We are running 2 clustered NW6SP5 servers with the latest NSS patches, latest TCP patches and the latest NIC drivers in a pure IP env. and we
have some issues with client file copy performance. The servers have 1Gbit connections each to the backbone.
Hardware is a so called "HP DL380G2 Packaged Cluster" with latest firmwares.
We have three clustered volumes running on these servers.
Performance when copying files TO the volumes is excellent no matter what protocol is used. Transferring files FROM the volumes using protocols
like FTP or HTTP is also excellent. Backup/restore performance is also quite good, around 200-250Mbit.
Using Windows XPSP2 with latest Netware client (4.91SP1) and latest patches and copying files from explorer or cmd prompt is terrible. About
20-30Mbit. When several clients copy files from the volumes every client still gets about 20-30Mbit so the server can obviously deliver more. I
have tried completely clean installations of XP without SP or with SP1/SP2 with only default 4.91SP1 client installed and there is no difference
I have ruled out that NSS can't deliver or the SCSI subsystem since performance is good except when copying files from the servers using the NCP
I have tried the ususal stuff, experimenting with different parameters and such. There is NO duplex or autosync issues, servers have been
connected to the same core switch for over 3 years and this issue is quite recent but to rule this out I have connected one of the servers to a
different switch and tried the same copying when client is directly connected to that switch in both gigabitmode, 100Mbit full duplex,
autosync.. you name it and I tried it. No difference in performance except when forcing a mismatched duplex connection.
Experimenting with different tcp parameters like "nagle" algorithm, delayed acknowledgement does make some difference but also have a tendency
to mess up backup performance. I have also tried the NSS /NOCOPYBUFFERSONXLATCH with no difference in performance at all.
The really strange thing is that performance differs on the different volumes but only when copying files from the volumes with the ncp protocol.
Running TSATEST against the volumes indicate that all of the volumes can deliver about 25MBytes/s
Any hints on what I can try next?