What's New on Java 7 Phaser

    1 Overview

   Java 7的并发包中推出了Phaser，其功能跟CyclicBarrier和CountDownLatch有些重叠，但是提供了更灵活的用法，例如支持动态调整注册任务的数量等。本文在Phaser自带的示例代码基础上进行一下简单的分析。
 
2 Glossary

2.1 Registration

    Phaser支持通过register()和bulkRegister(int
parties)方法来动态调整注册任务的数量，此外也支持通过其构造函数进行指定初始数量。在适当的时机，Phaser支持减少注册任务的数量，例如
arriveAndDeregister()。单个Phaser实例允许的注册任务数的上限是65535。


2.2 Arrival

    正如Phaser类的名字所暗示，每个Phaser实例都会维护一个phase
number，初始值为0。每当所有注册的任务都到达Phaser时，phase
number累加，并在超过Integer.MAX_VALUE后清零。arrive()和arriveAndDeregister()方法用于记录到
达，arriveAndAwaitAdvance()方法用于记录到达，并且等待其它未到达的任务。


2.3 Termination

    Phaser支持终止。Phaser终止之后，调用register()和bulkRegister(int
parties)方法没有任何效果，arriveAndAwaitAdvance()方法也会立即返回。触发终止的时机是在protected
boolean onAdvance(int phase, int
registeredParties)方法返回时，如果该方法返回true，那么Phaser会被终止。默认实现是在注册任务数为0时返回true（即
return registeredParties ==
0;）。此外，forceTermination()方法用于强制终止，isTerminated()方法用于判断是否已经终止。


2.4 Tiering

    Phaser支持层次结构，即通过构造函数Phaser(Phaser parent)和Phaser(Phaser parent,
int
parties)构造一个树形结构。这有助于减轻因在单个的Phaser上注册过多的任务而导致的竞争，从而提升吞吐量，代价是增加单个操作的开销。
 
3 Sample Usage

3.1 Sample 1

    在有些场景下，我们希望控制多个线程的启动时机：例如在并发相关的单元测试中，有时需要控制线程的启动时机，以期获得最大程度的并发，通常我们会使用CountDownLatch，以下是使用Phaser的版本。
import java.util.concurrent.Phaser;

public class PhaserTest1 {

    public static void main(String args[]) {
        //
        final int count = 5;
        final Phaser phaser = new Phaser();
        for(int i = 0; i < count; i++) {
            System.out.println("starting thread, id: " + i);
            phaser.register();
            final Thread thread = new Thread(new Task(i, phaser));
            thread.start();
        }
    }
   
    public static class Task implements Runnable {
        //
        private final int id;
        private final Phaser phaser;

        public Task(int id, Phaser phaser) {
            this.id = id;
            this.phaser = phaser;
        }
       
        @Override
        public void run() {
            phaser.arriveAndAwaitAdvance();
            System.out.println("in Task.run(), phase: " + phaser.getPhase() + ", id: " + this.id);
        }
    }
}
   以上例子中，由于线程是在一个循环中start，因此start的时机有一定的间隔。本例中这些线程实际开始工作的时机是在所有的线程都调用了phaser.arriveAndAwaitAdvance()之后。
   
此外，如果留心arriveAndAwaitAdvance()方法的签名，会发现它并没有抛出InterruptedException，实际上，即使
当前线程被中断，arriveAndAwaitAdvance()方法也不会返回，而是继续等待。如果在等待时希望可中断，或者可超时，那么需要使用以下
方法：
awaitAdvance(arrive())  // 等效于arriveAndAwaitAdvance()
awaitAdvanceInterruptibly(int phase)
awaitAdvanceInterruptibly(int phase, long timeout, TimeUnit unit)
 
3.2 Sample 2

    有些时候我们希望只有在某些外部条件满足时，才真正开始任务的执行，例如：
import java.io.BufferedReader;
import java.io.InputStreamReader;
import java.util.concurrent.Phaser;

public class PhaserTest2 {

    public static void main(String args[]) throws Exception {
        //
        final Phaser phaser = new Phaser(1);
        for(int i = 0; i < 5; i++) {
            phaser.register();
            System.out.println("starting thread, id: " + i);
            final Thread thread = new Thread(new Task(i, phaser));
            thread.start();
        }
       
        //
        System.out.println("Press ENTER to continue");
        BufferedReader reader = new BufferedReader(new InputStreamReader(System.in));
        reader.readLine();
        phaser.arriveAndDeregister();
    }
   
    public static class Task implements Runnable {
        //
        private final int id;
        private final Phaser phaser;

        public Task(int id, Phaser phaser) {
            this.id = id;
            this.phaser = phaser;
        }
       
        @Override
        public void run() {
            phaser.arriveAndAwaitAdvance();
            System.out.println("in Task.run(), phase: " + phaser.getPhase() + ", id: " + this.id);
        }
    }
}
    以上例子中，只有当用户按下回车之后，任务才真正开始执行。需要注意的是，arriveAndDeregister()方法不会被阻塞，并且返回到达时的phase number（arrive方法也是如此）。
 
3.3 Sample 3

    CyclicBarrier支持barrier action, Phaser同样也支持。不同之处是Phaser的barrier action需要改写onAdvance方法来进行定制。
import java.util.concurrent.Phaser;

public class PhaserTest3 {

    public static void main(String args[]) throws Exception {
        //
        final int count = 5;
        final int phaseToTerminate = 3;
        final Phaser phaser = new Phaser() {
            @Override
            protected boolean onAdvance(int phase, int registeredParties) {
                System.out.println("====== " + phase + " ======");
                return phase >= phaseToTerminate || registeredParties == 0;
            }
        };
       
