Re: Sun vs. Mac

On Tue, Jan 14, 2003 at 07:41:21AM -0800, CaptainX0r wrote:
>
> You're right, sorry.  PG 7.3.1 on Solaris 8.  I've got the
> default recommended /etc/system but with shmmax cranked way up

Ok, I have no experience with 7.3.1 in a production setting - we're
using 7.2.  But here are some things.

> which seems to have helped.  I don't have the system in front of
> me (and it's down, so I can't get to it), but from memory
> max_connections was increased to 64, shared_buffers up to 65536,
> sort_mem and vacuum_mem were doubled, and I think that's it.  I
> changed every seemingly relevant one, and spent a lot of time on

You'll need to increase the number of available semaphores more than
likely, if you add any connections.  You do indeed need to fix
shmmax, but if the postmaster starts, you're fine.

I would worry slightly about sort_mem.  I have managed to make
Solaris boxes with _lots_ of memory start swapping by setting that
too high (while experimenting).  Look for problems in your I/O.

> the *cost section trying various factors of n*10 on each, with
> no joy.

These are fine-tuning knobs.  You have a different problem :)

> > Did you split
> > the WAL onto its own filesystem?  You'll get a big win that
> > way.
>
> I have not.  What exactly do you by "own filesystem"?  Another
> filesystem?  I was planning on putting pg_xlog on the OS disk
> and moving $PGDATA off to a second disk.

That's what you need.  Without any doubt at all.  The xlog on the
same UFS filesystem (and disk) as the rest of $PGDATA is a nightmare.

Interestingly, by the way, there is practically _no difference_ if
you do this with an A5200 managed by Veritas.  I have tried dozens of
things.  It never matters.  The array is too fast.

> > Also, what fsync setting are you using (open_datasync is the
> > fastest in my experience).
>
> I've read that somewhere (maybe in the archives?) and I got no
> change with any of them.  But now I'm thinking back - do I need
> fsync=true for that to have an affect?  I'm not worried about
> the cons of having fsync=false at all - and I'm assuming that
> should be better than true and open_datasync.  Or am I confusing
> things?

Yes, if you change the fsync method but have fsync turned off, it
will make no difference.

> > Also, certain sort routines are abysmal.  Replace the
> > Solaris-provided qsort().
>
> I've read about this as well - but haven't even gotten that far
> on the testing/configuring yet.

If you're doing any sorting that is not by an index, forget about it.
Change it now.  It's something like a multiple of 40 slower.

A
--
----
Andrew Sullivan                         204-4141 Yonge Street
Liberty RMS                           Toronto, Ontario Canada
<andrew@libertyrms.info>                              M2P 2A8
                                         +1 416 646 3304 x110