I'm thinking about using MPI for that - it should minimize the fork/exec time. Another approach is the extensive use of queues (which is a good thing anyway) and process the queue messages using a more optimized backend.