NFS operations slow / may timeout

[Christopher Watford]

On OpenVMS 7.3-2 using the HP TCP/IP stack, NFS can be very slow.

However, the problem is not that reads/writes are slow, the problem is that testing for the existence of a file is ridiculously slow.

If a file exists and you check for its existence, mounted via UDP(Soft) it takes 0.0054s to come back saying it found the file.

If a file does not exist and you check for its existence, mounted via UDP(Soft) it takes anywhere from 25 to 90 seconds to come backing saying it could not find the file.

I have confirmed with ethereal that the SMB packets are actually taking this long, and it is not the host which is having the issue.

Sometimes the file existence check takes so long that the client thinks the server has timed out, causing a renegotiation and re-setup. This causes all open file handles in the same process to become stale and any operation on them after this point results in an error.

Is there a known issue with 7.3-2? We have a planned upgrade to 8.X soon, but I would like to make sure the issue isn't something else.

[Andy Bustamante · Accepted solution]

I would check the speed/duplex settings before any other troubleshooting. What

$ mcr lancp show device /char

will tell you current settings. Compare this against the switch your Alpha is connected to. You can set this at either at console _mode or with LANCP DEFINE

You should use

$ mcr lancp show device /count

to check for network errors.

Andy

[Christopher Watford]

That was our first guess, so we have already ensured the duplex and speed settings are correct.

We achieve the maximum (practical) throughput and minimum latency for all other I/O operations.

Just not testing for file existence...

[Andy Bustamante]

Do you have any patches install for TCPIP?

$ TCPIP SHOW VERSION

The latest is ECO 6, although there seems to be an intermittent host file issue, http://forums1.itrc.hp.com/service/forums/questionanswer.do?threadId=1057983. ECO 5 should be available from the support center, it's been removed from the ITRC download site. Each new ECO includes prior versions, so if you install ECO 5, you don't need 4,3,2 or 1. There are NFS updates included.

Andy

[Volker Halle]

Christopher,

welcome to the OpenVMS ITRC forum.

There is one sentence in the TCPIP V5.4 ECO 6 release notes regarding the NFS component, which may (or may not) apply:

...
(5) Readdir-plus takes too long
...

But this should already have been fixed with ECO 1. Let's see what patch level you're running.

How large are your directory files ? Does this problem only happen on specific directories ?

Volker.

[Christopher Watford]

Andy/Volker:

::>TCPIP SHOW VERSION

Compaq TCP/IP Services for OpenVMS Alpha Version V5.1 - ECO 4
on a AlphaServer ES40 running OpenVMS V7.3

Obviously we're a bit behind on the times. I can't upgrade the TCP stack on the machine, but should I assume that the Readdir-plus was slow back then too?

I believe Sunday we upgrade to 8.X with TCP/IP 5.4 (or the latest). I'll recommend ECO5 or higher for the TCP stack.

Thank you for your help!

[Andy Bustamante]

If you upgrade to 8.3, the latest TCPIP (5.5) will be included with operating system upgrade media. You can elect to have the o/s upgrade include the upgrade to TCPIP.

Andy

[Christopher Watford]

Andy:

Thanks.

I'll mark this thread as 'Closed/Answered' if the upgrade solves the problem.

[Bart Zorn_1]

To prevent confusion: OpenVMS V8.3 comes with TCP/IP Services for OpenVMS V5.6

Bart Zorn