Corion kindly advised my not to use Thread::Semaphore and pointed me to ... Thread::Queue.

I finally got around to putting together proof of the wisdom of Corion's advice.

This is a version of your Thread::Semaphore code, limiting to 4 concurrent threads and calculating factorials 1000! .. 2000!:

Total runtime: Took 47.511607 seconds

This version does the same calculations using the same number of concurrent threads, but ditches Thread::Semaphore in favour of Thread::Queue to queue the 1000 numbers to 4 reused threads, thus saving the startup and teardown costs of 996 threads:

Total runtime: Took 31.290966 seconds; giving a 33% saving of time.

But the biggest lesson of threading, is when not to use it. This version ditches threads altogether and uses the obvious optimisation:

Total runtime: Took 0.685503 seconds. Same results, but a 98.5% time saving over the first version above.