Thanks for the replies - I haven't checked out the doc links above yet, will take a look. I'm submitting jobs directly to the ORB listeners on either cluster node (localhost port on either). I'm taking a guess it seems to be time-based in that a bunch of jobs run on one node, then maybe after submitting another bunch of jobs (each submitted indepedently of the others) 10-20 minutes later they'll all run on the other.
[Message sent by forum member 'ocoro02' (ocoro02)]
http://forums.java.net/jive/thread.jspa?messageID=242531