Re: poor OCFS performance under Linux

The shared storage is a group of 7 Seagate 15K fibre drives, a Eurologic box, and Qlogic QLA2100 cards.

In further testing on a regular SCSI drive, my query mentioned earlier only ran about 20% slower on an OCFS partition compared to EXT2, nowhere near the 200% slower on the fibre drives.

I did find a bug on bugzilla.redhat.com relating to performance problems with a combination of RHEL 2.1/OCFS/Qlogic controllers, although it's not my exact configuration. The problem is still unsolved. http://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=103300

Forgive my lack of understanding of OCFS, but is there a chance that the direct-access nature of OCFS doesn't play well with software RAID?  In yet another timed test, a single fibre disk performed as quickly as a 7-disk software RAID 0 group of drives.

Daniel Morgan <damorgan_at_x.washington.edu> wrote in message news:<1079632317.127043_at_yasure>...
> Benjamin Wood wrote:
>
> > I'm having tough time getting good IO performance utilizing OCFS under
> > Linux.
> >
> > I have two x86 servers using QLA2100 cards against a 7 disk fibre
> > array configured with software raid-0. The servers are running RHAS
> > 3.0 2.4.21-4.EL, Oracle is 9.2.0.4, and OCFS is 1.0.10-1. OCFS was
> > created with:
> > mkfs.ocfs -F -b 128 -L /var/opt/oracle/oradata/orcl -m
> > /var/opt/oracle/oradata/orcl -u `id -u oracle` -g `id -g oracle` -p
> > 0775 /dev/md0
> >
> > I imported a 4 gig table and a count(*) utilizing a full table scan
> > against it executed in about 120 seconds. Other operations big IO
> > operations, (create datafiles, imports, intensive queries) ran slowly,
> > with top indicating about 80% IO wait time.
> >
> > I configured another database (single node) with identical
> > hardware/OS/Oracle/OCFS, utilizing EXT2 and the same query ran in 30
> > seconds (same execution plan) and everything was far quicker. For
> > kicks, I reformatted the datafile partition to OCFS and things were
> > back to running slow on this single node environment. OCFS seems to
> > be the weak link here, or at least my implementation of it.
> >
> > I must be missing something, but scouring the Metalink groups and
> > Google didn't come up with anything that's helped...
> >
> > Many thanks,
> > Benjamin
>
> I'm with Karsten but would be interested in what hardware is providing
> your shared storage? NetApp? EMC? Firewire? Hitachi? Other?
Received on Fri Mar 19 2004 - 14:29:23 CST