Given that the overhead of an Akka actor is a few hundred bytes on the Java heap, 18 million actors would consume about 5-7 GiB of heap… do you know how much heap you’re running this test on? The default JVM ergonomics are typically a quarter of system RAM, so running on a system with fewer than 16 GB would exhaust the heap before spawning the 18+ million actors.
It is possible that the CL implementation (especially if it’s local-only) is able to garbage collect actors as they go out of scope. Akka Actors must be manually stopped (or the ActorSystem terminated).
This hypothesis can be tested by modifying FibonacciActor to add context.stop(self) after originalSender ! ("result", x+y).