[ppml] [address-policy-wg] Those pesky ULAs again

[Paul_Vixie at isc.org]

> I'm curious about the opposition to EUI64.

i want to know when a host comes onto my network.  i want to have a chance
to set a particular address for it.  i want to be able to do an RFC2136 DNS
Dynamic Update to set the AAAA and/or PTR RR.  i want to be able to tell it
what the local name servers are.  and i want to be able to use non-sparse
addressing, such as a /112 (65536 possible hosts) rather than a /64 (which
is 18446744073709551615 possible hosts) per subnet.

> Maybe everyone was just too excited about plug-n-play hype back in 1995,
> and it seemed like a good idea to make IPv6 more dynamic, or maybe we're
> also so ingrained with IPv4 that we don't think about other approaches.

you're on the right track.  people who i trusted to know what the IETF was
and how it worked, told me in 1995 that IPv6 had "rapid renumbering" as one
of its fundamental design requirements.  the idea was to be able to add and
delete prefixes from a LAN as an enterprise added and deleted transit
providers.  so, lightweight multihoming would be a side effect of rapid
renumbering.  subnets and hosts would have more than one prefix, and each
prefix could have its own exit gateway (or share one).

none of that happened.  so the entire dynamic architecture that EUI64 was
supposed to be part of hasn't happened.  but unlike 8+8 and A6/DNAME, the
IETF hasn't abolished or deprecated EUI64 (yet?).  so we're doing DHCPv6
and letting the market sort out which approach is better.

> You've mentioned the split between EID/RID, and I don't know that that's
> *not* possible with IPv6 - I would think that v6 flow labels, header
> stacking and/or OSPF opaque LSAs and/or MPLS could all work here to keep
> routing scalable.

i hope i didn't say it wasn't possible.  i said the lack of an EID/RID split
was IPv4's major problem, and IPv6 either did nothing to solve it, or made it
worse, or made it harder to solve, or perhaps all three.  v6 flow labels were
put in to make it possible to renumber a TCP6 endpoint without teardown and
re-setup of sessions.  anybody got that working yet?

header stacking will probably go the way of RH0 if it hasn't already, but
maybe there's hope -- is there an RFC that shows how this would work for
mobile IP without requiring that i tunnel the traffic in still-open TCP
sessions through the gateway i was nearest to when i opened the session?

i guess i'm insufficiently imaginative since i can't think of how MPLS or
OSPF can be used to scale a global (exterior) internet routing system.

> Some consensus I got out of reading the IETF 68 was that there's no need for
> panic about routing in the short term, and on the other hand there's no
> clear solution long term - in other words the technology is still evolving.

:-).  In other words, it's like greenhouse gas, let the grandkids work it out,
we'll just continue to grow the current system incrementally so that when the
solution comes it can have a lot more of our processed output to deal with.

> Taking a guess here, I'm wondering if what you're thinking of is how to
> work E.164 or variable length NSAP addresses + Bind into IPv6 routing.

i wanted to use domain names as permanent identifiers and have network-wire
addresses be as irrelevant and ephemeral as IEEE 802.  i wanted to pay a setup
and in-flight renumberability complexity tax on every flow, in order to avoid
entrenchment and inertia in the topology.  there was talk at one time of an
ICMP6 message for "tell me your hostname" since it was expected that addresses
would change much faster than PTR RRs could be maintained.  IPv6 could have
been a dramatic unequivocal improvement over IPv4.  instead it's just bigger
addresses, and makes all non-address-size-related problems worse.

thanks for asking.