RE: Fwd: Minutes / Notes

[Erik Nordmark <Erik.Nordmark@sun.com>]

> I would contend that modern applications already separate identity and
> location. Take three different example, the web, SIP, and P2P file sharing
> networks. In web applications, the objects are identified by an http URL; in
> SIP, end points are identified by a SIP URL; in P2P networks files are
> identified by their names and attributes. All of these applications treat IP
> addresses strictly as locations -- some place from which you can get a web
> page, to which you can send voice packets, from which you can get a slice of
> a file. The specific IP addresses vary over time, depending upon load
> balancing in web farms, transient registrations in SIP, or which of the file
> publishers happens to be on-line with P2P file sharing networks. P2P and the
> web, combined, represent the bulk of current Internet traffic.

Christian,
I'm trying to reconcile the above with other things we are hearing:
1. the lack of a multihoming solution in IPv6 is a hindrance to deployment
2. the lack of PI space in IPv6 (whether or not multihoming) is
   a hindrance to deployment

If the folks concerned about #1 shared your view then presumably the only
thing we would need would be multiaddressing and a fix for the
interaction between source address selection and ingress filtering and we
would have solved #1.
However, there might be strong perception out there that IPv4 multihoming
handles applications that do not have their own identifier hence folks want 
to see IPv6 supporting the same thing, even if applications are moving towards
their own identifier space and application layer retries upon failures.

Even if we believed that a fix in source address selection + ingress filtering
was sufficient for #1 it doesn't seem to address the perception of #2.
I get the feeling that folks have been burned by renumbering when switching
ISPs in IPv4 and they never want to do that again, so convincing them
that the renumbering tools in IPv6 plus applications having their own
identifiers and retry logic is a good enough solution seems like a hard
sell. [Not that ID/loc separation makes everything use the PI identifiers; 
there will presumably still be filtering in firewalls etc applied to loctors.
But it sure would reduce the pain of renumbering.]

So do you think that fixing source address selection + ingress filtering
would be sufficient to address perception 1 and 2?



> The question is  not whether we want to separate the location and identity
> function of IP addresses. Clearly, application developers have voted on that
> one. They just use the location function, and rely on other systems for
> identity. IP addresses should be locators, period. The question is whether
> we want to pay an identifier tax at the IP layer. The tax will be
> significant: additional resolution procedures, additional overhead in the
> packets. The main justification of that tax would be to keep alive some long
> duration TCP connections, or some IPSEC sessions. This may benefit some
> applications, but the tax would have to be payed by everybody, whether they
> need the functionality or not. I would much rather not pay the identifier
> tax, and use a combination of application level sessions, TCP improvement,
> IPSEC fast rekeying, or maybe mobile-IP. There is no reason that everybody
> pays the tax when just a few benefit.

Good point.
Presumably applications that don't need the benefits of the identifiers
can operate on the locators and not incur any overhead at the IP layer.

  Erik