Re: The state of IPv6 multihoming development

[Iljitsch van Beijnum <iljitsch@muada.com>]

On Tue, 22 Oct 2002, RJ Atkinson wrote:

> > 4. Work on geographical aggregation, especially by getting an address
> >    allocation mechanism that supports this off the ground.

> 	At best, a different term needs to be found other than "geographical
> aggregation".  At worst, this is a bad idea.

Why a different term?

> 	The problem is that just because I am in city X, that still does not
> mean that any of the ISPs that I connect to will interconnect inside X.

Currently interconnection isn't a game played on the city level. In
Europe, there is usually a good interconnect for each country. In the US,
it's more difficult, but it should be possible to divide the country
roughly into three or four parts and have two interconnects between all
the major players in each of those parts.

Also, this isn't a binary thing: if you can't interconnect with a peer in
the target region, you simply don't aggregate routes learned from this
peer.

Finally, it doesn't matter where the interconnect location is, as long as
you can limit the presence of more specifics to a subset of all routers.
So a US-only network could very well have all the routes towards Asian
destinations inside routers on the west coast and routes towards European
inside routers on the east coast. This is a very long way from the target
region, but it still makes sense: routers in Chicago get to forward
packets in the right direction until they reach a router that has more
specific information.

As the routing tables grow, it might even be necessary to have several
routers in one interconnect location that each handle a different target
region. For instance, a network could have two routers at Palo Alto: one
that has domestic west coast routes and one that has Asia/Pacific routes.

I'll have a draft on this ready within a few days.

> 	In short, the IP routing topology is NOT closely congruent with the
> world's geographical topology.  If we assign addresses not congruent
> with
> the actual IP routing topology, this tends to increase the number
> of prefixes in the default-free-zone (which is generally understood
> to be a bad property).

Are you saying the IPv6 routing table should only contain PA routes (this
is current 6bone policy)? I don't think this is a sustainable position in
the long run. And if we are going to do PI, doing it in a way that at
least potentially supports aggregation would be good.

Iljitsch