The MPLS WG Archive

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

In message <200311172104.QAA03751@hank.bcentralhost.com>, mark seery writes:
> 
> 
> ---- Curtis Villamizar <curtis@workhorse.fictitious.org> wrote:
> 
> > > So bottom line is I would have to agree that two routers from different 
> > > vendors can not communicate in the same network using ECMP unless the 
> > > algorithm is known by both. So if ECMP is in an architecture or 
> > > framework document, then it is not a proprietary feature, and therefore 
> > > it should be standardized.
> > 
> > There is plenty of existance proof that contradicts the assertion in
> > your last paragraph.
> > 
> > Curtis
> 
> As soon as I sent this I knew the statement was too strong. My concern is tha
> t if two routers in the same network think the same packet is part of two dif
> ferent microflows because they use different algorithms, then undesireable re
> sults could emerge. I will add no more on this subject if it is the consensus
>  of the group that either:
> 
> a) this is incorrect from a technical perspective
> b) only known practical issues are worthy of discussion
> c) this is not the time or place for this discussion
> 
> thanks,
> mark


If two routers in a path do a split (happens in IGP, LDP or heirarchy
in RSVP/TE) and use an identical hash, then the downstream split will
always see microflows filtered by the first split.  If both are 2:1
splits, then the second becomes a NOOP because all of the packets
falling into one side of the hash have already been removed.

It is therefore a requirement that the hash be seeded independently by
each node in the topology.  This hash seeding is usually random with
input from addressing and things such as time of day in fine
granularity (usec or nsec) or other pseudorandom seed.  No node knows
how any other nodes have seeded the hash.  Nothing breaks as a result
of this.

[That would be a) above.]

I think enough has been said about Dave's "fundamental precept of
routing".  Next topic please.

Curtis