本节我们来研究下并发包中的Atomic类型。
AtomicXXX和XXXAdder以及XXXAccumulator性能测试
先来一把性能测试,对比一下AtomicLong(1.5出来的)、LongAdder(1.8出来的)和LongAccumulator(1.8出来的)用于简单累加的性能。
程序逻辑比较简单,可以看到我们在最大并发10的情况下,去做10亿次操作测试:
@Slf4j
public class AccumulatorBenchmark {
private StopWatch stopWatch = new StopWatch();
static final int threadCount = 100;
static final int taskCount = 1000000000;
static final AtomicLong atomicLong = new AtomicLong();
static final LongAdder longAdder = new LongAdder();
static final LongAccumulator longAccumulator = new LongAccumulator(Long::sum, 0L);
@Test
public void test() {
Map<String, IntConsumer> tasks = new HashMap<>();
tasks.put("atomicLong", i -> atomicLong.incrementAndGet());
tasks.put("longAdder", i -> longAdder.increment());
tasks.put("longAccumulator", i -> longAccumulator.accumulate(1L));
tasks.entrySet().forEach(item -> benchmark(threadCount, taskCount, item.getValue(), item.getKey()));
log.info(stopWatch.prettyPrint());
Assert.assertEquals(taskCount, atomicLong.get());
Assert.assertEquals(taskCount, longAdder.longValue());
Assert.assertEquals(taskCount, longAccumulator.longValue());
}
private void benchmark(int threadCount, int taskCount, IntConsumer task, String name) {
stopWatch.start(name);
ForkJoinPool forkJoinPool = new ForkJoinPool(threadCount);
forkJoinPool.execute(() -> IntStream.rangeClosed(1, taskCount).parallel().forEach(task));
forkJoinPool.shutdown();
try {
forkJoinPool.awaitTermination(1, TimeUnit.HOURS);
} catch (InterruptedException e) {
e.printStackTrace();
}
stopWatch.stop();
}
}
结果如下:
和官网说的差不多,在高并发的情况下LongAdder性能会比AtomicLong好很多。
AtomicReference可见性问题测试
在很多开源代码中我们有看到AtomicReference的身影,它究竟是干什么的呢?我们来写一段测试程序,在这个程序中我们定一了一个Switch类型,作为一个开关,然后写三个死循环的线程来测试,当开关有效的时候会持续死循环,在2秒后关闭所有的三个开关:
- 第一个是普通的Switch
- 第二个是使用了volatile声明的Switch
- 第三个是AtomicReference包装的Switch
@Slf4j
public class AtomicReferenceTest {
private Switch rawValue = new Switch();
private volatile Switch volatileValue = new Switch();
private AtomicReference<Switch> atomicValue = new AtomicReference<>(new Switch());
@Test
public void test() throws InterruptedException {
new Thread(() -> {
log.info("Start:rawValue");
while (rawValue.get()) {
}
log.info("Done:rawValue");
}).start();
new Thread(() -> {
log.info("Start:volatileValue");
while (volatileValue.get()) {
}
log.info("Done:volatileValue");
}).start();
new Thread(() -> {
log.info("Start:atomicValue");
while (atomicValue.get().get()) {
}
log.info("Done:atomicValue");
}).start();
Executors.newSingleThreadScheduledExecutor().schedule(rawValue::off, 2, TimeUnit.SECONDS);
Executors.newSingleThreadScheduledExecutor().schedule(volatileValue::off, 2, TimeUnit.SECONDS);
Executors.newSingleThreadScheduledExecutor().schedule(atomicValue.get()::off, 2, TimeUnit.SECONDS);
TimeUnit.HOURS.sleep(1);
}
class Switch {
