How it works is that the top-level invocation runs the job server, and the children feed requests to it. The top-level controls the parallelism.