
>> Note that the above paragraph by me captures a view that has meanwhile
>> evolved: I now think that there should probably be only one single
>> congestion control instance for everything (and then it wouldn't matter
>> much whether it's on the sender or receiver side). With that in mind, I'll
>> answer below:
>>
>>
>>  It is an interesting idea which had not really been considered before,
>>> and would obviously solve the congestion issue between rtcweb data channels
>>> and media channels.
>>>
>>> There is an efficiency argument about RTP/SCTP/DTLS/(ICE)UDP compared to
>>> plain SRTP/UDP (keyed with DTLS-SRTP or SDES, a whole other religious war
>>> over in rtcweb).  It's not a huge impact on video, especially at higher
>>> bandwidths, but could be significant for audio-only calls.
>>>
>>> In addition, some currently non-critical issues with SCTP would become
>>> more critical, such as an issue flagged about large reliable datagrams
>>> hogging the packet queues; and there might also be startup speed issues.
>>>
>>
>> I don't see why (a potentially adapted) SCTP couldn't be made to always
>> fragment packets beyond a certain size to avoid that hogging problem. And I
>> don't understand how that would become more critical?
>> As for startup speed, I think that this is not true (au contraire! no
>> need for slow starting into a network that's already used by another
>> stream) if we have only one congestion control instance.
>>
>>
>> On Apr 8, 2012, at 9:26 AM, Timothy B. Terriberry wrote:
>>
>>  Keep in mind that one of the arguments against the shim approach to
>>> multiple RTP sessions on a single port was that it would break all existing
>>> traffic management software at gateways, firewalls, etc., that tries to
>>> inspect RTP/SRTP packets. The same would be true of RTP/SCTP/DTLS/UDP,
>>> except moreso (as you would no longer even be able to get at the fields
>>> that are normally unencrypted in an SRTP session, regardless of your
>>> ability to upgrade software on the gateways and firewalls).
>>>
>>> To me this is at least as big a disadvantage as the efficiency argument,
>>> though adding that many more bytes of overhead to on the order of 100
>>> packets a second is already a pretty big disadvantage.
>>>
>>
>> I don't know much about and have no strong opinion about that first
>> argument, but regarding the overhead, note that there is significant
>> overhead in the other variant too: you have SCTP data traffic causing
>> receiver-side ACKs that convey pretty much the same information as RTCP
>> feedback, but both streams mutually ignore each other...
>>
>
> Not sure which thread these comments fit in, so I chose this one. It may
> be worth noting that:
>
> a. Delayed ACKs will make a send-side algorithm slower at reacting to
> increased delay upon congestion. This may be possible to get around by
> having some mechanism at the receive-side as well for detecting congestion
> and thereby force the ACK to be sent.
>
> b. Lost ACKs may introduce uncertainty for the send-side algorithm, which
> will never be a problem for a recieve-side algorithm. You can reduce the
> uncertainty by having something like an ACK sequence number in each ACK
> message, but you have still lost the delay information contained in the
> lost ACK. Maybe something like this already exists in
>

Also, how much extra overhead are we talking about when ACKing every packet
compared to having periodic REMB packets? I think there are a lot of
use-cases which won't have any significant amount of RTCWEB data streams,
and in those situations we only get the extra overhead from ACKing every
packet, as there are no significant data streams to share ACKs with.


>
>
>> Cheers,
>> Michael
>>
>> ______________________________**_________________
>> Rtp-congestion mailing list
>> Rtp-congestion at alvestrand.no
>> http://www.alvestrand.no/**mailman/listinfo/rtp-**congestion<http://www.alvestrand.no/mailman/listinfo/rtp-congestion>
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.alvestrand.no/pipermail/rtp-congestion/attachments/20120410/3d61f829/attachment-0001.html>