jBPM creates new Java classes out of binary process definition data stored in the database. This is needed for different versions of the process. It appears a special classloader is used for this. In SUN hotspot-class JVMs, parts of these load classes would be stored in perm gen rather than heap, which is not typically collected until absolutely necessary, potentially causing high cpu gc thrashing issues. If this was normal objects we were dealing with, the parallel scavenges or CMS fulls should collect them without problems. It appears this is not the case.

I believe the issue is that jBPM might be re-loading classes when not strictly necessary.