DomainKey-Signature: a=rsa-sha1; c=nofws; d=gmail.com; s=gamma;
	h=mime-version:sender:in-reply-to:references:from:date
	:x-google-sender-auth:message-id:subject:to:cc:content-type
	:content-transfer-encoding;
	b=vXE9fci5fcINqV756g0ZCjYrypA+jjRN2k7Dsu37U34vgLMVxBBgwWGXVJvSHiAIbL
	0mlj2/VFJZP8WJrkLgjvfNl5dE83y61o/bAxB2m5x94EHoS0qFBcGw+1kuaKUPfDV91X
	Dl0dI6WlH//Qd4fnEWBiWOBDTJ7AEDCruIOPU=
MIME-Version: 1.0
In-Reply-To: <875976BE-CAC0-4F10-B7CC-0D651B109967@engineyard.com>
References: <4A2EA121.2070004@gmail.com>
	<875976BE-CAC0-4F10-B7CC-0D651B109967@engineyard.com>
From: Greg Stark <gsstark@mit.edu>
Date: Sat, 13 Jun 2009 00:25:19 +0200
Message-ID: <407d949e0906121525w1e0f13f0vf41bd9f63c032ebe@mail.gmail.com>
Subject: Re: Taking the cache out of the equation?
To: Erik Jones <ejones@engineyard.com>
Cc: Rob Sargent <robjsargent@gmail.com>, pgsql-sql@postgresql.org
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: quoted-printable

On Sat, Jun 13, 2009 at 12:12 AM, Erik Jones<ejones@engineyard.com> wrote:
>
> On Jun 9, 2009, at 10:51 AM, Rob Sargent wrote:
>
>> Caching helps a *lot* and I'm thankful for that but I would like to take
>> it out of the picture as I massage my queries for better performance.
>> =A0Naturally the first invocation of the query cannot take advantage of =
the
>> cache and these queries would normally only be called once for the same
>> target data. =A0 =A0What tricks are there to flush, ignore, circumvent t=
he
>> caching boost? =A0(Especially in the production environment.)
>
> Why on earth would you want your queries to always go to disk?

I think he answered that in the original message -- to better
represent the real workload.

Unfortunately there isn't really a good answer. On Linux you can echo
1 > /proc/sys/vm/drop_caches but that doesn't affect the postgres
shared buffers and worse, it does affect other buffers that probably
would still be cached.

The best answer is usually to build a test configuration large enough
that it has similar cache effects as your production environment. Then
test random values and repeat the test many times to avoid any random
fluctuations.

--=20
greg
http://mit.edu/~gsstark/resume.pdf