The MPLS WG Archive

[Curtis Villamizar <curtis@workhorse.fictitious.org>]

In message <FFFC48AEAA5F7447929F4F0D93FCC12D02B927DE@zcard031.ca.nortel.com>, "
David Allan" writes:
> Tom:
> 
> I have to admit, the whole ECMP thing provides me with endless amusement.
> Pull a proprietary rabbit out of your hat and then paint those previously
> ignorant of your "feature" as clueless. If you'd been upfront with the fact
> that there was a problem literally "years ago", a lot of creative energy in
> OAM and PWs would not have been wasted.

We been here before.  ECMP has been around in everything from IGPs to
BGP to MPLS.  In IGPs it goes back 15 years.  In BGP 10 years or
more.  In RSVP/TE it was implemented very early on and was no secret.
And ECMP has also been stated as a customer requirement for LDP,
implemented, and deployed there too.

> Meanwhile we have the PW PID which seems to level the playing field across
> the board as an OAM channel for PWs that can carry just about anything
> without getting confused by already deployed ECMP mechanisms, we've also
> given ECMP a lot of thought and an approach is documented in 17fec-cv for
> the LDP layer. IMHO just about everything on the table at this point has
> taken ECMP into account or at least ECMP as well as it is "commonly"
> understood at this point.
> 
> I'm curious, is the ECMP that must be supported exclusively your
> implementation? T'would be nice if the vendors who've deployed ECMP
> documented what they've done in sufficent detail such that we could all
> innovate in this space and knew all variations we needed to comply with.
> Normally defining data plane OAM requires an agreed characterization of the
> data plane. The situation as it stands is intolerable.....
> 
> cheers
> Dave

There really is no secret to how ECMP is done.  There was discussion
on multipath in RFC2991 and RFC2992 and plenty of discussion on this
list and on PWE3.  The PWE3 architecuture
(draft-ietf-pwe3-arch-06.txt) might be insufficient for what you are
looking for because it just covers not breaking this functionality for
PW.  It does indicate that this ECMP stuff is a poorly guarded secret
at worst.

Its true that the only statement tying this to MPLS is "Some router
implementations also allow equal-cost multipath usage with RIP and
other routing protocols."  However, it is a very poorly kept industry
secret that every router from the PC-RT based NSF T1-NSS of late 1980s
vintage until now have used source/destination based hashing for IP
traffic no matter how it has been moved.  This includes concatonated
physical interfaces (a form of bundling), RSVP/TE from ingress, LDP,
and any form of RSVP/TE hierarchy (for example, LDP over RSVP/TE where
the underlying RSVP/TE uses multipath).

Currently draft-ietf-mpls-oam-requirements-02.txt references
draft-allan-mpls-loadbal-01.txt which has expired.  The problem may be
that draft-allan-mpls-loadbal-0x.txt doesn't reflect reality.  Recall
that the objections to this draft were because it did not cover
existing practice.  Perhaps we should start with a load balancing
internet-draft that referenced RFC2991 and RFC2992 and explicitly
referenced the definition of the PW PID (currently in
draft-ietf-pwe3-arch-06.txt).

Curtis