From: Date: February 1 2005 9:14am
Subject: Some experiments
List-Archive: http://lists.mysql.com/cluster/1443
Message-Id: <49185D4B-7429-11D9-913F-000A959312A2@mysql.com>
MIME-Version: 1.0 (Apple Message framework v619)
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: quoted-printable

I made some experiments to see how one could improve performance of=20
communication between MySQL Server and
the ndbd processes. In the process I stumbled over a performance=20
improvement that makes a very simple select query
fetching one record from ndbd via primary key and 2 integer attributes=20=

execute 40% faster (only accounts for time spent
in MySQL Server and ndbd, including client will decrease improvement in=20=

percent). This will be pushed into all
branches this week.

In addition I experimented with various priority settings. The=20
communication cost was decreased by 10% when I raised the
priority of the MySQL Server process (in the experiment it was actually=20=

an NDB API test program but this should not make a
big difference). This plays well  with experiments made a couple of=20
years ago on Solaris 7 where I was able to increase
performance by 50% by setting the API part in real-time scheduling mode.

The major reason for these improvements is due to the thread model and=20=

with a higher priority on the MySQL Server part, there
are less context switches.

The experiments were made on a Mac OS X laptop with all processes on=20
localhost. The results is only a hint of that a few percent
improvements can be derived through simple means. However before=20
actually used in practice they should be tested and
verified on the HW and OS used.

Rgrds Mikael
PS: I mainly used the renice shell command (renice -10 pid increases=20
the priority of process with number pid where -20 <=3D prio <=3D 20)

Mikael Ronstr=F6m, Senior Software Architect
MySQL AB, www.mysql.com

Jumpstart your cluster:
http://www.mysql.com/consulting/packaged/cluster.html