[Linux-cluster] RAW Devices performance issue

[Fajar A. Nugraha]

On Wed, Feb 16, 2011 at 5:04 PM, Shariq Siddiqui
<shariq.siddiqui at yahoo.com> wrote:
>
>
> Dear All,
>
> I am going to install Oracle RAC on two Servers, With shared SAN storage (Servers and Storage is IBM)
> OS = RHEL 5u5 x64 bit
>
> And we used multipathing mechanism and created multipathing devices.
> i.e. /dev/mapper/mpath1.
>
> Then I created raw device /dev/raw/raw1 of this /dev/mapper/mpath1 Block device as per pre-reqs for Oracle Cluster.
>
> Every thing looks good, But we faced the performance issue as under...
>
> when we run command :
> #dd if=/dev/zero of=/dev/mapper/mpath1 bs=1024 count=1000
> the writing rate is approx. 34 MB/s
>
> But If we run command
> #dd if=/dev/zero of=/dev/raw/raw1 bs=1024 count=1000
> the writing rate is very slow like 253 KB/s
>
> Please advice how to tune the performance.

Shouldn't you ask Oracle about that?

My GUESS is that in the first one the I/O is buffered, while in the
second /dev/raw/raw* is simply a block device opened with O_DIRECT
(thus bypassing buffer cache). You may want to retry dd with
"oflags=direct" and compare the results.

You might want to look at
http://en.wikipedia.org/wiki/Raw_device
http://download.oracle.com/docs/cd/B19306_01/relnotes.102/b15659/toc.htm#CJAICHEG

Again, better ask Oracle if you want to be sure.

-- 
Fajar