JL> I would like to hear anyone using anycast with TCP. What if two servers are JL> equal distance. Wouldn't you have a fair chance of equal 50% packets going JL> each way, killing tcp state connections. The more servers out there JL> advertising the same IP, the more likely you will have cases of equal JL> distance... or will packets typically go to the same server each time (or JL> at least for several minutes) even if the costs are the same?
It works surprisingly well - quite a few sites that are using it.
True packet balancing between pipes (ie: equal path) is generally going between the same source/destination routing gear anyways, so it arrives at the same place. The main routing decisions are done on a geographic basis in bgp, and it's very rare to have redundant connections to different geographic destinations in the same router.
There are a few corner cases to deal with, but that usually only shows up when your gear is very close together in the final network point in which you should probably be using a load balancer [like haproxy!] anyways.
Often enough to get around corner cases with people who keep long sessions going (ie: sites you log into), you use anycast to redirect load balancers which redirects the client to the actual endpoint server which stays active. eg: hit "www.example.com" (anycasted load balancer) which will redirect you to "server01.site02.example.com" (physical server/site).
--Received on 2009/01/26 18:02
This archive was generated by hypermail 2.2.0 : 2009/01/26 18:15 CET