Re: Finding bottleneck

"Merlin Moncure" <merlin.moncure@rcsonline.com> writes:
> One thing that might interest you is that the penalty in 8.1 for
> stats_command_string=true in this type of access pattern is very high: I
> was experimenting to see if the new cpu efficiency gave me enough of a
> budget to start using this.  This more than doubled the cpu load to
> around 70% with a runtime of 82 seconds.  This is actually worse than
> 8.0 :(.

That seems quite peculiar; AFAICS the pgstat code shouldn't be any
slower than before.  At first I thought it might be because we'd
increased PGSTAT_ACTIVITY_SIZE, but actually that happened before
8.0 release, so it shouldn't be a factor in this comparison.

Can anyone else confirm a larger penalty for stats_command_string in
HEAD than in 8.0?  A self-contained test case would be nice too.

            regards, tom lane