For smaller, but still reasonably large sites, having the multihoming
functionality in each host is still too much headache, that's why I
think it's important to be able to offload the necessary processing to a
relatively small number of boxes at the edges of the network. To the
internal network this will look like traditional IPv4-style multihoming,
to the service providers it looks like single homing. The only downside
is that the other end has to implement something similar. That's why we
also need to have it in the host stacks, even for hosts that don't
multihome themselves.
Can these boxes (which in essence are multihoming policy servers) exist
only at the edge of the enterprise network, or do they have to live at each
level of the internal routing hierarchy? A host with interfaces on
multiple subnets has to make the same decisions about address selection
when reaching an internal destination as when reaching an external one. I
guess the implication is that a solution which is routing-protocol agnostic
is more scalable than one which isn't.