private boolean enable = true;
public boolean get() {
return enable;
}
public void off() {
enable = false;
}
}
}
运行程序:
可以看到2秒后有一个开关卡住了,线程没有退出。这是一个可见性的问题,AtomicReference以及volatile可以确保线程对数据的更新刷新到内存。因为我们对于开关的关闭是在另一个定时任务线程做的,如果我们不使用volatile或AtomicReference来定义对象,那么对象的操作可能无法被其它线程感知到。当然,AtomicReference除了解决可见性问题还有更多AtomicXXX提供的其它功能。
AtomicInteger测试
下面我们来看一下AtomicInteger的compareAndSet()功能。首先说明这个程序没有任何意义,只是测试一下功能。在这个程序里,我们乱序开启10个线程,每一个线程的任务就是按照次序来累加数字。我们使用AtomicInteger的compareAndSet()来确保乱序的线程也能按照我们要的顺序操作累加。
@Slf4j
public class AtomicIntegerTest {
@Test
public void test() throws InterruptedException {
AtomicInteger atomicInteger = new AtomicInteger(0);
List<Thread> threadList = IntStream.range(0,10).mapToObj(i-> {
Thread thread = new Thread(() -> {
log.debug("Wait {}->{}", i, i+1);
while (!atomicInteger.compareAndSet(i, i + 1)) {
try {
TimeUnit.MILLISECONDS.sleep(50);
} catch (InterruptedException e) {
e.printStackTrace();
}
}
log.debug("Done {}->{}", i, i+1);
});
thread.setName(UUID.randomUUID().toString());
return thread;
}).sorted(Comparator.comparing(Thread::getName)).collect(Collectors.toList());
for (Thread thread : threadList) {
thread.start();
}
for (Thread thread : threadList) {
thread.join();
}
log.info("result:{}", atomicInteger.get());
}
}
执行结果如下:
11:46:30.611 [2c80b367-d80e-46b5-94f5-b7b172e79dad] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Wait 4->5
11:46:30.611 [7bccbb54-4573-4b77-979b-840613406428] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Wait 5->6
11:46:30.612 [c0792831-6201-4f6c-b702-79c1b798c3aa] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Wait 9->10
11:46:30.612 [949b0c26-febb-4830-ad98-f43521ce4382] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Wait 7->8
11:46:30.613 [ccc05b0f-11da-41fa-b8fc-59a90dfc2250] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Wait 6->7
11:46:30.611 [037e9595-73cb-4aa1-afee-4250347746c8] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Wait 3->4
11:46:30.611 [4f15d9ce-044e-4657-b418-4874d03e5d22] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Wait 1->2
11:46:30.611 [3a96c35c-bc4e-45f4-aae4-9fd8611acaea] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Wait 8->9
11:46:30.611 [94465214-27bf-4543-80e2-dbaeeb6ddc94] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Wait 0->1
11:46:30.611 [60f9cb50-21e6-45bc-9b4d-867783ab033b] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Wait 2->3
11:46:30.627 [94465214-27bf-4543-80e2-dbaeeb6ddc94] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Done 0->1
11:46:30.681 [4f15d9ce-044e-4657-b418-4874d03e5d22] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Done 1->2
11:46:30.681 [60f9cb50-21e6-45bc-9b4d-867783ab033b] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Done 2->3
