[Ocfs2-users] ocfs2 slow write performance on Linux 2.6.38

[fibreraid at gmail.com]

Hello all,

I am benchmarking OCFS2 in a single node environment to see how its
performance stacks up against other Linux file systems. My hardware is
dual CPU, 6-cores per CPU, 2.4GHz Westmere with 24GB RAM. My system
has 24 HDD's of which 23 are active drives in a md RAID 5 array with
64K chunk size, and one hot-spare. I have a LVM2 volume created out of
this md array. It is running Linux 2.6.38 kernel.

I am creating the OCFS2 file system as follows:

/sbin/mkfs.ocfs2 -N 2 -M local
--fs-features=sparse,refcount,xattr,indexed-dirs,discontig-bg,inline-data
-F /dev/md0/ocfs2max

I am then mounting it as:

/bin/mount /volumes/md0/ocfs2max

I am using fio 1.55 to measure performance. My benchmark script uses
uncached IO in 4K and 4M block sizes. I am running 8 fio jobs and 512
IO depth. I am using a 20GB file as the test file for this benchmark.

OCFS2 reads (both at 4M and 4K) are very good...within +-3% of the
other major Linux file systems. However, 4K writes are substantially
lower, about half the speed of other Linux FS's.

Can you provide some recommendations on how I can improve the write
speed with OCFS2? At the current write rate, the trade-off is too
much, but I really like the other features of OCFS2 so I am hopeful
that some tuning/tweaking will get the performance up where it needs
to be.

Thank you very much in advance!

- TG