Information on anycast network routing as it is applied to dns services is sometimes hard to find. It is also quite often misunderstood as somehow being superior to traditional unicast.
One assertion made is that anycast results in superior performance. This assertion is simply not true. This conclusion can be made by understanding the real behaviour of caching dns servers.
There have been some studies done by Duane Wessels to understand the behaviour of dns caches in selecting target servers in the face of multiple authoritative servers. The findings are published at DNS-OARC and were presented at NANOG.
The following is a slide from the NANOG presentation:
Iljitsch van Beijnum, 18 Sep 2003
BGP: Building Reliable Networks with the Border Gateway Protocol, O’Reilly Media
The original study and presentation can be found on the update note for PowerDNS which goes on to further say:
A careful reading of the presentation, and especially, the study results, should help the reader to understand that in terms of real world performance, it is far better to let the individual isp dns servers select the *best* server out of any given set of authoritative servers than to try to forcefeed them with an arbitrary choice by means of anycast routing.
As a matter of fact, of all the servers tested, bind is the most pronounced in this behaviour. It is by no means an obscure behaviour. It is also a design target because the behaviour increases as release version numbers increase.
DNS Operations, Analysis, and Research Center (DNS-OARC)
members include ICANN, ARIN, AFRINIC, APNIC, LACNIC, RIPE, isc, ultradns, afilias, nlnetlabs, verisign, nominet, microsoft, secure64, and a number of network carriers.
DNS-OARC MEMBERSHIP 2010