同步与 ReadWriteLock 性能

我试图证明当有很多读者而只有一些作者时同步会更慢。不知怎的，我证明了相反的情况。

以 RW 为例，执行时间为 313 ms：

package zad3readWriteLockPerformance;

import java.util.ArrayList;
import java.util.Collections;
import java.util.List;
import java.util.concurrent.locks.Lock;
import java.util.concurrent.locks.ReadWriteLock;
import java.util.concurrent.locks.ReentrantReadWriteLock;

public class Main {
    public static long start, end;

    public static void main(String[] args) {
        Runtime.getRuntime().addShutdownHook(new Thread(() -> {
            end = System.currentTimeMillis();
            System.out.println("Time of execution " + (end - start) + " ms");
        }));
        start = System.currentTimeMillis();
        final int NUMBER_OF_THREADS = 1000;
        ThreadSafeArrayList<Integer> threadSafeArrayList = new ThreadSafeArrayList<>();
        ArrayList<Thread> consumerThreadList = new ArrayList<Thread>();
        for (int i = 0; i < NUMBER_OF_THREADS; i++) {
            Thread t = new Thread(new Consumer(threadSafeArrayList));
            consumerThreadList.add(t);
            t.start();
        }

        ArrayList<Thread> producerThreadList = new ArrayList<Thread>();
        for (int i = 0; i < NUMBER_OF_THREADS/10; i++) {
            Thread t = new Thread(new Producer(threadSafeArrayList));
            producerThreadList.add(t);
            t.start();

        }



        //  System.out.println("Printing the First Element : " + threadSafeArrayList.get(1));

    }

}
class Consumer implements Runnable {
    public final static int NUMBER_OF_OPERATIONS = 100;
    ThreadSafeArrayList<Integer> threadSafeArrayList;

    public Consumer(ThreadSafeArrayList<Integer> threadSafeArrayList) {
        this.threadSafeArrayList = threadSafeArrayList;
    }

    @Override
    public void run() {
        for (int j = 0; j < NUMBER_OF_OPERATIONS; j++) {
            Integer obtainedElement = threadSafeArrayList.getRandomElement();
        }
    }

}
class Producer implements Runnable {
    public final static int NUMBER_OF_OPERATIONS = 100;
    ThreadSafeArrayList<Integer> threadSafeArrayList;

    public Producer(ThreadSafeArrayList<Integer> threadSafeArrayList) {
        this.threadSafeArrayList = threadSafeArrayList;
    }

    @Override
    public void run() {
        for (int j = 0; j < NUMBER_OF_OPERATIONS; j++) {
            threadSafeArrayList.add((int) (Math.random() * 1000));
        }
    }

}

class ThreadSafeArrayList<E> {
    private final ReadWriteLock readWriteLock = new ReentrantReadWriteLock();

    private final Lock readLock = readWriteLock.readLock();

    private final Lock writeLock = readWriteLock.writeLock();

    private final List<E> list = new ArrayList<>();

    public void add(E o) {
        writeLock.lock();
        try {
            list.add(o);
            //System.out.println("Adding element by thread" + Thread.currentThread().getName());
        } finally {
            writeLock.unlock();
        }
    }

    public E getRandomElement() {
        readLock.lock();
        try {
            //System.out.println("Printing elements by thread" + Thread.currentThread().getName());
            if (size() == 0) {
                return null;
            }
            return list.get((int) (Math.random() * size()));
        } finally {
            readLock.unlock();
        }
    }

    public int size() {
        return list.size();
    }

}

同步的例子，执行时间只有241ms：

package zad3readWriteLockPerformanceZMIENONENENASYNCHRO;

import java.util.ArrayList;
import java.util.Collections;
import java.util.List;

public class Main {
    public static long start, end;

    public static void main(String[] args) {
        Runtime.getRuntime().addShutdownHook(new Thread(() -> {
            end = System.currentTimeMillis();
            System.out.println("Time of execution " + (end - start) + " ms");
        }));
        start = System.currentTimeMillis();
        final int NUMBER_OF_THREADS = 1000;
        List<Integer> list = Collections.synchronizedList(new ArrayList<Integer>());
        ArrayList<Thread> consumerThreadList = new ArrayList<Thread>();
        for (int i = 0; i < NUMBER_OF_THREADS; i++) {
            Thread t = new Thread(new Consumer(list));
            consumerThreadList.add(t);
            t.start();
        }

