[c-nsp] throughput and QOS on 3750 switches

[Weston, Luke]

We are using 3750s as our access switches but have experienced
throughput issues when enabling QOS.

All we are doing is identifying traffic for qos using access lists and
then setting dscp values on ingress.

We are using the default egress queuing on all ports. There is no strict
priority queue configured.

 

The problem is when downloading from a gig attached server (either local
to the switch of further upstream) to a 100Mb desktop, download speed is
terrible. 

It doesn't matter what dscp the traffic has been marked to. I've tested
it with class-default and class-gold traffic with no other traffic on
the switch.

 

Turn mls qos off, and the download flys along. Switch the server port to
100Mb/s and it also flys along. So there is queueing happening when
there is a speed mismatch. 

 

A sniff of the traffic reveals LOADS of checksum errors with mls qos
enabled regardless of server connection speed. (but speed is ok at
100Mb/s)

Turn off mls qos and the checksum errors disappear. 

I've experimented with srr shaping and sharing and had some improvement
in speed, but no single improvement has been as effective as switching
mls qos off.

 Below is our very simple qos. The policy is applied on ingress to each
port.

 

mls qos

 

class-map match-any gold

  match access-group name qos-gold

class-map match-any bronze

  match access-group name qos-bronze

class-map match-any silver

  match access-group name qos-silver

!

!

policy-map csu-qos

  class gold

   set ip dscp cs5

  class silver

   set ip dscp af41

  class bronze

   set ip dscp cs3

  class class-default

 

 

We are using the same simple qos above with default egress queues on
6513s with no issues.

Can anyone suggest where we are going wrong?

Thanks 

Luke