        //
        for(int i = 0; i < count; i++) {
            System.out.println("starting thread, id: " + i);
            phaser.register();
            final Thread thread = new Thread(new Task(i, phaser));
            thread.start();
        }
    }
   
    public static class Task implements Runnable {
        //
        private final int id;
        private final Phaser phaser;

        public Task(int id, Phaser phaser) {
            this.id = id;
            this.phaser = phaser;
        }
       
        @Override
        public void run() {
            do {
                try {
                    Thread.sleep(500);
                } catch(InterruptedException e) {
                    // NOP
                }
                System.out.println("in Task.run(), phase: " + phaser.getPhase() + ", id: " + this.id);
                phaser.arriveAndAwaitAdvance();
            } while(!phaser.isTerminated());
        }
    }
}
   本例中的barrier action只是简单地打印了一条信息，此外在超过指定的迭代次数后终止了Phaser。


3.4 Sample 4

    在Smaple 3的例子中，主线程在其它工作线程结束之前已经终止。如果希望主线程等待这些工作线程结束，除了使用Thread.join()之外，也可以尝试以下的方式：
import java.util.concurrent.Phaser;

public class PhaserTest4 {

    public static void main(String args[]) throws Exception {
        //
        final int count = 5;
        final int phaseToTerminate = 3;
        final Phaser phaser = new Phaser() {
            @Override
            protected boolean onAdvance(int phase, int registeredParties) {
                System.out.println("====== " + phase + " ======");
                return phase == phaseToTerminate || registeredParties == 0;
            }
        };
       
        //
        for(int i = 0; i < count; i++) {
            System.out.println("starting thread, id: " + i);
            phaser.register();
            final Thread thread = new Thread(new Task(i, phaser));
            thread.start();
        }
       
        //
        phaser.register();
        while (!phaser.isTerminated()) {
            phaser.arriveAndAwaitAdvance();
        }
        System.out.println("done");
    }
   
    public static class Task implements Runnable {
        //
        private final int id;
        private final Phaser phaser;

        public Task(int id, Phaser phaser) {
            this.id = id;
            this.phaser = phaser;
        }
       
        @Override
        public void run() {
            while(!phaser.isTerminated()) {
                try {
                    Thread.sleep(500);
                } catch(InterruptedException e) {
                    // NOP
                }
                System.out.println("in Task.run(), phase: " + phaser.getPhase() + ", id: " + this.id);
                phaser.arriveAndAwaitAdvance();
            }
        }
    }
}
   如果希望主线程在特定的phase结束之后终止，那么可以在主线程中调用下述方法：
public static void awaitPhase(Phaser phaser, int phase) {
    int p = phaser.register(); // assumes caller not already registered
    while (p < phase) {
        if (phaser.isTerminated()) {
            break; // ... deal with unexpected termination
        } else {
            p = phaser.arriveAndAwaitAdvance();
        }
    }
    phaser.arriveAndDeregister();
}
    需要注意的是，awaitPhase方法中的if (phaser.isTerminated())
分支里需要能够正确处理Phaser终止的情况。否则由于在Phaser终止之后，
phaser.register()和arriveAndAwaitAdvance()方法均返回负值，那么上述方法可能陷入死循环。


3.5 Sample 5

    以下对Phaser进行分层的例子：
import java.util.concurrent.Phaser;

public class PhaserTest6 {
    //
    private static final int TASKS_PER_PHASER = 4;

    public static void main(String args[]) throws Exception {
        //
        final int phaseToTerminate = 3;
        final Phaser phaser = new Phaser() {
            @Override
            protected boolean onAdvance(int phase, int registeredParties) {
                System.out.println("====== " + phase + " ======");
                return phase == phaseToTerminate || registeredParties == 0;
            }
        };
       
        //
        final Task tasks[] = new Task[10];
        build(tasks, 0, tasks.length, phaser);
        for (int i = 0; i < tasks.length; i++) {
            System.out.println("starting thread, id: " + i);
            final Thread thread = new Thread(tasks[i]);
            thread.start();
        }
    }

    public static void build(Task[] tasks, int lo, int hi, Phaser ph) {
        if (hi - lo > TASKS_PER_PHASER) {
            for (int i = lo; i < hi; i += TASKS_PER_PHASER) {
                int j = Math.min(i + TASKS_PER_PHASER, hi);
                build(tasks, i, j, new Phaser(ph));
            }
        } else {
            for (int i = lo; i < hi; ++i)
                tasks[i] = new Task(i, ph);
        }
    }

    public static class Task implements Runnable {
        //
        private final int id;
        private final Phaser phaser;

        public Task(int id, Phaser phaser) {
            this.id = id;
            this.phaser = phaser;
            this.phaser.register();
        }

        @Override
        public void run() {
            while (!phaser.isTerminated()) {
                try {
                    Thread.sleep(200);
                } catch (InterruptedException e) {
                    // NOP
                }
                System.out.println("in Task.run(), phase: " + phaser.getPhase()    + ", id: " + this.id);
                phaser.arriveAndAwaitAdvance();
            }
        }
    }
}
    需要注意的是，TASKS_PER_PHASER的值取决于具体的Task实现。对于Task执行时间很短的场景（也就是竞争相对激烈），可以考虑使用较小的TASKS_PER_PHASER值，例如4。反之可以适当增大TASKS_PER_PHASER。