Would you believe that running the same code on the same system can be up to 100% faster depending on your choice of the JDK you use? (And mind you, I am talking standard Oracle JDKs here, nothing experimental or expensive!)
A couple of weeks ago I decided to run my standard performance tests for the Pentaho Reporting engine. I just upgraded from an six year old computer to a shiny new big, fat machine and was itching to give it a run. All the software freshly installed, the filesystem sparkly clean, lets try and see how the various reporting engine versions run.
And lo and behold, Pentaho Reporting 4.8 seemed to outperform Pentaho Reporting 5.0 by a staggering 100%. Digging into it with JProfiler, I could not find a clear target – comparing the ratios of time spent in all major subsystems was the same between both versions – just that 5.0 was twice as slow at every task than 4.8.
After nearly two weeks of digging, I finally found the culprit: For some arcane reasons I was using a 64-bit JDK for the 4.8 test, but a 32-bit JDK for the 5.0 test run. Using the exact same JDK fixed it.
The numbers: JDK7 – 32bit vs 64bit vs server vs client vm
I have a standard test, a set of reports that print 100k elements in configurations of 5000 rows/20 elements per row, 10000 rows and 10 elements, 20000 rows and 5 elements and 50000 rows and 2 elements.
Here are the results of running Pentaho Reporting 5.0 in different JDK configurations. All times are in seconds.
|32bit / -client||3.75||4.31||5.52||9.203|
|32bit / -server||2.2||2.5||3.2||5.3|
|64bit / -client||1.92||2.2||2.8||4.75|
|64bit / -server||1.9||2.18||2.78||4.75|
Running Pentaho Reporting 4.8 in the same configuration yields no different results (appart from statistical noise). So with all the major work that went into Pentaho Reporting 5 at least we did not kill our performance.
So the lesson learned is: Choose your JVM wisely, and if you happen to use a 32-bit system or JDK, make sure you use the ‘-server’ command line switch for all your Java tools, or you wait twice as long for your data to appear.