On 04.01.2012 07:58, Peter Geoghegan wrote:
> As part of the ongoing effort to reduce wake-ups when idle/power
> consumption, the attached patch modifies the background writer to
> hibernate in ten second bursts once the bgwriter laps the clock sweep.
> It's fairly well commented, so a description of how it works here
> would probably be redundant. The most controversial aspect of this
> patch is the fact that it adds a SetLatch() call to MarkBufferDirty(),
> though I went to some lengths to ensure that that call will very
> probably find the latch already set when it actually matters, so it
> only checks a single flag.
I think SetBufferCommitInfoNeedsSave() needs the same treatment as
MarkBufferDirty(). And it would probably be good to only set the latch
if the buffer wasn't dirty already. Setting a latch that's already set
is fast, but surely it's even faster to not even try.
> Thoughts? I can produce benchmarks, if that helps, but I think it's
> fairly unlikely that the patch introduces a measurable performance
> regression.
Yeah, I'd like to see a micro-benchmark of a worst-case scenario. I'm a
bit worried about the impact on systems with a lot of CPUs. If you have
a lot of CPUs writing to the same cache line that contains the latch's
flag, that might get expensive.
-- Heikki Linnakangas EnterpriseDB http://www.enterprisedb.com