[c-nsp] mitigate output drops on Cisco 2960G platform

Saku Ytti saku at ytti.fi
Sat Sep 21 06:29:33 EDT 2013


On (2013-09-20 20:11 +0200), Gert Doering wrote:

> If you don't need QoS, turning it *off* with "no mls qos" will help
> somewhat (as then all buffers are available for your packets, not only
> 1/4 which is allocated to that particular queue).  What we had to 

This common wisdom actually isn't true, not at least in 3560/3750.
Because there is some buffer which is unused completely if QoS is not
eabled, and enabling QoS and allocating everything to single queue will
give bit more buffers.

https://supportforums.cisco.com/docs/DOC-8093

I always bit wonder why Catalyst has such a weird QoS policy as default,
I wonder if anyone actually finds it desirable. I'm all for things
working out-of-the-box, but Catalyst QoS isn't such.
Saner out-of-the-box QoS for catalyst would be to have two classes, CoS
0 and CoS [1-7] and neither class should be policed, with BW distributed
50/50 to them. Even better if it could classify on frame size and have
classes for say <1400B and >=1400B.
But I can't imagine reaching any concensus on half-sane out-of-the-box
QoS, as people would then point out why it does not work in their
environment, so maybe 100% BE is best you can do.

-- 
  ++ytti


More information about the cisco-nsp mailing list