When running a jar-based job, I've noticed that the 1st run always takes the extra time to complete the job and consecutive runs take less time to finish the job. This behavior is reproducible on an interactive cluster. What's causing this? Is this expected?