Jignesh K. Shah wrote:
> I am running tests with PG8.3b2 on Solaris 10 8/07 and I still see IO
> flood when checkpoint happens.
> I have tried increasing the bg_lru_multiplier from 2 to 5 from default
> but I dont see any more writes by bgwriter happening than my previous
> test which used the default.
>
> Then I tried increasing checkpoint_completion_target=0.9 but still no
> spread of IO (checkpoint_timeout is set to default 5m)
>
> What am I missing?
Two things spring to mind:
Even though the write()s are distributed, the fsyncs are still going to
cause a spike. But it should be much less severe than before. How bad
are the spikes you're seeing? Compared to 8.2?
Are your checkpoints triggered by checkpoint_timeout or
checkpoint_segments? The calculation of "how much time do I have to
finish this checkpoint before the next one is triggered" takes both into
account, but the calculation wrt. checkpoint_segments is not very
accurate because of full page writes. Because of full page writes, we
write a lot more WAL right after checkpoint. That makes the load
distribution algorithm to think it's going to run out of
checkpoint_segments much sooner than it actually does. Raising
checkpoint_segments will help with that.
BTW, please turn on checkpoint_logging.
> How does PostgreSQL determine the Load distribution?
First, when the checkpoint starts, shared buffer pool is scanned, dirty
buffers are counted. Then bgwriter starts to write the buffers,
constantly estimating how much work it has left, and how much time it
has until next checkpoint. The write rate is throttled so that the
remaining writes are distributed evenly across the time remaining
(checkpoint_completion_target is a fuzz-factor applied to the estimate
of how much time is remaining).
Hope this helps..
-- Heikki Linnakangas EnterpriseDB http://www.enterprisedb.com
