SLOOOOOOOOOW File copy

[Alan Sielbeck]

I've got 3 ML350 servers running SBS 2003 R2. 4 GB Ram in each, 2.33 GHz Procs. Raid 5 system, 4 disks online, 1 online spare.

When copying files over the network from the server, I'm getting horrible results. A 700 MB file is showing anywhere from 75 minutes to 125 minutes to copy. If I copy the same file from another client PC, the file will copy in ~3 minutes. I've tried the SMB fix (disable smb signing from both the client and server side) but to no avail. All traffic is going through 1800-24g procurve switches. I've played with flow control on different ports. I did notice that if I locked the server to 100 MB HD (via the switch), I got the file transfer down to ~7 minutes. I've verified DNS / WINS settings. Client PCs with 1 GB cards do not have this problem. Only clients with 100 MB cards are having slow connection problems.

Any thoughts? Thanks in advance for any help you may provide!

[Alan Sielbeck]

One other note - I ran ethereal on the problem client while initiating a copy. I would get a flurry of TCP Dup ACK messages, some TCP Fast Retransmission, some TCP out of order. It looks like a whole lot of badness going on.

[Ron Kinner]

Duplex mismatch somewhere?

One side set at automatic and the other at 100 Full? That will usually be the cause of slow traffic in one way but not in the other.

Set both to automatic or both to 100 full.

Look for high collision rates on the Automatic side and high FCS rates on the Full side.

Ron

[Alan Sielbeck]

I was starting to go down that road. The servers are have 1 gb cards - I'm not having any problems whatsoever on 1 gb clients, only on 100 mb clients. It does appear to be happening on 90% of my 100 mb clients. I will double check all settings on the clients. Should I be able to run all clients and servers in auto mode (both set on the NIC and on the switch) and be happy? Lastly, the 1 pc I was troubleshooting I've played with all various settings on the client side from full auto to 100 HD. Same results with the 1 exception of forcing the server to 100 HD while the client was auto.

Thanks for your help. This is so frustrating.

[Ron Kinner]

You should be able to set the switch and the client both to auto and it should work fine. In the old days we used to worry about auto not working right but they fixed that years ago so auto is usually the best setting.

Ron

[Alan Sielbeck]

Thanks for the feedback Ron. Unfortunately it looks like all devices on the network are set to auto on both the switch side and device side. Same symptoms, 1%-2% utilization on 100 MB connections with 40% - 60% on 1 GB connections.

Most people have reported the problems starting around 2 months ago, approximately when SP2 was installed on the w2k3 servers...

[Ron Kinner]

There were some problems with RSS that started with SP2. Don't know if it would cause your problem but might be worth a shot to try and turn it off:

http://support.microsoft.com/?id=927695

Ron

[Alan Sielbeck]

From reading the article, it looks like this will not apply in our situation. All 3 SBS machines have a single nic, and are all on the internal network. We have a separate firewall in place than handles all NAT / firewall connections.

Thanks for the suggestion.

[Ron Kinner]

Can your 100M Full clients transfer file between themselves at a good speed? If so then it appears that the problem is in the switch. It appears to lack enough buffer capacity to compensate for the higher feed.

Odd tho. You would think that TCP/IP's congestion avoidance would come into play. I wonder if SP2 made changes to tcpip.sys and broke it?

If so, you might try replacing it with the older version that should still be on the machine somewhere. You would probably have to boot into Safe Mode without networking and replace the one in C:\windows\system32\dllcache first and then the one in C:\windows\system32\drivers.

Ron

[Alan Sielbeck]

Yes, client to client transfers are acceptable - also clients with GB cards have acceptable transfer rates. This evening I will be moving one of the client machines to the same switch hosting the server to see if that improves performance. Maybe I can isolate something in the chain and start narrowing down the problem.

(This is a network wide issue, happening between all 100 mb clients to respective servers in each of the 3 separate companies.)

[Alan Sielbeck]

Update: After moving the port used by one of the problem PCs to the same switch used by the server, I got reasonable throughput (45 - 65% link utilization). This led me to the trunk configuration on the switches. I had set up the trunks to use LACP, and two ports from each switch linked to a switch with the servers. I set the trunks to manual configuration, and I enabled flow control on both ends of the link. This seems to have solved the problems I was having. I'm still confused as to why I had fast copies with Gb clients, but horrible speed with 100 Mb clients. Everything is running much better now across the network. Thanks to all for reading / suggesting KB articles, etc.