On Tue, 2007-07-24 at 00:58 -0400, Tom Lane wrote:
> I wrote:
> > (BTW, in case you can't tell from the drift of my questions, I've
> > separated the patch into "add background wal writer" and "add async
> > commit", and am working on the first half.)
>
> I've committed the first half of that.
Cool
> Something that still needs
> investigation is what the default wal_writer_delay should be. I left
> it at 200ms as submitted, but in some crude testing here (just running
> the regression tests and strace'ing the walwriter) it seemed that I had
> to crank it down to 10ms before the walwriter was really doing the
> majority of the wal-flushing work.
There are two things to consider here.
If you are running solely async commit then walwriter should be at a
somewhat higher setting. The default is set at the "works well on crappy
hardware" value, for which my laptop is a good simulation. 50ms or below
reduces benefit considerably.
If you are not running async commits then walwriter can provide some
form of group commit. To do that you need to wind the time down. I think
that's what your seeing now.
My hope is to formalise that in the next release, so that walwriter can
autotune and to allow group commit
--
Simon Riggs
EnterpriseDB http://www.enterprisedb.com