CERC Performance Question

[Brice Figureau]

On Wed, 2006-07-26 at 15:42 +0200, Florent Gilain wrote:
> I have many PE1850 and 2800 with Perc4 raid card with 2 RAID 1 Hard disk,
> here is what i get :
> 
> [root at apache1 usr]# date; dd if=/dev/zero of=big_file bs=1024 count=5120000;
> date
> mer jui 26 15:33:31 CEST 2006
> 5120000+0 enregistrements lus.
> 5120000+0 enregistrements écrits.
> mer jui 26 15:35:59 CEST 2006
> [root at apache1 usr]# date; dd of=/dev/zero if=big_file bs=1024 count=5120000;
> date
> mer jui 26 15:36:41 CEST 2006
> 5120000+0 enregistrements lus.
> 5120000+0 enregistrements écrits.
> mer jui 26 15:38:31 CEST 2006
> 
> This seems very bad performances no ?

Use a greater blocksize. 
Nowadays hard drive have a default blocksize of 4kB.
Use oflag=direct if your version of dd supports it, otherwise you're
measuring also the page cache performance of the kernel (and not only
the disk subsystem).

I'm not that surprised by your figures, read performance seems greater
than write because of RAID1.

If you want to perform meaningfull measurements of the whole disk
subsystem, I'd recommend something like sysbench (or any other
filesystem benchmark).

If you just want to measure the PERC card performances, use the device
directly instead of a file (WARNING: you will __destroy__ all your data
if you write!!), otherwise you'll measure the filesystem, the page cache
and so on...

Hope this helps,
-- 
Brice Figureau <brice+dell at daysofwonder.com>