Routing

Hi,

I have a question on applying policer on LAG for traffic shaping on MX480.

Is there any issue with the accuracy of traffic shaping if the LAG consisted of interfaces from same FPC, but different MIC?

I have this concern after reading the following link.

https://groups.google.com/forum/#!topic/juniper-nsp-/OAufLF77etk

Thank you

MPC1 and MPC3 both use a single PFE for both MICs.  MPC2 has a PFE per MIC slot.

While it seems a bit odd that the PFE count goes down on the MPC3, it's true.  MPC1 and MPC2 both use the same 30 Gbps PFE.  The MPC3 uses a newer 130 Gbps PFE.

So, if you are on a MPC1 or MPC3 you only have a single PFE to worry about regardless if the LAG is on two different MICs in the same MPC.  On a MPC2 then you may see the behavior you linked to, where each interface on a different PFE could go up to the configured policer value resulting in a net of 2x policer value.

Does that help?

-Chad

Hi Chad,

Your response does help greatly.

As we are using MX-MPC2E-3D card, in order to obtain accurate traffic shaping, it is understood that I'll need to put all the interfaces in LAG within a single MIC.

Please correct me if I'm wrong.

Thank you

I haven't personally tested policing over multiple PFEs on either IChip or Trio so I can't say for certain.  However, based on what you linked to, yes.  The page matches up with the Trio processing path outlined in the O'Reilly MX book and I'm not aware of anything that would change that behavior.

That said, a possible alternative would be to configure the LAG for primary/backup, in which case traffic would only flow through a single interface, and thus a single PFE, at a time.  If you really need the additional capacity, or the LAG is more than two ports, then that isn't an option.

I dug a little bit further in the MX book in the Trio CoS section, specifically pages 421-423.  It touches on doing CoS scheduling for a LAG interface and two different modes that are available.  Scale mode takes the configured values, divides them by the number of interfaces in the LAG, and then applies that divided value to each physical interface.  Replicate mode simply takes the configured values and applies them to each interface in the LAG directly.  While basic policing / shaping doesn't have the same options, it does further support that the values you configure for policing will be sent to each PFE that supports an interface in the LAG, potentially resulting in a higher aggregate value over all of the LAG interfaces than you expected.

Cheers!

-Chad

Thank you Chad, you're providing great help!

As a point of information:

We faced a decision-set on this topic in the past. It was either satisfying the need for a fined-tuned policing implementation against the need for slot redundancy when provisioning the LAG link members - particularly, with customer-facing LAGs. In our case, we decided to give more importance to achieving slot diversity and be more tolerant on the LAG policing.

You may need to face the same decision-set...

...

I just ran across a KB article on this exact topic while looking for something else.

KB28129: Understanding policers when used with AE interfaces