Re: Network layer reqt? [was Re: Transport level multihoming]

[Daniel Senie <dts@senie.com>]

Thomas Narten wrote:
> 
> > In my experience it is usually only severe turbulance which triggers
> > an ICMP back to an endpoint that causes sessions to die; multi-minute
> > packet loss due to loops in response to re-homing is not altogether
> > usual, I think.
> 
> note: typical ICMP errors generated as a result of route flaps/hiccups
> that are sent back to an originating end system do not/should not
> cause sessions to die. RFC 1122 makes this clear. I've appended an
> excerpt from 1122 on this topic.  Note also, that pretty much all TCPs
> seem to do the right thing now, though it took many many years.

In the many cases I've witnessed and in the cases I live with every day,
I see no evidence of ICMP messages causing the connections to die, but
rather TCP timers firing after several minutes of retries. There's been
some sort of significant flapping several times a day midway between my
office and the colo space I use. The result is all my SSH sessions stop.
Ones I haven't been using generally will survive the switch to a new
path. Sessions I've been typing in (and thus are busy trying to
retransmit) generally do not survive.

It'd sure be nice if the tables settled down a bit faster, of course.
Settling time seems to be less of an issue the closer to the edge the
failure occurs. Where the problems I've been seeing are 5 or 6 hops in
from the edge, the latency on withdrawals is likely the culprit. I tell
my SSH client to reconnect way too often...

-- 
-----------------------------------------------------------------
Daniel Senie                                        dts@senie.com
Amaranth Networks Inc.                    http://www.amaranth.com