we have implemented a jBPM3 based system on a 4 node cluster driving roundabout 5000 long running processes a day. All nodes are using the same Oracle DB. We are driving the processes using asynchronous transacted JMS messages. So in case of crashes the JMS messages are not acknowledged and another node in our JBoss cluster will be responsible for the message. Because jBPM cannot handle concurrent updates we are doing pessimistic database locking on our main business object to garantee that only one node drives a process at certain moment in time.
Where are you located? I can also give you personal support on this topic.