Content-Type: text/plain; charset="utf-8"
Date: Tue, 30 Oct 2012 07:55:54 -0400
From: "Kevin Grittner" <kgrittn@mail.com>
Message-ID: <20121030115554.306900@gmx.com>
MIME-Version: 1.0
Subject: Re: How to keep queries low latency as concurrency increases
To: "Catalin Iacob" <iacobcatalin@gmail.com>,pgsql-performance@postgresql.org
Content-Transfer-Encoding: 8bit

Catalin Iacob wrote:

> Hardware:
> Virtual machine running on top of VMWare
> 4 cores, Intel(R) Xeon(R) CPU E5645 @ 2.40GHz
> 4GB of RAM

You should carefully test transaction-based pools limited to around 8
DB connections. Experiment with different size limits.

http://wiki.postgresql.org/wiki/Number_Of_Database_Connections

> Disk that is virtual enough that I have no idea what it is, I know
> that there's some big storage shared between multiple virtual
> machines. Filesystem is ext4 with default mount options.

Can you change to noatime?

> pgbouncer 1.4.2 installed from Ubuntu's packages on the same
> machine as Postgres. Django connects via TCP/IP to pgbouncer (it
> does one connection and one transaction per request) and pgbouncer
> keeps connections open to Postgres via Unix socket. The Python
> client is self compiled psycopg2-2.4.5.

Is there a good transaction-based connection pooler in Python? You're
better off with a good pool built in to the client application than
with a good pool running as a separate process between the client and
the database, IMO.

>  random_page_cost | 2

For fully cached databases I recommend random_page_cost = 1, and I
always recommend cpu_tuple_cost = 0.03.

-Kevin