zfs discuss - Dec 2008 - performance issues with os200811 and zfs

I have been attempting to set up a test server using OS 200811 (rc2)

The system is a 4 Gbyte Dell Optiplex 755, which has dual core 2 duo cpus.
I created four 10 Gbyte files and two 20Gbyte files under /home using dd
from /dev/zero.  I then created two zpools; the first I created a stripe of
two of the 10Gbyte files and then mirrored it with the remaining two files,
the second pool I simply created a mirror of two 20 Gbyte files.

I then started creating zfs file systems in the pools.  The target was to
create about 5400 files systems in each pool.  My script ran for almost two
days and only managed to create 3300 files systems in each pool at which
point I killed it.  The system is rebooting, or rather I should say trying
to reboot as it has been rebooting for over an hourand a half now.

Some months ago I tried this using the then current Solaris 10 release on
the same system with some minor differences:

Instead of two pools I had five.  One of the pools received all (n) of the
files systems, the other 4 pools got file systems in a random allotment
resulting in each pool having on average, n/4 files systems.

the Solaris 10 system did not use zfs for the root file system, so the
underlying files were on a UFS slice.

I did not have smb sharing  or compression enabled for the file systems

On the Solaris 10 system I had no difficulty creating 15000 zfs file
systems, and the system was responsive and rebooted in reasonable times.

I really would like to use the two pool setup, compression and smb sharing
in production so I am really hoping that this degradation in performance is
due to some interaction of hosting the block files on a zfs file system.  In
production I will be using iscsi volumes and a dual 64bit Xeon system with 8
Gbytes of RAM and will be sharing the file systems with nfs4 and smb.

Before I scramble about looking for another drive I can use for a USF file
system to test my hypothesis I thought I would see if anyone has any
insights.

Thanks, Alastair
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://mail.opensolaris.org/pipermail/zfs-discuss/attachments/20081212/553ae9f6/attachment.html>

>I have been attempting to set up a test server using OS 200811 (rc2)
>
>The system is a 4 Gbyte Dell Optiplex 755, which has dual core 2 duo cpus.
>I created four 10 Gbyte files and two 20Gbyte files under /home using dd
>from /dev/zero.  I then created two zpools; the first I created a stripe of
>two of the 10Gbyte files and then mirrored it with the remaining two files,
>the second pool I simply created a mirror of two 20 Gbyte files.
If you use this trick on UFS, the performance will not be stellar;
however, you have created nearly continguous files.

But in ZFS, the data of the two files will be randomly located;
everytime you write to the files, the data will move around.
This is not at all what ZFS expects.

You should try again using a "zvol" by using zfs create -V.

Casper