Data Centre QoS on 2960G and 3750

[Kevin Dorrell]

I am trying to tune the QoS on some data centre switches: 2960G and 3750. I have read the configuration guide chapter about QoS, and the relevant parts of Tim Szigeti's book, but there are still some points that are not clear to me. I just want to understand it right.

1. Are buffers allocated from a common pool across all the interfaces, or is there a "common pool" on each interface? That is, the "common pool" - is it a common pool in the sense that buffers in the pool are available to any interface that needs them, or is it a common pool in the sense that it belongs to the interface, but is common to all four queues on that interface.
   
2. Either way, is there any way that I can allow one interface to have more buffers than another interface? Clearly it would be useful if the (oversubscribed) switch uplinks could be allowed more output buffers than, say, a link down to a service console.
   
3. If mls qos is disabled, is there a single fifo? The reason I doubt this is that if I do show mls qos interaface Gi0/24 statistics, with mls qos disabled, nevertheless it still shows 4 queues with 3 thresholds each, and traffic in at least two of them (mainly in Q4T1 and Q2T1, mistakenly rendered in this command as Q3T1 and Q1T1).
   
4. Sharing the buffers between the queues. By deafult it is 25%, 25%, 25%, 25%. What happens if they don't add up to 100%? Is the pool shared proportionally?
   
5. The threshold points are express in % of the buffer allocation for the queue (I think). So, if a queue has, say, 20% allocation, am I right is saying there would be no point in setting a max-threshold (T3) level of more than 500%? The default is 25% of available, and T3=400%, making 100%. What if allocation * T3 is less than 100%, or more than 100%?
   

Finally, I would be interested to hear some thoughts about marking. I am using the usual baseline model as suggested by Tim Szigeti and others. I have a few Solaris 10 clusters spread across two data centres. Each cluster consists of two servers, one in each data centre. Between these, there are two interconnects that each run along a private VLAN, and they take different paths through the network. The traffic on these consists of heartbeats and synchronization. How would you treat this interconnect traffic? AF21? AF31? What do you do?

Kevin Dorrell

Luxembourg

[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 whatsoever (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

Kevin, have you seen this doc: https://supportforums.cisco.com/docs/DOC-8093 ?

#1 my understanding, those platforms support buffers reserved per interface and also a shared common pool.

#2 above reference shows how to do that, I recall.

#3 I understood with QoS disabled, there's only supposed to be a single FIFO queue.  When QoS is globally disabled, you see the individual queue counters increment?  What IOS version?

#4 and 5, the above reference might address.

For your cluster heartbeats and sync traffic, I would consider CS2, CS4 or CS5 but none of the AFxy classes, as you probably want to avoid dropping any of this traffic's packets.

[Kevin Dorrell]

Joseph, thank you for that.

Wow .. that is a really useful document, and I am still working my way through it.  I'll come back here if I need to discuss.

Section 2.1.1 is particularly interesting because it seems to imply that without enabling mls qos there are still the four queues, but that the traffic just uses one of them for user traffic.  Control plane traffic still goes on a different queue.  That seems to be consistent with what I observe ... that the counters for Q4T3 increment enormously (default traffic), but there are also incrementing counters on Q2T2 and Q2T3 (control plane traffic).

It also makes the pertinent point that just introducing mls qos will probably increase the number of drops, until you tune it properly.

Thanks again

Kevin