Re: Fwd: Re: note from the iesg plenary

[Margaret Wasserman <mrw@windriver.com>]

Hi Brian,

>I know that this subject has been discussed several times.  I believe
>the scenario that Margaret is concerned about is based on this
>diagram she posted to IPNG several months ago (correct me if I am
>wrong, Margaret):

This is one case that I am concerned about, yes.

You and Steve resolved this case by stating that a site must be
"convex", and I am sure that the text explaining this will be in the next 
version of the document -- I do understand what it means, and I agree
that it will solve the problem.

However, this does place several constraints on the concept of a site.
For example, consider two office of the same company, both connected to 
the Internet, that are connected by a high-cost link (higher cost than
traversing the Internet between the two locations).  These two 
locations could not be configured as a single site, although they 
would probably cross-pollinate their global addresses (for redundancy)
and might be treated as a single routing AS.  This adds some complexity
for routing purposes, as it clearly decouples the concept of sites
(which must be convex) and routing ASes (which do not have that
restriction).  

Also it is possible that all of the subnets of a single global address 
allocation may not be in the same site, right?  So, it would probably
not be acceptable for renumbering to be constrained by site boundaries.

Some other issues have been raised in my previous mail, but I forgot to
raise two additional issues:

Default Routes:  Will multi-interface-hosts and non-default-free site border 
routers potentially have several different default routes (one for global 
traffic, one for each site, etc.)?

And in general, will we allow multi-site hosts?  Or, can only a router have
interfaces in more than one site?

Just for the record, I think that I know the answers to many of the
questions that I have posed.  I'm sure that several of you know the answers
as well.  But, do we know the same answers?  

And, how can we standardize this behaviour well-enough for people to implement 
site-border, anycast-compatible, IPv6 hosts and routers without investing 
years of effort to understand IPv6?

One of the most disturbing issues (to me) is the widespread misconception
that it should be possible to route IPv6 packets by looking at only the
first 64-bits.  This may be true for the small number of addresses that
we've already defined, but the IESG has been quite clear that we shouldn't
rely on any hard internal boundaries in the addresses.  And this certainly
wouldn't work for host routes.

Also, site-local addressing and anycast addressing are the biggest changes in
IPv6 from IPv4.  Basically, addressing == routing.  And, I think that it
would be worth the investment of our time to make sure that we understand
all of the routing implications of site-local and anycast addressing.  Also,
we need to make sure that the appropriate semantics are included in the
current IPv6 routing protocols (OSPFv3, BGP-4+, RIPv6, ??) to handle site-local
addressing and the injection of anycast host routes (as we are suggesting in
some drafts).

I am not actually disagreeing with any of the work in this area, to date.
It just isn't complete.

Margaret