Hard to say based on provided information. A series of thread dumps taken every, say, 10 seconds, while flying into the problem area, then sitting there for a minute or so, might be useful. I'm not necessarily volunteering to look at the output but that'd be a good first set of data to start...