[lustre-discuss] Upgrading ZFS version for Lustre

[Mohr Jr, Richard Frank (Rick Mohr)]

> On Jul 27, 2018, at 1:56 PM, Andreas Dilger <adilger at whamcloud.com> wrote:
> 
>> On Jul 27, 2018, at 10:24, Mohr Jr, Richard Frank (Rick Mohr) <rmohr at utk.edu> wrote:
>> 
>> I am working on upgrading some Lustre servers.  The servers currently run lustre 2.8.0 with zfs 0.6.5, and I am looking to upgrade to lustre 2.10.4 with zfs 0.7.9.   I was looking at the manual, and I did not see anything in there that mentioned special steps when changing ZFS versions.  Do I need to run “zfs upgrade” on the underlying pools before upgrading the lustre version?
> 
> Running "zfs upgrade" will (potentially) enable new features in the filesystem, which means you cannot downgrade to ZFS 0.6.5 if you run into problems.  New features in ZFS 0.7.x related to Lustre include large dnodes, inode quota accounting, and Multi-Modifier Protection (MMP).  There are also a lot of performance improvements, so it is definitely a good upgrade.
> 
> I'd recommend doing the "zfs upgrade" step some time after the Lustre+ZFS (and probably also kernel) software update.  This gives you some time with the new Lustre and ZFS versions to ensure they are working well in your environment, then after this is done you can run the "zfs upgrade".  This also allows isolating any potential issues to the new ZFS features, instead of lumping it all together with the software updates.
> 
> You need to explictly enable the MMP support, by setting the pool property "multihost=on", and inode quota with "feature at userobj_accounting=enabled" when you are ready to use these features.

Thanks for all the info.  One of my main reasons for asking about the zfs upgrade step was related to the feasibility of downgrading if we ran into problems.  I will plan on postponing the zfs upgrade step until after we have been running the new lustre version for a while.

BTW, I didn’t find anything in the lustre manual about setting multihost=on or enabling the userobj_accounting feature.  Is that info documented elsewhere?  Or is it going to be added to the manual in the future?

--
Rick Mohr
Senior HPC System Administrator
National Institute for Computational Sciences
http://www.nics.tennessee.edu