On Wed, Mar 23, 2005 at 11:16:29AM -0600, Brandon Metcalf wrote:
>
> We moved from an HP-UX 10.20 box where the pgsql installation and data
> were on a vxfs fileystem.
My best guess, then, is that ufs tuning really is your issue. We
always used vxfs for our Sun database servers (which was a nightmare
all on its own, BTW, so I don't actually recommend this), so I don't
have any real ufs tuning advice.
The Packer Solaris database book (Packer, Allan N., _Configuring &
Tuning Databases on the Solaris Platform_. Palo Alto: Sun
Microsystems P, 2002. ISBN 0-13-083417-3) does suggest mounting the
filesystems with forcedirectio; I dimly recall using this for the wal
partition on one test box, and STR that it helped. Also, you want to
make sure you use the right fsync method; if it's still set to
"fsync" in the config file, you'll want to change that. I remember
finding that fsync was something like 3 times slower than everything
else. I don't have any more Solaris boxes to check, but I believe we
were using open_datasync as our method. You'll want to run some
tests.
You also should enable priority paging, but expect that this will
give you really strange po numbers from vmstat and friends. Priority
paging, I found, makes things look like you're swapping when you
aren't. Procmem is useful, but if you really want the goods on
what's going on, you need the SE toolkit. Just be careful using it
as root -- in some cases it'll modify kernel parameters behind the
scenes. In my case, I didn't have superuser access, so there wasn't
a danger; but I've heard sysadmins complain about this.
A
--
Andrew Sullivan | ajs@crankycanuck.ca
This work was visionary and imaginative, and goes to show that visionary
and imaginative work need not end up well.
--Dennis Ritchie
