File System vs Raw Device for Oracle

[Miguel Angel Aliaga_1]

Hello,
actually we are working with a 400GB OLTP Oracle database on raw device ( Oracle 8.1.7.x ) and we want to know about opinions of using raw devices or to migrate to file systems.


Regards.

[A. Clay Stephenson]

I've done these measurements and these are my conclusions ON MY DATA.

If you are running HP-UX 10.20 or 11.0, the best performance (essentially indistinguishable) from RAW/IO is obtained
using OnlineJFS mount options delaylog,nodatainlog,convosync=direct,mincache=direct for data files and indices. Redo and archive logs should be delaylog,nodatainlog.

The convosync=direct,mincache=direct option bypass the UNIX buffers and thus you are in essence raw/io. Typical maximum buffer cache size should be around 300MB. I find that setting a fixed buffer cache using bufpages is slightly better.

On 11.11, I've actually found that fully buffered I/O is the better performer. All files should be set to nodatainlog,delaylog AND increase buffer cache to about 900-1000MB.
The dynamic buffer cache routines are much improved in 11.11 so feel free to use them although for the present I'm stubbornly sticking to a fixed buffer cache but I've really not been able to measure any difference between reasonable dynamic buffer cache settings and good fixed buffer cache settings on my boxes.

---------------------------------------------

Of course, the real answer is to do the measurements yourself on your box and using your data.

----------------------------------------------
Don't overlook that maintenance (backups, restores, snapshots) is a very important factor as well; in that sense, filesystems win hands down.

[Leif Halvarsson_2]

Hi
With such large database this perhaps can be of interest:
http://veritas.com/products/category/ProductDetail.jhtml?productId=oracleeditionhp&_requestid=112719

[Sridhar Bhaskarla]

I agree that using Online JFS options particularly mincache=direct and convosync=direct bypass the buffers. 'mincache' takes care of non-synchronous writes while convosync takes care of 'synchronous' writes. I prefer to use filesystems with these options instead raw devices.

Quick IO and blocklevel backups from veritas seem to offer better performance but I haven't tried them yet.

Not to ignore, striping also can be considered to improve the performance and load balancing.

-Sri

[MANOJ SRIVASTAVA]

Hi


This is a nice thread :


http://forums.itrc.hp.com/cm/QuestionAnswer/1,,0x7e0d72106351d5118fef0090279cd0f9,00.html



Manoj Srivastava