The MPLS WG Archive

["James R. Leu" <jleu@mindspring.com>]

On Thu, Apr 27, 2000 at 11:28:05AM -0400, Curtis Villamizar wrote:
> 
> In message <20000427101728.C2129@doit.wisc.edu>, "James R. Leu" writes:
> > 
> > I had a design built around an ATM network that did just this.  A real
> > plus to this idea is if the tunnels to the others EBGP speakers were
> > created dynamically as soon at they were learned. (ie each time a new
> > BGP next hop is learned a tunnel is created to that BGP next hop)
> > 
> > Is this a new idea or is this the same idea as BGP short cuts?
> 
> 
> James,
> 
> In implementations that I am aware of, the tunnel goes to the peer and
> is intended to be configured according to a historic bandwidth
> statistic, typically 95 percentile originally suggested by Awduche.
> There seems to be some interest in using CR with the greater of a
> configured value and a measured value (filtered, of course), but I
> have not seem strong interest in using a measured value only and
> automatically setting up LSPs to either IBGP peers or as new IBGP next
> hops are learned (though I personally like the idea a lot, but more
> likely using IBGP peers rather than learned next hops).

The reason for using the BGP next hop is to solve the case when you use a
route reflector.  In which case each IBGP speaker only has one peer, but
many BGP next hops.

I agree that most "real" TE uses historical bandwidth to a particular node
when setting up a static tunnel.  In my design I envisioned categorizing the
next hops that a particular ingress will use.  An example of these categories
may be:

-High Density Customer POP
-Low Density Customer POP
-Major Exit Point
-Minor Exit Point

Each category would have a template associated with it.  The template would
have many generic parameters specified.  Each BGP next hop would be assigned
to one of these categories (by use of access lists configured by the admin).
Creation of tunnels to these next hops would be dynamic.

Any next hops that can not be classified into one of the categories would have
to be provisioned as static tunnels.

Any thoughts on this?

Jim
-- 
James R. Leu