In-Place Radix Sort

Question

This is a long text. Please bear with me. Boiled down, the question is: Is there a workable in-place radix sort algorithm?

---

Preliminary

Answer 1

First, think about the coding of your problem. Get rid of the strings, replace them by a binary representation. Use the first byte to indicate length+encoding. Alternatively, use a fixed length representation at a four-byte boundary. Then the radix sort becomes much easier. For a radix sort, the most important thing is to not have exception handling at the hot spot of the inner loop.

OK, I thought a bit more about the 4-nary problem. You want a solution like a Judy tree for this. The next solution can handle variable length strings; for fixed length just remove the length bits, that actually makes it easier.

Allocate blocks of 16 pointers. The least significant bit of the pointers can be reused, as your blocks will always be aligned. You might want a special storage allocator for it (breaking up large storage into smaller blocks). There are a number of different kinds of blocks:

• Encoding with 7 length bits of variable-length strings. As they fill up, you replace them by:
• Position encodes the next two characters, you have 16 pointers to the next blocks, ending with:
• Bitmap encoding of the last three characters of a string.

For each kind of block, you need to store different information in the LSBs. As you have variable length strings you need to store end-of-string too, and the last kind of block can only be used for the longest strings. The 7 length bits should be replaced by less as you get deeper into the structure.

This provides you with a reasonably fast and very memory efficient storage of sorted strings. It will behave somewhat like a trie. To get this working, make sure to build enough unit tests. You want coverage of all block transitions. You want to start with only the second kind of block.

For even more performance, you might want to add different block types and a larger size of block. If the blocks are always the same size and large enough, you can use even fewer bits for the pointers. With a block size of 16 pointers, you already have a byte free in a 32-bit address space. Take a look at the Judy tree documentation for interesting block types. Basically, you add code and engineering time for a space (and runtime) trade-off

You probably want to start with a 256 wide direct radix for the first four characters. That provides a decent space/time tradeoff. In this implementation, you get much less memory overhead than with a simple trie; it is approximately three times smaller (I haven't measured). O(n) is no problem if the constant is low enough, as you noticed when comparing with the O(n log n) quicksort.

Are you interested in handling doubles? With short sequences, there are going to be. Adapting the blocks to handle counts is tricky, but it can be very space-efficient.