[e2e] administrative domains and the network layer

[Jon Crowcroft]

In message <GOEHJEPMGDEICJALMGLDKENGCAAA.ben at layer8.net>, Ben Black typed:

 >>Much of the recent discussion on this list regarding what the end to end
 >>principle means in the context of the current Internet has overlapped in
 >>my mind with some private discussions on how to effectively multihome in
 >>IPv6 (and IPv4, to some extent).  Many paths seem to naturally lead back
 >>to a solution using NAT or one of its relatives, such as GSE, and I have
 >>seen others argue that such solutions violate the end to end principle.
 

I believe that GSE does NOT violate the end to end principle - it
actually re-enforces it and the security arguments used in the
anti-GSE draft are i think misguided in that they were arguments
against IP or mobility, not against GSE...some of the global
re-writing ideas in GSE, if applied right, would solve a WHOLE lot of
problems - it would give a nice opportunity to re-think inter-domain
routing properly too...

 >>Whether or not NAT actually violates the end to end principle is a 
 >>question I do not currently care to ask, but I have begun to wonder if
 >>part of the conflict might reside in the current OSI layering model,
 >>specifically in its definition of the network layer.  A key mechanism
 >>in managing a network as large as the Internet is the autonomous system.
 >>Autonomous systems are used within the routing system, but are completely
 >>ignored within the network layer.
>
 >>If the autonomous system concept were to be introduced into the 
 >>definition of the network layer, I see the opportunity to truly decouple
 >>host identity from network topology (this is not the same as trusting the
 >>host, as I believe issues of trust are well beyond the scope of the 
 >>network layer), which in turn opens the door to far simplified and
 >>scalable routing architectures (whether they use something akin to GSE,
 >>NAT, or a completely new approach).

 >>I am interested in whatever opinions you might have on the subject.

ok - so if the G Part of a GSE address is the current AS you are
"homed" to, then this works really well - it also works well for
scalalble multi-source multicast address allocation (same as the GLOP
address style, but in v6, there's plenty of bits to map an AS and
still leave a lot of AS specific addresses, and still have intra-AS
allcatio nschemes that scale:-)


GSE should be revisited - the MIT work on migrating state for TCP fits
in really well here and there;s copious security work that shows how to
do this safely - so long as you DONT tie the process state to the
inter-as path the way tcpv4 does, it all works really nicely - 


 cheers

   jon