There is a new interesting paper Flat-Combining NUMA Locks at David Dice's blog. The new algorithm improves performance in case of high contention by better exploitation of data locality. Lock handover is not done randomly but preferably between threads with the same locality (but avoiding starvation and fairness issues). Great reading.
Trivia - One of the co-authors is Nir Shavit. He is also co-author of the brilliant The Art of Multiprocessor Programming. If I had not read the book I would have strugled with the above paper much more...
