> Well, I'm about to try building a box with XFS over 6 IDE drives
> (using master+slave on two of the channels) doing software
> RAID-5. I want to test this setup as thoroughly as I possibly
> can, in as short a time period as I possibly can. (The IDE
> drives arrived today, and people want them in production by
> tomorrow morning, "if possible".)
>
> Anyone have any suggestions for tools to stress the system? To
> look for XFS+NFS and XFS+Samba bottlenecks and bugs? To make
> the drives fail as soon as possible, if they're so inclined?
>
> I've been using bonnie and iozone; any other suggestions?
>
> Thanks.
>
> Andrew Klaassen
For local access stress, run make in the cmd/xfstests directory (you will need
the various devel rpms installed, including rev 1.06 of the acl rpms if you
have a current cvs tree).
This will give you an fsstress binary in the xfstests/src directory,
doing something like:
fsstress -d /xfs-filesystem -p 8 -n 10000
will run 8 parallel threads doing 10000 random operations each in the
specified directory. You can bump the numbers, but you will chew up large
amounts of disk space.
Running xfsdump in parallel with nfs activity is a good combo. xfsdump will
access every file in the filesystem.
The reiserfs mongo benchmark will hammer a filesystem pretty well too.
As Robin Humble just pointed out, the NFS bandwidth is limited by the
network, so doing lots of local access in parallel is going to stress the
filesystem more, but there are paths which will only be exercised if you
use NFS.
Steve
|