On Wed, Jan 21, 2009 at 10:15 AM, Albe Laurenz <laurenz.albe@wien.gv.at> wrote:
>
> Tom Lane suggested in
> http://archives.postgresql.org/pgsql-general/2004-02/msg00471.php
> that it might be better to set shared_buffers "relatively
> small" and let the filesystem cache do the buffering, so that's
> another way you can go. His advice is usually good.
Note that for transactional databases that are too large to fit the
dataset into memory (think a 16Gig machine running a 100G
transactional database) this is very true. The shared_buffers seldom
get reused and you'll get faster throughput with lower shared_buffers.
In testing a 40Gig db on a 32Gig machine, I got the best pgbench
numbers with a shared_buffers setting in the hundreds of megs range.