Re: 9.4 regression

On Wed, Aug 7, 2013 at 4:18 PM, Kevin Grittner <kgrittn(at)ymail(dot)com> wrote:
> Thom Brown <thom(at)linux(dot)com> wrote:
>
>> pgbench -j 80 -c 80 -T 3600
>>
>> 269e78: 606.268013
>> 8800d8: 779.583129

I have also been running some tests and - as yet - they are
inconclusive. What I can say about them so far is that - at times -
they agree with these results and do show a performance hit.

I took the settings as posted and adjusted them slightly for my much
lower-powered personal laptop, changing checkpoint_completion_target
to 1.0 and checkpoint_timeout to 1min.

I am testing with the latest git HEAD, turning fallocate support on
and off by editing xlog.c directly. Furthermore, before each run I
would try to reduce the number of existing WAL files by making a "pre"
run with checkpoint_segments = 3 before changing it to 32.

For reasons I don't entirely understand, when WAL files are being
created (rather than recycled) I found a performance hit, but when
they were being recycled I got a slight performance gain (this may be
due to reduced fragmentation) but the gain was never more than 10% and
frequently less than that.

I can't explain - yet (if ever!) - why my initial tests (and many of
those done subsequently) showed improvement - it may very well have
something to do with how the code interacts with these settings.

--
Jon