I recently found out an interesting sorting algorithm called Smoothsort, originally created by the legendary comparion sorts. Naturally a thought came to me to study these and benchmark with STL sorting algorithms.
As you can see Leonardo number, a close cousin of the well-known Fibonacci number. Anyway please refer to Keith’s explanation for the ins and outs of this algorithm.
In contrast to the Smoothsort, which is mathematically sound and deep in theories (as Keith mentions, it’s hard for one(at least, for Keith and me) to imagine how Dijkstra originally came up with it), Timsort is rather a product of observations of real-word data and clever tricks to exploit common patterns found in them. Compared to Smoothsort, Timsort has a worse space usage of O(n), but it has a benefit of being a stable-sorter. Unfortunately, this algorithm is also far from simple. In my opinion, the best way to come to a good understanding of this is to read here.
Now, fun time. To see how these new kids stand their grounds against battle-tested standard sorting algorithms in STL, I adopted some benchmark code from the Timsort implementation above and modified a little. You can find the main code here.
Without furthur ado, here is the result:
|Partially sorted #0(int)||961||989||4699||2516||1478|
|Partially sorted #0(double)||1049||1098||5272||2774||1594|
|Partially sorted #1(int)||713||828||4798||2435||634|
|Partially sorted #1(double)||761||903||5077||2630||732|
- Data size: 100,000
- Data type: int or double
- Unit: miliseconds
- Partially sorted #0: each subarray of size 10 in 100,000 sequence is sorted
- Partially sorted #1: each subarray of size 1000 in 100,000 sequence is sorted
- The original implementation of Keith’s uses std:bitset. ‘Smoothsort(raw bit)’ is an modified one that uses raw bit operations instead
- Test hardware: Intel Core i7 920 / 6GB RAM
As you can see, both Timsort and Smoothsort didn’t cut the mustard. Smoothsort is worse than STL sorts in all cases(even with std:bitset replaced with raw bit operations). Timsort shows a trade-off. In random or almost random sequences, not as good as STL ones, but in partially sorted sequences like ‘Reversed’, ‘Sorted’ and ‘Partially sorted #1′, it shows impressive speed-up. Admittedly, apart from replacing an obvious culprit like std::bitset, I didn’t try any thorough optimization for each. So if you can see/suggest any optimization opportunities for both I missed, please leave a comment.
I was somewhat impressed with STL sorters at this point and found out I hadn’t known much about their internal implementions except the foggiest idea that it would be a variant of quick sort. So I digged into the source(STL in VS2012 RC, specifically). As you know, there are two, one which doesn’t maintain the relative order of records with equal keys(unstable) and the other which keeps it(stable).
std::sort is basically a quick sort with a median of medians algorithm to decide a partition pivot. Once the sequence becomes small enough(< 32), it switches to an insertion sort. Or if the recursion becomes too deep(> 1.5log2(N)), then it switches to a heap sort.
std::stable_sort is not a quick sort. It’s a sort(pun intended) of a merge sort. First, it insertion-sorts each chunk of size 32 and merge them hierarchically. Initially, it tries to get a temp storage of the half size of the original sequence and use it when merging. If the allocation fails, then it tries a smaller size, but this means more recursions and merges, so slower, of course. In the sense that it is a combo of merge sort and insertion sort, one can say it’s similiar to Timsort in essence, although the latter has much more complex tricks up its sleeve.
Codes for std::sort/std::stable_sort are relatively easy to follow(especially in comparison with understanding Smoothsort or Timsort), so I strongly recommend to take stock of them, if you haven’t done before.
- Asymptotic performance is not a whole story at all. The constant factor matters (an obvious thing, but still worth repeating)
- Timsort can be a viable alternative when data are not so random
- Smoothsort, though mathematically clever, doesn’t cut the mustard
- std::sort/std::stable_sort is pretty good in most of the cases
- For small data, insertion sort is very good. So it’s a good strategy to mix it with other algorithms to devise a good hybrid sorting algorithm