11:46:30.734 [037e9595-73cb-4aa1-afee-4250347746c8] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Done 3->4
11:46:30.780 [2c80b367-d80e-46b5-94f5-b7b172e79dad] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Done 4->5
11:46:30.785 [7bccbb54-4573-4b77-979b-840613406428] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Done 5->6
11:46:30.785 [ccc05b0f-11da-41fa-b8fc-59a90dfc2250] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Done 6->7
11:46:30.787 [949b0c26-febb-4830-ad98-f43521ce4382] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Done 7->8
11:46:30.838 [3a96c35c-bc4e-45f4-aae4-9fd8611acaea] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Done 8->9
11:46:30.890 [c0792831-6201-4f6c-b702-79c1b798c3aa] DEBUG me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - Done 9->10
11:46:30.890 [main] INFO me.josephzhu.javaconcurrenttest.atomic.AtomicIntegerTest - result:10
可以看到,Wait的输出是乱序的,最后Done的输出是顺序的。
AtomicStampedReference测试
AtomicStampedReference可以用来解决ABA问题,什么是ABA问题我们看这个例子:
线程1读取了数字之后,等待1秒,然后尝试把1修改为3。
线程2后启动,读取到数字1后修改2,稍等一下又修改回1。
虽然AtomicInteger确保多个线程的原子性操作,但是无法确保1就是原先读取到的那个1,没有经过别人修改。
可以再换一个例子来说,如果我们现在账上有100元,要修改为200元,在修改之前账户已经被操作过了从100元充值到了150然后提现到了100,虽然最后还是回到了100,但是这个时候严格一点的话,我们应该认为这个100不是原先的100,这个账户的版本发生了变化,如果我们使用乐观行锁的话,虽然余额都是100但是行锁的版本肯定不一致,AtomicStampedReference就是类似行乐观锁的概念。
@Test
public void test() throws InterruptedException {
AtomicInteger atomicInteger = new AtomicInteger(1);
Thread thread1 = new Thread(() -> {
int value = atomicInteger.get();
log.info("thread 1 read value: " + value);
try {
TimeUnit.SECONDS.sleep(1);
} catch (InterruptedException e) {
e.printStackTrace();
}
if (atomicInteger.compareAndSet(value, 3)) {
log.info("thread 1 update from " + value + " to 3");
} else {
log.info("thread 1 update fail!");
}
});
thread1.start();
Thread thread2 = new Thread(() -> {
int value = atomicInteger.get();
log.info("thread 2 read value: " + value);
if (atomicInteger.compareAndSet(value, 2)) {
log.info("thread 2 update from " + value + " to 2");
try {
TimeUnit.MILLISECONDS.sleep(100);
} catch (InterruptedException e) {
e.printStackTrace();
}
value = atomicInteger.get();
log.info("thread 2 read value: " + value);
if (atomicInteger.compareAndSet(value, 1)) {
log.info("thread 2 update from " + value + " to 1");
}
}
});
thread2.start();
thread1.join();
thread2.join();
}
看下运行结果:
11:56:20.373 [Thread-1] INFO me.josephzhu.javaconcurrenttest.atomic.AtomicStampedReferenceTest - thread 2 read value: 1
11:56:20.381 [Thread-1] INFO me.josephzhu.javaconcurrenttest.atomic.AtomicStampedReferenceTest - thread 2 update from 1 to 2
11:56:20.373 [Thread-0] INFO me.josephzhu.javaconcurrenttest.atomic.AtomicStampedReferenceTest - thread 1 read value: 1
11:56:20.483 [Thread-1] INFO me.josephzhu.javaconcurrenttest.atomic.AtomicStampedReferenceTest - thread 2 read value: 2
