|Worst-case space complexity|
Timsort is a hybrid stable sorting algorithm, derived from merge sort and insertion sort, designed to perform well on many kinds of real-world data. It uses techniques from Peter McIlroy's "Optimistic Sorting and Information Theoretic Complexity", in Proceedings of the Fourth Annual ACM-SIAM Symposium on Discrete Algorithms, pp. 467–474, January 1993. It was implemented by Tim Peters in 2002 for use in the Python programming language. The algorithm finds subsequences of the data that are already ordered, and uses that knowledge to sort the remainder more efficiently. This is done by merging an identified subsequence, called a run, with existing runs until certain criteria are fulfilled. Timsort has been Python's standard sorting algorithm since version 2.3. It is also used to sort arrays of non-primitive type in Java SE 7, on the Android platform, and in GNU Octave.
Timsort was designed to take advantage of runs of consecutive ordered elements that already exist in most real-world data, natural runs. It iterates over the data collecting elements into runs, and simultaneously merging those runs together. When there are runs, doing this decreases the total number of comparisons needed to fully sort the list.
Timsort iterates over the data looking for natural runs of at least two elements that are either non-descending (each element is greater than or equal to its predecessor) or strictly descending (each element is less than its predecessor). Descending runs are later blindly reversed, so the strict order maintains the algorithm's stability; i.e., equal elements won't be reversed. Note that any two elements are guaranteed to be either descending or non-descending.
A reference to each run is then pushed onto a stack.
Minimum size (minrun)
Timsort is an adaptive sort, using insertion sort to combine runs smaller than the minimum run size (minrun), and merge sort otherwise.
Minrun is selected so most runs in a random array are, or become, minrun in length. It also results in a reasonable number of function calls in the implementation of the sort.
Because merging is most efficient when the number of runs is equal to, or slightly less than, a power of two, and notably less efficient when the number of runs is slightly more than a power of two, Timsort chooses minrun to try to ensure the former condition.
Minrun is chosen from the range 32 to 64 inclusive, such that the size of the data, divided by minrun, is equal to, or slightly less than, a power of two. The final algorithm takes the six most significant bits of the size of the array, adds one if any of the remaining bits are set, and uses that result as the minrun. This algorithm works for all arrays, including those smaller than 64; for arrays of size 63 or less, this sets minrun equal to the array size and Timsort reduces to an insertion sort.
Concurrently with the search for runs, the runs are merged with merge sort. Except where Timsort tries to optimise for merging disjoint runs in galloping mode, runs are repeatedly merged two at a time, with the only concerns being to maintain stability and merge balance.
Stability requires non-consecutive runs are not merged, as elements could be transferred across equal elements in the intervening run, violating stability. Further, it would be impossible to recover the order of the equal elements at a later point.
In pursuit of balanced merges, Timsort considers three runs on the top of the stack, X, Y, Z, and maintains the invariants:
- |Z| > |Y| + |X|
- |Y| > |X|
If the invariants are violated, Y is merged with the smaller of X or Z and the invariants are checked again. Once the invariants hold, the next run is formed.
Somewhat inappreciably, the invariants maintain merges as being approximately balanced while maintaining a compromise between delaying merging for balance, and exploiting fresh occurrence of runs in cache memory, and also making merge decisions relatively simple.
On reaching the end of the data, Timsort repeatedly merges the two runs on the top of the stack, until only one run of the entire data remains.
Timsort performs an almost in-place merge sort, as actual in-place merge sort implementations have a high overhead. First Timsort performs a binary search to find the location in the first run of the first element in the second run, and the location in the second run of the last element in the first run. Elements before and after these locations are already in their correct place, and may be removed from further consideration. This not only optimises element movements and running time, but also allows the use of less temporary memory. Then the smaller of the remaining runs is copied into temporary memory, and elements are merged with the larger run, into the now free space. If the first run is smaller, the merge starts at the beginning; if the second is smaller, the merge starts at the end.
Say, for example, two runs A and B are to be merged, with A being the smaller run. In this case a binary search examines A to find the first element aʹ larger than the first element of B. Note that A and B are already sorted individually. When aʹ is found, the algorithm can ignore elements before that position when merging. Similarly, the algorithm also looks for the first element bʹ in B greater than the last element of A. The elements after bʹ can also be ignored when merging. This preliminary searching is not efficient for highly random data, but is efficient in other situations and is hence included.
An individual merge keeps a count of consecutive elements selected from the same input set. The algorithm switches to galloping mode when this reaches the minimum galloping threshold (min_gallop) in an attempt to capitalise on sub-runs in the data. The success or failure of galloping is used to adjust min_gallop, as an indication of whether the data does or does not contain sufficient sub-runs.
In galloping mode, the algorithm searches for the first element of one array in the other. This is done by comparing that initial element with the (2k − 1)th element of the other array (first, third, seventh, and so on) so as to get a range of elements between which the initial element will lie. This shortens the range for binary searching, thus increasing efficiency. In cases where galloping is found to be less efficient than binary search, galloping mode is exited.
