同步与ReadWriteLock性能

Question

我试图证明，当有许多读者而只有一些作者时，同步是比较慢的。不知何故我证明了相反的情况。

RW示例的执行时间为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();
    }

}

同步示例中，执行时间仅为241 is：

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));
        }
    }

}

为什么当我拥有比作者多十倍的读者时，同步收集会更快。如何显示我在许多文章中读到的RW锁的进展？

Answer 1

获取一个ReadWriteLock的实际成本通常比获取一个简单互斥体的成本慢得多。javadoc for ReadWriteLock将讨论以下内容：

读写锁是否会比使用互斥锁提高性能取决于与被修改相比数据被读取的频率、读和写操作的持续时间以及对数据的争用-即尝试同时读取或写入数据的线程数。例如，一个集合最初填充了数据，之后很少被修改，同时经常被搜索(例如某种目录)，是使用读写锁的理想选择。但是，如果更新变得频繁，那么数据大部分时间都是独占锁定的，并发性几乎没有增加。此外，如果读操作太短，读写锁实现的开销(本质上比互斥锁更复杂)可以控制执行成本，特别是因为许多读写锁实现仍然通过一小部分代码序列化所有线程。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，这可能意味着性能取决于实际获取锁所花费的时间。

您的基准测试也有一个问题，就是Math.random是同步的。从它的javadoc

此方法被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，则可以减少对每个线程拥有自己的伪随机数生成器的争用。

因此，即使您的并发读取器在获得ReadWriteLock后没有相互阻塞，他们可能仍然在争夺Math.random中获得的锁，从而挫败了使用ReadWriteLock的一些好处。您可以通过使用ThreadLocalRandom来改进这一点。

此外，正如assylias所指出的，不考虑JIT编译和其他运行时怪癖的天真Java基准是不可靠的。您应该使用Java微基准管理(JMH)进行类似的基准测试。