Re: A DBA philosopical question

jared.hecker_at_gmail.com wrote:
> I am consulting at an all-Windows shop, running mostly 9.2.0.3 -
> 9.2.0.6 under Windows 2003 Server. The one no-no I've encountered so
> far is that their SAN array is all RAID 5. As they have grown by
> acquisition, this is an old, old StorageTek unit that supposedly does
> not support RAID 1,0.
>
> I am looking into I/O constraint issues :-). What I found is pretty
> surprising, and thus generates the "philosophical question". Their
> Oracle servers are configured with one internal drive and two logical
> volumes from the SAN array. The DBA's here rightly multiplexed their
> control files, one on each drive. Looking at the physical disk access
> screens via OEM, I see that the C: drive takes far longer to write. It
> contains control01.ctl. So the question is: Is it better to multiplex
> to two spindles (the faster SAN volumes) to gain the I/O, or is it
> better to multiplex to three drives?
>
> Kind regards,
> jh
>

Regarding the RAID5 question, I recently ran some tests on a midrange Dell server with its onboard RAID, which can be configured as either RAID5 or RAID10. The system had (4) 15KRPM SCSI drives that were used for each test, which was run using "iozone" and various file and record sizes. Large files (>20GB) were used to ensure that buffering was exhausted.

The results were not quite what I expected. On all but one of the heaviest continuous write activities, the RAID5 won hands down, and where the RAID10 had the advantage, it was very slight (<5%).

I think the reason behind this is the constant number of drives used in each test. In RAID10 you have 2 groups of mirrors which are being striped, so you are essentially striping across (2) spindles.

With the RAID5 and the same number of disks the data is striped across more drives (3, if you ignore the parity), increasing the bandwidth, even when the parity calculation is being done. At least thats the only explanation that I see.

Now if you can spend the $ for additional drives, for equal storage and/or striping, the story will probably change...

It probably pays to test your specific storage system, as they seem to all have different I/O profiles.

--Peter Received on Tue Dec 06 2005 - 17:49:01 CST