switch uplinks issues

[ccnak1ing]

Anyone able to help with this one?

2 access switches 3750 and 2960 have resilient connections to 6500 distribution switches.  Clients on each access switch download a 15MB file from a server in the server farm.  Client on 2960 takes 8minutes, Client on 3750 takes 60minutes, switch clients around and the client on the 2960 still takes 8 mins

Observations:

2960 uplinks are 100MB, 3750 uplinks are 1Gig

Clients are 100 Full duplex

Server is on 1000 full duplex

Any help or ideas are welcomed

[siddhartham]

It shouldn't definately take 60 min but it should even take 8 min to download a 15 MB file on the LAN unless there is some limitation on the server side.

Can you do a tracert from the clients that are connected to both the switches to the server and paste the output, this will tell us the latency in the network.

Can you also explain how your network is configured, this will help us to troubleshoot the issue.

[Leo Laohoo]

2960 uplinks are 100MB, 3750 uplinks are 1Gig

How?  Fibre?  Copper?

If they are going via copper to the 6500 what line card are they going to?  What line card is the server going to?

Is the line card a 6748 or a 61XX or 63XX or 65XX?

[Jason Dance]

Is the issue isolated to just the transfer rate between the server and clients? 

How many other devices are on the access switches?

How busy are the uplinks, and is there any trunking or etherchannel protocols involved? 

Are there any errors logged on the interfaces involved?

What file transfer method are you using?

What OS is on the server and clients?

[ccnak1ing]

We are just using transfer to simulate conversation between the endpoints as the client authenticate and talk to the server through the transaction. We bypassed the network and placed a test pc on the same switch as the server and got the same result.

I have attached a pic of the tests we carried out, in the end we put in a workaround on 3750 to use queue-set 2 with a larger buffer.

mls qos queue-set output 2 buffers 10 10 10 70

mls qos queue-set output 2 threshold 4 3100 3100 100 3200

switch(config)# int FastEthernet1/0/1
switch(config-if)# queue-set 2

Not convinced with the solution though, I was thinking if the server is sending too much (over 100mb) then the pc should send a backoff message, but pc acts differently when on 3750 - output drops on the switchport interface slow transaction,  when on 2960 - no output drops on the switchport interface fast transaction