On Wed, 2003-10-15 at 12:57, Tom Lane wrote:
> "Stephen" <jleelim@xxxxxx.com> writes:
> > Is it normal for plain VACUUM on large table to degrade performance by over
> > 9 times? My database becomes unusable when VACUUM runs. From reading
> > newsgroups, I thought VACUUM should only slow down by 10% to 15%.
>
> We have heard reports of very significant slowdowns from people who were
> already nearly saturating their disk I/O bandwidth, and then VACUUM
> pushed their systems over the knee of the response curve.
Makes sense but, wouldn't vacuum saturate the I/O bandwidth by
definition? Especially with modern CPU's when vacuum runs it's going to
be reading from disk as fast as the disk can possibly supply the data.
For this reason, I think the suggestion you made recently of putting in
a small delay in the main vacuum loop might be helpful.
I also acknowledge that SCSI makes a big difference here, but LOTS of
people run PG on cheap linux boxes with IDE drives, so if there is
something we can do to help this setup, it would be "a good thing" IMHO.