java - 堆排序与插入排序 JMH 基准测试 : why my insertion impl. 花费的时间更少？

Question

我已经实现了插入排序和堆排序。理论上，堆排序的时间复杂度为nlogn，插入的时间复杂度为n^2。那么为什么我的插入实现要快 6 倍才能对 100,000 个长数组进行排序？

我使用 JMH 对每种排序算法的平均时间进行基准测试。这是我的基准代码:

import java.util.concurrent.ThreadLocalRandom;
import java.util.concurrent.TimeUnit;
import java.util.stream.IntStream;

import org.openjdk.jmh.annotations.Benchmark;
import org.openjdk.jmh.annotations.BenchmarkMode;
import org.openjdk.jmh.annotations.Mode;
import org.openjdk.jmh.annotations.OutputTimeUnit;
import org.openjdk.jmh.runner.Runner;
import org.openjdk.jmh.runner.RunnerException;
import org.openjdk.jmh.runner.options.Options;
import org.openjdk.jmh.runner.options.OptionsBuilder;

public class MyBenchmark {

// setup the benchmark - create a new array for each iteration
    @State(Scope.Thread)
    public static class MyState {
        int[] array = null;

        @Setup(Level.Iteration)
        public void doSetup() {
            array = createArray(100000, 0, 100);
        }
    }

    @Benchmark
    @BenchmarkMode(Mode.AverageTime)
    @OutputTimeUnit(TimeUnit.SECONDS)
    public void insertionSort(MyState state) {
        int[] array = state.array;

        for (int i = 1; i < array.length; i++) {
            int element = array[i];
            for (int j = i - 1; j >= 0; j--) {
                if (element < array[j]) {
                    int temp = array[j];
                    array[j] = element;
                    array[j + 1] = temp;
                } else {
                    break;
                }
            }
        }
    }

    @Benchmark
    @BenchmarkMode(Mode.AverageTime)
    @OutputTimeUnit(TimeUnit.SECONDS)
    public void heapSort(MyState state) {
        int[] array = state.array;
        sort(array, array.length);
    }

    public static void sort(int[] arr, int size) {

        for (int i = 0; i < size;) {
            maxHeapify(size, arr);
            int temp = arr[0];
            arr[0] = arr[size - 1];
            arr[size - 1] = temp;
            size--;
        }
    }

    private static void maxHeapify(int size, int[] arr) {
        int nonLeafs = size / 2;
        for (int i = nonLeafs; i > 0; i--) {
            int arrayPos = heapToArrayPos(i), leftChild = heapToArrayPos(leftChild(i)),
                    rightChild = heapToArrayPos(rightChild(i));
            if (rightChild < size) {
                if (arr[rightChild] < arr[leftChild]) {
                    if (arr[arrayPos] < arr[leftChild]) {
                        switchWithLeftChild(arrayPos, arr);
                    }
                } else if (arr[arrayPos] < arr[rightChild]) {
                    switchWithRightChild(arrayPos, arr);
                }
            } else if (arr[arrayPos] < arr[leftChild]) {
                switchWithLeftChild(arrayPos, arr);
            }
        }
    }

    private static int heapToArrayPos(int heap) {
        return heap - 1;
    }

    private static int rightChild(int pos) {
        return pos * 2 + 1;
    }

    private static int leftChild(int pos) {
        return pos * 2;
    }

    private static void switchWithRightChild(int pos, int[] arr) {
        int father = arr[pos];
        int childPos = heapToArrayPos(rightChild(pos + 1)), child = arr[childPos];
        arr[childPos] = father;
        arr[pos] = child;
    }

    private static void switchWithLeftChild(int pos, int[] arr) {
        int father = arr[pos];
        int childPos = heapToArrayPos(leftChild(pos + 1)), child = arr[childPos];
        arr[childPos] = father;
        arr[pos] = child;
    }

    public static void main(String[] args) throws RunnerException {
        Options opt = new OptionsBuilder().include(MyBenchmark.class.getSimpleName()).forks(1).build();

        new Runner(opt).run();
    }

    public static int[] createArray(int length, int minValue, int maxValue) {
        return IntStream.generate(() -> ThreadLocalRandom.current().nextInt(minValue, maxValue)).limit(length)
                .toArray();
    }

    public static int[] createArray(int length) {
        return createArray(length, 0, 10);
    }

    public static int[] createArray(int minValue, int maxValue) {
        return createArray(10, minValue, maxValue);

    }
}

这是基准测试输出:

