On 8/28/06, Jeff Ross <[EMAIL PROTECTED]> wrote:
On Sat, 26 Aug 2006, Jeff Ross wrote:
> I followed with great interest the recent thread on misc@
> http://marc.theaimsgroup.com/?l=openbsd-misc&m=115620585301672&w=2
> aboout bad write performance with the MegaRAID 320-2 card, since I was also
> experiencing what I perceived to be slow write performance with the same
> card.
>
> Unlike Robert Urban, my set up is three sets of RAID 1 drives, using Hitachi
> 10K disks. (dmesg follows).
>
> Here are the results of a couple of tests that I ran with an LSI MegaRAID
> 320-2X with and without an onboard battery. I'd love to figure out why I got
> no significant change with or without a battery.
>
> bsd.mp with battery installed but not connected, write through, adaptive
> read, direct I/O:
>
> [EMAIL PROTECTED]:/home/jross $ dd if=/dev/zero of=/backup/test_file bs=64k
> count=102400
> 102400+0 records in
> 102400+0 records out
> 6710886400 bytes transferred in 130.387 secs (51468680 bytes/sec)
>
>
> bsd.mp with battery installed and connected, write back, adaptive read,
> direct I/O:
>
> [EMAIL PROTECTED]:/home/jross $ dd if=/dev/zero of=/backup/test_file bs=64k
> count=102400
> 102400+0 records in
> 102400+0 records out
> 6710886400 bytes transferred in 132.209 secs (50759331 bytes/sec)
>
> dmesg:
<snipped>
This is not a goood benchmark. Here seems to be a better one:
dukkha and nirvana are virtually identical servers, both running the same
kernel. dukkha's battery unit for its LSI card hasn't been installed yet.
Here is dukkha doing its altroot thing with no battery backup on the LSI
card:
OpenBSD 4.0-beta (GENERIC.MP) #20: Thu Aug 24 17:08:00 MDT 2006
[EMAIL PROTECTED]:/usr/src/sys/arch/i386/compile/GENERIC.MP
9:19AM up 2 days, 18:24, 5 users, load averages: 0.18, 0.19, 0.16
Removing scratch and junk files:
Backing up root filesystem:
copying /dev/rsd0a to /dev/rsd1f
31995+1 records in
31995+1 records out
262103552 bytes transferred in 33.298 secs (7871297 bytes/sec)
** /dev/rsd1f
** Last Mounted on /
And here is nirvana doing the altroot thing with the battery backup on the
card:
OpenBSD 4.0-beta (GENERIC.MP) #17: Thu Aug 24 16:48:07 MDT 2006
[EMAIL PROTECTED]:/usr/src/sys/arch/i386/compile/GENERIC.MP
9:29AM up 1 day, 15:47, 4 users, load averages: 3.12, 2.88, 2.67
Removing scratch and junk files:
Backing up root filesystem:
copying /dev/rsd0a to /dev/rsd1g
31995+1 records in
31995+1 records out
262103552 bytes transferred in 11.636 secs (22524909 bytes/sec)
** /dev/rsd1g
** Last Mounted on /
Or, on the order of three freaking times faster!
The last dd test I tried was from /dev/zero to the raw device, so when
done I had to newfs the partition and was shocked to see newfs fly by,
where before it was very slow.
So, the short version is that the battery backup does make a BIG
difference.
FWIW for SATA 150-6 and RAID1:
No battery, with writeback set:
[EMAIL PROTECTED]:/home/ethant# sudo dd if=/dev/zero of=/dev/rsd1c bs=64k
count=4096
4096+0 records in
4096+0 records out
268435456 bytes transferred in 7.021 secs (38229520 bytes/sec)
[EMAIL PROTECTED]:/home/ethant# sudo dd if=/dev/zero of=/dev/rsd1c bs=64k
4096+0 records in
4096+0 records out
268435456 bytes transferred in 5.812 secs (46179131 bytes/sec)
No battery, with writethru set:
[EMAIL PROTECTED]:/home/ethant# sudo dd if=/dev/zero of=/dev/rsd2c bs=64k
4096+0 records in
4096+0 records out
268435456 bytes transferred in 6.554 secs (40953595 bytes/sec)
[EMAIL PROTECTED]:/home/ethant# sudo dd if=/dev/zero of=/dev/rsd2c bs=64k
4096+0 records in
4096+0 records out
268435456 bytes transferred in 6.641 secs (40418772 bytes/sec)
