Question

我需要进行1000计算，每个计算基于一个随机数。整个运行需要可重现。

在单线程环境中，我只是根据种子创建随机数，并将其用于每次计算：

Random random = new Random(37);
for (Calculation c : calculations) {
    c.doCalculation(r.nextInt());
}

在多线程环境中，我有10个线程，我有一个种子随机种子线程Random's：

Random initRandom = new Random(37);
for (List<Calculation> p : calculationPartitions) {
    final Random threadRandom = new Random(initRandom.nextLong());
    executorService.submit(() -> {
        for (Calculation c : p) {
            c.doCalculation(threadRandom.nextInt());
        }
    });
}
// Merge the calculation results back in the same order as the calculationPartitions
...

这是一个好主意吗？一般来说它仍然是均匀分布的随机数吗？每个threadRandom由initRandom播种的事实是否会破坏统一分布？

由于可重复的原因，我无法共享1个全局随机数，因为某些线程在某些运行中可能比其他线程运行得更快，因此它们不会总是以相同的顺序调用全局随机数（并且争用可能会导致性能下降）。 / p>

Answer 1

您将通过Random获得均匀的数字分布，但是不保证列表中每个Calculation对象列表的运行顺序，因此如果这些顺序将结果传递给共享列表，则顺序可能会因运行而异。

Answer 2

特别是关于均匀分布，似乎每10个像素有1个随机位置：

哦等等，每1个像素有1个随机位置，种子线程方法看起来更好！？在一个完美的均匀分布（尺寸250000），它将全黑：

左：

public class SingleRandomProof extends JFrame {

    public static final int WIDTH = 500;
    public static final int HEIGHT = 500;
    public static final int SIZE = WIDTH * HEIGHT;

    public static void main(String[] args) {
        SingleRandomProof proof = new SingleRandomProof();
        proof.pack();
        proof.setVisible(true);
        proof.doCalc();
    }

    private JLabel panel;

    public SingleRandomProof() throws HeadlessException {
        super("1 random");
        BufferedImage image = new BufferedImage(WIDTH, HEIGHT, BufferedImage.TYPE_INT_ARGB);
        panel = new JLabel(new ImageIcon(image));
        setContentPane(panel);
    }

    private void doCalc() {
        BufferedImage image = new BufferedImage(WIDTH, HEIGHT, BufferedImage.TYPE_INT_ARGB);
        Graphics g = image.getGraphics();
        g.setColor(Color.WHITE);
        g.fillRect(0, 0, WIDTH, HEIGHT);
        g.setColor(Color.BLACK);
        Random r = new Random(37);
        for (int i = 0; i < SIZE; i++) {
            int position = r.nextInt(SIZE);
            g.fillRect(position % HEIGHT, position / HEIGHT, 1, 1);
        }
        panel.setIcon(new ImageIcon(image));
    }

}

右：

public class SeededThreadRandomProof extends JFrame {

    public static final int WIDTH = 500;
    public static final int HEIGHT = 500;
    public static final int SIZE = WIDTH * HEIGHT;

    public static void main(String[] args) {
        SeededThreadRandomProof proof = new SeededThreadRandomProof();
        proof.pack();
        proof.setVisible(true);
        proof.doCalc();
    }

    private JLabel panel;

    public SeededThreadRandomProof() throws HeadlessException {
        super("10 seeded randoms");
        BufferedImage image = new BufferedImage(WIDTH, HEIGHT, BufferedImage.TYPE_INT_ARGB);
        panel = new JLabel(new ImageIcon(image));
        setContentPane(panel);
    }

    private void doCalc() {
        BufferedImage image = new BufferedImage(WIDTH, HEIGHT, BufferedImage.TYPE_INT_ARGB);
        Graphics g = image.getGraphics();
        g.setColor(Color.WHITE);
        g.fillRect(0, 0, WIDTH, HEIGHT);
        g.setColor(Color.BLACK);
        Random initRandom = new Random(37);
        for (int j = 0; j < 10; j++) {
            Random r = new Random(initRandom.nextLong());
            for (int i = 0; i < SIZE / 10; i++) {
                int position = r.nextInt(SIZE);
                g.fillRect(position % HEIGHT, position / HEIGHT, 1, 1);
            }
        }
        panel.setIcon(new ImageIcon(image));
    }

}

每个线程一个随机实例，由原始Random实例播种以获得可重复性

2 个答案: