NFS performance problems - large file writes slow system severely

[Barry C. Dowell]

Ooops, back with another update. Things are somewhat better, but our developers are still finding the performance much slower than it should be.

The systems behave better, but not up to speed. Better in that the reading/writing IO doesn't bring the system to it's knees now, but still not right in that it still takes 40 minutes to read and write a 117mb file.

I'm continuing to work on the issue (with help along the way), and will certainly come back with an update once this issue is resolved. Thanks again to everyone for their help and suggestions and hopefully things will be resolved soon.

[Barry C. Dowell]

Ok, I think we finally have the real solution to this problem.

Switching to UDP instead of TCP had helped the situation (or seemed to), but still the performance wasn't good.

So, we continued plugging away trying to resolve the issue with Dave Olker pushing in the direction (along with some others above, thanks again) of the problem being network related.

We went with one last hail Mary type play today and opted to change one of the machines from manually locked 100Base-TX,FD (full duplex) to auto-negotiation. Though the box was reporting the proper speed, it wasn't performing as if the speed was actually as reported.

Some more information here real quick -- these boxes were formerly connected to managed switches. Our networking group had instructed us to lock the port speeds and duplex settings. Though the switches work up to Gig speed, the systems were only 100Base-TX. Anyway, when running through the managed switches, things were (by memory) performing just fine.

Along the way we moved these boxes from those switches to smaller Cisco 8 port unmanaged switches. Those switches plug into a Netscreen firewall which uplinks back to the bigger switches.

Apparently with the HP systems running on these smaller Cisco 8 port switches the key is going auto-negotiation.

I knew that our Windows NFS server needed autonegotiation, but that is because it runs at Gig speed, rather than 100Base-TX. I was assuming (as was our regular HP sys admins/hardware support folks) that when the boxes were reporting 100Base-TX,FD that things were functioning correctly but until the system was flipped to auto-negotiation things were definitely not right.

Switching to auto-negotiation got things back up to snuff. After that was done we were able to switch back to TCP for the transport for these mounts and again things were zooming along.

We still have a minor issue (hopefully minor) with filelocking apparently not working correctly, but I'll get that resolved outside of this original issue. (Maybe another forum thread, and already have an e-mail in Dave's pile incase he had a quick answer for that one).

Thanks again all for your suggestions, tips, patience, and help. Much appreciated by this cross-training Windows systems administrator :-)