[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [MMUSIC] ICE-TCP Issue #1: Will S-O work on various operating systems?
Le Monday 25 February 2008 19:51:50 ext Jonathan Rosenberg, vous avez écrit :
> The approach described in ICE-tcp is equivalent to the "P2PNAT" solution
> described in that paper, without port prediction. Note that, this
> approach only works about 45% of the time according to Figure 8 in that
> paper (that Figure is the main result of the paper). Adding port
> prediction helps. STUNT#2 works better; that mechanism, however, assumes
> port prediction. Without port prediction, it has a bit better
> effectiveness than P2PNAT (around 50% according to a discussion I had
> with Saikat).
It would seem to me that all solutions are bad (compared to ICE-UDP). I am
keen to believe that TCP will always work better than UDP from NAT to
non-NAT, and UDP will always work better than TCP from NAT to NAT.
Maybe, we need to look back at the use-cases. As far as I understand there are
two classes of use-cases for ICE-TCP:
- fallback to ICE-UDP when there is an anti-UDP firewall (e.g. for RTP),
- as a bulk peer-to-peer transport (e.g. for file transfer).
The first scenario is a firewall rather than NAT issue. I don't know if we
want to venture there. Nevertheless, I expect that, if UDP is black-holed,
most certainly, so is TCP simultaneous open, i.e. only normal outgoing TCP
works. If people know of significant counter-examples, I'd like to hear about
it. Otherwise, specifying TCP-SO for this scenario might be a big waste of
time, especially considering the big implementation and deployment hurdles.
The second scenario is very different because:
1/ plain UDP cannot be used even if it works,
2/ being a *bulk* transport, relaying (TURN) may be very inadequate.
Under these assumptions, I would suspect that neither 25, 45 nor even 50%
success rate will "cut it". Lack of support from deployed operating systems
is a mere aggravating factor at this point. Then, if there is no satisfactory
TCP-based solution, some UDP encapsulation may be required, though I am not
very confident whether that will fly in IESG.
--
Rémi Denis-Courmont
_______________________________________________
mmusic mailing list
mmusic at ietf.org
http://www.ietf.org/mailman/listinfo/mmusic