One thing that might explain the discrepancy in the total memory numbers could be that if you leave the Compute drop down at default it will average the metrics across all nodes in the cluster. Make sure you're selecting just the driver node when you want to see metrics for just that node.