shaping delay

davey.red · ‎03-23-2009

I am having trouble finding the information I am looking for. Example: on a 10Mbit Ethernet access circuit into an IP VPN (mpls) with a CAR/CIR of 4Mb, the standard configuration would be a hierarchical/nested policy-map which would "shape average".

What would be the impact on EF/priority traffic of the shaping delay incurred by this shaping?

Would there be any difference on e.g. 100Mb port with 40Mb CAR/CIR?

Hopefully someone can point me in the right direction?

Joseph W. Doherty · ‎03-30-2009

"What would be the impact on EF/priority traffic of the shaping delay incurred by this shaping? "

I have not come across clear documentation, either. I would suspect the Bc/Tc would have an impact much like an interface's hardware FIFO buffer might. Further, I've noticed (some?) shapers implement their own WFQ, and it's unclear how a subordinate CBWFQ interacts with the natural shaper's queues.

To fully guarantee good servicing of priority traffic, you might place it in the parent policy rather than the child policy.

e.g.

policy-map parent1

class class-default

shape average 4000000

service-policy child1

policy-map child1

class real-time

priority percent 30

.

vs.

policy-map parent2

class real-time

priority 500000

class class-default

shape average 3500000

service-policy child2

policy-map child2

.

Of course the latter, doesn't allow non-real-time traffic to use all the bandwidth.

"Would there be any difference on e.g. 100Mb port with 40Mb CAR/CIR? "

In theory, wouldn't think so, but in practice, I wouldn't assume actions would scale exactly the same.

davey.red · ‎03-30-2009

Would that work? I don't have deep enough understanding of how Tx Q (which I have been told plays no part in this scenario), Shaper and CBWFQ interact in a hierarchical structure. It would seem sensible that the shaper simply pushes the queue back the the LLQ/CBWFQ queues and takes packets as required as the Tx ring would.

In your configuration, how would the 2 policy maps be processed; in serial? in parallel? I can't find documentation that explains this all very clearly.

I did find another article which is interesting. If these tests are correct then shaping delay is ave. 14ms (pre-12.4(20)T) or 8.5ms (post-12.4(20)T), or thereabouts.

http://www.gossamer-threads.com/lists/cisco/nsp/93730

Hopefully someone can advise further on this.

Joseph W. Doherty · ‎03-30-2009

"Would that work?"

Unclear what you're referring to.

". . . Shaper and CBWFQ interact in a hierarchical structure. It would seem sensible that the shaper simply pushes the queue back the the LLQ/CBWFQ queues and takes packets as required as the Tx ring would."

I would agree, but that's not been well documented. A Cisco reference in your reference notes the many CBWFQ changes in 12.4(20)T. One is, shaper now defaults to FIFO instead of WFQ. That probably accounts for the decreased shaping delay.

NB: The new QoS model in 12.4(20)T was promised several years ago. Makes all platforms' QoS function more like the 7500s and FlexWAN boards. I think it's great except for the sudden change of QoS processing with existing pre-12.4(20)T configs. I think it would have been better to have some config option to denote old vs. new QoS processing model.

"In your configuration, how would the 2 policy maps be processed; in serial? in parallel? "

What I showed was two different hierarchical partial examples. Either would apply the "parent" policy on the interface.