Galloping is beneficial only when the initial element of one run is not one of the first seven elements of the other run. This implies an initial threshold of 7. To avoid the drawbacks of galloping mode, the merging functions adjust the threshold value. If the selected element is from the same array that returned an element previously, min_gallop is reduced by one. Otherwise, the value is incremented by one, thus discouraging a return to galloping mode. In the case of random data, the value of min_gallop becomes so large that galloping mode never recurs.
When merging is done right-to-left, galloping starts from the right end of the data, that is, the last element. Galloping from the beginning also gives the required results, but makes more comparisons. Thus, the galloping algorithm uses a variable that gives the index at which galloping should begin. Timsort can enter galloping mode at any index and continue checking at the next index which is offset by 1, 3, 7, …, (2k − 1)… and so on from the current index. In the case of right-to-left merging, the offsets to the index will be −1, −3, −7, …
Galloping is not always efficient. In some cases galloping mode requires more comparisons than a simple linear search. While for the first few cases both modes may require the same number of comparisons, over time galloping mode requires 33% more comparisons than linear search to arrive at the same results.
In the worst case, Timsort takes comparisons to sort an array of n elements. In the best case, which occurs when the input is already sorted, it runs in linear time, meaning that it is an adaptive sorting algorithm.
In 2015, Dutch and German researchers in the EU FP7 ENVISAGE project found a bug in the standard implementation of Timsort.
Specifically, the invariants on stacked run sizes ensure a tight upper bound on the maximum size of the required stack. The implementation preallocated a stack sufficient to sort 264 bytes of input, and avoided further overflow checks.
However, the guarantee requires the invariants to apply to every group of three consecutive runs, but the implementation only checked it for the top three. Using the KeY tool for formal verification of Java software, the researchers found that this check is not sufficient, and they were able to find run lengths (and inputs which generated those run lengths) which would result in the invariants being violated deeper in the stack after the top of the stack was merged.
As a consequence, for certain inputs the allocated size is not sufficient to hold all unmerged runs. In Java, this generates for those inputs an array-out-of-bound exception. The smallest input that triggers this exception in Java and Android v7 is of size 108864. (Older Android versions already triggered this exception for certain inputs of size 67536) 65
The Java implementation was corrected by increasing the size of the preallocated stack based on an updated worst-case analysis. The article also showed by formal methods how to establish the intended invariant by checking that the four topmost runs in the stack satisfy the two rules above. This approach was adopted by Python and Android.
- Peters, Tim. "[Python-Dev] Sorting". Python Developers Mailinglist. Retrieved 24 February 2011.
[Timsort] also has good aspects: It's stable (items that compare equal retain their relative order, so, e.g., if you sort first on zip code, and a second time on name, people with the same name still appear in order of increasing zip code; this is important in apps that, e.g., refine the results of queries based on user input). ... It has no bad cases (O(N log N) is worst case; N−1 compares is best).
- "[DROPS]". Retrieved 1 September 2018.
TimSort is an intriguing sorting algorithm designed in 2002 for Python, whose worst-case complexity was announced, but not proved until our recent preprint.
- Chandramouli, Badrish; Goldstein, Jonathan (2014). Patience is a Virtue: Revisiting Merge and Sort on Modern Processors. SIGMOD/PODS.
- "[#JDK-6804124] (coll) Replace "modified mergesort" in java.util.Arrays.sort with timsort". JDK Bug System. Retrieved 11 June 2014.
- "Class: java.util.TimSort<T>". Android Gingerbread Documentation. Archived from the original on 16 July 2015. Retrieved 24 February 2011.
"liboctave/util/oct-sort.cc". Mercurial repository of Octave source code. Lines 23-25 of the initial comment block. Retrieved 18 February 2013.
Code stolen in large part from Python's, listobject.c, which itself had no license header. However, thanks to Tim Peters for the parts of the code I ripped-off.
- "listsort.txt". Python source code. 10 February 2017.
- MacIver, David R. (11 January 2010). "Understanding timsort, Part 1: Adaptive Mergesort". Retrieved 2015-12-05.
- de Gouw, Stijn; Rot, Jurriaan; de Boer, Frank S.; Bubel, Richard; Hähnle, Reiner (July 2015). "OpenJDK's Java.utils.Collection.sort() Is Broken: The Good, the Bad and the Worst Case" (PDF). Computer Aided Verification: 273–289. doi:10.1007/978-3-319-21690-4_16.
- de Gouw, Stijn (24 February 2015). "Proving that Android's, Java's and Python's sorting algorithm is broken (and showing how to fix it)". Retrieved 6 May 2017.
- Python Issue Tracker – Issue 23515: Bad logic in timsort's merge_collapse
- Auger, Nicolas; Nicaud, Cyril; Pivoteau, Carine (2015). "Merge Strategies: from Merge Sort to TimSort". hal-01212839.
- timsort.txt – original explanation by Tim Peters revised branch version
- listobject.c:1910@7b5057b89a56 – Python's Tree implementation
- Python's listobject.c – the C implementation of Timsort used in CPython
- OpenJDK's TimSort.java – the Java implementation of Timsort
- GNU Octave's oct-sort.cc – the C++ implementation of Timsort used in GNU Octave
- Sort Comparison – a pure Python and Cython implementation of Timsort, among other sorts
- Gee.TimSort - Vala implementation of Timsort