Subject: Re: Questions on query planner, join types, and work_mem
From: Hannu Krosing <hannu@2ndquadrant.com>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: Robert Haas <robertmhaas@gmail.com>, Peter Hussey <peter@labkey.com>,
	pgsql-performance <pgsql-performance@postgresql.org>
In-Reply-To: <22961.1280944836@sss.pgh.pa.us>
References: <AANLkTikMwVZKDYLitnsK31R=VNOuTH8NtmDHDGkBR7k8@mail.gmail.com>
	<18551.1280275502@sss.pgh.pa.us>
	<AANLkTimXMR0jMdw+66E4Xx=SqNV37Ne9WfkFjPFZVjbV@mail.gmail.com>
	<25534.1280498632@sss.pgh.pa.us>
	<AANLkTik3oY1=F_uC843c84LxBLNdHxZaWdfjxrAtyBeu@mail.gmail.com>
	<1280819031.14579.38.camel@hvost>
	<AANLkTik6UBt2ezVFzdCM8iaUPkra_7f-Hs_AmGZgvFhh@mail.gmail.com>
	<1280944268.3735.12.camel@hvost>  <22961.1280944836@sss.pgh.pa.us>
Content-Type: text/plain; charset="UTF-8"
Date: Wed, 04 Aug 2010 21:41:27 +0300
Message-ID: <1280947287.3735.56.camel@hvost>
Mime-Version: 1.0
Content-Transfer-Encoding: 7bit

On Wed, 2010-08-04 at 14:00 -0400, Tom Lane wrote:
> Hannu Krosing <hannu@2ndquadrant.com> writes:
> > Of course there are more variables than just *_page_cost, so if you nail
> > down any other one, you may end with less than 1 for both page costs.
> 
> > I have always used seq_page_cost = 1 in my thinking and adjusted others
> > relative to it.
> 
> Right, seq_page_cost = 1 is sort of the traditional reference point,
> but you don't have to do it that way.  The main point here is that for
> an all-in-RAM database, the standard page access costs are too high
> relative to the CPU effort costs:
> 
> regression=# select name, setting from pg_settings where name like '%cost';
>          name         | setting 
> ----------------------+---------
>  cpu_index_tuple_cost | 0.005
>  cpu_operator_cost    | 0.0025
>  cpu_tuple_cost       | 0.01
>  random_page_cost     | 4
>  seq_page_cost        | 1
> (5 rows)
> 
> To model an all-in-RAM database, you can either dial down both
> random_page_cost and seq_page_cost to 0.1 or so, or set random_page_cost
> to 1 and increase all the CPU costs.  The former is less effort ;-)
> 
> It should be noted also that there's not all that much evidence backing
> up the default values of the cpu_xxx_cost variables.  In the past those
> didn't matter much because I/O costs always swamped CPU costs anyway.
> But I can foresee us having to twiddle those defaults and maybe refine
> the CPU cost model more, as all-in-RAM cases get more common.

Especially the context switch + copy between shared buffers and system
disk cache will become noticeable at these speeds.

An easy way to test it is loading a table with a few indexes, once with
a shared_buffers value, which is senough for only the main table and
once with one that fits both table and indexes,


> 			regards, tom lane


-- 
Hannu Krosing   http://www.2ndQuadrant.com
PostgreSQL Scalability and Availability 
   Services, Consulting and Training