Kevin Brown <kevin@sysexperts.com> writes:
> You don't have any real confidence in pgbench. Fair enough. What
> *would* you have confidence in?
Measurements on your actual application?
In fairness to pgbench, most of its problems come from people running
it at tiny scale factors, where it reduces to an exercise in how many
angels can dance on the same pin (or, how many backends can contend to
update the same row). And in that regime it runs into two or three
different Postgres limitations that might or might not have any
relevance to your real-world application --- dead-index-row references
used to be the worst, but I think probably aren't anymore in 7.3.
But those same limitations cause the results to be unstable from run
to run, which is why I don't have a lot of faith in reports of pgbench
numbers. You need to work quite hard to get reproducible numbers out
of it.
No, I don't have a better benchmark in my pocket :-(
regards, tom lane