原文: http://whitesock.iteye.com/blog/1135457
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(count);
- for(int i = 0; i < count; i++) {
- System.out.println("starting thread, id: " + i);
- 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(count) {
- @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);
- 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(count) {
- @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);
- 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。