Re: Planner constants for RAM resident databases

Emil Briggs wrote:

>I'm working with an application where the database is entirely resident in RAM
>(the server is a quad opteron with 16GBytes of memory). It's a web
>application and handles a high volume of queries. The planner seems to be
>generating poor  plans for some of our queries which I can fix by raising
>cpu_tuple_cost. I have seen some other comments in the archives saying that
>this is a bad idea  but is that necessarily the case when the database is
>entirely resident in RAM?
>
>Emil
>
>
>

Generally, the key knob to twiddle when everything fits in RAM is
random_page_cost. If you truly have everything in RAM you could set it
almost to 1. 1 means that it costs exactly the same to go randomly
through the data then it does to go sequential. I would guess that even
in RAM it is faster to go sequential (since you still have to page and
deal with L1/L2/L3 cache, etc). But the default random_page_cost of 4 is
probably too high for you.

John
=:->