Since your CPU is at 100%, there is little that you can do in the
appserver to tune things. I assume you've done the recommended things
for basic glassfish production tuning (using the server compiler,
setting the request-processing threads to a suitable value for your
platform, using proper GC settings) -- if not, look for Jeanfrancois
Arcand or Scott Oaks blogs on that subject.
But I suspect your bigger issue is with openwebload. There are a couple
of reasons why it may not be the best load generator for you. First,
none of its requests use keepalive, so really what you're measuring is
the speed of the TCP stack on your OS; the TCP stack handles the
connections. You might check various OS groups on how to adjust OS
parameters for a system that accepts lots of connections, which may
improve your CPU usage somewhat.
The second problem with openwebload is that it is a single-threaded,
no-think-time architecture (at least as far as I can tell -- though I
could be wrong about that). If so, it's just like Apache's ab load
generator, which is not really suitable for throughput measurements. For
a full explanation of why not, see
http://weblogs.java.net/blog/sdo/archive/2007/03/ab_considered_h.html
-Scott
On Mon, 2008-02-25 at 02:57, glassfish_at_javadesktop.org wrote:
> hello,
>
> so my small test-case is doing the following (sorry i don't have the code here):
>
> - JSP takes username and password and calls a "validate" method on a SLSB
> - SLSB validates the given strings against hardcoded values and sends back a status (accept,decline)
>
> for testing i use the performance tool openwebload (http://openwebload.sourceforge.net) which gives the possibility to simulate concurrent access by more than one users.
>
> i ramp up from 300 to 600 concurrent users and notice the maximum TPS of about 400 - 450 TPS.
>
> another thing i realized is that the cpu on the server machine is utilized 100% during testing but the memory usage does not increase with the amount of users (stays at 50%). imho that is a very strange behaviour.
>
> if you want i can post the code (which i think is not that interesting) and the detailed openwebload output later when i am at home again.
>
> thanks for help,
>
> mario
> [Message sent by forum member 'the666pack' (the666pack)]
>
> http://forums.java.net/jive/thread.jspa?messageID=260669
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: users-unsubscribe_at_glassfish.dev.java.net
> For additional commands, e-mail: users-help_at_glassfish.dev.java.net
>