Re: [Detnet] IP dataplane

["Pascal Thubert (pthubert)" <[e-mail]>]

Hello Toerless

> > We already spent quite a bit of effort on that. Would you kindly
> comment on https://sandbox.ietf.org/doc/html/draft-pthubert-detnet-
> ipv6-hbh-06 ?

Not sure where that comes from, please use https://datatracker.ietf.org/doc/draft-pthubert-detnet-ipv6-hbh/

> 
> wht is the teal with snadbox.ietf.org, it seems not to work for me,
> "rendering fails". Is this something for collaborative
> commenting/editing, or is there a github too ?

Yes, that's the link that was memorized in my browser and it seemed to work. 

> 
> As i mentioned in DetNet, i think it would be great if we could try to
> see if/how much unification across different data-planes we can get.
> With that thought in mind, it might be useful to split any work (like
> this
> document) into
> 
> a) an encoding idependent part describing use-cases/scenarios,
>    AND information model. Where information model describes
>    which variables we want, and where/how each of them needs to be
>    examined.
> 
> b) encoding spec. Which may be as lame as an IPv6 HbH heder, or
> multiple,
>    or maybe an extension header that can be shared across IPv6 and
> future
>    MPLS.

Makes sense to me; do you think the architecture did not dig deep enough?

> 
> > f) Aggregation/disaggregation is i think a problem we can punt up
> from IP layer
> >   to overall DetNet layer. Aka: We have no architectural design or
> solution of
> >   this for MPLS AFAIK. But we would need reference scenarios to work
> against them.
> >
> > IPv6 offers the capability to aggregate by encapsulating multiple
> streams in one tunnel.
> > The key for DetNet is that the outer tunnel must signal the needed
> information for the aggregation including a new sequencing information
> and packet treatment, independent of the inner signaling when tunnels,
> OAM or application flows are merged.
> 
> Sure, but how would that be fundamentally different from aggregation in
> MPLS ?
> Which AFAIK, we have also not yet tackled in DetNet.

Fundamentally the same thing. But we need to stuff all the DetNet info early in the packet, and minimize the cost of encapsulation? This rules out having to add UDP and stuff and point to small HbH EHs.


> 
> > The trick is to determine how wider the aggregation tunnel must be vs
> the sum of inner flows to adapt the bounded latency, a problem that
> might be related to the wide area discussion.
> 
> Sure, but thats not an encap issue, unless you are thinking only about
> the added packet header impact.

We agree

> 
> And we also have the option to agregate even without tunneling. Just
> think about all the traffic to one destination, and you just instantice
> a per-destination traffic shaper on every hop, but let the controller
> calculate the correct shapper parameter for the aggregate at that hop.
> Not good for SP networks, but maybe for non-SP networks.

Think about a large pipe between 2 sites where you funnel multiple streams.
That pipe will encaps between egress of site one and ingress of site 2. 
It can be very stable and share dynamically by shorter lived DetNet streams.


> 
> Aka: I am not saying that we shouldn't figure out the information model
> for aggregation, but its a big new building block.


For now I just mean we need a native IPv6 solution based on EHs, so we can use encapsulation for aggregation. Do we agree?

Keep safe,

Pascal

> 
> > Enjoy the rest of the week and remember the 6MAM v6ops on Friday!
> 
> Indeed!
> 
> You to.
> 
> Cheers
>     Toerless
> >
> > Keep safe,
> >
> > Pascal
> >
> > Cheers
> >    Toerless
> >
> > On Wed, Nov 10, 2021 at 01:48:29PM +0000, Pascal Thubert (pthubert)
> wrote:
> > Thanks to both Lou and David for rewording very correctly what I was
> trying to say at the meeting today.
> >
> > Effectively, the question is whether we need for a pure IP dataplane
> that can carry the DetNet information for Forwarding and Service
> sublayer.
> >
> > Arguments on the table:
> >
> > - Not all DetNet environments employ MPLS and pseudowires as a basic
> tool. Forcing such concepts beyond their domain may hinder adoption in
> pure IP (IT) and industrial (OT) spaces. A native IP solution is
> desirable. Note that IPv4 can be encapsulated in IPv6 so arguably we
> can live with IPv6 signaling.
> > - Hardware operation (common ASICs) need the information very early
> in the packet. Digging after UDP may not be feasible in all cases. The
> DetNet information cannot be missed and should be very early after the
> IP header (before UDP if present).
> > - An IP solution is bound to other rules than MPLS, e.g., use of EH
> and encapsulation vs. stack of label. The properties of the solution
> might be different and IP may possibly express richer semantics. But as
> of now, the IP data plane is more limited than the MPLS one.
> > - There's a need with IPv6 to encapsulate when playing with merging,
> > re-sourcing (for duplication and network coding), altering the
> > destination (to an intermediate elimination node) or changing packet
> > header information (e.g., sequencing); such encapsulation will hinder
> > with the visibility of deep information (UDP and application data),
> > which cannot be used for DetNet processing
> > - the elimination or decapsulation node may not be the destination,
> but it still needs to understand the DetNet signaling; the DetNet
> signaling should be 100% L3, fully independent of L4 and above, and
> should not imply that the transport is USP.
> > - the 5-6 tuple points on upper layer information for a flow. DetNet
> may aggregate flows (several times leading to multiple
> reencapsulations) disaggregate flows (decapsulations and separation)
> and OAM packets. We want the information that signals the dataplane
> processing independent of which application flow / already merged and
> encasulated application flows / and or OAM is transported.
> > - it would make sense to build a solution that integrates well with
> IPv6 state of the art, especially SRv6, and very possibly HbH which is
> getting traction, see the discussion at 6MAN / v6ops on Friday.
> >
> > Bottom line: I disagree with adopting a document that would seem to
> indicate that we're done with IP. On the contrary I believer that
> there's a rich ground that we need to plow to deliver the best for IP
> networks.
> >
> > What do others think?
> >
> > Keep safe,
> >
> > Pascal Thubert
> >
> > _______________________________________________
> > detnet mailing list
> > [e-mail]
> > https://www.ietf.org/mailman/listinfo/detnet
> >
> > --
> > ---
> > [e-mail]
> >
> > _______________________________________________
> > detnet mailing list
> > [e-mail]
> > https://www.ietf.org/mailman/listinfo/detnet
> 
> --
> ---
> [e-mail]