William Herrin wrote:
The point of TCP use IP address for identification is hosts can
confirm IP address is true by 3 way handshaking.
Yeah, but that touches one of the central flaws of the design of IP,
v4 and v6.
We are talking about design of TCP, not IP.
No part of identifying and authenticating communication
should reside at layer 3.
That's why we have port numbers for TCP, though you may
call something equivalent to them SPIs for IPsec.
The IP address shouldn't identify anything.
It should reflect only the host's current position in the network.
You are saying IP address should identify current position in
the network.
The address should be as ephemerally attached to the endpoint as the
layer 2 MAC address and as quickly changeable. Without disrupting
upper layer communication. It would be a crying shame to replace the
layer 4 protocols without doing something about that flaw.
Just say "IP mobility". And it's layer 3 issue.
I actually came up with a solution to BGP scalability. If you
abandon stability of the layer 3 address, just throw it out the
window, it turns out to be relatively easy to build a routing
protocol which constructs ephemeral address hierarchies that
represent the current state of connections in the network even though
the physical network itself is still a general graph. The ephemeral
hierarchies aggregate well reducing the worldwide routing table to a
few tens of thousands of routes.
Then, you need two sets of IP addresses, one for physical network
another for virtual network. Former needs large routing table.
With IP mobility, the latter needs no routing table or BGP.
Only to replace well known port numbers by well known connection
IDs and port scanning by connection ID scanning?
Easy to make this impractical. QUIC has.
It can be made so by sparsely populated port number space.
So, when all what needed are more bits for address and port, don't
try to put all the complicated features someone might have thought
useful.
Masataka Ohta
