MergeShuffle: A Very Fast, Parallel Random Permutation Algorithm
arxiv.org
arxiv.org
The number of random bits used by both algorithms is within a few percent of the theoretically optimal value (4.6% more for Rao-Sandelius and 5.5% more for MergeShuffle) so I don't think that's really a significant concern.
Does anyone know of a practical use for a very fast, parallelizable shuffle algorithm that uses few random bits? All the shuffling I've done has used small enough N that Fisher-Yates was just fine.
According to the table of experimental results, the algorithm described in this paper and two of its competitors can all do better than n * log2(n) bits.