[address-policy-wg] Discrepancy Between RIPE Policies on IPv4 and IPv6 Provider Independent (PI) Address Space

> - With IPv6 we would like to offer a client (collocation) at 
> least a /112 (with PI this isn't allowed if I'm correctly informed)

This points out a material difference between IPv6 and IPv4. 

IPv6 is classful. Normally, ISPs get a /32 class allocation and
give their customers /48 class assignments. Currently the PI policy
comes from wanting to allow organizations to get a /48 class
assignment from RIPE directly. But if the organization is involved
in the hosting business, it is conceivable that customers will
expect to receive a /48 class block from the data centre operator.

Forget whether this is allowed or not because that is irrelevant.
The point is that in the IPv6 world there is the expectation that
no network will ever get less than a /64. In a data centre environment
one would expect that people renting a server or even a rack, would
be satisfied with a /64. But people renting a room, or a row of racks,
might justifiably demand a /48 for their own internal subnetting and
routing, even if much of it happens on virtual routers and switches.
With the increasing number of cores per chip, even an organization
renting a full rack might need more than a /64 because they need
to have a network hierarchy within their rack.

The problem is that /48 class blocks can't be stretched. In IPv4,
there are no classes and everyone allocates according to strict needs
now and in the very near future. But IPv6 has classes and allocates
so that 99% of allocations can be permanent and never have to be
expanded.

We would not want organizations asking for a /40 PI allocation this
year, then returning next year for another /40, and the year after that.

The only way out that I can see is to treat a data centre like an ISP
and offer organizations one /32 for each data centre that they operate.
This would not apply to WAN operators since they can ask RIPE for a /30 
or a /28 to cover all of their WAN and all data centres, then allocate
internally at whatever bit boundary is appropriate.

One thing that RIPE should never accept is any plan that assigns less
than
/64 to a network. If such a plan is submitted, it should be returned and
the applicant should be requested to expand any longer prefixes to a
full
/64, then resubmit. In other words, if you ask for too few addresses,
RIPE
should deny the request, because it is not in line with IPv6
architecture.


--Michael Dillon