Hard vs soft NFS mounts

[Olivier Masse]

There are endless philosophical debates here about whether we should use hard or soft NFS mounts with our SAP installation.

Considering that our NFS server is highly available, I'm reconsidering it.

I dead somewhere in an official MC/ServiceGuard documentation that we should use hard NFS mounts. But it didn't give much details as to why.

Dave Olker also seems to lean towards hard mounts when in an HA environment in one of his documents, i.e. "forcibly unmounting NFS filesystems"

However, me and some colleagues all hate the hard mounts because we've seen lots of gotchas in the past with them, but things seem different now. For instance, we're supposed to be able to kill processes hung on an NFS call while in the past a reboot as needed when this happend.

What do you use? Are hard mounts really better than before?

Thanks

[Pete Randall · Accepted solution]

Olivier,

For what it's worth, I've always used hard mounts except for a brief and unsatisfactory experiment with soft mounts way back in the days that we used to have problems. Changing over to soft mounts didn't make much difference for us so we went back to hard mounts and kept patching NFS. In my opinion, current, well patched versions of NFS have few problems with hard mounts and that's what we'll stick with.


Pete

[Dave Olker]

Hi Olivier,

Yes, I do like hard mounts. :)

The only reason I've ever advised against hard mounts is when customers have NFS servers that are crashing frequently and their clients are stuck with non-responsive mount points. Since there is no way to forcibly unmount filesystems on HP-UX (yet, but keep reading) soft mounts allow the client's requests to eventually time out and return control to the application - assuming the application doesn't crash when it receives the I/O error resulting from the failed NFS request.

-- HOWEVER --

In a very short time (like October 2004) HP will be releasing HP-UX 11i v2 update 2, and one of the new features of this release is forcible NFS unmounts! This feature will allow an NFS client to unmount a hung NFS filesystem even if the kernel thinks there are outstanding I/O's pending on the filesystem. 11i v2 update 2 will run on most PA-RISC and Itanium systems, so anyone running 11i v1 on these systems today can get this feature very soon.

There are other NFS features arriving in 11i v2u2, like direct I/O over NFS, which will greatly aid applications like Oracle, Sybase, Informix that like to manage their own data caches and want direct access to the underlying filesystems with read-ahead and write-behind.

Bottom line - once 11i v2u2 comes out and forcible NFS unmounts are available, one of the last remaining reasons to use soft mounts will be history. (Just my opinion - as you said, hard vs. soft is a philosophical debate, so if you're a "soft" proponent - I mean no offense. :)

Regards,

Dave

[Olivier Masse]

Dave,

I'll check if SAP has a note in OSS about hard vs. soft mounts and we should be set until 11iv2. Even though our HA NFS servers should be 99.5% reliable, I really don't want to end up with locked hard mounts if there's a glitch.

BTW: Thanks for your book, it's a great reference which is concise and well vulgarized.

Olivier