Message-ID: <4492D882.6030109@kaltenbrunner.cc>
Date: Fri, 16 Jun 2006 18:12:50 +0200
From: Stefan Kaltenbrunner <stefan@kaltenbrunner.cc>
User-Agent: Thunderbird 1.5.0.2 (X11/20060516)
MIME-Version: 1.0
To: Bruce Momjian <pgman@candle.pha.pa.us>
CC: Qingqing Zhou <zhouqq@cs.toronto.edu>,  pgsql-hackers@postgresql.org
Subject: Re: Test request for Stats collector performance improvement
References: <200606161603.k5GG3MS27858@candle.pha.pa.us>
In-Reply-To: <200606161603.k5GG3MS27858@candle.pha.pa.us>
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: 7bit

Bruce Momjian wrote:
> OK, based on reports I have seen, generally stats_query_string adds 50%
> to the total runtime of a "SELECT 1" query, and the patch reduces the
> overhead to 25%.

that is actually not true for both of the platforms(a slow OpenBSD
3.9/x86 and a very fast Linux/x86_64) I tested on. Both of them show
virtually no improvement with the patch and even worst it causes
considerable (negative) variance on at least the Linux box.

> 
> However, that 25% is still much too large.  Consider that "SELECT 1" has
> to travel from psql to the server, go through the
> parser/optimizer/executor, and then return, it is clearly wrong that the
> stats_query_string performance hit should be measurable.
> 
> I am actually surprised that so few people in the community are
> concerned about this.  While we have lots of people studying large
> queries, these small queries should also get attention from a
> performance perspective.
> 
> I have created a new test that also turns off writing of the stats file.
> This will not pass regression tests, but it will show the stats write
> overhead.

will try to run those too in a few.


Stefan