Re: 8.4 open item: copy performance regression?

On Sat, 20 Jun 2009, Simon Riggs wrote:

> I would suggest that we check how much WAL has been written. There may
> be a secondary effect or a different regression hidden in these results.

What's the easiest way to do that?  My first thought was to issue a 
checkpoint before the test (which is a good idea to make each test 
consistent anyway), save the output from pg_controldata, test, checkpoint, 
and look at the control data again.  This seems kind of clunky though, but 
still better than trolling through the OS statistics for the data.  Any 
clever ideas for a better way to measure bytes of WAL written during a 
particular chunk of code?  We may need some sort of checkpoint/sync after 
the test to get correct results, because I've noticed that the tests I run 
sometimes continue writing out buffers for a few seconds after the test 
time is finished.

--
* Greg Smith gsmith@gregsmith.com http://www.gregsmith.com Baltimore, MD