MLS QoS

[dlongpre]

Hi QoS Expert :-)

I know that "mls qos queue-set output 1 buffers 25 55 10 10" share buffer between interface queues (for queue-set 1). Per example, Queue 2 and 4 will get each 10% of the buffer interface, right ? Yes!

My first question is about the scheduler "srr-queue bandwidth share weight1 weight2 weight3 weight4" which is the relative queue weight:

Per example on 1000Mbps interface:  srr-queue bandwidth share 30 20 25 25

Queue1 will get 30/(30+20+25+25)* 1000Mbps = 300Mbps

Queue2 will get 20/(30+20+25+25)* 1000Mbps = 200Mbps

Queue3 will get 25/(30+20+25+25)* 1000Mbps = 250Mbps

Queue4 will get 25/(30+20+25+25)* 1000Mbps = 250Mbps

Question 1: Does the buffer queue allocated ratio should be aligned with the srr-queue bandwidth share. My concern is what about if I allocate 1% of buffer for queue4 and I srr-queue it with  97% of the bandwidth?  Does scheduler (SRR) will be able to drive 97%  of the bandwidth for output queue-4 if only 1% of the interface buffer has been assigned to queue-4? You following me ? :-)

mls qos queue-set output 1 buffers 25 55 19 1

srr-queue bandwidth share 1 1 1 97

Question 2: I know with priority-queue out, the queue-1 weight value is ignored during srr-queue bandwitdh share calculation. What about srr-queue bandwitdh shape . I read somewhere that even queue-1 srr-queue bandwitdh shape was ignored and queue-1 will be served no matter whether srr share/shape are configured. However I read also somewhere that priority-queue out will use the srr-queue bandwitdh shape queue-1 value and will never use higher bandwidth than the configured shape value.

So, with priority-queue out, does SRR SHAPE is ignored for output queue-1 or NOT?

Thanks for this clarification,

Dominic.

[Mark Malone]

Hi

If your taking this from like a 3750/3560 etc

Queue 1 is shaped and queue 2,3,4 are shared by default , if you enable priority queue out you enable strict priority queuing which removes queue 1 from its default of shaped/shared process , any traffic that then gets assigned to the pq will be serviced first before other queues, this can introduce latency so be careful if using this command

You should only tweak the queues/buffers as last resort really when you see large amounts of drops in 1 queue which are incrementing total output drops ,  mls on certain switches due to  buffers can cause drops , I don't think they released it in a doc but the buffers on a 3750 I believe is only around 6mb compared to newer 3850s which is 12mb , that 6mb say when mls is enabled is split across the 4 queues rather than 1 so you increase the chance of buffer filling up and dropping packets when mls is enabled on certain platforms so that's when you might have to tweak them depending on your traffic load , I wouldn't give 1 queue 97% of your interface bandwidth I would run it as default leave it for few days see the build up in queues and then if necessary tweak them

Good explanation and examples

https://supportforums.cisco.com/document/31581/egress-qos#_Toc191205668

http://www.cisco.com/c/en/us/support/docs/switches/catalyst-3750-series-switches/116089-technote-switches-output-drops-qos-00.html

[Joseph W. Doherty]

Disclaimer

The Author of this posting offers the information contained within this posting without consideration and with the reader's understanding that there's no implied or expressed suitability or fitness for any purpose. Information provided is for informational purposes only and should not be construed as rendering professional advice of any kind. Usage of this posting's information is solely at reader's own risk.

Liability Disclaimer

In no event shall Author be liable for any damages wha2tsoever (including, without limitation, damages for loss of use, data or profit) arising out of the use or inability to use the posting's information even if Author has been advised of the possibility of such damage.

Posting

I don't think they released it in a doc but the buffers on a 3750 I believe is only around 6mb compared

You might find of interest: http://www.cisco.com/c/en/us/support/docs/switches/catalyst-3750-series-switches/116102-qanda-egress-00.html

that 6mb say when mls is enabled is split across the 4 queues rather than 1

True - On a 3560/3750, the default buffer settings reserves half the buffers to interface egress queues, so the default does often lead to more drops than when QoS is disabled.  However, I believe I've been able to obtain lower drops, with QoS enabled, over when QoS is disabled, with buffer tuning.

[Mark Malone]

Thanks Joseph for that doc could never find one where it was actually specified the buffers on 3750s while the newer data sheets for the newer models all seem to specify exactly what there capable of

[Joseph W. Doherty]

Disclaimer

The Author of this posting offers the information contained within this posting without consideration and with the reader's understanding that there's no implied or expressed suitability or fitness for any purpose. Information provided is for informational purposes only and should not be construed as rendering professional advice of any kind. Usage of this posting's information is solely at reader's own risk.

Liability Disclaimer

In no event shall Author be liable for any damages wha2tsoever (including, without limitation, damages for loss of use, data or profit) arising out of the use or inability to use the posting's information even if Author has been advised of the possibility of such damage.

Posting

Q1 - There are two general recommendations.  Either assign buffers to align with bandwidth allocations or assign buffers to inversely align with buffer allocations.

My recommendation is you allocate what you need to accomplish your QoS goals.

What SRR can drive the queue to isn't artificially limited by either the (share) bandwidth setting or buffer allocation.  For example, a share queue with 1% bandwidth, and minimum buffers, could run the interface to100%.  Of course, what happens when you have multiple active queues, depends on the queue settings  and the actual traffic.

Q2 - I recall PQ, for Q1, ignores both share and shape settings.