Robert Haas <robertmhaas@gmail.com> wrote:
> Hmm. That looks a lot like a profile with no lock contention at
> all. Since I see XLogInsert in there, I assume this must be a
> pgbench write test on unlogged tables? How close am I?
Not unless pgbench on HEAD does that by default. Here are the
relevant statements:
$prefix/bin/pgbench -i -s 150
$prefix/bin/pgbench -T $time -c $clients -j $clients >>$resultfile
Perhaps the Intel cores implement the relevant primitives better?
Maybe I didn't run the profile or reports the right way?
> I was actually thinking it would be interesting to oprofile the
> read-only test; see if we can figure out where those slowdowns are
> coming from.
I'll plan on doing that this weekend.
>> tps = 21946.961196 (including connections establishing)
>> tps = 22911.873227 (including connections establishing)
>>
>> For write transactions, that seems pretty respectable.
>
> Very. What do you get without the patch?
[quick runs a couple tests that way]
Single run with -M simple:
tps = 23018.314292 (including connections establishing)
Single run with -M prepared:
tps = 27910.621044 (including connections establishing)
So, the patch appears to hinder performance in this environment,
although certainty is quite low with so few samples. I'll schedule
a spectrum of runs before I leave this evening (very soon).
-Kevin