        ArrayList<Thread> producerThreadList = new ArrayList<Thread>();
        for (int i = 0; i < NUMBER_OF_THREADS / 10; i++) {
            Thread t = new Thread(new Producer(list));
            producerThreadList.add(t);
            t.start();
        }

        //  System.out.println("Printing the First Element : " + threadSafeArrayList.get(1));

    }

}

class Consumer implements Runnable {
    public final static int NUMBER_OF_OPERATIONS = 100;
    List<Integer> list;

    public Consumer(List<Integer> list) {
        this.list = list;
    }

    @Override
    public void run() {
        for (int j = 0; j < NUMBER_OF_OPERATIONS; j++) {
            if (list.size() > 0)
                list.get((int) (Math.random() * list.size()));
        }
    }

}

class Producer implements Runnable {
    public final static int NUMBER_OF_OPERATIONS = 100;
    List<Integer> threadSafeArrayList;

    public Producer(List<Integer> threadSafeArrayList) {
        this.threadSafeArrayList = threadSafeArrayList;
    }

    @Override
    public void run() {
        for (int j = 0; j < NUMBER_OF_OPERATIONS; j++) {
            threadSafeArrayList.add((int) (Math.random() * 1000));
        }
    }

}

为什么当我的读者比作者多十倍时，同步收集会更快。如何显示我在很多文章中读到的 RW 锁的进度？

获取 ReadWriteLock 的实际成本通常比获取简单互斥锁的成本慢得多。这javadoc https://docs.oracle.com/javase/8/docs/api/java/util/concurrent/locks/ReadWriteLock.html对于 ReadWriteLock 来说：

读写锁是否会比使用互斥锁提高性能取决于数据读取与修改的频率、读写操作的持续时间以及数据的争用 - 即是，同时尝试读取或写入数据的线程数。例如，最初填充了数据，此后很少修改但经常搜索的集合（例如某种目录）是使用读写锁的理想候选者。然而，如果更新变得频繁，那么数据大部分时间都被独占锁定，并且并发性几乎没有增加。此外，如果读取操作太短，读写锁实现的开销（本质上比互斥锁更复杂）可能会主导执行成本，特别是因为许多读写锁实现仍然通过一个序列化所有线程。小部分代码。最终，只有分析和测量才能确定读写锁的使用是否适合您的应用程序。

因此，您的线程正在执行非常简单的操作，这一事实可能意味着性能取决于实际获取锁所花费的时间量。

你的基准也有一个问题，那就是Math.random已同步。从其javadoc https://docs.oracle.com/javase/8/docs/api/java/lang/Math.html#random--:

该方法经过正确同步，以允许多个线程正确使用。但是，如果许多线程需要以很高的速率生成伪随机数，则可能会减少每个线程拥有自己的伪随机数生成器的争用。

因此，即使并发读取者在获得 ReadWriteLock 后不会互相阻塞，他们仍可能会争夺在Math.random，抵消了使用 ReadWriteLock 的一些优点。您可以通过使用来改进这一点线程局部随机 https://docs.oracle.com/javase/8/docs/api/java/util/concurrent/ThreadLocalRandom.html.

此外，正如 assylias 指出的那样，不考虑 JIT 编译和其他运行时怪癖的简单 Java 基准测试是不可靠的。您应该使用Java 微基准测试工具 (JMH) http://openjdk.java.net/projects/code-tools/jmh/对于这样的基准。