Re: Concurrent I/O (AIX 5.2)

  • From: "Gene Sais" <Gsais@xxxxxxxxxxxxxxxxxxx>
  • To: <oracle-l@xxxxxxxxxxxxx>, "Stuart Clowes" <stuart.clowes@xxxxxxxxx>
  • Date: Thu, 29 Jun 2006 08:08:23 -0400

Have you looked at vmo parameters that affect page stealing.  I have
mine set to:
 
vmo -p -o minperm%=5            # Was 20
vmo -p -o maxclient%=10         # Was 80
vmo -p -o maxperm%=10           # Was 80

This alone is the first place I would start.

>>> "Stuart Clowes" <stuart.clowes@xxxxxxxxx> 06/28/06 4:06 PM >>>

We bit the bullet and implemented concurrent I/O (AIX 5.2 ML6, 9.2.0.5,
JFS2, ESS storage).

It seemed to make operations doing table scans slower (exports in
particular took 50% longer). We have 8K block size,
DB_FILE_MULTIBLOCK_READ_COUNT=8. I suspect that not having the benefit
of JFS2 readahead may be hurting us here.

There was no noticable difference on operations performing random IO.

We did this because of a perceived problem with I/O (intermittent
periods of high wait I/O with apparently low actual physical I/O;
variations in response time of standard queries, even at apparently
'quiet' times). An IBM guru told us that the wait I/O may be due to the
stresses of managing page replacement in the JFS2 cache, rather than
problems at the disk layer. Howerver, moving to CIO made no obvious
difference to our wait I/O or disk throughput..

I'd be interested in the experiences of others who have moved from JFS2
cached I/O to concurrent I/O. Did you have a positve, negative, or mixed
experience? 

Other related posts: