Learning Curve of Netflix Prize

After couple days experiments, I thought my "Basic" SVD hit the limitation, but I still have some questions for SVD-like model approach for Netflixe prize.
1.Why in my experiment, 100 features didn't get the better score than 32 features, is it over-fitting(100 features)? or some of my parameters didn't set well for 100 features?
2.How to decide how many feature is best for different problem, how to set the best prediction score of different parameter(feature, learning rate, K) setting? or just try try and try?

By PragmaticTheory's link, it is time to move on kNN algorithm.
My "Basic" SVD-like model's best score is RMSE=0.9137 w/ 32 features, K=0.015, Lrate=0.01