java - 多线程与单线程

Question

这个问题已经有答案了:

Why single thread is faster than multithreading in Java? (5 个回答)

已关闭 7 年前。

我测试了一些简单的条件:
考虑一个长度为 10 000 000 的 int 数组。填充:

1. 使用单(主)线程。
2. 使用双工作线程并将它们连接起来直到完成。第一个从数组的开头一直填充到数组的中间。从末尾到中间的第二个。
3. 使用 ExecutorService 固定池 (2)，调用执行并等待终止。
4. 使用具有默认工作线程数量(可用处理器数量)的 ForkJoinPool

<小时/>

import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.ForkJoinPool;
import java.util.concurrent.RecursiveAction;
import java.util.concurrent.TimeUnit;

public class PerformanceTest {
    private static final int ARRAY_LENGTH = 10_000_000;
    private static int[] array;
    private static final int ITERATIONS = 10;

    public static void main(String[] args) throws InterruptedException {
        for (int i = 0; i < ITERATIONS; i++) {
            array = new int[ARRAY_LENGTH];
            long millis = System.currentTimeMillis();
            singleWorkerFill();
            System.out.println("Single worker: " + (System.currentTimeMillis() - millis));

            array = new int[ARRAY_LENGTH];
            millis = System.currentTimeMillis();
            doubleWorkerFill();
            System.out.println("Double worker: " + (System.currentTimeMillis() - millis));

            array = new int[ARRAY_LENGTH];
            millis = System.currentTimeMillis();
            forkJoinWorkersFill();
            System.out.println("Executor workers: " + (System.currentTimeMillis() - millis));

            array = new int[ARRAY_LENGTH];
            millis = System.currentTimeMillis();
            executorWorkersFill();
            System.out.println("ForkJoin workers: " + (System.currentTimeMillis() - millis));

            System.out.println("---------------------------------------------");
            Thread.sleep(1000);
        }
    }

    private static void singleWorkerFill() {
        for (int i = 0, len = array.length; i < len; i++) {
            array[i] = i;
        }
    }

    private static void doubleWorkerFill() throws InterruptedException {
        Thread worker1 = new Thread(new HeadArrayFiller());
        Thread worker2 = new Thread(new TailArrayFiller());
        worker1.start();
        worker2.start();
        worker1.join();
        worker2.join();
    }

    private static void executorWorkersFill() throws InterruptedException {
        ExecutorService executorService = Executors.newFixedThreadPool(2);
        executorService.execute(new HeadArrayFiller());
        executorService.execute(new TailArrayFiller());
        executorService.shutdown();
        executorService.awaitTermination(Long.MAX_VALUE, TimeUnit.MILLISECONDS);
    }

    private static void forkJoinWorkersFill() throws InterruptedException {
        ForkJoinPool pool = new ForkJoinPool();
        pool.invoke(new HeadArrayFiller());
        pool.invoke(new TailArrayFiller());
        pool.shutdown();
        pool.awaitTermination(Long.MAX_VALUE, TimeUnit.MILLISECONDS);
    }

    private static class HeadArrayFiller extends RecursiveAction implements Runnable {
        @Override
        public void run() {
            for (int i = 0, middle = array.length / 2; i <= middle; i++) {
                array[i] = i;
            }
        }

        @Override
        protected void compute() {
            run();
        }
    }

    private static class TailArrayFiller extends RecursiveAction implements Runnable {
        @Override
        public void run() {
            for (int i = array.length - 2, middle = array.length / 2; i > middle; i--) {
                array[i] = i;
            }
        }

        @Override
        protected void compute() {
            run();
        }
    }
}

我预计单线程模型与其他模型相比没有机会，但事实并非如此。以下是以毫秒为单位的测试结果:

ITERATION #1Single worker: 7Double worker: 10Executor workers: 11ForkJoin workers: 6ITERATION #2Single worker: 6Double worker: 4Executor workers: 5ForkJoin workers: 4ITERATION #3Single worker: 4Double worker: 4Executor workers: 5ForkJoin workers: 4ITERATION #4Single worker: 5Double worker: 5Executor workers: 5ForkJoin workers: 4ITERATION #5Single worker: 5Double worker: 5Executor workers: 4ForkJoin workers: 5ITERATION #6Single worker: 4Double worker: 4Executor workers: 5ForkJoin workers: 4ITERATION #7Single worker: 4Double worker: 4Executor workers: 4ForkJoin workers: 5ITERATION #8Single worker: 4Double worker: 4Executor workers: 4ForkJoin workers: 5ITERATION #9Single worker: 4Double worker: 4Executor workers: 4ForkJoin workers: 5ITERATION #10Single worker: 5Double worker: 4Executor workers: 4ForkJoin workers: 4

正如您注意到的，单线程模型在启动时比多线程双线程模型更快。 Fork-join 模型似乎是最好的，同样 ExecutorService 也是如此。我建议对迭代进行一些 JIT 编译器优化。测试结束时它们都非常相似。

无论如何，主要问题是为什么双线程模型性能与单线程相同(甚至在冷启动时更慢)。我如何才能达到预期速度近两倍的性能？

谢谢

Answer 1

对于现代计算机来说，初始化 10M 整数是一项非常快的任务，并且在两个独立的内核上并行执行操作的好处不会补偿(或只是补偿)启动线程、上下文切换的开销它们之间、协调它们等等。

开始在每次迭代中做更多的工作(例如 hibernate 5 毫秒)，多线程的优势将开始显现。