Tomasz Ostrowski <tometzky@batory.org.pl> writes:
> I'd also set
> log_checkpoints=on
> to get an idea how it behaves.
Yeah, that's really the *first* thing to do. You need to determine
whether the episodes of slowness are correlated with checkpoints
or not; there's no point fooling with the checkpoint parameters if
not.
Another thing you could try in order to get a handle on what's
happening is to watch the output of "vmstat 1" or a similar tool,
and see if there are spikes in CPU or I/O load when things are
slow. If not, I'd next look into PG's pg_locks view to see if
queries are getting blocked on locks.
regards, tom lane