At 12:40 PM 10/23/04 +0800, Gu huaping wrote:
I have read RFC2373 where the authors give definition for Anycast Address. But I still cannot find the detail ROUTING materials. Can some one give me some hints to find some material relating to Anycast?

An anycast address is simply an address, and is routed the same way any other address is routed.


What makes an anycast address interesting is the fact that several servers respond to it. For example, many of the DNS roots distribute their load by deploying instances of their service in a variety of places. Standard routing will take a given request to one or another of them, which will reply. This offloads a central device and the communication paths leading to it, and provides faster service to the initial requester.

There are two fundamental problems with anycast. One is that if one of those devices fails and routing is not changed to remove it from the picture, the service is effectively black-holed in the region. To solve this, folks implement routing processes on the server itself - it advertises to the network that it has routing connectivity to the address, not that it *is* the address. Should it fail, the putative router fails, and routing is recalculated without it. Voila, the service is now offered via only the remaining servers.

The other problem is that routing can change, with the result that a session to an anycast server migrates in mid-communication from one server to another. For this reason, anycast is useful today for simple applications like DNS, but isn't very useful for more general services like www. There has been a proposal in the past, which never got traction I guess, to use Mobile IP with optimized routing. One treats the anycast address as the home address, the client of the service as the correspondent node, and the address of the individual server as the care-of-address of the mobile node. When the anycast server replies to the initial SYN, the client starts talking to the actual server's address, meaning that the session doesn't migrate even if routing changes. It would be good, I think, if this could be standardized and deployed, as it would then allow us to use distributed services in an improved way.

I should say that we can implement distributed services today, using DNS - we list several server addresses in the a record. The problem is that the use of these is random, rather than being distributed rationally using routing. The Mobile IP based scenario I just described enables us to do so in a way that predictably improves things.


-------------------------------------------------------------------- IETF IPv6 working group mailing list [EMAIL PROTECTED] Administrative Requests: https://www1.ietf.org/mailman/listinfo/ipv6 --------------------------------------------------------------------

Reply via email to