Well, in the dialup example given, his IP address would be constant
within a session, just not constant across sessions.
true. and if the dialup host only had intermittent connectivity anyway,
having the DNS name dynamically bound to the IP address (when the host
is connected) is about as good as one can hope for.
(OTOH, for hosts with stable connectivity I'd argue that dynamic DNS
isn't nearly good enough. but that's not what VK was talking about.)
The more interesting case would be enhancing TCP to allow endpoint IP
addresses to change during the life of the connection.
sort of. you still need an endpoint identifier, it just needs to be
seperable from the identifier for the network attachment point. it's
not clear that TCP is a good place to make the change, since
(a) TCP *wants* to deal in endpoint identifiers rather than network
location identifiers, and (b) there are other protocols besides
TCP that run over IP that also need this service.
but breaking the assumption that IP header address == TCP header
address would be a good thing - assuming you can maintain an
accurate and efficient mapping between the two.