Re: WG next steps

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

On Mon, 18 Nov 2002, J. Noel Chiappa wrote:

>     >> Do we think it is worthwhile to continue down the network/routing
>     >> based solutions? The lack of feedback on my draft suggests people
>     >> aren't very interested in working on this.

> Oh, Iljitsch, I thought you got lots of feedback - but it was all of the form
> "geographic addressing is a waste of time"! :-)

That's like objecting to the film Star Wars by objecting "the stars are
not at war!" and then refuse to see the flick...

>     >> If we want to do this at the routing level, we have start exploring
>     >> less obvious stuff. For instance, using the flow label or diffserv
>     >> code points to swim across the default zone towards a network that
>     >> knows more specific routing information.

> Just to prevent confusion, I wouldn't call that a "routing" solution (which,
> to me, means having the routing mechanisms keep track of multi-homed sites).

> I think it's better to call it an internetwork-layer solution, one which
> uses a secondary "locator" in the packet header.

Didn't some dead English guy four odd centuries ago have something to
say about the relative value of naming schemes?

Actually I wouldn't necessarily want to include a second full locator.
The extra field could/would be used to rearrange the IPv6 address in
such a way that the top of the aggregation tree is moved to another
place. Using this it should be possible to kind of select a new
hierarchy rather than make sure the one you end up with always works.
(I'm not saying I'm 100% convinced this can work, though, but if we can
borrow a header field we can see how far we can come.)

>     > In the long run that will be the only thing that scales

> It's not clear exactly what you're referring to when you say "that ...

I didn't say this. On the contrarry, I'm convinced that routing can only
scale if we allow people to be present in two or more places in the
hierarchy only when those places are relatively close together. (See
GAPI. There you must connect to ISPs within the same
country/state/province.) Note also that current IPv6 address allocation
mechanisms aren't hierarchical so they should be considered
non-scalable. We can see this in the v4 global routing table which is
110k entries big, while the number of multihomers can't be much more
than 10k. In v6 we have a one time gain and the growth curve will be
slightly flatter, but no fundamental differences.

I was reading the Kleinrock/Kamoun paper by the way, but it doesn't
apply very directly to what we're trying to do here. One of the most
problematic assumptions they make is that there is a single network. In
reality, we're dealing with many networks, and shifting traffic from one
to the other because information is lost in aggregation is not something
we can simply do because it makes engineering sense.

Iljitsch