I just read this updated paper on MDX performance hints. The section Calculation Non Empty Behavior is very interesting because it clarified me some doubt I had using the Non Empty Behavior property for a measure.
The lesson I learnt is that it is better to specify only one measure (or one single tuple) in Non Empty Behavior, instead of a list of measures. The paper suggests that a different (optimization?) behavior is implemented in SP2 against previous versions, but I didn’t found a detailed reference about applied changes. If someone has info, please let me know, thanks!