On 17 May 2004, at 04:11, Iljitsch van Beijnum wrote:
This is pretty much impossible with Cisco equipment, as in order to load balance, the different routes must all be learned from the same neighboring AS. (So if everyone used Cisco equipment the AS paths would have to be identical.)
It may be worth noting here that with the style of Internet-scope anycast deployment we're doing with F (ISC-TN-2003-1) each anycast node operates as a globally-unique AS, so even if a single AS was peering with F-root nodes in different places it ought not to be possible for traffic from a single TCP transaction to be split between nodes.
But as long as we're dissing anycast root DNS servers: how many of the root servers are being anycast now or in the future, and how many won't? As we've seen a few times with .org, and thought we saw with .org many more times, over-enthusiastic anycasting isn't without its dangers.
It seems to me that prudence and diversity are best served through a mixture of anycast and unicast nameservers in an NS set, the latter to provide path diversity and local stability to topologically distant regions and the former to accommodate any practical issues that might arise with the latter.
Joe
_______________________________________________ Ietf@xxxxxxxx https://www1.ietf.org/mailman/listinfo/ietf