zfs discuss - Dec 2008 - zfs_nocacheflush, nvram, and root pools

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

hi,

i have a system connected to an external DAS (SCSI) array, using ZFS.  the
array has an nvram write cache, but it honours SCSI cache flush commands by
flushing the nvram to disk.  the array has no way to disable this behaviour.  a
well-known behaviour of ZFS is that it often issues cache flush commands to
storage in order to ensure data integrity; while this is important with normal
disks, it''s useless for nvram write caches, and it effectively disables
the
cache.

so far, i''ve worked around this by setting zfs_nocacheflush, as
described at
[1], which works fine.  but now i want to upgrade this system to Solaris 10
Update 6, and use a ZFS root pool on its internal SCSI disks (previously, the
root was UFS).  the problem is that zfS_nocacheflush applies to all pools,
which will include the root pool.

my understanding of ZFS is that when run on a root pool, which uses slices
(instead of whole disks), ZFS won''t enable the write cache itself.  i
also
didn''t enable the write cache manually.  so, it _should_ be safe to use
zfs_nocacheflush, because there is no caching on the root pool.

am i right, or could i encounter problems here?

(the system is an NFS server, which means lots of synchronous writes (and
therefore ZFS cache flushes), so i *really* want the performance benefit from
using the nvram write cache.)

	- river.

[1]
http://www.solarisinternals.com/wiki/index.php/ZFS_Evil_Tuning_Guide#Cache_Flushes
-----BEGIN PGP SIGNATURE-----

iD8DBQFJNRJVIXd7fCuc5vIRAgDlAJ0boVf5zmvkRySeIHVumsKm3VSVhACffyOK
POEMyzG8U2yQYeZr01uJ74Q=9eBp
-----END PGP SIGNATURE-----

On 12/02/08 03:47, River Tarnell wrote:
> hi,
> 
> i have a system connected to an external DAS (SCSI) array, using ZFS.  the
> array has an nvram write cache, but it honours SCSI cache flush commands by
> flushing the nvram to disk.  the array has no way to disable this
behaviour.  a
> well-known behaviour of ZFS is that it often issues cache flush commands to
> storage in order to ensure data integrity; while this is important with
normal
> disks, it''s useless for nvram write caches, and it effectively
disables the
> cache.
> 
> so far, i''ve worked around this by setting zfs_nocacheflush, as
described at
> [1], which works fine.  but now i want to upgrade this system to Solaris 10
> Update 6, and use a ZFS root pool on its internal SCSI disks (previously,
the
> root was UFS).  the problem is that zfS_nocacheflush applies to all pools,
> which will include the root pool.
> 
> my understanding of ZFS is that when run on a root pool, which uses slices
> (instead of whole disks), ZFS won''t enable the write cache itself.
i also
> didn''t enable the write cache manually.  so, it _should_ be safe
to use
> zfs_nocacheflush, because there is no caching on the root pool.
> 
> am i right, or could i encounter problems here?
Yes you are right and this should work. You may want to check that
the write cache is disabled on the root pool disks
using ''format -e'' + cache + write_cache + display.
> 
> (the system is an NFS server, which means lots of synchronous writes (and
> therefore ZFS cache flushes), so i *really* want the performance benefit
from
> using the nvram write cache.)
Indeed, performance would be bad without it.
> 
> 	- river.
Neil.