[c-nsp] Need help w/ output drops on 7613 WS-X6748-GE-TX

[Peter Rathlev]

On Wed, 2011-01-05 at 12:47 +0200, Hank Nussbacher wrote:
> At 10:56 05/01/2011 +0100, Peter Rathlev wrote:
> >Do you have QoS enabled? What does "show queueing interface Gi9/29" tell
> >you?
...
> gp#show queueing interface Gi9/29
> Interface GigabitEthernet9/29 queueing strategy:  Weighted Round-Robin
...
>      WRR bandwidth ratios:  100[queue 1] 150[queue 2] 200[queue 3]
>      queue-limit ratios:     50[queue 1]  20[queue 2]  15[queue 3]  15[Pri Queue]
...
>    Packets dropped on Transmit:
> 
>      queue     dropped  [cos-map]
>      ---------------------------------------------
>      1                  1590686  [0 1 ]
>      2                      250  [2 3 4 ]
>      3                        0  [6 7 ]
>      4                        0  [5 ]
...
> How would you recommend adjusting the interface mls queues?

Queue 1 has 50% of the buffers and the most drops. You could increase
queue 1 buffer size but that would of course be at the expense of the
other queues.

We've chosen to combine queues 1 and 2, since we don't really use a lot
of classes. We use the following interface commands:

interface GigabitEthernet4/1
 wrr-queue cos-map 1 2 0 1 2 3 4 
 wrr-queue queue-limit 70 0 15 
!

This gives 70% of the buffer space to queue one, and no space at all to
queue 2. The cos-map-command puts CoS 0-4 in queue one, so queue 2 isn't
used.

Caveat #1: The "wrr-queue cos-map" command propagates to all other ports
on same ASIC, typically blocks of 12 ports. So you can't have different
CoS maps on ports on the same ASIC.

Caveat #2: "wrr-queue queue-limit 70 0 15" reserves no space for queue
2, so any traffic happening to end up in that queue for any reason is
dropped.

Instead of starving queue 2 completely you could just adjust the
partitioning. Default as you can see is

50% queue 1 (CoS 0 + 1, typically Best Effort and Scavenger)
20% queue 2 (CoS 2 + 3 + 4, typically various Assured Forwarding)
15% queue 3 (CoS 6 + 7, "network" traffic (IGP etc))
15% queue 4 (priority/EF, CoS 5, voice and jitter sensitive traffic)

So 60/10/15/15 might also work. Or if you don't use EF much (or don't
need buffers for it) then 65/10/15/10.

Adjusting WRED threholds might also give good results, letting TCP back
off gracefully.

I don't know of any way to list interface buffer utilization, to trial
and error seems to be the only way.

-- 
Peter