Re: [dnsext] Re: I-D ACTION:draft-vandergaast-edns-client-ip-00.txt

[Colm MacCárthaigh <[e-mail]>]

On Fri, Jan 29, 2010 at 9:05 AM, Martin Barry <[e-mail]> wrote:
> "Anycast TCP is also in use by a few CDNs and works at least as well as any
> DNS-layer solution. Anycast is stable for minutes or hours at a time, so
> it's rare for two TCP packets to the same destination to reach different
> anycast contributors."

This statement is not supported by data, and does not seem credible.
While anycast is stable, there is no reason to believe that it works
at least as well as a DNS-layer solution.  At my previous employer
(Joost.com) we ran a small private CDN for P2P seeding, video and
image and distribution, using both anycast and DNS tricks. DNS tricks
proved to be at least twice as good at lowering network latencies,
though granted we were not particularly well-peered and had only a
handful of transit providers.

Anycast directionality is essentially out of the hands of the service
provider and the behaviour depends on such things as upstream transit,
filtering and peering policies. In contrast, DNS tricks are
controllable on a very granular basis. The biggest problem with DNS
tricks is that there may be a anti-correlation between the resolver's
locality and its users - which this I-D is specifically an attempt to
fix.

> Conversely "dns-tricks" can be, and are, used by those who don't meet the
> above criteria. The only limitations are having authorative servers capable
> of applying the "tricks".

+1

-- 
Colm