Solaris ZFS and cooked files

All,

Has anyone any experience of using ZFS with cooked files? I have some systems where I use UFS and mount it with FORCEDIRECTIO to bypass the unix buffering (easier than setting up SVM) which works well ans wondered how ZFS compared with this


BTW ZFS only exists in Solaris 10 06/06 or later




Regards

Colin


There are 10 types of people in the world, those that understand binary and those that don't


__________________________________________________ _______________
Catch up on all the latest celebrity gossip
http://clk.atdmt.com/GBL/go/115454061/direct/01/

On Oct 23, 2:05*am, Colin Dawson wrote:
> All,
>
> Has anyone any experience of using ZFS with cooked files? I have some systems where I use UFS and mount it with FORCEDIRECTIO to bypass the unix buffering (easier than setting up SVM) which works well ans wondered how ZFS compared with this
>
> BTW ZFS only exists in Solaris 10 06/06 or later
>
> Regards
>
> Colin
>
> There are 10 types of people in the world, those that understand binary and those that don't
>
> __________________________________________________ _______________
> Catch up on all the latest celebrity gossiphttp://clk.atdmt.com/GBL/go/115454061/direct/01/

We initially wanted to go with ZFS due to ease of admin, we got very
inconsistent performance. Sometimes it was lightning quick, other
times our checkpoint times were minutes not sub second. To cut short,
we decided it was too unstable and were having to many i/o related
problems as well as convincing ourselves we were having CPU spikes.

We then tried UFS which was much better and always consistent but not
as good in performance as it should have been - I left the Unix people
alone so they knew we must be onto something.
Somehow I managed to pester the Unix guys to try raw (which we did
raid 5, 1+0, 1 unfortunately not in active/active we found out the
hard way). We now use LOFS loopback filesystem for our solaris 10
containers and it seems to work fine. 25% better than cooked (but we
didn't have the advantage of direct i/o). KIO seems to make it easier
to manage AIO vps though (which if a lot of fun when using Sun's T2
chips).
One major trick seems to be getting the correct stripe / segment size
for what you are doing - 64kb in our case (32k was better in theory
but not real world). Only problem is that now and again the fabric
saturates if I try really hard.

Obviously try it and see, but I won't go near ZFS again even if
performance isn't an issue.

My current disk conclusions - raid 1 if active/active SAN or 1+0, play
around with stripe size and buy a load of 15k 147gb drives before you
can't get em anymore - more spindles seems to be good no matter how
tempting 400gb/1tb SAS may appear. And stick with EMC, they are not
completely clueless when tuning SAN's / support in Aus anyhow, which
probably saves a lot.