Going through some very back email on this list... In message <yt7mcel11t.fsf@cesium.clock.org>, "Sean M. Doran" writes:
P.S.: Curtis Villamizar had another interesting approach which involved pushing content far afield to machines with the same transport-layer (IP) addresses, relying upon closest-exit routing to connect one to the topologically-closest replication machine. Unfortunately, while this could be really cool for NSPs to offload stuff towards peering points (public or private), it also has some poor scaling properties and is uncomfortably reliant upon the stability of routing.
If he's done any more thinking about the idea, I'd love to hear about it though.
I'd like to hear what the scaling problems were. The replicated routes are advertised IBGP only and aggregated at the borders so I'm not sure what is "uncomfortably reliant upon the stability of routing" either. If you point traffic toward an ANS aggregate some of it is probably going to web servers served by this scheme. We have at least one fairly large customer using this but not one of the major sources of web content. We see their replicated web server subnet as a /27 internally and after aggregation you see it only as part of a /14. Those a bit more timid announce the same subnet from multiple sites with us setting differing local-pref and use DNS based load split, using the replicated prefix for fast fallback. This does nothing to improve proximity of access but means a down tail circuit is mended in under a minute rather than after DNS times out. Curtis