Nathan Scott wrote:
On Wed, Aug 11, 2004 at 03:39:44PM -0500, James Foris wrote:
I have been using XFS in both IRIX and Linux for many years now, and
overall I am
quite happy with it - recommend it to everyone, run it on my home systems,
But, recently I ran into something that I need some help in
The title says it all, really, but the details follow below.
Before I forget completely, from talking to Christoph a few
days back he suggested that the problems you were seeing may
have been resolved by some fixes that have gone into the tree
in the last week or two (they certainly touch this area) - if
you could retry your tests and let us know whether you still
see this degredation on Linus' current -bk tree, that would
be much appreciated!
Sorry for not getting back myself, thanks.
I have retried these tests on 2.6.9-rc1-bk17 and the results were basically
the same for XFS, some improvement for JFS, significant improvement for
EXT3, and even worse degredation of REISERFS.
At this point, JFS with O_DIRECT is the fastest file system per my tests;
previously, XFS w/o was the best.
(This is, of course, full of caviots; default mkfs tuning with internal journals
on all file systems, single 10K RPM U320 SCSI disk, etc.).
If you want, I can resend the test system description and the actual numbers.
I looked at the 2.6.5 and 2.6.6 versions of the file fs/direct-io.c, and
as I do not use bitkeeper (even CVS is blocked by our corporate firewalls),
the differences were too much to easily overcome.
What I will be doing (tomarrow - this has become a hot-potatoe that has
landed in my lap) is testing 2.6.5, 2.6.6-rc1, 2.6.6-rc2, 2.6.6-rc3 and
2.6.6 to narrow down when this happened. I will let you know what I