You can resolver the Spill to memory by increasing the shuffle partitions, but 16 GB of spill memory should not create a major impact of your job execution. Could you share more details on the actual source code that you are running?