Re: [Isis-wg] Request for WG adoption of draft-xu-isis-mpls-elc-00

["Nobo Akiya (nobo)" <[e-mail]>]

Hi Robert,

Please see in-line.

> Hi Nobo,
> 
> Inline ...
> 
> > In MPLS, there's no standards in load balancing. Meaning an
> > implementation may load balance only on the top label, only on the
> > bottom label or set of labels which may or may not be the entire label
> > stack.
> 
> Practically I do not know any implementation which would ignore top label
> as an input for hash, but you are correct that there is no standard and that
> such implementation is possible in theory.
> 
> > To take this into account, every SID advertised will need to be
> > advertised as a "range", even service SIDs.
> 
> Nope. I do not think this needs to be extended to service SIDs at all.
> After all services require transport and transport SIDs is all we care about for
> IP/MPLS loadbalancing.

If (and only if) there is an implementation that load balances only on the bottom label, then service SID is what gets used for load balancing.

> 
> > And what you are suggesting is that the "range" needs to be 7 or 8
> > bits out of 20 bits.
> 
> 5 bits gives option for 32 ECMP paths, 6 - 64, 7-128 that should do it :)

That (7 bits) will realistically be sufficient, although this will be break if someone has >128 hash buckets.

> 
> > Yes even SID advertisements can be optimized to take the "range" into
> > account, but still a very big link/service label space & SRGB is
> > required for this.
> 
> I think for v6 forwarding this is clearly beneficial. Hence such optimization in
> SR architecture is welcome.

For v6, this sounds good.

> 
> If MPLS is going to benefit from it remains to be seen or agreed up on. I was
> just trying to share quick thought which just came up when reading this
> thread.

Yes and it's great that this topic is finally getting much-needed discussions ... although we need a lot more :)

One other thing that I'd like to point out is that EL is the first thing that has come close to standardizing the load balancing on MPLS data plane, and it is setting up the "infrastructure" for improved OAM techniques (ex: CC per ECMP path). Whether we go with EL or something else on MPLS based SR, I think it'll be very beneficial to really think about the OAM angle.

Thanks for interesting discussion!

-Nobo

> 
> >> Is this still more appealing then say mpls over UDP draft ;-) ?
> >> http://tools.ietf.org/html/draft-ietf-mpls-in-udp-05
> >
> > I prefer not to get into this topic for now :)
> 
> :-)
> 
> Cheers,
> R.