To:
Brad Knowles <brad.knowles@skynet.be>
cc:
Bill Woodcock <woody@pch.net>, <dnsop@cafax.se>
From:
Bruce Campbell <bruce.campbell@ripe.net>
Date:
Tue, 5 Nov 2002 14:50:44 +0100 (CET)
In-Reply-To:
<a05200d00b9ed68000da8@[10.0.1.3]>
Sender:
owner-dnsop@cafax.se
Subject:
Re: quibbles about what is anycast.
On Tue, 5 Nov 2002, Brad Knowles wrote:
> At 2:35 AM -0800 2002/11/05, Bill Woodcock wrote:
>
> > Brad Knowles wrote faster than he thought:
> > > True anycast for protocols other than UDP? I sincerely doubt it.
> > > If you have evidence for this, I'd love to see it.
> >
> > Works fine for TCP and ICMP. What protocols do you have doubts about?
>
> I have doubts about TCP. Can you explain in more detail how it
> works for TCP, especially for the case where the route from IP
> address A to IP address B changes to a different machine that serves
> IP address B, while in the middle of a connection?
If the route changes in the middle of a connection, then the TCP stream is
reset as the new back-end server has no knowledge of the ongoing TCP
stream. This of course excludes smart backends which pass knowledge of
TCP state around and can continue with the end-user being none-the-wiser.
However, we're discussing TCP as used in DNS. That means either very
short TCP streams (simply because it was larger than UDP), or mildly long
(AXFRs).
Statistically (and I have pretty graphs to back this up), short-lived TCP
queries are not interrupted due to changes in the underlying routing
infrastructure.
Long-lived TCP streams are an altogether seperate matter, and are more
likely to be interrupted by such changes over a larger path, however this
probability is still low.
> And how often do
> route changes occur?
In terms of your average DNS TCP connection, hardly ever.
--
Bruce Campbell RIPE
Systems/Network Engineer NCC
www.ripe.net - PGP562C8B1B Operations/Security
#----------------------------------------------------------------------
# To unsubscribe, send a message to <dnsop-request@cafax.se>.