Here's float_sort, which requires a RightShift functor. For this functor RightShift(x, 0) >> n must equal RightShift(x, n), which is stricter than integer_sort is. float_sort sorts floats as integers, but flips the negative bins around to get the exact same sorting order as std::sort. Positives floats sort fine via integer_sort. A functor is required because of the cast operation; I don't see any generic way to code a general cast operation. The RightShift functor looks like this: struct rightshift { int operator()(const DATATYPE &x, const unsigned offset) const { return *((int *)&x) >> offset; } }; Note the cast. NaNs get dumped in the last bucket by default, without special-casing. std::sort dumps them randomly in the file because they evaluate as equivalent to everything, which is a little odd, and behavior I have no intention to emulate. In my testing, float_sort takes 61% of the time that std::sort takes on 40MB files, so the speedup is substantial. float_sort takes about 10% longer than integer_sort, while std::sort takes about 20% longer on floats than ints. I'll pretty this up a bit and add a version that takes a comparison functor, then move on to string::sort, which will take more work. That should complete the sorting library I intend to propose adding to Boost. Steve