[ale] Disk testing app wanted

[Neal Rhodes]

M-NPHW64<Jmills at motorola.com> wrote:
> ALErs -
>
> Is there a simple read/write test or script I can turn loose on a
suspect
> SATA drive and let it run? There's no data on the disk I care to
preserve.
>
> TIA.
> ?- Mills

Yes.  iozone is both a performance exerciser and a diagnostic benchmark.
It has a -d option which causes it to read after write and compare the
pattern it put down to what it read back. 

It also has dozens of write and read patterns to really beat the snot
out of the disk. 

We used it extensively a couple of years back when a new Compaq server
running Enterprise Redhat kept trashing our Progress DB.    Of course
Compaq blamed Progress.     We talked with the iozone developer at the
time and he either was in the process of finishing up the feature;
basically we were an early tester.   Guess what?  iozone's test barfed
on a 16GB test file.    Eventually we found that flavor of Linux was not
properly handling the buffers when the server had more than 4GB of RAM;
it put data in the buffers beyond 4GB; it just didn't write them
reliably to disk and eventually forgot about them.    The final straw
was when I created a directory for something, shutdown and restarted the
server, and the directory wasn't there after reboot.    When we cut the
server back to 4GB of RAM it worked fine. 

Now our SOP on building a new server is to always run 

        nohup iozone -a -g 16G -r 8 -+d &
        

on all filesystems of any new server before we put databases on it.
Note that since linux is buffering i/o, you should make the test file
size as large as you have time for.

Now, it sounds like you really suspect some specific portion of the
drive surface has become suspect.    If so, as other have suggested, the
smartctl diagnostics should show you records of prior sector failures
that have been mapped out.    That would be more useful than what I'm
suggesting.  It would take weeks for iozone to run a 250GB test on a
256GB drive. 

Neal Rhodes
MNOP Ltd
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.ale.org/pipermail/ale/attachments/20090819/04bac9e1/attachment.html