Re: [GLOBAL-V6] IPV4 to IPv6 migration

[bill fumerola <billf@opendns.com>]

On Tue, 03 Jun 2008, Brian E Carpenter wrote:
> I'm toying with the idea of positively advocating a ULA
> prefix as the *preferred* prefix for all internal purposes
> (especially network management). Then, adding and removing
> PA prefixes would not impact internal operations (especially
> network management) in any way.

[ speaking as an end-site administrator - not a NSP/backbone type. ]

i've used this model twice, both with small networks (5-10 locations,
some with direct connectivity, some with tunnels). as long as you get
the address selection logic right, it has all the benefits of RFC1918 w/o
the collision drawbacks. i used PI space both times - but that represents
the current reality of IPv6 routing, not any hard rule.

one ULA /48 for infrastructure, a /48 per-site/per-provider, and a PI
allocation large enough to cover the entire network (internally allocating
/48s and /56s to sites, depending on inter-site connectivity) would be
my ideal layout. it allows for me to push more policy decisions down
to the host level ("bottom barrel transit? source w/ this address.",
"top-tier transit? source w/ this address.", and similar. these tricks
can be done w/ filter based forwarding - at a cost to infrastructure.

there are a few things that would force me to break away from ULA
addressing on infrastructure. if enough operators repeat the same mistakes
often made in deploying IPv4 - blocking all RFC1918 (even important ICMPs)
and/or all ICMP - translated to blocking all ULA and/or all ICMPv6. maybe
as a community can get this one right this time, but probably not.

-- bill