bonnie++ uninterruptible under heavy I/O load

[Simone Piunno <simone.piunno@wseurope.com>]

Hello,

I'm testing a pair of new servers we just bought.
They are HP DL585 dual Opteron 844 with 8G RAM, RAID1 over 2x72G SCSI disks 
(HP CISS driver) and running 2.6.11.  In /proc/driver/cciss/cciss0 we have:

cciss0: HP Smart Array 5i Controller
Board ID: 0x40800e11
Firmware Version: 2.56
IRQ: 18
Logical drives: 1
Current Q depth: 0
Current # commands on controller: 0
Max Q depth since init: 6
Max # commands on controller since init: 190
Max SG entries since init: 31
cciss/c0d0:       72.83GB       RAID 1(1+0)

As a test, I'm running bonnie++ while two md5sum are checksumming /dev/zero.
I see 98% CPU is correctly allocated on the two md5sum processes, but when 
bonnie is in the "Writing intelligently phase" the system appears noticeably 
unresponsive (high latency) and bonnie and pdflush are uninterruptible almost 
all the time (D in top output).  If I try to kill -9 bonnie++, it goes on 
consuming I/O bandwith for several tens of seconds, before finally dieing.

Unresponsiveness is not 2.6.11 specific (we've seen the same thing on 2.6.10 
and 2.6.8), not I/O scheduler specific ("as" and "deadline" behave the same)
and not CPU/SMP specific (reproduced on single P4 HT and single P3), but only 
on these two DL585 servers we've seen bonnie++ resisting kill -9 for tens of 
seconds.

Of course on request I can provide any other useful info.
Any help is appreciated.

TIA,
  Simone Piunno

-- 
Simone Piunno, chief architect
Wireless Solutions SPA - DADA group
Europe HQ, via Castiglione 25 Bologna
web:www.dada-ws.com tel:+39512966811 fax:+39512966800