RE: [PWE3] Proposed new 5.4.3 for arch draft

["Yaakov Stein" <yaakov_s at rad.com>]

> > > Seems more generic to just mention ECMP instead of MPLS, or
mention MPLS as an 
> > > example.
> > 
> > But this is an MPLS specific problem we are addressing.

> NH=> Don't want to labour this one but.....if you think about 
> what you are saying here Stewart, this is yet another example 
> of why MPLS is not the same as IP.
> > 
> > ECMP in an IP network works just fine (provided you ECMP per flow
and 
> > not per packet).

> NH=> I agree with Stewart here.  Load balancing is fine 
> within a single layer network (any)....I believe its called 
> TE ;-).  The problems occur when folks try and load balance 
> over nested layered networks.  So if you start peeking inside 
> the payload area of *any* layer network (I could apply this 
> to SDH or OTNs in principle) to see what the clients are then 
> you can expect problems.  This another consequence of mp2p 
> LDP due to this just doing SPF routing.  Even our IP design 
> teams wish ECMP did not exist in MPLS (esp when one considers 
> all the fault-management problems it creates too).

Isn't it refreshing to see Neil agree with Stewart ?

Stewart - It isn't completely clear to me why and how an IP network
performs load balancing on a per-flow rather than per-packet basis.
Is this done by IP source and destination address?
What if we have several PWs between these two IP addresses?

Neil - One of the initial justifications of MPLS was to solve
the IP equal-cost problem. Of course this could be done
by only allowing MPLS-TE solutions, and that may be what operators
like yourself need. 

But there are a lot of best-effort uses for MPLS as well, 
and these shouldn't have to suffer from overloaded links
when there are underloaded ones available. 
Admittedly peaking into client layer headers is a hack,
but it is such a useful one that it may qualify as one of those
rare exceptions where the rules can be broken
(or better, the MPLS rules should be modified by adding a type
identifier).

Y(J)S


_______________________________________________
pwe3 mailing list
pwe3@ietf.org
https://www1.ietf.org/mailman/listinfo/pwe3