@Christian to mitigate the effects of the hotspot compiler I always made sure to execute the test multiple times (easy enough since it is a karaf command) and only used the number once it stabilized (typically after the 3rd run with 5000 invocations).
The only exception was ECF generic, where it did not stabilize for yet to clearify reasons. There the best run was usually the 2nd, so I used that one. Our test is passing a random String[] as the parameter and receiving a random String[] as the return value to emulate some payload. Given that SOAP is less compact than binary RMI I think 50% of the RMI performance is actually a great result and as you can see CXF was a lot faster in our test than the other implementations (except for the redhat one). I have to look up the exact number again, but I think it was in the area of 2200ms for 5000 invocations in 5 threads, or 2272 calls/s which - depending on the payload - is not that far from your result of 4000 calls/s - for which you used 20 threads if I understood that correctly. As already promised to Scott, I will push the test to a github repo as soon as possible. -- View this message in context: http://karaf.922171.n3.nabble.com/Proposal-Lightweight-standalone-remote-OSGi-implementation-for-karaf-cellar-tp4045343p4045404.html Sent from the Karaf - Dev mailing list archive at Nabble.com.
