Not really. I ran some numbers, and it was more or less between the clock and the LRU. The thing that makes it difficult to evaluate is that this algorithm effectively keeps the double number of entries, although it does not allow to access to all of them, but that messes with the comparison. So, I would say, if you have limited memory and your results are really big (just to make sure that data structures overhead is negligible), then it performs worse. Otherwise, it performs better.