zfs discuss - Nov 2009 - uneven usage between raidz2 devices in storage pool

We had an existing zfs storage pool comprised of 2 raidz2 devices (or arrays)
and we just added a 3rd raidz2 device of equal size to the first two. While
making use of the extra capacity is mindlessly simple, we also want to take
advantage of the performance benefits.

How do we spread the data that is stored on the first two raidz2 devices across
all three so that when we continue to write data to the storage pool, we will
get the added performance of writing to all three devices instead of just the
empty new one?

Thanks and more details can be provided as needed.
-- 
This message posted from opensolaris.org

Erin wrote:
> How do we spread the data that is stored on the first two raidz2 devices
> across all three so that when we continue to write data to the storage
pool,
> we will get the added performance of writing to all three devices instead
of
> just the empty new one?
All new writes will be spread across the 3 vdevs.  Existing data stays where it
is for reading, but if you update it, those writes will be balanced across all 3
vdevs.  If you are mostly concerned with write performance, you don''t
have to do
anything.

Regards,
Eric

> All new writes will be spread across the 3 vdevs. Existing data stays where
it
> is for reading, but if you update it, those writes will be balanced across
all 3
> vdevs. If you are mostly concerned with write performance, you
don''t have to do
> anything.
> 
> Regards,
> Eric
The issue that we have is that the first two vdevs were almost full, so we will
quickly be in the state where all writes will be on the 3rd vdev. It would also
be useful to have better read performance, but I figured that solving the write
performance optimization would also help the read performance.

Thanks
-- 
This message posted from opensolaris.org

Erin wrote:
> The issue that we have is that the first two vdevs were almost full, so we
> will quickly be in the state where all writes will be on the 3rd vdev. It
would
> also be useful to have better read performance, but I figured that solving
the
> write performance optimization would also help the read performance.
Because of ZFS dynamic striping, there will likely be a bias toward the new vdev
for writes, but I don''t think there will ever be a case when writes
*only* go to
the new vdev.  Someone with deep internal ZFS knowledge could probably fill in
more details.

You didn''t mention how wide your raidz2 vdevs are, but I would imagine
that even
with a larger proportion of writes going to the new vdev, your overall write
performance (particularly on concurrent writes) will improve regardless.

Eric