Re: [MMUSIC] Merging ICE aggressive and regular nomination (was Re: [tram] Comment on draft-williams-peer-redirect-01: might it not converge?)

[Justin Uberti <[e-mail]>]

On Wed, Jul 30, 2014 at 3:20 PM, Jonathan Lennox <[e-mail]> wrote:

>
>  On Jul 30, 2014, at 5:59 PM, Justin Uberti <[e-mail]> wrote:
>
>   On Wed, Jul 30, 2014 at 2:14 PM, Jonathan Lennox <[e-mail]>
> wrote:
>
>>  Thus the suggestion I’m making — do regular nomination, but allow media
>> to be sent on any valid pair prior to selection.  (Yes, re-reading 5245, I
>> should have said “Valid” rather than “Confirmed”.  I mis-remembered the
>> terminology.)
>>
>
>  Understood. But controlled side needs some way to know what pair it
> should use to send media, so we need some way to indicate this. Would this
> be done via USE-CANDIDATE? If so, we could end up with multiple
> USE-CANDIDATEs as we figure out the ideal path to use, which doesn't fit
> exactly with regular nomination as defined in 5245, but otherwise seems
> exactly like what I am proposing. Basically, controlled side uses the pair
> that it most recently received USE-CANDIDATE on as its selected pair.
>
>
>  Oh, that’s true, it has to pick something.
>
>  In the semantic I’m suggestion, I think it doesn’t *matter* what it
> picks — it could choose any valid pair, just like the controlling agent
> does, since it knows that’s a working round-trip path.
>
>  For both sides, if you receive media on any candidate (from an IP and
> port from which you’ve received a valid connectivity check), accept it.
>  USE-CANDIDATE is then just an optimization that lets you close down the
> unselected candidates.
>
>  I don’t think you want a rule of “most recently received USE-CANDIDATE”
> to determine the selected pair.  Checks will race each other, especially
> when the paths’ RTTs are very different.
>
>  In my model, there is only ever one USE-CANDIDATE sent per component —
> i.e., it works like pure regular nomination, except for the “early media”.
>

OK, I get it now. I think the main issues you face are:
1) Inability to selectively free candidates - all candidates except the
selected pair can be freed by the controlled side after USE-CANDIDATE.
2) No guarantee of symmetric RTP - remote side can use a different pair
than local side - and you can't resolve this without USE-CANDIDATE (see #1).
3) Potential issues when media arrives on non-nominated pair.

Given #3, I think I'd prefer to find a solution that also addresses #1/#2.

>
>       5245 says that an ICE endpoint MUST be prepared to receive media on
>>> any candidate, but only sends on the Selected candidate pair.  I don’t know
>>> if this means, in practice, that ICE-bis could change to the behavior I’ve
>>> described above and still be backward compatible, or if there’d need to be
>>> some sort of ice-option negotiation to indicate that this early media (oh,
>>> dread phrase) is acceptable.
>>>
>>
>>  Yes, I think we need a new ICE option, or else old ICE impls in the
>> controlled role will be surprised when they think the selected pair is A
>> but media arrives on B. We've already seen a number of implementations have
>> issues with this.
>>
>>
>>  Do they have issues receiving media if there is no selected pair yet?
>>  This would be nonconformant with 5245, but nevertheless wouldn’t surprise
>> me.
>>
>
>  Since the controlled side sends the ClientHello in the typical case,
> this isn't usually a problem. The issue that is most often seen is that the
> ServerHello comes from B instead of A and is ignored.
>
>
> These are all in cases when the controlling endpoint is using aggressive
> nomination, though, right?  Do you have any data for regular nomination,
> i.e. media received before any USE-CANDIDATE attribute?
>

Not at this time.