[Nice] Trickle ICE

[Lorenzo Miniero]

Hi everybody,

I'm using libnice in an open source WebRTC gateway I've published recently,
called Janus (https://github.com/meetecho/janus-gateway) and it has worked
great so far!


In the latest few days I've started studying how to implement Trickle ICE,
though, and I'm a bit stuck. I guess you already know what Trickle ICE is,
but if you don't, it's basically a recent proposal (
http://tools.ietf.org/html/draft-ietf-mmusic-trickle-ice-01), especially
conceived (at least in the beginning) for WebRTC, to speed up the ICE
process. The idea is not to wait for all the candidates to be gathered to
send your SDP, but to first send a candidates-less SDP, and then send one
by one (trickle) each candidate as soon as you're able to gather it.

Ideally, this should be able to speed up the whole ICE process, as you may
find a suitable pair before your last candidate has been gathered. In the
worst case scenario, it's as slow as regular ICE.


The problem with this, though (and here comes why I'm posting here), is
that it's not backwards compatible, or at least not 100%, with regular ICE.
Specifically, as explained in Section 3 of the current specification (
http://tools.ietf.org/html/draft-ietf-mmusic-trickle-ice-01#section-3), the
main problem is that having just a few candidate pairs to check time after
time could easily empty the current list of candidates an agent is aware
of, thus leading to a failure in the ICE process. For instance, you get a
local candidate, send that one to your legacy-ICE peer using an out-of-band
channel, and while you're gathering one more candidate to send later on,
the validation and pairing of the candidate you've just sent has already
been failed by the peer: since the peer is only aware of that single
candidate, and there's no other candidate to try in the list, for what the
peer is concerned ICE just failed.

Since I'm trying to understand how I can make use of libnice to implement
trickle in my gateway nevertheless, do you think there's any trick one
could use for the purpose, without needing support for trickling to be
added to the library itself? I was thinking especially of some ways to
"delay" the failure of the pairing of some candidates, for instance, so
that to allow further candidates coming later to arrive before the whole
process for the agent fails. In general, I've noticed libnice does not fail
right away when no pair can be found, but that some seconds pass, probably
to take into account timeouts and the like, but since that may not be
sistematic (or is it?) I'm not sure that could be depended on for trickling
scenarios.


Do you have any suggestion on how this could be achieved? Are there plans
to include support for actual Trickle in the library in the future, by the
way?

Thanks!
Lorenzo
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.freedesktop.org/archives/nice/attachments/20140422/46675585/attachment.html>