Tom Lane wrote:
> Alexander Staubo <alex@purefiction.net> writes:
> > No, fsync=on. The tps values are similarly unstable with fsync=off,
> > though -- I'm seeing bursts of high tps values followed by low-tps
> > valleys, a kind of staccato flow indicative of a write caching being
> > filled up and flushed.
>
> It's notoriously hard to get repeatable numbers out of pgbench :-(
>
> A couple of tips:
> * don't put any faith in short runs. I usually use -t 1000
> plus -c whatever.
> * make sure you loaded the database (pgbench -i) with a scale
> factor (-s) at least equal to the maximum -c you want to test.
> Otherwise you're mostly measuring update contention.
> * pay attention to when checkpoints occur. You probably need
> to increase checkpoint_segments if you want pgbench not to be
> checkpoint-bound.
While skimming over the pgbench source it has looked to me like it's
necessary to pass the -s switch (scale factor) to both the
initialization (-i) and the subsequent (non -i) runs. I'm not sure if
this is obvious from the documentation but I thought it may be useful to
mention.