Pchangba,

does 'it depends' count as an answer? In most cases CIO would be more benefitial for your database but its really hard to answer as long as you don't tell much about the OS and Oracle version, the kind of workload, the other specifics of your system and much more.

IBM says that direct I/O (DIO) tends to benefit heavily random access workloads while CIO tends to benefit heavily update workloads.

Generally:
Not all workloads benefit from the use of CIO (i.e. they get more benefit from filesystem caching). In some cases, it may be desirable to create multiple filesystems and allocate files based on their anticipated workload characteristics.

Therefore, individual filesystems could be mounted in CIO mode, or default caching mode, depending on the I/O characteristics of the files within the particular filesystem.


When CIO is used, the standard filesystem level I/O serialization mechanisms are disabled, which avoids inode contention issues.

If planning to use JFS2 CIO, be sure to isolate online redo logs and control files in a separate file system that was created with agblksize=512.

File systems which contain Oracle data (.dbf) files should be created with agblksize=4096 if db_block_size >= 4K or agblksize=2048 if db_block_size=2K.

On CIO-mounted filesystems that contain datafiles, the database block size needs to be multiple of agblksize.

Make sure that your oracle binaries, archivelogs and dumpfiles are NOT in the filesystems running on cio/dio.

If you are not sure what kind of DB you have, try both - cio and dio - and see what is more performant. Try as well using cio or dio for your 'dbf' filesystem while still using filesystem buffering for your redo's.

If you have AIX 5.3/6.1 and oracle 10.2.0.4 - try to use other mount options for your archivelogs: we always always use rbrw for archivelogs and in many cases for the redo's with great success (rbr = release behind sequential read, rbrw = release behind sequential read and write).

Please be aware all this will be of not much benefit at all when your data is not spread across as much devices as you can if its on SAN space and your hot tables moved to separate disk. A good practice is as well, to use inline logs for your datafilesystems instead of separate external logs.

If you are running AIX 6.1 you even might want to switch off logging completely for your archivelogs and dumps (when you create your filesystems with the nolog capability) - but this obviously is a risk in case of a server crash ... what isn't logged, cannot be restored. Anyways, worth to mention in case you want more perfomance when dumping out or restoring from backups. You can for example switching off the logging before you start the task and switch it on once completed.

In any case - if you decide for cio, make sure you go away from the default async io settings that AIX 5.3 has as a default since they are way too low.
Choose minservers = amount of your lcpus / threads, choose maxreqs = 65536 and try out what value of maxservers seems optimal for you - on most of our trading boxes, 200 maxservers seem to be perfect but some OLTP boxes were much better off after setting 300 maxservers.

Again - there are no settings that suit all ... you have to try what suits best to your database.

Hope this helps,
kind regards
zxmaus

.