This is based on a discussion with an advanced JPPF user, who pointed out a number of weaknesses and inconsistencies and possible enhancements in the way NodeExecutionManagerImpl handles the execution of the tasks within a node. Here's the list:

1) in the current version we have two maps that maintain the association between each task and the corresponding future for the node's executor:

protected final Map<Long, NodeTaskWrapper> taskMap = new Hashtable<Long, NodeTaskWrapper>();
protected final Map<Long, Future<?>> futureMap = new Hashtable<Long, Future<?>>();

The link is made by the common Long value. This is very inefficient and the 2 maps should be merged into a single one:

protected final Map<NodeTaskWrapper, Future<?>> futureMap = new IdentityHashMap<NodeTaskWrapper, Future<?>>();

This will also simplify the code maintaining the associations.

2) We use an Executor to execute the tasks but still use Object.wait() and Object.notify() to synchronize on the completion of a task bundle. This could be much simplified by using an ExecutorCompletionService on top of the executor and receiving the results asynchronously until they're all recieved.

3) The synchronization within NodeExecutionManagerImpl is too broad. Instead of using synchronized method we shoudl use more granular synchronization on specific objects on which cocurrency applies:

synchronized(futureMap) {
  // ... do some stuff ...
}