[Lustre-discuss] Unbalanced load across OST's

[Brian J. Murrell]

On Thu, 2009-03-19 at 14:33 -0400, Aaron Everett wrote:
> Hello all,

Hi,

> We are running 1.6.6 with a shared mgs/mdt and 3 ost’s. We run a set
> of tests that write heavily, then we review the results and delete the
> data. Usually the load is evenly spread across all 3 ost’s. I noticed
> this afternoon that the load does not seem to be distributed.

Striping as well as file count and size affects OST distribution as
well.  Are any of the data involved striped?  Are you writing very few
large files before you measure distribution?

> OST0000 has a load of 50+ with iowait of around 10%
> 
> OST0001 has a load of <1 with >99% idle
> 
> OST0002 has a load of <1 with >99% idle

What does lfs df say before and after such a test that produces the
above results?  Does it bear out even use amongst the OST before, and
after the test?

> df confirms the lopsided writes:

lfs df [-i] from a client is usually more illustrative of use.  As I say
above, if you can quiesce the filesystem for the test above, do an lfs
df; lfs df -i before the test and after.  Assuming you were successful
in quiescing, you should see the change to the OSTs that your test
effected.

> OST0000:
> 
> Filesystem            Size  Used Avail Use% Mounted on
> 
> /dev/sdb1             1.2T  602G  544G  53% /mnt/fortefs/ost0

What's important is what it looked like before the test too.  Your test
could have, for example, wrote a single object (i.e. file) of nearly
300G for all we can tell from what you've posted so far.

b.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 197 bytes
Desc: This is a digitally signed message part
URL: <http://lists.lustre.org/pipermail/lustre-discuss-lustre.org/attachments/20090319/7e7805bd/attachment.pgp>