On Mon, 18 Aug 2008, Tom Lane wrote:
> the seldom-used log-each-transaction option, which pretty obviously
> is a drag on performance anyway
I always recommend that people run with log each transaction turned on,
beause it's the only way to gather useful latency information. I think
runs that measure TPS without even considering that are much less useful.
The fact that it's seldom used is something I consider a problem with
pgbench.
> I repeat my concern that transaction rates measured with this patch will
> be significantly different from those seen with the old code
Last time I tried to quantify the overhead of logging with timestamps on I
couldn't even measure its impact, it was lower than the usual pgbench
noise. That was on Linux. Do you have a suggestion for a platform that
it's a problem on that I might be able to use? I'd like to do some
measurements.
--
* Greg Smith gsmith@gregsmith.com http://www.gregsmith.com Baltimore, MD