I'm happy to anounce that a paper called "Distributed Matrix Factorization with MapReduce using a series of Broadcast-Joins" written by me and my colleagues at TU Berlin has been accepted for publication at the ACM Conference on Recommender Systems 2013.
The paper discusses Mahout's latest distributed ALS implementation and contains experiments on the Netflix dataset (100M ratings), the Yahoo Music dataset (> 700M ratings) and a synthetic dataset generated from Netflix with 25M users and more than 5 billion ratings. Best, Sebastian