RE: The state of IPv6 multihoming development

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

On Wed, 23 Oct 2002, Greg Maxwell wrote:

> Contentential boundries are the only ones that make sence.. Any tighter
> and there is too much movement and interconnection.

That's too bad for the movee. If someone wants to move from Nova Scotia
to Costa Rica, they'll just have to renumber. I can live with that.

> But.. They don't buy us much.. At best they would reduce the routing
> table by a factor of 4 or so, probably less..

50% for the US because that's where most of the crap is today. But in
the long run it will be better as other regions start to do
more multihoming.

> Heres the thing.. If a simple linear reduction like that will suffice,
> then lets just flood the DFZ and let hardware improvments take care of it.

Three or so rounds of a factor 10 improvement should be able to take
care of it for a good long time. We hear "exponential growth" but of
course the growth will only be exponential for so long: in the end, it
will be an S curve. So a good solution would be to go for continental
aggregation in a couple of years, country/state/province aggregation in
five years and metropolitan aggregation in ten years or so. But then we
have to start assigning addresses based on metro *now* if we want to
avoid renumbering later.

> If we do not believe it will suffice, then we must have a solution that
> does not increase the routing tables for networks who do not have a
> direct relationship with the multihomed network.

The only way you need to see all routes is if you are a "tier 1" ISP and
don't pay anyone for transit. If having all these routes in your tables
is too expensive you can simply pay someone for transit and use a
default. It's not like we're forcing innocent bystanders to upgrade
their routers, only those who do routing as their core business.

> Lets work on answering these questions: Is it approiate for network
> operators to have to carry the burden created by other peoples
> multihoming?

Yes. The alternative is becoming telcos and spend a cent a minute to
charge someone for a service that costs only a tenth of a cent to
provide in the first place. All ISPs share the costs of multihoming, but
they also all have multihomed customers so who cares.

> and Can we expect routers handle the future growth of the
> IPv6 internet without aggregation?

Maybe it will turn out that they can, but counting on it is too
dangerous.

> My thought on scaling: I don't think that it's unreasonable to say that
> simple enhanced versions of todays routing hardware/software could handle
> a million routes.. But what happens when we want links that move many 10s
> or hundreds of gigabits a second... and we're forced into using optical
> packet-level switching..

Switching isn't the problem as routing table lookups scale at O(log(n)).
However, processing the updates scales at O(n*log(n)), assuming longer
prefixes are as stable as short ones, which they aren't.

So routing table lookups should use about 17% more processing for a
factor 8 increase (~1M routes) while processing routing updates requires
9.4 times more CPU power. On the other hand, routing table lookups have
timing constraints (at least if you want to forward at line rate) while
routing updates can be delayed.