11:56:20.484 [Thread-1] INFO me.josephzhu.javaconcurrenttest.atomic.AtomicStampedReferenceTest - thread 2 update from 2 to 1
11:56:21.386 [Thread-0] INFO me.josephzhu.javaconcurrenttest.atomic.AtomicStampedReferenceTest - thread 1 update from 1 to 3
下面我们使用AtomicStampedReference来修复这个问题:
@Test
public void test2() throws InterruptedException {
AtomicStampedReference<Integer> atomicStampedReference = new AtomicStampedReference<>(1, 1);
Thread thread1 = new Thread(() -> {
int[] stampHolder = new int[1];
int value = atomicStampedReference.get(stampHolder);
int stamp = stampHolder[0];
log.info("thread 1 read value: " + value + ", stamp: " + stamp);
try {
TimeUnit.SECONDS.sleep(1);
} catch (InterruptedException e) {
e.printStackTrace();
}
if (atomicStampedReference.compareAndSet(value, 3, stamp, stamp + 1)) {
log.info("thread 1 update from " + value + " to 3");
} else {
log.info("thread 1 update fail!");
}
});
thread1.start();
Thread thread2 = new Thread(() -> {
int[] stampHolder = new int[1];
int value = atomicStampedReference.get(stampHolder);
int stamp = stampHolder[0];
log.info("thread 2 read value: " + value + ", stamp: " + stamp);
if (atomicStampedReference.compareAndSet(value, 2, stamp, stamp + 1)) {
log.info("thread 2 update from " + value + " to 2");
try {
TimeUnit.MILLISECONDS.sleep(100);
} catch (InterruptedException e) {
e.printStackTrace();
}
value = atomicStampedReference.get(stampHolder);
stamp = stampHolder[0];
log.info("thread 2 read value: " + value + ", stamp: " + stamp);
if (atomicStampedReference.compareAndSet(value, 1, stamp, stamp + 1)) {
log.info("thread 2 update from " + value + " to 1");
}
value = atomicStampedReference.get(stampHolder);
stamp = stampHolder[0];
log.info("thread 2 read value: " + value + ", stamp: " + stamp);
}
});
thread2.start();
thread1.join();
thread2.join();
}
运行结果如下:
11:59:11.946 [Thread-1] INFO me.josephzhu.javaconcurrenttest.atomic.AtomicStampedReferenceTest - thread 2 read value: 1, stamp: 1
11:59:11.951 [Thread-1] INFO me.josephzhu.javaconcurrenttest.atomic.AtomicStampedReferenceTest - thread 2 update from 1 to 2
11:59:11.946 [Thread-0] INFO me.josephzhu.javaconcurrenttest.atomic.AtomicStampedReferenceTest - thread 1 read value: 1, stamp: 1
11:59:12.053 [Thread-1] INFO me.josephzhu.javaconcurrenttest.atomic.AtomicStampedReferenceTest - thread 2 read value: 2, stamp: 2
11:59:12.053 [Thread-1] INFO me.josephzhu.javaconcurrenttest.atomic.AtomicStampedReferenceTest - thread 2 update from 2 to 1
11:59:12.053 [Thread-1] INFO me.josephzhu.javaconcurrenttest.atomic.AtomicStampedReferenceTest - thread 2 read value: 1, stamp: 3
11:59:12.954 [Thread-0] INFO me.josephzhu.javaconcurrenttest.atomic.AtomicStampedReferenceTest - thread 1 update fail!
