How I would do it... I wouldn't do Catzilla... as was said already it is not consistent so that is a bad test as you do not know if changes you made were from core parking or run variance.
This is what I would do, I would test 2D (Wprime and SuperPi, multi threaded/single threaded) I would make 3 runs of each to get an average. Then unpark the cores, and retest the same way.
Try it on 3D benchmarks as well.. ones that respond to CPU (Vantage, 11, Firestrike)... same process. Here though, write down all scores.
EDIT: It should go without saying (saying it anyway), that you settings must be exactly the same across the board. Change ONLY the parking thing.
EDIT2: Again though, last I recall hearing about this, there were not many(any?) improvements in benchmarks. If it was a big deal, someone would have heard about it, and it also would more than likely be in the Little Black Book...that said, some worthwhile testing to put it to bed again would be great.