Date: Sun, 23 Jul 2006 15:28:48 +1200
From: Mark Kirkwood <markir@paradise.net.nz>
Subject: Re: Forcing using index instead of sequential scan?
In-reply-to: <26827.1153585341@sss.pgh.pa.us>
To: Tom Lane <tgl@sss.pgh.pa.us>, robin.c.smith@bt.com
Cc: pgsql-performance@postgresql.org
Message-id: <44C2ECF0.1040701@paradise.net.nz>
MIME-version: 1.0
Content-type: text/plain; format=flowed; charset=ISO-8859-1
Content-transfer-encoding: 7bit
User-Agent: Thunderbird 1.5.0.4 (X11/20060704)
References: 
 <D012B9600FFB984AB18C0D53C475A4A081F51B@E03MVZ3-UKDY.domain1.systemhost.net>
	<26827.1153585341@sss.pgh.pa.us>

Tom Lane wrote:
> <robin.c.smith@bt.com> writes:
>> I have been testing the performance of PostgreSQL using the simple tool
>> found at http://benchw.sourceforge.net however I have found that all the
>> queries it run execute with sequential scans. The website where the code
>> runs has examples of the execution plan using indexes.
> 
> The reason the website gets indexscans is that he's fooled with the
> planner cost parameters.  In particular I see that...(snipped)
> 

Indeed I did - probably should have discussed that alteration better in 
the documentation for the test suite!

In addition I was a bit naughty in running the benchmark using size 1 
(i.e about 1G) an a box with 2G ram - as this meant that (on the machine 
I was using then anyway) indexscans on query 0 and 1 were *always* 
better than the sequential options.

A better test is to use the size factor at 2 x physical ram, as then the 
planners defaults make more sense! (unless or course you *want* to model 
a data mart smaller than physical ram).

Best wishes

Mark