3 Replies Latest reply on Jan 22, 2009 12:16 PM by jesper.pedersen

    reasonable defaults for CORBA configuration?

    jhalliday

      I'm having issues when running org.jboss.test.jca.test.TxConnectionManagerStressTestCase with the CORBA based transaction service (JBossTS JTS). I can make the test pass, but to do so I need to mess with various combinations of:

      ulimit -s
      ulimit -u
      java -Xss
      jacorb.poa.queue_max
      jacorb.poa.thread_pool_max

      and even then I still need to turn down the log4j threshold or the disk writes slow things down to the point that the test times out.

      We already recommend JTS users increase jacorb.poa.thread_pool_max (from 8 to 32), but that alone is not sufficient to get the test to pass.

      So, my question is: is it ok for the test to fail with the default settings for the above parameters. If not, what default parameters can we reasonably ship JBossAS with given typical user hardware specs, and/or how far can we reasonably cut back the test before it becomes meaningless.

        • 1. Re: reasonable defaults for CORBA configuration?
          dimitris

          I suppose we can use a combination of the changes you propose: bump up the ORB defaults a little bit, cut back the stress test, or even alter the test execution based on the target (JTA vs JTS) doing some property checking.

          • 2. Re: reasonable defaults for CORBA configuration?
            jhalliday

            Also setting jacorb.poa.queue_wait=on may reduce the number of Threads we need in the pool to keep up with the test, which in turn may reduce the need to mess with -Xss and ulimit in order to avoid the 'java.lang.OutOfMemoryError: unable to create new native thread' problem.

            • 3. Re: reasonable defaults for CORBA configuration?
              jesper.pedersen

              I took a look at the test case in question.

              The success/failure has a lot to do with timing (which should be fixed) and the parameters used for the setup. The JTS setup shows > 10x longer running period than the JTA counterpart.

              So I would say that the *StressTestCase should be excluded for now from the run - and we should look into how better test cases could be implemented in this area.