In the last days, I was able to run some syntetical load tests on a Sparc V490 with 4 dual-core cpus, having 16GB or RAM and running Solaris 9. As the target system will be JBoss 3.2.6, we tested against this one.
Tests were conducted with JDK 1.4 and JDK 5.0, having a separate machine firing http requests at the JBoss.
JBoss was rock solid, even when the test app first had an error, not removing stateful session beans, thus passivating thousands of beans.
As expected, garbage collector settings have a huge impact on the overall performance. With bad settings we were only able to saturate two cpus, while with the right ones we kept all 8 cores busy.
Also JDK 5.0 performs a lot better than JDK 1.4 (which was expected anyway).
We also found out, that top on solaris has a bug, only displaying at most 999 lwp, even if a process has much more of them.
For someone usually only having one to two cpus at his hands, this was a nice occasion to play with a "big iron" (yeah, I know that this still is a small box).