[RTW] Review of draft-holmberg-rtcweb-ucreqs-00 (Web Real-Time Communication Use-cases and Requirements)

[Harald Alvestrand]

On 03/10/2011 07:05 PM, Ted Hardie wrote:
> On Thu, Mar 10, 2011 at 4:57 AM, Christer Holmberg
>
>>> A5: The web application MUST be able to control the media
>>> format (codec) to be used for the streams sent to a peer. I
>>> think the MUST is that the sender and recipient need to be
>>> able to find a common codec, if one exists; I'm not sure I
>>> see a MUST for the webapp actually picking one.
>> First, the sender and recipient of need to be able to perform
>> codec negotiation, in order to find the common codecs.
>>
>> If the codec negotiation is handled by the web application
>> (i.e. JavaScript based) the API must support this.
>>
>> If the codec negotiation is handled by the browser, then the app
>> might not need to not have as much control.
>>
>> We try to cover that in the note associated with A5.
>>
> So, I think we're all in agreement that rtc-web must specify a
> mechanism that allows for codec negotiation.  But I think we may need
> some more discussion on the expected mechanics.  The options include:
I numbered them....
> 1) Web app queries browser/host system via API for available codecs and
> sends selected codecs to rendezvous server, which runs the selection
> algorithm.  All peers acknowledge the selection.
>
> 2) Web app queries browser/host system via API for available codecs and
> sends selected codecs to peers, which answer the offer.  The original
> app acknowledges the answer, and things move on from there.
>
> 3) Web app requests browser/host system to select candidate codecs based
> on some set of characteristics; it then sends the selected codecs to
> rendezvous server, which runs the selection algorithm.  All peers
> acknowledge the selection.
>
> 4) Web app requests browser/host system to select candidate codecs based
> on some set of characteristics; it sends selected codecs to peers,
> which answer the offer.  The original app acknowledges the answer, and
> things move on from there.
Don't forget the option of "Web app tells the browser/host system to 
negotiate with another system about codecs, never passing the 
information about codecs to where the Web app sees it". This of course 
requires the negotiation protocol to be pretty browser-embedded.

The difference between 1+2) and 3+4) is that the Web app tells the 
system something about what it requires; there's some basic part of this 
going on in all reasonable cases, since the Web app is going to tell the 
system whether it wants audio or video codecs, but it can be extended 
with more parameters. For instance, for Opus, it's important to tell the 
system whether it's going to be used for music or for voice (in this 
case, the codec is the same, but when media starts flowing, the 
parameters are different).

With parameter-rich codecs such as H.264, the enumeration of all the 
possible parameter combinations that the hardware/OS/browser might 
support might be an unreasonable task, and it's not certain the 
necessary interfaces are even available.
> The two axes which vary in that set of choices are:  whether the web
> app makes the selection from among candidate codecs or the
> browser/host system makes the selection based on info provided;
> whether the negotiation takes place in the rendezvous server or in a
> peer-base offer/answer/acknowledgement set.   An obvious consequence
> of these choices is that the logic for condec selection moves around.
> An additional consequence of these choices will be what element in the
> system needs to know about the possibility of network-provided
> transcoding.
>
> I think some discussion of which negotiation method is expected would
> be useful.  If, for example, we rule out the negotiation server acting
> as the agent for negotiation, we can re-use the same protocol
> mechanics for offer-answer-acknowledgement, no matter whether the web
> app or browser/host system provides the codec selections.
I'm not sure the Web app can actually tell the difference between the 
rendezvous server acting as the agent and the negotiation being 
forwarded by the rendezvous server to a third party (the destination); 
in both cases, the Web app sends an offer and gets an answer back (in 
the O/A model).

If we can't detect it, it's hard to rule it out.