Originally posted by: salmanucit77

Hi,
I have gone through a problem just now and want to share and need suggestions. I installed a new oracle 11G Release 1 database on my AIX6.1 server where i mounted the datafiles file system with CIO option as per recommendation to achieve the performance but then i realized that if i use cp command to copy and data on this file systems, it is very very very slow (around 2 MB per second). I also tried to restore the backup of my database from another files system (without CIO option) to this data files file system and restoration also took long time. It means, if we mount or datafiles file system with CIO options, we can gain performance, but in case of data restoration, we will be facing worst performance and only workaround it to mount the file system wihout CIO option to restore the data on this file system.
So any one with any suggestion on the scenario?
Thanks

Salman

Originally posted by: orphy

You must be doing something different. Most real-world DB setups that I know of don't do this type of restore/copy often enough to worry about issues like this. You of course could still use CIO/DIO to gain the performance that you expect. When you need to restore/copy, there is no reason you couldn't simply umount/mount the filesystems without CIO/DIO, do your restore/copy, and umount/mount the same FSs with CIO/DIO.
Orphy

Originally posted by: j.gann

use dd with a large blocksize for this kind of copy jobs: dd bs=4M
do not run copy jobs in parallel to avoid fragmentation

Originally posted by: MarkTaylor

This is exactly the reason why you should NOT set CIO on the filesystem. Filesystem operations with a small block size are up to 7 x times less efficient on a CIO mounted filesystem .. the optimal block size is somewhere around 1MB .. Most decent software vendors provide a way to enable CIO within the application i.e. You can do this in Oracle by setting

Oracle init.ora: filesystemio_options = SETALL

Ref: http://www.dba-oracle.com/t_aix_cio.htm

HTH
Mark Taylor

Originally posted by: niella

Hi Mark,

In the the link you gave it states:

"The IBM documentation also suggests that some Oracle workloads may benefit more from full JFS than from CIO, but in my experience, only the Oracle redo logs should be on a JFS filesystem."

So this implies the DBA actually uses CIO for his data-files, right? I cannot find conclusive "switch to DIO or CIO in this scenario" type documents from Metalink, and in our own tests we see significant CPU usage drops, but either a slight degradation or fairly similar throughput for sequential operations. When I asked the DBA to increase the Oracle block-size beyond 8K (since page-ahead no longer has an effect) I was told that this would be impractical on an existing system since the database tables would have to be recreated or something along that lines. So I wonder if CIO or DIO really should only be considered during database creation?

It would be really awesome if you or anyone else knew of a way to determine more clearly in which Oracle setups it would help to help to switch to either DIO or CIO and what sort of improvements they saw...

Thanks,
Niel