zfs discuss - Sep 2006 - Some questions about how to organize ZFS-based filestorage

Hi all,

I am trying to organize our small (and the only one) filestorage using and
thinking in ZFS-style )

So I have SF x4100 (2 x DualCore AMD Opteron 280, 4 Gb of RAM, Solaris 10 x86
06/06 64 bit kernel + updates),  Sun Fiber Channel HBA card (Qlogic-based) and
Apple Xraid 7Tb (2 raid controllers with 7 x 500 Gb ATA disks per each
controller). Two internal SAS drives are in RAID1 mode using built-in LSI
controller.

Xraid is confugured like the following - 6 disks in HW RAID 5 and one spare disk
per controller.

So I have :

# format
Searching for disks...done


AVAILABLE DISK SELECTIONS:
       0. c0t2d0 <DEFAULT cyl 8872 alt 2 hd 255 sec 63>
          /pci at 0,0/pci1022,7450 at 2/pci1000,3060 at 3/sd at 2,0
       1. c4t6000393000017312d0 <APPLE-Xserve RAID-1.50-2.27TB>
          /pci at 1,0/pci1022,7450 at 2/pci1077,10a at 1/fp at 0,0/disk at
w6000393000017312,0
       2. c5t600039300001742Bd0 <APPLE-Xserve RAID-1.50-2.27TB>
          /pci at 1,0/pci1022,7450 at 2/pci1077,10a at 1,1/fp at 0,0/disk at
w600039300001742b,0


I need a place to keep multiple builds of the products (a huge number of small
files). This will take about 2 Tb - so it''s quite logical to give the
whole "1." or "2." from the output above. What will be the
best block size that I need to supply to "zfs create" command to get
the most from filesystem that has a huge number of small files?

The other tank will host users'' homes, projects'' files and
other files.

Now I am thinking to create two separate ZFS pools. "1." and
"2." will be the only physical devices in both pools.

Or I''d better go and create one xfs pool that includes both
"1." and "2."?

Later on I will use NFS to share this filestorage between Linux, Solaris,
OpenSolaris and MacOSX hosts.
 
 
This message posted from opensolaris.org

Sergey wrote:
> I am trying to organize our small (and the only one) filestorage using and
> thinking in ZFS-style )
..
> AVAILABLE DISK SELECTIONS: 0. c0t2d0 <DEFAULT cyl 8872 alt 2 hd 255 sec
63>
>  /pci at 0,0/pci1022,7450 at 2/pci1000,3060 at 3/sd at 2,0 1.
c4t6000393000017312d0
> <APPLE-Xserve RAID-1.50-2.27TB> 
> /pci at 1,0/pci1022,7450 at 2/pci1077,10a at 1/fp at 0,0/disk at
w6000393000017312,0 2.
> c5t600039300001742Bd0 <APPLE-Xserve RAID-1.50-2.27TB> 
> /pci at 1,0/pci1022,7450 at 2/pci1077,10a at 1,1/fp at 0,0/disk at
w600039300001742b,0
> 
> 
> I need a place to keep multiple builds of the products (a huge number of
> small files). This will take about 2 Tb - so it''s quite logical to
give the
> whole "1." or "2." from the output above. What will be
the best block size
> that I need to supply to "zfs create" command to get the most
from
> filesystem that has a huge number of small files?
You''re not quite thinking ZFS-style yet. With ZFS you do not have to
worry about block sizes unless you want to - the filesystem handles
that for you.


cheers,
James C. McPherson
--
Solaris kernel software engineer, system admin and troubleshooter
               http://www.jmcp.homeunix.com/blog
Find me on LinkedIn @ http://www.linkedin.com/in/jamescmcpherson

Sergey wrote:
> What will be the
> best block size that I need to supply to "zfs create" command to
get
> the most from filesystem that has a huge number of small files?
You do not need to specify a recordsize.  Since your workload is not 
record-structured (ie. a database), you should not specify a recordsize.
> The other tank will host users'' homes, projects'' files
and other
> files.
> 
> Now I am thinking to create two separate ZFS pools. "1." and
"2."
> will be the only physical devices in both pools.
> 
> Or I''d better go and create one xfs pool that includes both
"1." and
> "2."?
It will most likely be better to create one zfs pool with both your 
devices.  This will provide better performance and easier management.
You should create different filesystems for each project, home 
directory, etc.
> Later on I will use NFS to share this filestorage between Linux,
> Solaris, OpenSolaris and MacOSX hosts.
That should work just fine.

--matt