On Tue, 2007-03-13 at 17:17 +0000, Simon Riggs wrote:
> On Tue, 2007-03-13 at 10:08 -0700, Jeff Davis wrote:
> > On Tue, 2007-03-13 at 12:53 -0400, Tom Lane wrote:
> > > Jeff Davis <[EMAIL PROTECTED]> writes:
> > > > I agree that ss_report_loc() doesn't need to report on every call. If
> > > > there's any significant overhead I agree that it should report less
> > > > often. Do you think that the overhead is significant on such a simple
> > > > function?
> > >
> > > One extra LWLock cycle per page processed definitely *is* a significant
> > > overhead ... can you say "context swap storm"? I'd think about doing it
> > > once every 100 or so pages.
> > >
> > No lock is needed to store the hint. If somehow the hint (which is
> > stored in a static table, no pointers) gets invalid data due to a race
> > condition, the new scan will simply consider the hint invalid and start
> > at 0.
> > I did this precisely to avoid causing a performance regression for usage
> > patterns that don't benefit from sync scans.
> Shared memory access is still a performance/scalability concern because
> so many people want access to it at the same time.
> There really is no need to do this after each block. 8 CPUs ought to be
> able to do 8 scans without tripping over each other. Especially if they
> are on separate tables.
Ok, I'll do it every 100 pages.
---------------------------(end of broadcast)---------------------------
TIP 6: explain analyze is your friend