Re: old GSE idea

[Peter Tattam <peter@jazz-1.trumpet.com.au>]

You don't need to break into TCP stacks if the variable bits are zeroed at the
borders and if advertised addresses (i.e. DNS, IPSEC) always have these bits
zeroed.  Existing Ipv6 stacks should be able to work right away if the zeroing
rules are adhered to.

Even so, you still have to solve the issue of binding the strongly aggregated
routing gook to the site unique ID without degenerating back to what we have in
IPv4.

Restating my earlier observation - if the core is going to be strongly
aggregated, the MH information to any particular site is absent at the core.
Ti does actually still exist in the network, so it needs to be reconstituted
somehow.   DNS based approaches to the binding problem would not be rugged or
timely enough, nor would they deal adequately with load sharing.  Something
akin to the connectivity cache would still be required.  The question is to use
a push approach (broadcast or BGP like) or a pull approach (DNS like).

here's another idea (again apologies if it's already been thought of).  why not
apply a few more rules to the aggregated BGP tree.  For example, each leaf BGP
node can only be N (e.g. N=2) BGP hops from the core.  If that were the case if
might not be unreasonable to extend BGP so that a transient query could be made
to the nearest node to determine the MH paths of the other end.  This
information would travel through the core, but not need not remain in the core. 
The choice as to whether to cache the information would a matter of policy.  A
core BGP router would typically only pass information but not cache it whereas
an edge BGP router would be free to cache what it desired - perhaps within the
contraints of its CPU and memory.

Peter

On Wed, 16 Apr 2003, Brian E Carpenter wrote:

> Actually the requirement isn't that the bottom 80 bits be globally
> unique, but that they be mutually unique among the set of hosts
> involved (which may of course be more than two, due to referrals).
> I don't know if that constraint is easier to meet.
> 
> Some would say that the answer to your question is HIP.
> 
> I do agree we should agree on a systematic breakdown and comparison
> of approaches.
> 
>    Brian
> 
> Iljitsch van Beijnum wrote:
> > 
> > On woensdag, apr 16, 2003, at 15:40 Europe/Amsterdam, Brian E Carpenter
> > wrote:
> > 
> > > I think we should fly up one level and discuss a hypothetical
> > > world in which addresses in A000::/3 are deemed to be mutable
> > > in flight between bits 3 and 47 inclusive. See what it does
> > > to TCP, SCTP and IPSEC for example.
> > 
> > Well, break them... The TCP/UDP checksum should be easy enough to fix,
> > IPsec AH not much harder. The real problem is that if I have a session
> > with a001::1 and suddenly packets start coming in from a002::1, how do
> > I know these belong to the same session? This can be fixed by making
> > the bottom 64/80 bits should be globally unique, or by informing the
> > other side of all possible values that may appear in those 45 bits
> > beforehand.
> > 
> > As long as we're flying up levels, why not go up one more and compare
> > different multiple-PA approaches?
> 
> 

--
Peter R. Tattam                            peter@trumpet.com
Managing Director,    Trumpet Software International Pty Ltd
Hobart, Australia,  Ph. +61-3-6245-0220,  Fax +61-3-62450210