JMH 1.12 (released 51 days ago) VM version: JDK 1.8.0_65, VM 25.65-b01 VM invoker: C:\Program Files\Java\jdk1.8.0_65\jre\bin\java.exe VM options: -Dfile.encoding=UTF-8 -Xbootclasspath:C:\Program Files\Java\jdk1.8.0_65\jre\lib\resources.jar;C:\Program Files\Java\jdk1.8.0_65\jre\lib\rt.jar;C:\Program Files\Java\jdk1.8.0_65\jre\lib\jsse.jar;C:\Program Files\Java\jdk1.8.0_65\jre\lib\jce.jar;C:\Program Files\Java\jdk1.8.0_65\jre\lib\charsets.jar;C:\Program Files\Java\jdk1.8.0_65\jre\lib\jfr.jar;C:\Program Files\Java\jdk1.8.0_65\lib\tools.jar
Warmup: 20 iterations, 1 s each
Measurement: 20 iterations, 1 s each
Timeout: 10 min per iteration
Threads: 1 thread, will synchronize iterations
Benchmark mode: Average time, time/op
Benchmark: org.sample.MyBenchmark.heapSort

Run progress: 0.00% complete, ETA 00:01:20
Fork: 1 of 1
Warmup Iteration 1: 17.651 s/op
Warmup Iteration 2: 16.004 s/op
Warmup Iteration 3: 14.640 s/op
Warmup Iteration 4: 14.699 s/op
Warmup Iteration 5: 14.836 s/op
Warmup Iteration 6: 14.900 s/op
Warmup Iteration 7: 14.758 s/op
Warmup Iteration 8: 15.084 s/op
Warmup Iteration 9: 15.652 s/op
Warmup Iteration 10: 15.121 s/op
Warmup Iteration 11: 15.315 s/op
Warmup Iteration 12: 15.299 s/op
Warmup Iteration 13: 15.234 s/op
Warmup Iteration 14: 14.822 s/op
Warmup Iteration 15: 15.078 s/op
Warmup Iteration 16: 15.565 s/op
Warmup Iteration 17: 15.509 s/op
Warmup Iteration 18: 15.189 s/op
Warmup Iteration 19: 14.748 s/op
Warmup Iteration 20: 14.902 s/op
Iteration 1: 14.888 s/op
Iteration 2: 15.381 s/op
Iteration 3: 16.099 s/op
Iteration 4: 15.536 s/op
Iteration 5: 15.635 s/op
Iteration 6: 16.446 s/op
Iteration 7: 16.034 s/op
Iteration 8: 15.828 s/op
Iteration 9: 15.666 s/op
Iteration 10: 16.071 s/op
Iteration 11: 15.962 s/op
Iteration 12: 15.777 s/op
Iteration 13: 15.757 s/op
Iteration 14: 15.424 s/op
Iteration 15: 15.449 s/op
Iteration 16: 15.920 s/op
Iteration 17: 14.609 s/op
Iteration 18: 14.651 s/op
Iteration 19: 14.661 s/op
Iteration 20: 14.607 s/op

Result "heapSort": 15.520 ±(99.9%) 0.486 s/op [Average] (min, avg, max) = (14.607, 15.520, 16.446), stdev = 0.560 CI (99.9%): [15.034, 16.006] (assumes normal distribution)

JMH 1.12 (released 51 days ago) VM version: JDK 1.8.0_65, VM 25.65-b01 VM invoker: C:\Program Files\Java\jdk1.8.0_65\jre\bin\java.exe VM options: -Dfile.encoding=UTF-8 -Xbootclasspath:C:\Program Files\Java\jdk1.8.0_65\jre\lib\resources.jar;C:\Program Files\Java\jdk1.8.0_65\jre\lib\rt.jar;C:\Program Files\Java\jdk1.8.0_65\jre\lib\jsse.jar;C:\Program Files\Java\jdk1.8.0_65\jre\lib\jce.jar;C:\Program Files\Java\jdk1.8.0_65\jre\lib\charsets.jar;C:\Program Files\Java\jdk1.8.0_65\jre\lib\jfr.jar;C:\Program Files\Java\jdk1.8.0_65\lib\tools.jar Warmup: 20 iterations, 1 s each Measurement: 20 iterations, 1 s each Timeout: 10 min per iteration Threads: 1 thread, will synchronize iterations Benchmark mode: Average time, time/op Benchmark: org.sample.MyBenchmark.insertionSort

