Re: [mpls] Reference Augmented Forwarding - MPLS RAF

[Robert Raszuk <[e-mail]>]

Hi Tony,


> I was assuming that you were supporting all of the actions that MNA is
> planning on supporting.  My bad. What are you supporting?
>

My goal was to support selective (in few network nodes) processing on
packets, ideally all operator's defined -  more or less like in SFC
manner, but OAM postcard, trace with separate responses, .

I noticed that where MNA and RAF differ it was not only in the model of
data plane vs control plane action messaging. Fundamentally I was not
really focusing on supporting anything which needs to be executed on all
nodes.

Where support of ECMP or slices needs to be executed on all LSRs my
assumption was (and still is) that is it much better to use existing
encoding for it.

So I do admit I wrote this draft as an alternative to MNA. However now
after our discussion on the list (for good or for bad :) I think we have
different objectives and honestly MNA and RAF are (or can be) complimentary
and independent.

MNA could be used as an envelope to carry actions which are needed to be
executed on all LSRs. RAF can focus on action chains which need to be
executed selectively on few nodes (of a given type ex: ABR), differ from
area to area etc ...

Assuming that ECMP is one of our network actions, then we want some entropy
>> label (or similar value) to be considered when forwarding so that we
>> achieve multi-path traffic separation.
>>
>
> IMO network programming is not really aiming for such usecase(s).
>
>
>
> If our 5G brethren are correct, and carrying a GISS is important, then it
> seems painful to carry MNA and EL separately.
>

Yes I see this now. I also now realize why you reacted to encoding slice ID
into EL the way you did. By "you" I mean group of people - nothing about
personally yourself Tony.


But the same scaling question exists even if you’re just doing GISS.  Let’s
> assume that I need 8 slices in my network.  Do I need 8 RSVs?
>

Sorry I do not know what is GISS ... geographic information systems  ? But
yes if you really want to classify packets into a slice at hop you need to
carry it explicitly (in ISD via RFV pointer or embed it somewhere in the
stack).

But for slices I would rather take a network wide view, focus on slice to
slice forwarding difference and build end to end slice plane - not a hop by
hop classification based mapping. But this is just me.

Kind regards,
R.














>
> Tony
>
>
>
>