Excerpts from Noel Chiappa on Thu, Dec 04, 2008 11:05:56AM -0500:
> What's troubling me much more (wearing both engineering and
> architectural hats) is that we are, in effect, building a second
> packet-switching system on top of the first, but separate from it 

As argued elsewhere ... in order for it to be a different "packet
switching system", it would have to have different addressing,
routing, and switching.

In the scope of the globally routed Internet, addressing is the same
-- a BGP neighbor in the core BGPs with an ETR just as with any other
node.  Routing is the same -- good old BGP.  The switching is the same
-- every packet is switched by the same core routers in the same way.

All that's different is that packets are prepared for crossing the
globally routed Internet at a new point, which is to say that when
they implement map-and-encap, edge sites are separated away from the
core, but the core is the same.  There is no new switching system,
only a new adaptation, at its edges, to the existing switching system,
once a prefix is removed from it.  

Excerpts from Tony Li on Thu, Dec 04, 2008 09:20:10AM -0800:
> Ok, but the underlying system is effectively a set of
> connection-oriented tunnels.  Yes, you're using packets to emulate
> that connection, but you're definitely treating the tunnel as a
> point-to-point link.

At the last three IETF meetings I've told you that encapsulation does
not imply "tunnel", and it certainly doesn't imply a network-layer
connection.  There is no setup phase, and there is no shared state
necessary for a decapsulator to receive a packet.  If I were a
decapsulator, you could encapsulate a packet and throw it at me and I
would do the right thing whether I had ever heard of you before or
not.  Paul Francis and crew called a "funnel", as opposed to a
"tunnel".

> Agreed.  Moreover, if there is a way out, why do we need this kludge
> as a stepping-stone...

Ignoring "kludge" for the moment because everything is equally grubby
as far as I can tell, this is the big question.  This is why it's
important to evaluate the various approaches to locator/identifier
separation.  Can we solve the routing scaling problem indirectly while
solving locator/identifier split for session control?  

I'm not saying map-and-encap is the most elegant thing we can do.
However, we do have time constraints.  You can only change all the
hosts in a few special markets.  For others you have to do some kind
of "separation", whether NAT, proxying, or map-and-encap.  NAT doesn't
work for sites that are multihomed unless you change the hosts (socket
bindings break when routing changes).  Proxying is desperate.  What
are you left with?  Also, see Lixia's slides from two IETFs ago where
she said that investing in developing map-and-encap loses us little
and may gain us everything.


Scott
_______________________________________________
rrg mailing list
[email protected]
https://www.irtf.org/mailman/listinfo/rrg

Reply via email to