可以看到,现在我们修改数据的时候不仅仅是拿着值来修改了,还要提供版本号,读取数据的时候可以读取到数据以及版本号。这样的话,虽然数值不变,但是线程2经过两次修改后数据的版本从1变为了3,回过头来线程1再要拿着版本号1来修改数据的话必然失败。
一个有趣的问题
本文比较短,我们再来看网友之前问的一个有意思的问题,程序如下。
@Slf4j
public class InterestingProblem {
int a = 1;
int b = 1;
void add() {
a++;
b++;
}
void compare() {
if (a < b)
log.info("a:{},b:{},{}", a, b, a>b);
}
@Test
public void test() throws InterruptedException {
new Thread(() -> {
while (true)
add();
}).start();
new Thread(() -> {
while (true)
compare();
}).start();
TimeUnit.MILLISECONDS.sleep(100);
}
}
这位网友是这么问的,他说见鬼了,不但能看到日志输出,而且我发现之前判断过一次a<b,之后输出a>b居然是成立的,结果里可以看到true,JVM出现Bug了可能:
他觉得a和b不是静态的,为啥会出现并发问题呢于是问了同事:
- 同事A说是肯定多线程问题,加volatile可以解决,但是他发现为a和b加上volatile也不行
- 同事B说是AtomicInteger可以解决并发性问题,但是把a和b都用上了AtomicInteger也没用
- 同事C貌似看出了问题说需要锁,为add()方法增加synchronized关键字锁一下,但是也没用
这位网友其实是没有搞清楚多线程情况下,可见性问题、原子性问题解决的事情,同事也把各种并发的概念混淆在一起了。
我们这么来看这段代码,这段代码里一个线程不断操作a和b进行累加操作,一个线程判断a和b,然后输出结果。出现这个问题的原因本质上是因为a<b是三步操作,取a,取b以及比较,不是原子性的,在整个过程中可能穿插了add线程的操作a和b。如果先获取a,然后a++ b++,然后获取b,这个时候a<b,如果先a++,然后获取a,获取b,最后b++,这个时候a>b。我们来看一下compare()方法的字节码,可以很明显看到ab的比较分明是4行指令,我们不能以代码行数来判断操作是否是原子的,不是原子意味着操作过程中可能被穿插了其它线程的其它代码:
0 aload_0
1 getfield #2 <me/josephzhu/javaconcurrenttest/atomic/InterestingProblem.a>
4 aload_0
5 getfield #3 <me/josephzhu/javaconcurrenttest/atomic/InterestingProblem.b>
8 if_icmpge 67 (+59)
11 getstatic #4 <me/josephzhu/javaconcurrenttest/atomic/InterestingProblem.log>
14 ldc #5 <a:{},b:{},{}>
16 iconst_3
17 anewarray #6 <java/lang/Object>
20 dup
21 iconst_0
22 aload_0
23 getfield #2 <me/josephzhu/javaconcurrenttest/atomic/InterestingProblem.a>
26 invokestatic #7 <java/lang/Integer.valueOf>
29 aastore
30 dup
31 iconst_1
32 aload_0
33 getfield #3 <me/josephzhu/javaconcurrenttest/atomic/InterestingProblem.b>
36 invokestatic #7 <java/lang/Integer.valueOf>
39 aastore
40 dup
41 iconst_2
42 aload_0
43 getfield #2 <me/josephzhu/javaconcurrenttest/atomic/InterestingProblem.a>
46 aload_0
47 getfield #3 <me/josephzhu/javaconcurrenttest/atomic/InterestingProblem.b>
50 if_icmple 57 (+7)
53 iconst_1
54 goto 58 (+4)
57 iconst_0
58 invokestatic #8 <java/lang/Boolean.valueOf>
61 aastore
62 invokeinterface #9 <org/slf4j/Logger.info> count 3
67 return
所以这位网友的理解有几个问题:
- 多线程操作的对象安全不安全和对象是否静态没关系,即使不是static的也可能会并发被多个线程来操作
- 不能根据代码行数或代码是否简单来判断代码是否原子的,别说Java代码了,就是字节码也不行
我们再来看看他三位同事的说法:
- 同事A可能没有彻底理解volatile的作用,现在是两个线程操作a和b相互交错干扰,加上了volatile只会让问题更严重(你可以写一段代码对比下加上和不加上volatile最后出现true的概率),这不是可见性问题
- 同事B可能也没仔细考虑AtomicInteger的作用,AtomicInteger是用来实现多线程情况下原子性操作Integer,现在并没有多个线程来并发修改a和b,使用AtomicInteger不能解决问题
- 同事C貌似是看到了问题的所在,但是他也没理清楚,add()方法仅仅只有一个线程在执行为这个方法加上锁是没有用的,现在的问题在于add()和compare()的干扰,它们需要串行执行才能确保a和b整体的完整
所以要进行简单修复这个问题的话就是为add()和compare()都加上synchronized关键字,除了这个锁的方式有没有其它方式呢?你可以想想。
小结
本文简单测试了一下java.util.concurrent.atomic包下面的一些常用Atomic操作类,最后分享了一个网友的问题和疑惑,希望文本对你有用。
同样,代码见我的Github,欢迎clone后自己把玩,欢迎点赞。
欢迎关注我的微信公众号:随缘主人的园子