Daniel Kalchev <daniel@digsys.bg> writes:
>>> I don't think I'd go as far as to lower random_page_cost to 1.0, but
>>> certainly there's a case for using an intermediate value.
> The question is: how does one find the proper value? That is, is it
> possible to design planner benchmarking utility to aid in tuning
> Postgres?
The trouble is that getting trustworthy numbers requires huge test
cases, because you have to swamp out the effects of the kernel's own
buffer caching. I spent about a week running 24-hour-constant-disk-
banging experiments when I came up with the 4.0 number we use now,
and even then I didn't feel that I had a really solid range of test
cases to back it up.
My advice to you is just to drop it to 2.0 and see if you like the plans
you get any better.
regards, tom lane