Re: Configuring Veritas Quick I/O & Oracle 8

Hi,

while Veritas state that there is a 5% performance loss (compared to RAW devices) when using Quick I/O on OLTP systems, I assume that this number might even be higher for DSS.
What kind of RAW do you mean: are you directly going through to the physical disks or are you using the Veritas rdsks? Have you any red spots on disk performance? Are the results really based on I/O bottlenecks or did other (database?) bottlenecks appear?
How where the tables created for the second test? Did the execution plan changed? If yes, then why?

There is a number of things to check when comparing a database based on raw devices (thoses numbers are only rough estimates and should be adopted to your environment):

        database block size: 8 or 16k

	stripe width when using RAID 0+1 (don't mix it with RAID 1+0
	which will be supported by Veritas 3.X and higher)

	interleave (<g>, they really call it this way): 64k or 128k

	async I/O support (check the kernel!!!!)
	
	use of parallel query option
	
	multi_block_read_io_count (also influences the boarder line, 
	when the optimizer chooses an full table scan over an indexed 
	approach)

	size of the raw devices (I still think something around 2 GB 
	should be the maximum.)

	and so on...

It's hard to judge from the distance what the cause of the problems is.

Regards
Meinolf
On Sat, 15 May 1999 08:41:51 +0200, "Herbert Zarb" <herbert.zarb_at_bov.com> wrote:

>Hi,
>
>We are currently performing a number of tests to evaluate the viability of
>using Veritas Quick I/O instead of raw devices for our Oracle databases. Our
>system is an SUN Enterprise 4500, with Solaris 2.6 and Oracle 8.0.5. We
>tried a long DSS process several times. Whilst using normal UNIX UFS
>filesystems for the data files, this process took around 13 hours to
>complete. However, when we retried the test using Veritas file systems and
>Quick I/O, it is taking MUCH longer - as a matter of fact it's been going
>on for the last 28 hours and it has not completed yet. Also, we had tried
>the run using raw devices and this took around 40 hours to complete -
>initially we discarded this result but now it is starting to look less
>unrealistic.
>
>From the UNIX side of it, there does not seem to be any particular re-tuning
>to be done when using Veritas file systems & Quick I/O (neither when using
>raw devices if we come to that). Hence I am suspecting that it is either the
>type of test being done which is accentuating the benefits of normal file
>systems over direct I/O to files, or that there is some Oracle tuneables
>that have not been properly set.
>I would really appreciate any help or advice regarding the matter.
>
>Herbert
>
>PS. We've been reading a lot about Veritas with Quick I/O providing the same
>performance as raw. I'd like comments about that as well please <G>.
>
>
>
Received on Tue May 18 1999 - 13:23:00 CDT