> -----Original Message-----
> From: Testbed Support for GridPP member institutes [mailto:TB-
> [log in to unmask]] On Behalf Of Alessandra Forti
>
> Reading a bit more about it [1] what is not clear to me is if we really
> need spaces bigger than /64 for single clusters.
>
I think you do, but possibly not all that much larger.
> I mean /64 is only 1 subnet but has 18,446,744,073,709,551,616 IPv6
> addresses.
>
This is true, but not really the key point - generally you don't get to
use all of those addresses, what that gives you is the scope to allow
your nodes to allocate their own addresses within that space. The
assumption is that a /64 block is a single flat network. You can't
subnet that space into (say) two /65 subnets without breaking things,
or at the very least, breaking a lot of assumptions and making life
difficult and complicated. It's more meaningful to think about that as,
(as you say) one network. The 'lots of addresses' is essentially an
implementation detail.
In particular, I think you'd want at least two /64 subnets for a reasonable
cluster - one to use for stateless autoconfiguration, and one to use for
manually assigned logical service addresses. A /56 allocation gives you
256 subnets, which is probably more than you'd need for something like a
grid cluster, but is sensible for something like a department, where you
might want to divide up (say) central services on one, client desktops
on another, server management cards on another, or have divisions between
different buildings or whatever.
> One argument can be that we do fit the decirption if we don't go through
> the University routers - Manchester doesn't for example - even though
> administratively we belong to the University.
>
IPv6 addressing should normally reflect routing layouts, so if you're
separately routed, I'd have thought you should have a separate address
allocation. If you had a subnet of the University space, upstream routers
would have a routing table entry for Manchester's address range that pointed
to the university routers, and that would be wrong for you. It is possible
to route something like that, but it's a bit awkward. However, I think you'd
just need to have the discussion with Janet/NNW and see how they'd prefer
to handle it in practice. It would/will be interesting to see what they say.
> I think that so far it was all very accademic and people are rushing to
> get huge address spaces whether they need it or not.
>
I think there's an element of that, but it is important to think in terms
of how many networks you have, not how many addresses, and there's also the
argument that we can afford to hand out lots of space - as John notes,
Janet have a /32 which is fairly spacious. If everyone were kept to the
'small institution' block of a /48, then that would allow 2^16 blocks. There's
plenty of scope for giving major universities a more generous /44 without
being too concerned about running out. That said, if everyone got that, we'd
have a problem because there's only room for 2^12=4096 of those, and that's
probably a bit tight - OK for universities, fatal if you start looping (say)
schools in as well (a quick Google suggests that according to the DfE, there
are ~24,000 schools in England). A mixture of allocation sizes at that
level is probably sensible. At this point the balance being struck is between
the liklihoods of running out of networks at the top level, and running out
near the edges - it's worth avoiding giving people excessively stingy blocks
and having to renumber them in the near future if you can avoid it.
Ewan
|