Re: [MMUSIC] Seriously, USE-CANDIDATE!!! in 5245bis

[Emil Ivov <[e-mail]>]

On 23.07.14, 18:21, Ari Keränen wrote:
> On 7/23/14 6:00 PM, Emil Ivov wrote:
>>
>>
>> On 23.07.14, 12:30, Ari Keränen wrote:
>>> On 7/22/14 7:12 PM, Emil Ivov wrote:
>>>>
>>>> On 22 Jul 2014 6:11 PM, "Ari Keränen" <[e-mail]
>>>> <mailto:[e-mail]>> wrote:
>>>>  >
>>>>  > On 7/22/14 11:06 AM, Emil Ivov wrote:
>>>>  >>
>>>>  >> On 22.07.14, 11:02, Ari Keränen wrote:
>>>>  >>>
>>>>  >>> Hi,
>>>>  >>>
>>>>  >>> On 7/21/14 4:45 PM, Emil Ivov wrote:
>>>>  >>>>
>>>>  >>>> In London we mentioned the option of adding a new attribute that,
>>>> during
>>>>  >>>> aggressive nomination, the controlling agent can use to notify
>>>> its
>>>> peer
>>>>  >>>> that ICE processing is over and no more changes are going to be
>>>>  >>>> attempted. This way controlled agents will no they can now
>>>> release
>>>>  >>>> resources.
>>>>  >>>>
>>>>  >>>> This could either be a new attribute (e.g.,
>>>> USE-CANDIDATE-CONFIRMED) or,
>>>>  >>>> as Justin suggested in an off-list discussion, we could have the
>>>>  >>>> controlling agent send a second USE-CANDIDATE once aggressive
>>>> nomination
>>>>  >>>> converges.
>>>>  >>>
>>>>  >>>
>>>>  >>> Did you consider using the updated offer to indicate that you're
>>>> really
>>>>  >>> done? This would require that you always send it (i.e., a slight
>>>>  >>> modification to 5245), but that's something that has been already
>>>>  >>> proposed to be updated for other purposes.
>>>>  >>
>>>>  >>
>>>>  >> The point is to have all that information without involving
>>>> signalling,
>>>>  >> as is the case with regular nomination.
>>>>  >>
>>>>  >> Let me ask the question the other way. Is there any reason not
>>>> to do
>>>> this?
>>>>  >
>>>>  >
>>>>  > Just the usual: are the benefits higher than the added complexity.
>>>> Here, depending how it's done, the added complexity should be fairly
>>>> low
>>>> though.
>>>>
>>>> The complexity of having to handle that over signalling and the need to
>>>> implement it on every sig protocol that ever dares to use ICE is
>>>> arguably much higher than that of an extra attribute
>>>
>>> Actually I meant the complexity of implementing something like this, and
>>> why to do/not to do this (in any way).
>>
>> Yes, so did I. We are talking about the "complexity" of adding one ICE
>> attribute as opposed to intertwining the states of ICE and signalling
>> and making them dependent on each other.
>>
>>>>  > For the benefits, currently with SIP you are allowed to release the
>>>> not selected candidate resources after 3 seconds from concluding ICE
>>>> (http://tools.ietf.org/html/rfc5245#section-8.3.1).
>>>>
>>>> And that is the whole point because currently with aggressive
>>>> nomination
>>>> you simply don't know exactly when ICE has concluded.
>>>
>>> So if you assume aggressive ICE has concluded after you've got your
>>> first USE-CANDIDATE checks done (as the text currently says), and you
>>> stop doing keep-alives on all except the selected candidate pair and let
>>> your relayed and reflexive candidates die away. You still have some time
>>> to switch between the candidates since the candidates are still alive
>>> for at least some (tens of?) seconds. Isn't that OK?
>>
>> If I assume ICE has concluded after my *first* USE-CANDIDATE and let
>> everything die away then, with aggressive nomination, how am I ever
>> going to get a second?
>
> It takes a while before your server reflexive and relayed candidates die
> out so you have some time to get more checks through. Of course you
> shouldn't close the local sockets immediately. However, if you want to
> do that considerably (> tens of second) later, and perhaps multiple
> times, it may get more complicated (something like mobility with ICE).

These are exactly the kinds of hacks that implementations try to do 
today. Magic timers and proprietary logic. This is exactly what I was 
hoping we could remove by replicating the same feature in aggressive 
that we already have for regular nomination. As Justin put it:  giving a 
way to the controlling agent to communicate it's selection to the 
controlled.

It simply doesn't make sense not to have it and it would simplify things 
a lot so the argument "it would add complexity" doesn't really hold.

Emil

-- 
https://jitsi.org