[lustre-discuss] MGTMDT device getting full

Christopher J. Morrone morrone2 at llnl.gov
Fri Oct 16 10:37:12 PDT 2015

Hi Torsten,

There is no reason to suspect that space usage on the MDT will be the 
same as the average space usage on the OSTs.

Your MDT is storing the metadata about _all_ of the files in your Lustre 
filesystem.  You can think of this metadata as a whole bunch of 
zero-length files with some extended attributes, because under the 
covers that is basically what the MDT is storing.

So space usage on the MDT will be directly proportional to the total 
number of files in your Lustre filesystem.  The size of those files 
doesn't really matter much, because the contents of the files are stored 
on the OSTs.

So your issue is that, for your filesystem, on average your files are 
too small to allow filling your OSTs.  Some possible solutions are:

* Increase the size of your MDT
* Encourage/require your users to start using larger average file sizes

Granted, neither approach is terribly easy.

I think that you can pretty safely ignore the changelogs suggestion.


On 10/16/2015 07:31 AM, Torsten Harenberg wrote:
> Am 16.10.2015 um 16:01 schrieb Ben Evans:
>> Looks like you¹ve got some really large changelogs built up.  Did you have
>> robin hood, or some other consumer running at some point that has since
>> stalled?
> Don't think so, as I never heard about "Robin Hood" in the context of
> Lustre.
> The setup is pretty simple, devices were created with
> mkfs.lustre --fsname=lustre --mgs --mdt --backfstype=ext4
> --failnode= at tcp --verbose /dev/mapper/MGTMDT
> and
> mkfs.lustre --fsname=lustre --ost --backfstype=ext4
> --failnode= at tcp --mgsnode= at tcp
> --mgsnode= at tcp --verbose /dev/mapper/OST0000
> respectively
> The only "add-on" we have is quota support. Back in 2.1.5 that was
> enabled with:
> lfs quotacheck –ug /lustre
> lfs quotaon /lustre
> The file system is mounted on about 200 nodes and accessed by cluster
> users.
> Best regards,
>    Torsten

More information about the lustre-discuss mailing list