Things are going to be solved.
I share information here for others that would run into the same troubles.

1/ Changing the BIOS back to an older version (AMI 1.1 instead of AMI 2.0) masked the issue (I/O management somewhere in the controller ?) . But this is not satisfying, as an older BIOS may not handle correctly my hardware. An older version may crash the box due to hardware/software incompatibility.
However, no warning/error message from the kernel: from its POV, everything is fine. So I switched back to the recent version.

2/ I had set very hard values for page cache:
    vm.dirty_ratio = 3
    vm.dirty_background_ratio = 0
In my case, on a 6GB server, this lets 184 MB to the page cache. This is really low, but it's done willingly to avoid caching too much and have the kernel start flushing too much. The counter-part is that when my filesystem needs to flush a lot a meta-data pages, then the page cache is filled and the whole application is frozen, waiting for those I/Os to be completed.
With those parameters:
    vm.dirty_ratio = 20
    vm.dirty_background_ratio = 5
The small writes are amortized in the stream of data writes from the application, and the application is not frozen.
(so you were right: there was a page cache issue)

The question stands: my does XFS generates such a bunch of small I/O writes throughout the disk at around 688GB ?

My fstab mount options are classical ones:
"defaults,nobarrier,noatime,nodiratime"

Maybe the software RAID 0 has helped triggering the problem, too: I don't know if writes on a RAID can generate more I/O than on direct disk. I guess so (I/O fragmentation), but that's only a guess.

--
Mathieu Avila

Le 22/09/2010 22:41, Stan Hoeppner a écrit :

Mathieu AVILA put forth on 9/22/2010 5:26 AM:

I have run my test again with default parameters for mkfs.
I still have this issue. For 20 seconds, the writes are either stalled,
or very slow.
I have run "vmstat" at the same time than "dd", and it appears that the
block device continues to receive write requests, while "dd" is blocked
in the kernel.
With blktrace, I can see that during this period of time, the block
receives a lot of small write requests throughout the volume ranging
from the start till the point where the file has stopped writing. During
the other periods of time, the volume is written normally, starting at
offset 0 and filling the disk continuously.

What happens with "dd if=/dev/zero of=/DATA/big oflag=direct"?  You said
the copy is hanging in the kernel.  Maybe a buffer cache issue?

What fstab mount options are you using for this filesystem?

Could this be an effect of tree rebalancing for extents management (both
inode of big file and free space trees) ? Can it be a hardware problem ?
Have you ever seen that issue before ?

WRT tree rebalancing, that's beyond my knowledge level and someone else
will need to jump into this thread.  If it's a hardware problem you
should be seeing something in dmesg or the kernel log, or both.  If
you're not seeing controller or device errors it's probably not a
hardware problem.  Have you tried this same test with only one of those
two 500GB drives, no mdraid stripe?  That would eliminate any possible
issues with your mdraid implementation.  Speaking of which, could you
please share your mdraid parameters for this stripe set?  That could be
a factor as well.

--
Mathieu Avila
IT & Integration Engineer
mathieu.avila@opencubetech.com

OpenCube Technologies http://www.opencubetech.com
Parc Technologique du Canal, 9 avenue de l'Europe
31520 Ramonville St Agne - FRANCE
Tel. : +33 (0) 561 285 606 - Fax : +33 (0) 561 285 635