[PVFS2-developers] Underlying filesystem benchmarks

Nathan Poznick poznick at conwaycorp.net
Sat Aug 21 01:57:31 EDT 2004


A while back there was a discussion involving the strange server pauses,
which ended up being attributed to the fact that on some filesystems,
an fsync() is equivalent to a sync().  At the time I ran the test which
exhibited this behavior on a number of different filesystems.  I decided
today to do some very simple benchmarking of different filesystems when
used as the filesystem underlying the PVFS2 storage area.

Disclaimer:  These are unscientific tests and make no claims about the
tunability of the filesystems tested (things may be improved by tweaking
settings), the stability of them (I had no issues during my tests, but
they only lasted a few hours), nor the completeness of the testing that
I've done.

The filesystems tested:
ext2
ext3 (data=ordered)
ext3 (data=writeback)
ext3 (data=journal)
jfs
xfs
reiserfs
reiser4

These tests were run with the server on a Dell 2650 running a
2.6.8.1-mm1 kernel on SLES 9.  The client was a HP DL360 G2 running
Redhat 7.3, accessing the filesystem via the kernel module on a
2.4.18-19.xsmp kernel.

There was a single pvfs2-server configured, using defaults from
pvfs2-genconfig.  The storage space was set to be on a partition which
was recreated freshly for each test run.  On the server side, the
sequence of events went like:
mkfs
mount
create storage space
start server
(run test)
kill server
umount

On the clients, the sequence of events was,
load module
start client
mount
(run test)
umount
kill client
remove module

So basically, I tried to make sure that everything started from scratch
with each individual test run.  And now that the explanation is done,
here come the numbers!

For the first test, I used 'touch' through the kernel module on the
client to create 10,000 empty test files.

Results: http://wang-fu.org/pvfs2_create.gif

For the second test, I used dd with bs=16M to write a single 4GB file to
the PVFS2 filesystem.

Results: http://wang-fu.org/pvfs2_data.gif


-- 
Nathan Poznick <poznick at conwaycorp.net>

Because everyone uses language to talk, everyone thinks he can talk
about language. - Johann Wolfgang von Goethe

-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
Url : http://www.beowulf-underground.org/pipermail/pvfs2-developers/attachments/20040821/3a416f86/attachment.bin


More information about the PVFS2-developers mailing list