2.6.2.

Introsort

Section authors: Juerg Moser, Kyle Ross, and WeiChia Su.

 
Sequence
Algorithm 2.1
 

 
 
Comparison
Permuting 2.5

Based 2.2 
 

 
Sorting Algorithm 2.6
 

 
Introsort
 

Refinement of: Sequence Sorting Algorithm (2.6), therefore of Com-

parison Based (2.2), Permuting (2.5), Sequence Algorithm (2.1).
Prototype:
template<class RandomAccessIterator>
void introsort(RandomAccessIterator first,
RandomAccessIterator last)

Effects: Standard effects of a Sequence Sorting Algorithm (2.6). In

brief: the elements in [first, last) after execution are a permutation
of the original elements in the range, and they are in nondecreasing
order according to the comparison operator.

Asymptotic complexity: Let N = last first.

Average case (random data): O(N log N )

Worst case: O(N log N )

Complexity in terms of operation counts:

Average case (random data):

Value comparisons:
0.95N log2 N + 14.1N 13.0
Value assignments:
0.77N log2 N + 8.9N + 2.7
Iterator operations:
4.41N log2 N + 54.2N 10.7
Integer operations:
0.9N + 0.1
 Total operations:
6.14N log2 N + 78.0N 20.9
Worst case (median-of-3 killer sequence):
Value comparisons:
3.00N log2 N + 29.4N 0.7
Value assignments:
1.00N log2 N + 17.6N 0.6
Iterator operations:
16.00N log2 N + 182.3N 6.6
Integer operations:
10.39N log2 N + 123.1N 6.0
Total operations:
30.39N log2 N + 352.5N 14.0
See also Sorting Algorithm Operation Counts (2.91) for sam-
ple counts on random data for introsort and other sorting al-
gorithms.

Observations:

Influence of size threshold:

We noticed a slight increase in performance on random data
when the threshold at which the algorithm switches to inser-
tionsort is reduced from the value of 16 used in SGI STL to
13.
 Recursion switch:
One might expect a smaller function call overhead if the re-
cursive invocation of the introsort loop is performed on the
smaller partition. However, our experiments showed that for
random data this gain is more than outweighed by the ad-
ditional integer and comparison operations that need to be
performed.
2.6.3. Performance Comparison

Table 1 shows a comparison of the performance of introsort, quicksort,

and heapsort on random sequences. The counts are in multiples of 1,000.
1

Table 1: Performance Comparison for Random Data.

Size Algorithm Comp- Assign- Iterator Integer Total
arisons ments Ops Ops Ops
1 Introsort 11.933 9.369 53.055 0.983 75.340
Quicksort 9.369 9.369 53.264 0.958 72.960
Heapsort 10.317 15.435 137.052 122.743 285.547
4 Introsort 56.549 43.357 247.393 3.778 351.077
Quicksort 56.549 43.357 248.220 3.747 351.873
Heapsort 49.301 69.756 644.529 574.988 1338.574
16 Introsort 272.258 194.542 1135.54 15.049 1617.389
Quicksort 272.258 194.542 1138.87 15.010 1620.680
Heapsort 229.188 310.908 2961.02 2634.81 6135.926
64 Introsort 1219.03 871.73 5127.39 60.232 7278.38
Quicksort 1219.03 871.73 5140.76 60.193 7291.71
Heapsort 1044.79 1371.66 13381.5 11884.5 27682.45
256 Introsort 5623.13 3835.56 22956.0 240.46 32655.15
Quicksort 5623.13 3835.56 23009.4 240.43 32708.52
Heapsort 4690.92 5998.23 59665.6 52909.1 123263.85
1024 Introsort 24442.9 16893.9 100246 961.8 142544.6
Quicksort 24442.9 16893.9 100460 961.8 142758.6
Heapsort 20813.3 26042.5 263255 233154 543264.8

1
See also Sorting Algorithm Operation Counts (2.91) for a similar table.
Table 2 shows the same comparison for median-of-3 killer sequences as
input data.

Table 2: Performance Comparison for Median-of-3 Killer Sequence

Data.
Size Algorithm Comp- Assign- Iterator Integer Total
arisons ments Ops Ops Ops
1 Introsort 28.78 17.07 176.62 118.01 340.48
Quicksort 199.12 6.40 420.93 2.92 629.38
Heapsort 10.39 15.57 137.94 122.34 286.23
4 Introsort 140.83 77.73 849.76 568.73 1637.04
Quicksort 3063.08 28.26 6224.92 11.75 9328.01
Heapsort 49.67 70.35 648.53 573.84 1342.40
16 Introsort 662.022 345.08 3934.16 2630.13 7571.39
Quicksort 48334.2 124.51 97097.8 46.92 145603.44
Heapsort 231.14 313.83 2982.58 2635.72 6163.26
64 Introsort 3035.45 1510.69 17812.3 11881.8 34240.24
Quicksort 769724 545.44 1541310 187.73 2311767
Heapsort 1052.57 1383.29 13467.2 11889.2 27792.26
2.6.4. Introsort iterator trace plot

Using the iterator tracing tool we have generated plots that illustrate
the workings of introsort on different input sequences of size 1024. The
version of introsort that we used was the one implemented in SGI STL.

The first plot shows the iterator traces for a random input sequence. Up
to time 43000 we basically have the behavior of quicksort, which sorts
sequences larger than a certain minimum size. When only sequences
smaller than this size remain, the algorithm switches to insertionsort,
which is applied once to the entire sequence.
1200
vector_sort.log

1000

800

600

400

200

0
0 10000 20000 30000 40000 50000 60000 70000
The behavior for an already sorted sequence is similar except that the
median-of-three strategy always finds the perfect pivot and that inser-
tionsort at the end does not really have anything to do.

1200
vasc1024.log

1000

800

600

400

200

0
0 5000 10000 15000 20000 25000 30000 35000 40000 45000
For the median-of-three killer sequence the picture is quite different how-
ever. Quicksort fails bitterly and after reaching a maximum number of
recursions, the algorithm switches to heapsort to finish the work.

1200
vm31024.log

1000

800

600

400

200

0
0 50000 100000 150000 200000 250000 300000 350000

Compare with other algorithms: Heapsort (2.6.2), Mergesort (??)

See appendix (2.6.5) for operation count plots.