Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
> I think that would be reasonable provided someone can come up
> with some appropriate wording. My understanding is that if you
> have a really small system then you might need >25% and if you
> have a really big system you might need <25%, but I'm not sure
> where the edges are.
Yeah, I remember a study which showed 40% as optimal, but I think
that was on a server with 2GB RAM, which is smaller than my desktop
workstation. (Heck, digital cameras with that much aren't that
rare.)
We might want to advise that if there are periods of irregular
response time, particularly if they are related to the checkpoint
cycle, if adjustments to the checkpoint and background writer
settings don't completely resolve it, they might want to try
reducing shared_buffers. The improvements to the checkpoint and
background writer areas in 8.3 helped a lot with this issue, but I
don't believe it's been totally eliminated (yet). Perhaps that's
now infrequent enough that it's not necessary to mention it. Dunno.
I guess I'd be interested to hear Greg Smith weigh in on this one.
I've generally stopped tweaking when our web support folks say I've
got it to the point where we're not getting any timeouts against our
20 second limit for queries which normally run in less than 1 ms.
-Kevin