8 Replies Latest reply on Mar 31, 2005 11:36 AM by clebert.suconic

    Hypothesis: JBoss is slow compared to BEA

    rupfsw

      Hi,
      We have spent 13 months to migrate from WL 5.1 to JBoss and is it almost production ready. However, the speed tests show that JBoss is about 2-3 times slower than BEA WL 5.1.

      I could put a lot of technical detail here, but right now my concern is if I should buy BEA 8.1 instead.

      If you have any test results then pls. share it in less than 10 lines for me to understand if this is a general experience or if we are just bad at tuning. The DB is Oracle and is the same for the tests.

      Looking forward to hear experience from you and perhaps if JBoss itself are planning to release something faster in the near future, which for our company could mean that we could go with Jboss AS.

      Thanks,
      Rasp

        • 1. Re: Hypothesis: JBoss is slow compared to BEA
          triathlon98

          You really need to give more technical details about the application and the kind of tunings you have already applied. Also check the wiki, forums and there is an entry in the blogs about tuning.

          When comparing, please compare using the same version JVM. Sun 1.5 is faster than 1.4, JRockit is supposed to be faster than Sun etc.

          Have you trimmed the JBoss configuration? What JVM switches have you appled (you should at least increase available memory) etc.

          Anothe important question is what classloader isolation you are using. There are huge performance benefits to be gained by not having to serialize invocations between layers in your application etc, etc.

          Joachim

          • 2. Re: Hypothesis: JBoss is slow compared to BEA

            Sounds like you may not have understood all of the JBoss configuration and fine-tuning information. I suggest you get in touch with sales and get yourself expert help to ensure a smooth transition to production.

            • 3. Re: Hypothesis: JBoss is slow compared to BEA
              rupfsw

              Thanks Joachim for your kind reply and suggestions and thanks Juha we are thinking of that.

              Our application is a big enterprise application where there can be upto 500 or more user transaction simultaneously.

              We have done the test using JMeter,Load Runner with 500 virtual users.

              Tuning
              We have done the best use of the wiki
              http://www.jboss.org/wiki/Wiki.jsp?page=JBossASTuningSliming
              But we are using JBoss 3.2.3. Do everybody think that JBoss 3.2.6 can be a best option.

              32bit & 64 bit
              At present we are using 32bit machines with Windows platform.
              How best it can be with 64 bit.Do anybody have some sample results.


              Trimming
              We have avoided unused configuration settings and war files.
              and other than default configuration we are using run -c all for clustering.

              VM
              We have tried jdk1.3,jdk1.4 and jrockit in which the performace of jrockit found faster,
              For weblogic 5.1 we have used jdk1.3 only.

              This is the vm configuration we applied for development purpose.
              -Xms386m -Xmx786m -XX:+UseParallelGC -XX:+UseAdaptiveSizePolicy
              The GC methods not found much useful.
              The same -Xms386m -Xmx786m we have applied for Weblogic 5.1
              We are testing with -Xms786m -Xmx1024m and we will post the results of JBoss3.2.3 vs WL5.1


              Class loader isolation
              Thanks for suggesting this option. We were yet to try optimisation here.
              to Joachim & Juha :If you are having more suggestion regarding this please let us know,
              So lets this topic be an interesting one for all JBoss fans.

              Thanks
              Rasp





              • 4. Re: Hypothesis: JBoss is slow compared to BEA
                triathlon98

                Something that can affect performance a lot (I was bitten by this one, but still not 100% sure I got it fixed) is jndi lookup performance.

                In our config it happened that jndi was sometimes dead slow (5 to 10s), and other times instantenious.

                This was presumably caused by the RMI stubs including server name in their data but the hostname not being known on the client side. Easily fixed either by forcing the server to only return the ip address, or by setting the hostname/ip in the hosts file on the client side.

                The easy way to check this is by doing a telnet to the server on port 1099.

                Joachim

                • 5. Re: Hypothesis: JBoss is slow compared to BEA
                  rupfsw


                  A jndi call can be made like this inside a deployment descriptor
                  <jndi-name>jnp://localhost:1099/HelloJndi</jndi-name>

                  Is this the way u mentioning.

                  Thanks
                  Rasp

                  • 6. Re: Hypothesis: JBoss is slow compared to BEA
                    triathlon98

                    no, I mean a lookup from client code using
                    ctx=new Initialcontext
                    ctx.lookup(foo)

                    Joachim

                    • 7. Re: Hypothesis: JBoss is slow compared to BEA
                      chrisdutz

                      At our university we did about 2 Years of performance analysis and comparisons based upon the SPECjAppServer2002 and 2004. The results all quite went in the same direction: At the moment it seems that the web-layer (Tomcat) is the bottleneck in all tests. Everything works fine untill you realy hit tomcat with work. I'll hope that since tomcat is now dealt with by the jboss guys that these problems will be resolved in the near futurre, since we managed to optimize JBoss to knock-out a lot of other commercial appservers (they all have problems with their web-layer) especially regarding stability on realy long high load tests, but it seems that the bea guys simply have a better web-container, since this was the only appserver we tested to scale allmost linear.

                      • 8. Re: Hypothesis: JBoss is slow compared to BEA
                        clebert.suconic

                        We are using Specj2004 now to optimize our container, and we are looking at this seriously. If you wanna help, you are more than welcome.

                        I'm technically cordinating Specj2004 efforts here at JBoss, and if you want a contact, I have my e-mail set at my profile.


                        Clebert