Message-ID: <4E6E519B.9050709@squeakycode.net>
Date: Mon, 12 Sep 2011 13:38:19 -0500
From: Andy Colson <andy@squeakycode.net>
User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64;
	rv:6.0.2) Gecko/20110902 Thunderbird/6.0.2
MIME-Version: 1.0
To: Robert Schnabel <schnabelr@missouri.edu>
CC: "sthomas@peak6.com" <sthomas@peak6.com>,
	pgsql-performance <pgsql-performance@postgresql.org>
Subject: Re: Allow sorts to use more available memory
References: <4E6E426D.7060405@missouri.edu> <4E6E45CA.20606@squeakycode.net>
	<4E6E4826.8050002@peak6.com> <4E6E4DFB.2070605@missouri.edu>
In-Reply-To: <4E6E4DFB.2070605@missouri.edu>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit

On 9/12/2011 1:22 PM, Robert Schnabel wrote:
>
> On 9/12/2011 12:57 PM, Shaun Thomas wrote:
>> On 09/12/2011 12:47 PM, Andy Colson wrote:
>>
>>> work_mem is not the total a query can use. I believe each step can
>>> use that much, and each backend can use it for multiple bits. So if
>>> you had two backends, each doing 2 sorts, you'd use 2*2 = 4 * 2GB =
>>> 8GB.
>> Exactly. Find a big query somewhere in your system. Use EXPLAIN to
>> examine it. Chances are, that one query has one or more sorts. Each one
>> of those gets its own work_mem. Each sort. The query have four sorts? It
>> may use 4*work_mem. On a whim a while back, I doubled our 8MB setting to
>> 16MB on a test system. During a load test, the machine ran out of
>> memory, swapped out, and finally crashed after the OOM killer went nuts.
>>
>> Set this value *at your own risk* and only after *significant* testing.
>> Having it too high can have rather unexpected consequences. Setting it
>> to 1 or 2GB, unless you have VERY few threads, or a TON of memory, is a
>> very, very bad idea.
>>
> Yep, I know. But in the context of the data warehouse where *I'm the
> only user* and I have a query that does, say 4 large sorts like
> http://explain.depesz.com/s/BrAO and I have 32GB RAM I'm not worried
> about using 8GB or 16GB in the case of work_mem = 4GB. I realize the
> query above only used 1.9GB for the largest sort but I know I have other
> queries with 1 or 2 sorts that I've watched go to disk.
>
> Bob
>
>
>
>

Wow, you are getting close to the limits there.  Another thing you can 
do is mount tmpfs in ram and then just let it spill.

-Andy