On Tue, 21 Aug 2007, ITAGAKI Takahiro wrote:
> Does anyone have a way to measure the performance difference by
> bgwriter_lru_xxx ? I have no performance results not only of the patch
> but also of those parameters. I'd like to use those test cases to compare
> manual and automatic tunings of lru parameters for 8.4.
The version of this patch I submitted at
http://archives.postgresql.org/pgsql-patches/2007-05/msg00142.php puts
statistics into the pg_stat_bgwriter structure so you can compare how much
work is being done by the background writer vs. the backends. Even if
there is no explicit change in the BW behavior, I would very much like to
see that part get committed so people can actually tune more easily by
hand using the stock PG in 8.3. Heikki didn't really like the way I
passed that data around internally, but I never got a suggestion for doing
it a better way I thought was an improvement.
Heikki's "Bgwriter strategies" thread used that to compare the various
approaches, which I did quite a bit of as well but didn't bother
publishing the results as they weren't any more interesting than his:
http://archives.postgresql.org/pgsql-hackers/2007-07/msg00144.php
I have a set of scripts to automate pgbench to collect this data
automatically and compare various settings. Now that I see all the big
patches are settling down, I'll collect all that up and pass it along.
I've been mulling over options here for a month now, and I'm not done with
this patch yet; I'll take Bruce's message as a call to urgent action to
finish and submit my final results ASAP.
--
* Greg Smith gsmith@gregsmith.com http://www.gregsmith.com Baltimore, MD