On Sun, Feb 19, 2012 at 4:35 PM, Simon Riggs <simon@2ndquadrant.com> wrote:
> We don't need to wait until nobody has it set, we just need to wait
> for the people that had it set when we first checked to be out of that
> state momentarily.
I've just finished doing some performance analysis on various aspects
of hint bit setting for this patch.
I've seen as high as 14% of transactions writing full pages during a
pgbench run. That sounds quite bad, but on pgbench at least all of
those are associated with UPDATEs, which would dirty the page anyway,
so there aren't any more full page writes overall.
Checkpoints would be delayed only until a virtual transaction ends or
a virtual transaction comes out of DelayCkpt state. If a virtual
transaction was long running it wouldn't spend much time in the
delaying state, especially if we take into account I/O requirements.
So although I'm not exactly happy with the overheads, they don't seem
to be as big a problem as they sound. Plus this is all optional and
avoidable.
--
Simon Riggs http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services