Re: Swappiness setting on a linux pg server

On Thu, Oct 19, 2006 at 06:53:49PM +0200, Tobias Brox wrote:
> [Jim C. Nasby - Thu at 11:45:32AM -0500]
> > > > The issue with pg_xlog is you don't need bandwidth... you need super-low
> > > > latency. The best way to accomplish that is to get a battery-backed RAID
> > > > controller that you can enable write caching on.
> > >
> > > Sounds a bit risky to me :-)
> >
> > Well, you do need to understand what happens if the machine does lose
> > power... namely you have a limited amount of time to get power back to
> > the machine so that the controller can flush that data out. Other than
> > that, it's not very risky.
>
> We have burned ourself more than once due to unreliable raid controllers
> ...

Well, if you're buying unreliable hardware, there's not much you can
do... you're setting yourself up for problems.

> > quantities of memory. So in your case, 600M wouldn't be pushing things
> > much at all. Even 1G wouldn't be that out of the ordinary. Also remember
> > that the more memory for shared_buffers, the less for
> > sorting/hashes/etc. (work_mem)
>
> What do you mean, a high value for the shared_buffers implicates I
> can/should lower the work_mem value? Or just that I should remember to
> have more than enough memory for both work_mem, shared_buffers and OS
> caches? What is a sane value for the work_mem? It's currently set to
> 8M.

The key is that there's enough memory for shared_buffers and work_mem
without going to swapping. If you're consuming that much work_mem I
wouldn't worry at all about OS caching.

What's reasonable for work_mem depends on your workload. If you've got
some reporting queries that you know aren't run very concurrently they
might benefit from large values of work_mem. For stats.distributed.net,
I set work_mem to something like 2MB in the config file, but the nightly
batch routines manually set it up to 256M or more, because I know that
those only run one at a time, and having that extra memory means a lot
of stuff that would otherwise have to spill to disk now stays in memory.
--
Jim Nasby jim(at)nasby(dot)net
EnterpriseDB http://enterprisedb.com 512.569.9461 (cell)