[Gluster-devel] excessive inode consumption in namespace?

[Rhesa Rozendaal]

Anand Avati wrote:
> 
> 
>     I'm having a bit of a puzzle here.
> 
>     I've setup a 1 server, 1 client test. The server exports 4 bricks and a
>     namespace volume, and the client unifies those (see specs at the
>     bottom).
> 
>     The namespace directory actually sits on the first partition, so I would
>     expect that partition to show twice the number of consumed inodes
>     compared to
>     the other partitions. But what I'm seeing instead is a threefold
>     consumption:
> 
> 
> That is a bit of wrong math.
> 
> No of inodes in NS = No of directories in any child + No of files in all 
> children.
> It is not directly proportional to any one child.

You're right. I did the math, and it works out exactly. I was just surprised 
that my ns+brick used almost exactly 3 times as much inodes as the other 
bricks, but that's just a coincidence, due to my directory/file ratio.

> The namespace should be put on a partition where you can create LOT of 
> small files. Generally XFS or reiserfs will do a good job. If you are 
> using ext3, format it with a very small block size.

Yes, I will be moving the ns volume to a dedicated disk once I get the space 
for it. For now, the 488M inodes should be sufficient.

I'm actually moving back from xfs to ext3, with smaller partitions. We had big 
problems with xfs after a severe power outage. Its repair requirements are 
unrealistic: for a 8TB partition, you'd need a box with something like 24GB of 
ram, plus 2 weeks downtime. We can't afford either, which is why we're 
switching to glusterfs on ext3.

Putting the namespace on xfs is still a viable option, I guess. It depends on 
how fast the amount of content grows.

Anyway, thanks for making me look closer, and understand what actually goes on!

Rhesa