On Tue, Oct 13, 2015 at 2:32 AM, Heikki Linnakangas <hlinnaka@iki.fi> wrote:
> 80% of the CPU time is spent in the b-tree comparison function.
In the logs, my duration per COPY command increases from about 1400ms
for one process to about 3800ms when I have four running concurrently.
That's really my only data point, unfortunately. Strace isn't super
helpful because it just says 70-ish% of the time is wait4, but that's
not significantly different than the results using one process.
Everything else is bog standard default. I bumped up
checkpoint_segments to avoid checkpoints during the test, and
increased work_mem and maintenance_work_mem to avoid disk affecting
results, and still got the same behavior.
I wasn't blaming the patch. :) I thought it didn't get merged or
something, and after learning that wasn't the case, my only theory
went out the window. I'm a bit stuck, because this seems incredibly
wrong. I'll keep digging.
--
Shaun Thomas
bonesmoses@gmail.com
http://bonesmoses.org/