It may also be useful to include a message queue (RabitMQ / SQS are good candidates) implemented using the pub/sub model to allow arbitrary scaling of the compute nodes.
I have some other ideas about something like JMX / MBeans here to allow arbitrary instrumentation of methods. This would provide more flexibility than the benchmarking module already included.