The MPLS WG Archive

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

My ideas within may be old, naive, or impractical.  If so, just say so and
I'll go back to lurking.

On Thu, Mar 30, 2000 at 12:18:02PM -0500, Curtis Villamizar wrote:
> 
> In message <38E28A7D.134BD731@fore.com>, David Charlap writes:
> > "Abes, Andi" wrote:
> > > 
> > > Now my question is, why can't this model be followed under traffic
> > > engineering conditions ?

[snip]

> End to end LSPs across AS boundaries just isn't going to happen.  Each
> ingress might have to originate 10,000s of LSPs.  In the core there
> could potentially be 70,000 * 70,000 LSPs.  Even with these LSPs
> inside of aggregated traffic engineered LSPs and backbone LSRs that
> could handle that amount of state, the 20 bit label number space could
> be exceeded.  End-to-end LSPs is true connection oriented QoS routing
> which doesn't work.  It isn't what MPLS was designed for and for good
> reason.

I agree that true end to end LSPs across AS bounds will not scale on the
Internet.  BGP is a great example of why the piece by piece model is a
"good thing" (TM).

By using piece by piece LSPs with end to end label stacking (forming logical
end to end LSPs) we could achieve a good approximation of a true end to end
LSP.  A corollary to this would be that the more signalling you introduce to
this process, the closer the approximation.

Low level of signalling, low level of confidence
------------------------------------------------
Using existing BGP info each ingress node would form a deep label stack solely
on the existing BGP info.  The nodes in the path of course would have to
support this form of labeling, but this could be advertised via BGP as well.

High level of signalling, high level of confidence
--------------------------------------------------
Using a request oriented model in which each AS allocates a mapping across
peering links which are pushed on to intra AS LSPs resulting in relatively
shallow label stacks.  This would require an end to end signalling protocol.

Compromise
----------
A mix of these would be that via BGP each AS would advertise what levels of
CoS it will support across it's core.  The ingress node would then try to choose
a deep label stack that would carry it through AS's that support the correct
CoS.  This has the benefit of not requiring end to end signalling, but would
attempt to reach the level of confidence that the end to end signalling would
offer.

-- 
James R. Leu