[ipv6-wg at ripe.net] Re: [lir-wg] IXP networks routing

Måns Nilsson wrote:

> --On Wednesday, March 05, 2003 17:53:00 +0100 Jeroen Massar
> <jeroen at unfix.org> wrote:

<SNIP>

> > I also am wondering how people define "multihoming".
> > Do they define it as:
> >  - Multiple prefixes over multiple cables from multiple upstreams.
> >  - One prefix over multiple cables from multiple upstreams.
> >  - One prefix over one cable from multiple upstreams.
> > 
> > If you are talking about the last case, where one only has one
> > physical upstream... one shouldn't call that multihoming.
> > I guess the second one is where people are talking about.
> > And the first one is what I would call real multihoming,
> > though one needs SCTP to make that work.
> 
> Cases one and two. One or more prefixes advertised over 
> multiple upstreams.

Case one means having more than one prefixes on the cable,
or in diagram style:

+-----+   +-----------+   2001:db8:11:22::/64  +------+
| the +---+ Cable ISP +------------------------+ eth0 |
|  N  |   +-----------+                        |      |
|  E  |                                        | YOU  |
|  T  |   +-----------+   3ffe:ffff:33:42::/64 |      |
|     +---+ ADSL ISP  +------------------------+ eth1 |
+-----+   +-----------+                        +------+

Thus YOU gets 2 different prefixes from 2 different upstreams.
Ofcourse YOU could also get two /48's routed to him from these
upstreams or more ofcourse. The fact is that a 'server' eg
a web server will have 2 IP's; eg:
www.example.com AAAA 2001:0db8:11:22::80
                AAAA 3ffe:ffff:33:42::80

Ofcourse replace Cable and ADSL with the bigger lines,
this is just to illustrate that the definition of 'multihoming'
is quite a different thing for most people. As I said this
will only work when using SCTP to 'multihome' when one of the
two uplinks fail.

In case two both upstreams carry "your" prefix, eg 2001:db8:11:22::/64
to the rest of the world, diagram style:

+-----+   +-----------+   2001:db8:11:22::/64  +------+    +------+
| the +---+ Cable ISP +------------------------+  R   |    |      |
|  N  |   +-----------+                        |  o   |    |      |
|  E  |                                        |  u   +----+ eth0 |
|  T  |   +-----------+   2001:db8:11:22::/64  |  t   |    |      |
|     +---+ ADSL ISP  +------------------------+  er  |    |      |
+-----+   +-----------+                        +------+    +------+

Note that the moment you only have one physical path to the rest of
the world you should not be talking about 'multihoming' any more (case
3)

I always wonder why people want L3 level 'multihoming' even though
their L1 and L2 path aren't.

> Although my idea is to give people so much space in their initial
> allocation that only the biggest networks will need more than 
> one prefix, thus keeping the routing table as close to
> "nPrefixes == nAS-numbers" as possible. 

Every ISP that can match up to the current requirements for address
space
can get a /32 from all three RIRs. If you can't, you simply are not big
enough and you won't have any multiple links either.

If one really wants 99.9% certainty that their inet works one either
needs
to do it themselves and thus get multiple L1+L2 paths and the hardware
along
with it and then most of the times you are a big enough customer to get
a
/32 too. If you aren't you should just get a better upstream.
Yes, I know, it all sounds quite harsh.

Greets,
 Jeroen