Re: Tyan 2885 and Sil 3114 performance problem
On Wed, Mar 01, 2006 at 11:58:40AM +0100, Joost Kraaijeveld wrote:
> I experiencing (unworkable) performance problems after a change in my
> hardware configuration. I added a second drive (SAMSUNG SP2504C) to my
> on-board Sil 3114 SATA controller). Largish I/O operations (e.g. start
> applications of moderate size (Openoffice, Mozilla, Synaptic) , running
> bonnie, hdparm -t /dev/sda, copying large files) lock up my Gnome
> desktop and make my music stutter ,the first one being the nastiest ;-).
> Running hdparm or bonnie do not show any performance problems in terms
> of actual throughput (see below) but the unresponsive machine is quit
> annoying. None of the perfomance problems appear if I do the same
> operations on my 3Ware 9500 RAID 5 array.
>
> Does anyone recognize these problems and/or does that anyone have a
> sollution? If not, could you give me an advise as to what I could do to
> diagnose the probleme better?
>
> I have the following hard/software configuration:
>
> Stock Debian 2.6.15-1-amd64-k8-smp kernel
> - / and swap-partion on /dev/sda1 (sil 3114)
> - /var on /dev/sdb1 (sil 3114)
> - /home on /dev/sdc1 (3ware)
>
> Dual Opteron Tyan 2885 board with 2 prcs (248), 4 GB
> Sil 3114 on-board SATA controler
> with 1 WDC WD2000JD-00H and 1 SAMSUNG SP2504C, no raid
> 3Ware 9500S-8 RAID5 with 5 WDC WD1200SD disks
Sorry for the late reaction. We have a similar board and found that a
Promise TX4 has a higher throughput. DOn't know about the problems with
sound stuttering or so cause this is a server without sound configured.
FYI, lspci says: "Promise Technology, Inc. PDC20718 (SATA 300 TX4)".
Another thing to try is to use a different IO scheduler. To see the
available and used IO schedulers, use:
cat /sys/block/sda/queue/scheduler
To select a different one (cfq, for example), use:
echo cfq > /sys/block/sda/queue/scheduler
This can also be selected at boot time using the "elevator=" parameter
(see Documentation/kernel-parameters.txt for more information). The
default is anticipatory, but in my experience cfq distributes IO more
equal to processes, especially when used in combination with ionice
(see Documentation/block/ioprio.txt for code).
Erik
--
+-- Erik Mouw -- www.harddisk-recovery.com -- +31 70 370 12 90 --
| Lab address: Delftechpark 26, 2628 XH, Delft, The Netherlands
Reply to: