[Gluster-users] Usage Case: just not getting the performance I was hoping for

[Brian Candler]

On Thu, Mar 15, 2012 at 12:51:39PM +0000, Paul Simpson wrote:
>    On 15 March 2012 12:49, Fabricio <[1]fcannini at gmail.com> wrote:
> 
>      Em 15-03-2012 09:24, Sabuj Pattanayek escreveu:
> 
>      Striped volumes are unfortunately broken on top of XFS at the
>      moment:
>      [2]http://oss.sgi.com/archives/xfs/2012-03/msg00161.html
> 
>      Yea, I found this out after copying several million files into a
>      stripe that XFS doesn't report the size of sparse files correctly.

Oh it *reports* the size of sparse files correctly. Unfortunately, due to
its aggressive pre-allocation strategy, it allocates and zero-fills space
which glusterfs intended to leave as a hole.

I pointed out the problem on the list, and was told that no one strategy can
fit every usage case.  This one is particularly tuned to Samba (which writes
128K past the end of the file then back-fills with data).  It was suggested
that glusterfsd could avoid this problem by using ftruncate before writing.

However, I don't know if Red Hat's kernel has a version of XFS without this
preallocation, because it seems they're not affected by it (or else it just
remains broken and nobody has noticed)

>      Lost a week of time on the copy after which I switched all my bricks
>      to ext4. XFS, never again!
> 
>      XFS saved my ass more than once when loss was certain, Sabuj. I
>      would not write it off so fast.

XFS does rock (IMO), in particular the tools like xfs_bmap and the like are
peerless.  It just doesn't work well with this particular (and unusual)
usage pattern.

Regards,

Brian.