[CentOS] Performance problem with CentOS and XFS on EMC

Fri Mar 9 21:16:54 UTC 2007
Mark Hull-Richter <mhull-richter at datallegro.com>

Good point (use a good XFS), but it seems I misspoke myself:

Whether we use JFS or XFS, when the cache is enabled, we get less than
50% of the performance we see when we use direct i/o.  Direct i/o is not
an option for general use, so we need to understand why cached i/o is so
slow.

We write huge files sequentially on the whole and as a rule in our app,
so that is probably a factor.  We believe that our i/o overwhelms the
block level elevator algorithm, and that the cache flush algorithm isn't
particularly concerned with sequential i/o.

Can anyone shed some light on this?

Thanks.