Tom,
> > I do have bgwriter_delay increased to 10, per previous
> > recommendation, which did smooth out the throughput graph
> > considerably. I can continue to adjust those settings.
>
> Please try a variety of settings and post your results. It would give
> us some hard data to help in deciding what the defaults ought to be.
> (I really doubt that the current defaults are reasonable...)
Doing a systematic progression test with bgwriter has been on my TODO list for
a while. Unfortunately, the Scalable Test Platform, which is designed for
short runs with no vaccuum isn't the best place for it. Mark and OSDL
allocated us a dedicated machine but I've been tied up with paid work for the
last 4 weeks and unable to finish the setup.
The idea is to have a machine that we can do 4-hour runs, with vacuum, of
different bgwriter, checkpoint, etc. settings. I've also built a 200GB DSS
database for doing some significant testing with DSS workloads after some of
the discussions around shared_buffers in November.
If anyone here has a great desire to write database-backed GUC test scripts
(pref in Perl) then please give me a buzz on the testperf project
(www.pgfoundry.org/testperf).
--
Josh Berkus
Aglio Database Solutions
San Francisco