Although it might be better to make a thread pool, where each Arc-side thread is a "task" (which will reschedule itself until it ends), much like how Erlang does it and how SNAP will eventually do it. This will allow you to easily launch and destroy thousands of Arc-side threads on a machine with far fewer cores.
Of course there's the minor problem of I/O then... you'll have to use nonblocking I/O in the VM and emulate blocking I/O on the Arc-side by setting up a poll for the Arc-side task.
