On 21 May 2007, at 10:03, Lars Eggert wrote:
On 2007-5-17, at 20:53, ext Phelan, Tom wrote:
Aren't the default NAT timeouts for UDP and TCP different -- UDP
has to
be short because there's no other way to tell that a connection is no
longer needed -- TCP can be longer because what the timeout is
doing is
just cleaning up broken connections.
Yep. I was going to make the same comment to Ian (a NAT can't treat
DCCP like UDP *and* TCP, because the timeouts are usually very
different).
At any rate, DCCP would be more like TCP in this regard than UDP, I'd
expect.
That'd be my expectation as well. As a datapoint, do we know what
timer the Linux NAT code uses for DCCP? (And if we know of other
NATs that translate DCCP, do we know their timers?)
This sounds like an issue for a BEHAVE DCCP draft, no?
For this draft, I propose to leave the keep-alive interval at 15
seconds, to match ICE. In practice, I don't think this timer is
critical, since implementations should send periodic RTCP reports
(typically every 5 seconds on average), preventing these keep-alive
messages from being sent.
Colin
