Re: [V3] Thoughts on scope for ript

[Justin Uberti <[e-mail]>]

On Wed, Mar 25, 2020 at 12:58 PM Eric Rescorla <[e-mail]> wrote:

>
> On Wed, Mar 25, 2020 at 12:52 PM Justin Uberti <juberti=
> [e-mail] <[e-mail]>> wrote:
>
>>
>>
>> On Wed, Mar 25, 2020 at 12:38 PM Stephan Wenger <[e-mail]> wrote:
>>
>>> Hi Justin, all:
>>>
>>>
>>>
>>> *From: *V3 <[e-mail]> on behalf of Justin Uberti <juberti=
>>> [e-mail]>
>>> *Date: *Wednesday, March 25, 2020 at 12:32
>>> *To: *Spencer Dawkins at IETF <[e-mail]>
>>> *Cc: *Victor Vasiliev <[e-mail]>, Jonathan Rosenberg <
>>> [e-mail]>, "[e-mail]" <[e-mail]>
>>> *Subject: *Re: [V3] Thoughts on scope for ript
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>> On Wed, Mar 25, 2020 at 12:26 PM Spencer Dawkins at IETF <
>>> [e-mail]> wrote:
>>>
>>> Replying to Justin, after reading Victor's reply - thank you both for
>>> replying ...
>>>
>>>
>>>
>>> On Wed, Mar 25, 2020 at 11:55 AM Justin Uberti <[e-mail]>
>>> wrote:
>>>
>>>
>>>
>>>
>>>
>>> On Wed, Mar 25, 2020 at 9:28 AM Spencer Dawkins at IETF <
>>> [e-mail]> wrote:
>>>
>>> This is interesting ...
>>>
>>>
>>>
>>> On Wed, Mar 25, 2020 at 10:31 AM Victor Vasiliev <vasilvv=
>>> [e-mail]> wrote:
>>>
>>> I would like to second Justin's point here about the value of a
>>> client-server RTP streaming by itself (e-f, though in-band c-d are also
>>> valuable).  There's a lot of value in the world where I can take an HTTP
>>> URL to some endpoint in the cloud and give it to, say, an IoT camera device
>>> to use as a realtime media sink, or pass it to an off-the-shelf streaming
>>> library to show a realtime media source on screen.
>>>
>>>
>>>
>>> I know most of my own concerns about RIPT media have started out with
>>> "this won't be as good as RTP for media", but what I've been getting from
>>> RIPT discussions is, that may not matter as much as I expected.
>>>
>>>
>>>
>>> For a use case like Victor's, great media quality would be great, but
>>> good enough media quality may be ... good enough.
>>>
>>>
>>>
>>> Am I putting words in the mouths of the proponents?
>>>
>>>
>>>
>>> Some applications may tolerate somewhat lower quality in fallback
>>> scenarios, but I think we should be aiming for RTP performance with HTTP
>>> ease of deployment.
>>>
>>>
>>>
>>> I was actually hoping that wasn't the case, because ISTM that the
>>> question whether RTP performance over HTTP is possible/likely is going to
>>> suck up a LOT of the discussion time during the BOF ...
>>>
>>>
>>>
>>> To Eric's point, "H3 is not widely deployed yet, and so we should think
>>> about how we want things to look". I think it's entirely possible to
>>> achieve RTP performance over H3, and hopefully also have a reasonable
>>> fallback when only H2 is available.
>>>
>>>
>>>
>>> [Stephan]: many video conferencing technologies (including Vidyo, Zoom),
>>> have previously, or are currently supporting http-based media transport,
>>> often for years or even decades.  They are almost never enabled by any
>>> professional deployer, because the resulting QoS is user-perceived as so
>>> bad that the marketing guys decided that offering no video/voice (and
>>> instead make people pick of the phone) is better for product reputation
>>> than offer the degraded QoS.
>>>
>>> Do any of you guys have hard numbers what sort of delay h3-based
>>> transport can achieve on today’s infrastructure?
>>>
>>
>> With H3 we have lots of options; there are lots of details to work
>> through here but given that the substrate is QUIC there is no theoretical
>> reason why RTP performance isn't possible.
>>
>> Regarding HTTP media transport in existing products (i.e., non-H3), you
>> are correct that there are indeed limitations and as such providers do what
>> they can to avoid this option, but it is nowhere near as grim as you make
>> it out to be. Recall that the first web version of Zoom exclusively used
>> websockets for transport
>> <https://webrtchacks.com/zoom-avoids-using-webrtc/>.
>>
>
> I mostly concur with Justin here: I would expect QUIC (which is to say
> probably H3 + Datagram) to be roughly comparable to RTP [0]. Anything over
> TCP will obviously be non-awesome, but we already do have some nontrivial
> fraction of media going over TCP to TURN relays, so it's not like it
> doesn't work at all, especially in good network conditions.
>
> -Ekr
>
> [0] There are some open issues with how QUIC congestion control and the
> media rate control interact, but I imagine they are mostly soluble.
>

Yes, the CC stuff is one of the details I was thinking of. The other main
one is the right abstractions in cloud providers to allow more direct
access to the underlying transport.