Thanks very much for your prompt response.
Either the JVM is tenuring objects immediately or the Eden-Survivor-Old cycle is so fast it does not get logged in the 5-second period jstat is collecting info. Also, I did not know how the JVM decides whether to do a full or minor collection - that explains a lot.
I have already taken heap dumps during "good" and "bad" runs, but have not been able to spot anything interesting. The "good" run has about 650 threads total, while the "bad" about 1000, and the main difference was a lot more Object.wait() threads during the "bad" run. Not much to work on, really.
Have also used jmap to create a histogram, but most of the memory seemed (not surprisingly) to be allocated to primitive arrays and other JDK classes, so no useful information there either. I was hoping to find a large chunk of the heap allocated to some of our application's objects, which would indicate a memory leak, but this was not the case.
Anyway, I'll change NewRatio, read the papers you suggested and post back when I have some results.
C.