IPv6 Addressing Question

[Mike Jones]

Hi guys,

I have been following everything I can find for a while about IPv6
deployments trying to learn all I can from them, And I have a question for
you. Why do network operators often seem to assign global addresses to
interfaces instead of just to routers? Most of the Tier 1/2 network router
addresses on traceroutes seem to be global per-interface addresses.

My 'proposed' (?) example 'core' router: Assigned 2001:db8::A1:34:56:42/128
to loopback. All interfaces requiring some kind of static configuration
assigned address fe80::A1:34:56:42/64 to point static routes etc at*^, and
other routers addressed in config as fe80::<router>%interface (i believe
that syntax is required by the standards? for what that's worth).

*^ i assume static routes are few and far between, so perhaps
fe80::2001:db8:cafe could also be 'next hop' for 2001:db8:cafe::/48 on each
segment to make it easier to move between routers.

IPv6 routing protocols seem in some cases to exclusively use automatic link
local addresses. Even for manual configuration, link locals deal with the
ND exhaustion attack problem in the core quite nicely, while also
simplifying address management.

Are there practical reasons for global addresses on router interfaces? My
networks have them where there are devices connected, but I'm not sure if
anything uses them for routing purposes. I have been ignoring the 'tunnel
address' on my ltunnelbroker 6in4 tunnels and leaving my other layer 3 VPNs
etc unnumbered without issue so far. When routing across ethernet networks
I also use link-local addresses* to avoid another configuration to manage.

>From when the discussion has come up it has been reported that assigning a
global address to the loopback interface and having ICMP
generation/management use that is widely supported by cisco/juniper/etc
gear with 1/2 lines and works out of the box on linux? Distributing these
loopback addresses in your IGP the same way you distribute the interface
addresses now would seem to reduce IGP table sizes while still functioning
the same. In theory with only a single address per device the backbone
could use a single aggregate for addressing the core with scalable logical
address assignment**.

- Mike

[Sorry it turned into a long post, but i'm hoping to spark a discussion on
the pros/cons of different approaches - "96 more bits, no magic... but, way
simpler?"]

*I use fe80::1 as the 'default gateway' in static server configs. fe80:: as
(deprecated?) 'link-local any-router' address seemed ideal until I quickly
found 'any' includes devices with no 'off' switch, hence the :1.

** Example: 2001:db8::XX:YY:ZZ:ID/128 XX=Region,YY=PoP, ZZ=Suite/Rack?,
ID=Router, each can be 0-FFFF. This 'large network' example has a single
aggregate /64*** you can either route without ND worries or drop at the
edge as desired.

*** routing performance of longer than /64 shouldn't matter for packets
going to low traffic interfaces? especially if filtered from inbound
external traffic
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.cluenet.de/pipermail/ipv6-ops/attachments/20130406/b8d7eb61/attachment.html