To: Jesper Krogh <jesper@krogh.cc>
cc: pgsql-performance@postgresql.org
Subject: Re: planner costs in "warm cache" tests 
In-reply-to: <4C040488.30906@krogh.cc> 
References: <4C02A342.40801@krogh.cc> <14644.1275244491@sss.pgh.pa.us>
	<4C040488.30906@krogh.cc>
Comments: In-reply-to Jesper Krogh <jesper@krogh.cc>
	message dated "Mon, 31 May 2010 20:48:40 +0200"
Date: Mon, 31 May 2010 15:55:33 -0400
Message-ID: <18310.1275335733@sss.pgh.pa.us>
From: Tom Lane <tgl@sss.pgh.pa.us>

Jesper Krogh <jesper@krogh.cc> writes:
> On 2010-05-30 20:34, Tom Lane wrote:
>> Well, hmm, I really doubt that that represents reality either.  A page
>> access is by no means "free" even when the page is already in cache.
>> I don't recall anyone suggesting that you set these numbers to less
>> than perhaps 0.01.
>> 
> Thank you for the prompt response. Is it a "false assumption" that the
> cost should in some metric between different plans be a measurement
> of actual run-time in a dead-disk run?

Well, the default cost parameters (seq_page_cost=1, random_page_cost=4)
are intended to model the non-cached state where most page fetches
actually do require a disk access.  They are definitely too large
relative to the cpu_xxx_cost parameters when you have a fully-cached
database, but what I've seen people recommending for that condition
is to set them both to the same value in the vicinity of 0.1 or 0.01
or so.  If it's only mostly cached you might try intermediate settings.

			regards, tom lane