Run progress: 50.00% complete, ETA 00:10:15 Fork: 1 of 1 Warmup Iteration 1: 1.726 s/op Warmup Iteration 2: 1.636 s/op Warmup Iteration 3: 1.968 s/op Warmup Iteration 4: 1.970 s/op Warmup Iteration 5: 1.961 s/op Warmup Iteration 6: 1.966 s/op Warmup Iteration 7: 1.962 s/op Warmup Iteration 8: 1.961 s/op Warmup Iteration 9: 1.959 s/op Warmup Iteration 10: 1.965 s/op Warmup Iteration 11: 1.966 s/op Warmup Iteration 12: 1.970 s/op Warmup Iteration 13: 1.964 s/op Warmup Iteration 14: 1.952 s/op Warmup Iteration 15: 1.955 s/op Warmup Iteration 16: 1.956 s/op Warmup Iteration 17: 1.972 s/op Warmup Iteration 18: 1.966 s/op Warmup Iteration 19: 1.954 s/op Warmup Iteration 20: 1.956 s/op
Iteration 1: 1.969 s/op
Iteration 2: 1.963 s/op
Iteration 3: 2.050 s/op
Iteration 4: 2.019 s/op Iteration 5: 1.934 s/op
Iteration 6: 1.953 s/op
Iteration 7: 1.961 s/op
Iteration 8: 1.972 s/op
Iteration 9: 1.957 s/op
Iteration 10: 1.956 s/op
Iteration 11: 1.975 s/op
Iteration 12: 1.950 s/op
Iteration 13: 1.965 s/op
Iteration 14: 1.961 s/op
Iteration 15: 1.950 s/op
Iteration 16: 1.956 s/op
Iteration 17: 1.975 s/op
Iteration 18: 1.966 s/op
Iteration 19: 1.959 s/op
Iteration 20: 1.965 s/op

Result "insertionSort":
1.968 ±(99.9%) 0.022 s/op [Average] (min, avg, max) = (1.934, 1.968, 2.050), stdev = 0.025 CI (99.9%): [1.946, 1.990] (assumes normal distribution)

Run complete. Total time: 00:09:55

Benchmark Mode Cnt Score Error Units
MyBenchmark.heapSort avgt 20 12.692 ± 0.282 s/op
MyBenchmark.insertionSort avgt 20 2.024 ± 0.020 s/op

编辑:由于我已经发布了问题，所以我在基准测试之前添加了 @setup 来设置数组，因此数组创建操作不会成为一个因素。我再次运行基准测试，插入排序的结果几乎相同。堆排序基准测试平均快了 3 秒。我只发布了更新后的结果摘要。

Answer 1

您的堆排序实现不正确。您发布的代码似乎正在进行选择排序。也就是说，对于每个项目，它调用 maxHeapify，获取堆中的第一个项目，将其放在末尾，并减少计数。因此 maxHeapify 被调用 size 次，每次的大小都会减小。 maxHeapify 中内部循环的迭代次数最终类似于 (n^2)/4。

您已经实现了复杂度为 O(n^2) 的优化选择排序。

进行就地堆排序的技巧是首先构建堆(一次)，然后重新排列它以进行排序。您调用 maxHeapify 一次:

maxHeapify(size, arr);

完成后，您将获得一个有效的最大堆，其中最大的项位于 arr[0] 等。这需要 O(n) 时间。

您想要的是一个按升序排列的数组。为此，您需要构建一个循环，从堆中复制最大的项(即 arr[0])并临时保存。然后，取出堆中的最后一个项目，将计数减一，然后在顶部重新插入该项目，根据需要向下筛选。最后，将前一个最大的项目放置在之前被最后一个项目占据的位置。当 count 达到 0 时，您就有了一个排序数组:

int count = size;
while (count > 0)
{
    int save = arr[0];      // save the largest item
    arr[0] = arr[count-1];  // move last item to top
    arr[count-1] = save;    // and place the largest item
    count = count - 1;      // reduce the count
    SiftDown(0);            // sift item into place
}

您所做的就是在堆上连续调用removeMax，并将结果存储回数组中空出的位置。

SiftDown 与将项目插入堆时使用的方法相同。

请参阅我的博文，A Simple Heap of Integers ，查看使用 O(n) heapify 方法构建堆的完整示例。它是用 C# 编写的，但我认为很简单，如果你了解 Java，你就能理解它。我没有展示如何进行排序部分，但是使用该代码和上面的几行，您应该做得很好。