[Lustre-discuss] Resizing OSTs?

[Andreas Dilger]

On 2010-11-09, at 06:16, Brian J. Murrell wrote:
> On Tue, 2010-11-09 at 14:06 +0100, Roy Dragseth wrote: 
>> 
>> It would be nice to make use of the extra volume and I'm wondering if we can 
>> just take down the OST, extend the logical partition on the array, run 
>> resize2fs on the filesystem and be good to go?  Will the MDT just pick up the new OST size automatically?
> 
> As for resizing, as far I understand, it is something that should work,
> but we don't test it -- _at_all_.  If you have enough OSTs that you need
> to do this with, it might be worth your while investing in backing one
> up and resizing it to see if it works.  Of course, you can use your
> backup to verify the operation.  If you do this, please report your
> findings here.

Speaking with my non-Oracle hat on - I have done offline resizing of OSTs on top of LVM many times w/o problems (subject to other OST size limitations of course).  As suggested elsewhere, using the latest Lustre e2fsprogs is important.  Also, as Brian mentions, having a backup is really a good idea.  If you only resize a single OST at a time, it would not need a huge amount of space for the backup.

> But also, if you have the space to back one up (per the above) you could
> simply use the information in the other thread I mentioned to go through
> your OSTs rebuilding them one by one on the larger disks, utilizing all
> of the space when you do the initial formatting of them.

Though this would be slower, I've done this on occasion as well if I want to change the configuration of the filesystem (e.g. creating fewer inodes, or chaning other format-time options).  It is possible to use lfs_migrate script to empty out the OST first, if this is possible/safe in your environment, so that the OST can be take offline without impacting the rest of the filesystem.

Also, in bug 14489 att 15794 there is an untested patch to pass ioctls down from the lustre mountpoint (e.g. /mnt/mdt/lustre-mdt0000) to the underlying ldiskfs filesystem, so that 

Cheers, Andreas