[Lustre-discuss] maximum MDT inode count

[Andreas Dilger]

For future filesystem compatibility, we are wondering if there are any
Lustre MDT filesystems in existence that have 2B or more total inodes?

This is fairly unlikely, because it would require an MDT filesystem
that is > 8TB in size (which isn't even supported yet) and/or has been
formatted with specific options to increase the total number of inodes.

This can be checked with "dumpe2fs -h /dev/{mdtdev} | grep 'Inode count'"
on the MDT.

If there are issues with privacy, please email me directly. Otherwise,
I think people would be interested to know what the larger MDT sizes
are (even if they aren't as high as 2B inodes).

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

_______________________________________________
Lustre-discuss mailing list
Lustre-...@lists.lustre.org
http://lists.lustre.org/mailman/listinfo/lustre-discuss

[Jakob Goldbach]

> Otherwise,
> I think people would be interested to know what the larger MDT sizes
> are (even if they aren't as high as 2B inodes).
>

I have three clusters of 400M inodes each. MDT formated with an inode
size of 256 and one inode per 1024 byte: '-I 256 -i 1024'.

The cluster is split in three primarily to keep an eventual restore time
low but ideally this would be one big cluster with 1.2G inodes.

To other users: Note the inode size is less than recommended. My
stripecount is one so this is feasible without a performance hit but
Andreas warned me about future features which might add more user
attributes.

/Jakob

[Daniel Leaberry]

We have ~1.4 billion inodes and we're using about 50 million. We use
lots of small files so I took our highest utilization ever and doubled
it. Better safe than sorry.

We did format with an inode every 1024 bytes because we use no striping.

--
Daniel Leaberry
Senior Systems Administrator
iArchives Inc.

[Daniel Leaberry]

We have ~1.4 billion inodes and we're using about 50 million. We use
lots of small files so I took our highest utilization ever and doubled
it. Better safe than sorry.

We did format with an inode every 1024 bytes because we use no striping.

--
Daniel Leaberry
Senior Systems Administrator
iArchives Inc.

[D. Marc Stearman]

On our filesystem that was built to support BGL, we have 2B inodes
with 2T of capacity.

Inode count: 2097315840
Inodes per group: 32768
Inode blocks per group: 4096
Inode size: 512

We explicitly chose this number and put it in the XML when the
filesystem was built. Then the users changed their I/O pattern and
did not end up creating nearly as many files as we estimated. Since
then we have created file systems with a smaller number of inodes to
reduce fsck time. (Not really an issue as much now, though with the
fast fsck feature).

-Marc

----
D. Marc Stearman
LC Lustre Administration Lead
ma...@llnl.gov
925.423.9670
Pager: 1.888.203.0641

On Jun 17, 2008, at 10:10 AM, Andreas Dilger wrote:

> For future filesystem compatibility, we are wondering if there are any
> Lustre MDT filesystems in existence that have 2B or more total inodes?
>
> This is fairly unlikely, because it would require an MDT filesystem
> that is > 8TB in size (which isn't even supported yet) and/or has been
> formatted with specific options to increase the total number of
> inodes.
>
> This can be checked with "dumpe2fs -h /dev/{mdtdev} | grep 'Inode
> count'"
> on the MDT.
>
>
> If there are issues with privacy, please email me directly.
> Otherwise,
> I think people would be interested to know what the larger MDT sizes
> are (even if they aren't as high as 2B inodes).
>
> Cheers, Andreas
> --
> Andreas Dilger
> Sr. Staff Engineer, Lustre Group
> Sun Microsystems of Canada, Inc.
>
> _______________________________________________

> Lustre-devel mailing list
> Lustre...@lists.lustre.org
> http://lists.lustre.org/mailman/listinfo/lustre-devel