Lustre discuss - Apr 2009 - size on mds,

I just patched a version of e2scan to also return the uid, and gid of  
the owner of a file.
Our goal was to find a quick way to find disk utilization by uid/ 
gid.  As running a du, on the filesystem takes 24 hours.  While  
e2scan (on the mds) takes only an hour.

We found out though, obviously that the size returned by e2scan is  
the size on the MDS which is very small. It does not represent the  
size of the objects the file points at.

I think it was at SC last year there was talk of ''size on the
mds''  I
assume this would not be in inode.i_size as that is the underlying  
ldiskfs size, not the sum of the objects.

Is the full size of all the objects, being located in metadata on the  
mds still expected? If so which release?  Just curious.

We will instead turn on quotas, and not enforce them.  We just have  
to wait till our next down time to do that.


Brock Palen
www.umich.edu/~brockp
Center for Advanced Computing
brockp at umich.edu
(734)936-1985

On Apr 09, 2009  10:59 -0400, Brock Palen wrote:
> I just patched a version of e2scan to also return the uid, and gid of  
> the owner of a file.
> Our goal was to find a quick way to find disk utilization by uid/ 
> gid.  As running a du, on the filesystem takes 24 hours.  While  
> e2scan (on the mds) takes only an hour.
> 
> We found out though, obviously that the size returned by e2scan is  
> the size on the MDS which is very small. It does not represent the  
> size of the objects the file points at.
In 1.6.7 an estimate of the file size will be stored on the MDS
for files that have been recently changed.  It doesn''t modify
exsiting files though.
> I think it was at SC last year there was talk of ''size on the
mds''  I
> assume this would not be in inode.i_size as that is the underlying  
> ldiskfs size, not the sum of the objects.
The full (accurate) size on MDS will be available in a post-2.0 release.

Cheers, Andreas
--
Andreas Dilger
Sr. Staff Engineer, Lustre Group
Sun Microsystems of Canada, Inc.