In our case, we have a good reason to want to cause one - we have been seeing very occasional native OOM errors with part of our application. They are so rare that they are proving very difficult to diagnose. We would like to improve our monitoring to capture better data for diagnosis but to do so, we need to be able to test our monitoring.
Causing a Java heap OOM is pretty easy - just give your virtual machine a smallish heap size using the -Xmx switch and write some Java code that allocates lots of big Strings, populates them with random content (to defeat optimisations like interning) and stores them in a collection so that there is always a reference to them. As I've said previously, diagnosing these issues is a fairly well understood process, the usual solution being to get a heap dump using a switch such as -XX:+HeapDumpOnOutOfMemoryError and then using a tool such as Eclipse MAT to analyse the contents of the dump.
Causing a native OOM is less straightforward but there are ways to do it. A very comprehensive article on the subject of Native OOM errors can be found on the IBM site. This article also has some sample code which causes a native OOM by using some native code and JNI to allocate space in the native heap using malloc. It also suggests another way to chew up the native heap which doesn't need native code, so this is the way I did it. The java.nio.ByteBuffer class includes a method which allows the caller to allocate a 'direct' buffer which exists in the native heap. Here is a fragment of my code which does this...
private ByteBuffer createBuf() {
ByteBuffer bb = ByteBuffer.allocateDirect(BUFFSIZE);
int idx = 0;
while (idx < BUFFSIZE) {
Long l = new Long(r.nextLong()); // r is a java.util.Random
bb.putLong(idx,l.longValue());
idx += 1024;
}
return bb;
}
Call this enough times in a loop and keep the resulting references in a collection and you have created your own native memory leak. I found when testing this that I need to be careful to run it only in a 32 bit JVM on a machine with plenty of RAM, otherwise it won't hit the OOM error but it will use up all of your system's memory to the point where the system starts page thrashing. Don't try this on a production box!
Armed with this I was able to test questions like whether the -XX:+HeapDumpOnOutOfMemoryError switch will produce a heap dump for a native OOM. The answer is 'no', at least on Java 1.5.0_22.
Posts
