[rescue] Trimming down Solaris 10

[Francois Dion]

On Sun, Dec 7, 2008 at 3:50 AM, Meelis Roos <mroos at linux.ee> wrote:
>> like failmode. Also if running the storage for oracle on a SAN,
>> disabling the flush cache command (or having the SAN ignore it) is a
>> good idea, as long as the cache is truely protected.
>
> Why would that be good idea - what replaces the integrity and ordering
> guarantees?

ZFS was designed to deal with single disks or jbods and as such, to
ensure data integrity, it doesn't consider the hard disk cache a
stable storage, so to make sure it is commited to the platter, ZFS
issues a cache flush on every transaction.

This is the safe way and right way when dealing with single disks or
jbods. Most other OSes/FS dont do that and performance might be higher
in this config, until of course you get in a scenario where your power
supply dies etc.

There is one case where it is not advised to flush cache to disk.
Where a single disk might have 8MB of cache, a SAN might have 4GB of
cache or more. So that in itself would affect performance drastically.
The other part of it is that the SAN cache _is_ stable storage. It is
backed up with batteries. If that is the case, and the SAN doesn't
already ignore the flush cache command (some do, some dont), then you
can either config the SAN to ignore the command, or add set
zfs:zfs_nocacheflush = 1 to /etc/system.

Having said that, I'm told that Solaris 10 U6 doesn't need this
anymore, but I'll have to research this.

Francois