To: Scott Carey <scott@richrelevance.com>
cc: Jesper Krogh <jesper@krogh.cc>,
	"pgsql-performance@postgresql.org" <pgsql-performance@postgresql.org>
Subject: Re: planner costs in "warm cache" tests 
In-reply-to: <31D5250B-7392-45DD-A234-8A5BBFF8F49A@richrelevance.com> 
References: <4C02A342.40801@krogh.cc> <14644.1275244491@sss.pgh.pa.us>
	<4C040488.30906@krogh.cc> <18310.1275335733@sss.pgh.pa.us>
	<31D5250B-7392-45DD-A234-8A5BBFF8F49A@richrelevance.com>
Comments: In-reply-to Scott Carey <scott@richrelevance.com>
	message dated "Tue, 01 Jun 2010 00:13:22 -0700"
Date: Tue, 01 Jun 2010 10:03:38 -0400
Message-ID: <2193.1275401018@sss.pgh.pa.us>
From: Tom Lane <tgl@sss.pgh.pa.us>

Scott Carey <scott@richrelevance.com> writes:
> It is still best to have random_page_cost to be slightly larger (~50%)
> than sequential_page_cost, because even when entirely in RAM,
> sequential reads are faster than random reads.  Today's CPU's do
> memory prefetching on sequential access.

Do you have any actual evidence of that?  Because I don't believe it.
Neither PG nor any kernel that I've ever heard of makes any effort to
ensure that logically sequential blocks occupy physically sequential
buffers, so even if the CPU tries to do some prefetching, it's not
going to help at all.

Now, if the database isn't entirely cached, then indeed it's a good
idea to keep random_page_cost higher than seq_page_cost.  But that's
because of the actual disk fetches, not anything that happens in RAM.

			regards, tom lane