The Swing SQE team has reported that their SwingBasher stress testing application crashes after about 12 hours of running. I've been examining the problem using SwingMark (which is similar and has some nice memory testing features). The problem appears to be that AWT leaks about 10k of native heap space per frame that is created.

Here is how to reproduce this:

1. Get SwingMark from: http://javaweb.eng/engineering/jdk-performance/swingmark.html
2. Run SwingMark with the following options:

java SwingMark -r 30 -m memout.txt

This will run SwingMark 30 times in the same VM. It will also track Java Heap usage and put the info into the memout.txt file. It measures the current heap size and usage after each run and writes it in XML format.

When SwingMark starts up, note the memory using a tool such as the NT Task Manager. After the first few runs (allowing heap size to stablize) note the total memory size. Now watch as it grows by about 10k (on average) each run.

When the run is complete you should look at the memout.txt (which shows that Heap size remains constant). This means the leak is on the native side.