[c-nsp] QoS best practices

[Raymond Lucas]

Peter,

On Fri, 2009-10-16 at 12:13 +0200, Peter Rathlev wrote:
> I have previously read the QoS SRND, but I must say I'm not impressed. I
> know by testing that e.g. Auto-QoS doesn't solve the buffering problems
> (it actually seems to worsen it) and I find the SRND somewhat
> mechanically written. I may just have had a bad day though. :-)

I was more mentioning the SRND in the context of what markings to use for
what traffic and how the different queues should be treated.

You are right that some of the implementation recommendations aren't great.
We also saw the buffering issue I think you are referring to.  For us it
was 3750MEs acting as WAN routers.  The same QoS model as 3560s and
'normal' 3750s.  Much testing and tweaking of buffer sizes followed and we
ended up with good enough results for us using:

Queue     :       1       2       3       4
----------------------------------------------
buffers   :      20      30      45       5
threshold1:     100     700     700      40
threshold2:     100     800     800     100
reserved  :      50     100     100     100
maximum   :     400    1000    1000     100

Egress Priority Queue : enabled
Shaped queue weights (absolute) :  3 0 0 0
Shared queue weights  :  1 60 35 5

This is for an enterprise scenario.  We never had problems with q1,
presumably since it's priority and there shouldn't be too much need for
queuing anyway.  q4 is bulk traffic in our scenario so we aren't worried
about drops there anyway.  It is also worth noting these interfaces were
being rate limited by setting the interface speed to 10/100 as appropriate
and using "srr-queue bandwidth limit x" to get values between 2 and 64 Mbps
which I don't think helped.

We never used Auto-QoS so I can't comment on that.

I could discuss further with you off list if you want.

Cheers,
Ray