Re: [tlug] Apache MPM performance

["Dimitar Dimitrov" <dimitar.dimitrov@example.com>]

Hi Ngyen,

As you noted, the main advantage of the UltraSPARC CPUs is that you
can run a lot of parallel tasks with low power consumption for
reasonable price (for CPUs that can run linear code fast, you need to
look at NetBurst). In your case, as you are using ab, which is a
single-threaded benchmark you are not testing any of the advantages of
the MDM module or the machine you bought.

See the following two articles for explanations:

http://weblogs.java.net/blog/sdo/archive/2007/03/ab_considered_h.html
http://weblogs.java.net/blog/sdo/archive/2007/04/simple_benchmar.html

Ideally you should use something like Testmaker (pushtotest.com) or
OpenSTA  to create test scripts for your own site and expected
scenarios. You also want to use a couple of load-generating machines
(different than the server) on a switched network. If you can arrange
for multiple network cards on the serve - do it. Make sure that you
warm the caches before running the real benchmark. Use the tools to
benchmark and graph the change in response time relative to the
variation of the number of users from each type. It can easily take
you a day or two (measuring scalability has always been difficult).

Solaris 10 does have vmstat; no valgrind (at least on my machine);
check DTrace - the built in performance measurement tool for Solaris10
there are a number of tutorials and cool scripts on the Internet.

cheers,
Dimitar