Re: [mpls] MPLS Network Actions Framework

[Jeff Tantsura <[e-mail]>]

Hi Tony,

 

I’m pretty much with Robert here, most vendors have been supporting context labels (driven by MVPN and egress protection) for quite some time, distributing context (binding of value to action) over control plane is well understood.

 

Cheers,

Jeff

 

From: Robert Raszuk
Sent: Thursday, April 7, 2022 1:12 AM
To: Tony Li
Cc: mpls; [e-mail]
Subject: Re: [mpls] MPLS Network Actions Framework

 

Hi Tony,

 

So I am not so much worried about control plane state - especially that distribution of those actions mappings could be as simple as configuration push or pub-sub model. If you think that control plane memory associated with keeping state of local actions would be an issue then we clearly would have a different opinion on this. 

 

But please observe that even if you carry action ID in a label stack you still need to have your hardware understand it (state) and execute such action (no difference). 

 

The only difference I could think of would be in parameters of actions. But looking at list of use cases I do not see that it contains such requirement to embed parameters of executed actions also in each packet. 

 

So in control plane assisted model those parameters would indeed occupy data plane memory while in packet embedded case (example when they are part of SRH) they would not. 

 

To the question of dereferencing AIL I am not clear what exactly you are describing. In my model transit node capable of recognizing AIL is doing a fixed size 20 bit lookup (label lookup :) which either reveleas local action or not. 

 

Many thx,
R.

 

 

 

 

 

 

On Thu, 7 Apr 2022 at 02:13, Tony Li <[e-mail]> wrote:

 

Hi Robert,

 

Well no more than any other proposal - assuming we are talking about data plane state.  

 

 

Ummm… Other proposals are putting everything into the data plane, with no reliance on the control plane.  You’re adding a level of indirection from the data plane into control plane state. Not unthinkable, but as you note, it does present scale questions and setup questions.

 

See when control plane signals action to AIL mapping such action can be a targeted action. So the state in the data plane will be only installed in the nodes which actually need to perform special actions. 

 

 

Almost agreed. Wouldn’t every node along the path need to have state to at least dereference the AIL mapping? Yes, it might be a non-action, but the node can’t know that without performing the dereference.

 

And that is going to happen irrespective on how we carry the actions - explicitly in the packet or implicitly in control plane signalling and only mapping to it in the packet. 

 

 

Agreed, you can put your actions in the data plane or put them in the control plane with a level of indirection. It’s a tradeoff.

 

Tony