Build a Spark job generating a lot of TMP data. The idea is to bring the cluster to a limit and see if nm-local-dir will bring better performance.