On Fri, 9 Mar 2007, Jim Nasby wrote:
> I'm wondering if pg_bench is a good test of this stuff. ISTM it's
> unrealistically write-heavy, which is going to tend to not only put a
> lot of dirty buffers into the pool, but also keep them pinned enough
> that you can't write them.
Whether it's "unrealistically" write-heavy kind of depends on what your
real app is. The standard pgbench is a bit weird because it does so many
updates to tiny tables, which adds a level of locking contention that
doesn't really reflect many real-world situations. But the no-branch mode
(update/select to accounts, insert into history) isn't too dissimilar from
some insert-heavy logging applications I've seen.
The main reason I brought this all up was because Itagaki seemed to be
using pgbench for some of his performance tests. I just wanted to point
out that the LRU background writer specifically tends to be very
underutilized when using pgbench.
--
* Greg Smith gsmith@gregsmith.com http://www.gregsmith.com Baltimore, MD