本節我們來研究下併發包中的Atomic型別。
AtomicXXX和XXXAdder以及XXXAccumulator效能測試
先來一把效能測試,對比一下AtomicLong(1.5出來的)、LongAdder(1.8出來的)和LongAccumulator(1.8出來的)用於簡單累加的效能。
程式邏輯比較簡單,可以看到我們在最大併發10的情況下,去做10億次操作測試:
@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();
}
}
結果如下:
和官網說的差不多,在高併發的情況下LongAdder效能會比AtomicLong好很多。
AtomicReference可見性問題測試
在很多開原始碼中我們有看到AtomicReference的身影,它究竟是幹什麼的呢?我們來寫一段測試程式,在這個程式中我們定一了一個Switch型別,作為一個開關,然後寫三個死迴圈的執行緒來測試,當開關有效的時候會持續死迴圈,在2秒後關閉所有的三個開關:
- 第一個是普通的Switch
- 第二個是使用了volatile宣告的Switch
- 第三個是AtomicReference包裝的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;
}
}
}
執行程式:
可以看到2秒後有一個開關卡住了,執行緒沒有退出。這是一個可見性的問題,AtomicReference以及volatile可以確保執行緒對資料的更新重新整理到記憶體。因為我們對於開關的關閉是在另一個定時任務執行緒做的,如果我們不使用volatile或AtomicReference來定義物件,那麼物件的操作可能無法被其它執行緒感知到。當然,AtomicReference除了解決可見性問題還有更多AtomicXXX提供的其它功能。
AtomicInteger測試
下面我們來看一下AtomicInteger的compareAndSet()功能。首先說明這個程式沒有任何意義,只是測試一下功能。在這個程式裡,我們亂序開啟10個執行緒,每一個執行緒的任務就是按照次序來累加數字。我們使用AtomicInteger的compareAndSet()來確保亂序的執行緒也能按照我們要的順序操作累加。
@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());
}
}
執行結果如下:
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的輸出是亂序的,最後Done的輸出是順序的。
AtomicStampedReference測試
AtomicStampedReference可以用來解決ABA問題,什麼是ABA問題我們看這個例子:
執行緒1讀取了數字之後,等待1秒,然後嘗試把1修改為3。
執行緒2後啟動,讀取到數字1後修改2,稍等一下又修改回1。
雖然AtomicInteger確保多個執行緒的原子性操作,但是無法確保1就是原先讀取到的那個1,沒有經過別人修改。
可以再換一個例子來說,如果我們現在賬上有100元,要修改為200元,在修改之前賬戶已經被操作過了從100元充值到了150然後提現到了100,雖然最後還是回到了100,但是這個時候嚴格一點的話,我們應該認為這個100不是原先的100,這個賬戶的版本發生了變化,如果我們使用樂觀行鎖的話,雖然餘額都是100但是行鎖的版本肯定不一致,AtomicStampedReference就是類似行樂觀鎖的概念。
@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();
}
看下執行結果:
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
下面我們使用AtomicStampedReference來修復這個問題:
@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();
}
執行結果如下:
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,現在我們修改資料的時候不僅僅是拿著值來修改了,還要提供版本號,讀取資料的時候可以讀取到資料以及版本號。這樣的話,雖然數值不變,但是執行緒2經過兩次修改後資料的版本從1變為了3,回過頭來執行緒1再要拿著版本號1來修改資料的話必然失敗。
一個有趣的問題
本文比較短,我們再來看網友之前問的一個有意思的問題,程式如下。
@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);
}
}
這位網友是這麼問的,他說見鬼了,不但能看到日誌輸出,而且我發現之前判斷過一次a<b,之後輸出a>b居然是成立的,結果裡可以看到true,JVM出現Bug了可能:
他覺得a和b不是靜態的,為啥會出現併發問題呢於是問了同事:
- 同事A說是肯定多執行緒問題,加volatile可以解決,但是他發現為a和b加上volatile也不行
- 同事B說是AtomicInteger可以解決併發性問題,但是把a和b都用上了AtomicInteger也沒用
- 同事C貌似看出了問題說需要鎖,為add()方法增加synchronized關鍵字鎖一下,但是也沒用
這位網友其實是沒有搞清楚多執行緒情況下,可見性問題、原子性問題解決的事情,同事也把各種併發的概念混淆在一起了。
我們這麼來看這段程式碼,這段程式碼裡一個執行緒不斷操作a和b進行累加操作,一個執行緒判斷a和b,然後輸出結果。出現這個問題的原因本質上是因為a<b是三步操作,取a,取b以及比較,不是原子性的,在整個過程中可能穿插了add執行緒的操作a和b。如果先獲取a,然後a++ b++,然後獲取b,這個時候a<b,如果先a++,然後獲取a,獲取b,最後b++,這個時候a>b。我們來看一下compare()方法的位元組碼,可以很明顯看到ab的比較分明是4行指令,我們不能以程式碼行數來判斷操作是否是原子的,不是原子意味著操作過程中可能被穿插了其它執行緒的其它程式碼:
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
所以這位網友的理解有幾個問題:
- 多執行緒操作的物件安全不安全和物件是否靜態沒關係,即使不是static的也可能會併發被多個執行緒來操作
- 不能根據程式碼行數或程式碼是否簡單來判斷程式碼是否原子的,別說Java程式碼了,就是位元組碼也不行
我們再來看看他三位同事的說法:
- 同事A可能沒有徹底理解volatile的作用,現在是兩個執行緒操作a和b相互交錯干擾,加上了volatile只會讓問題更嚴重(你可以寫一段程式碼對比下加上和不加上volatile最後出現true的概率),這不是可見性問題
- 同事B可能也沒仔細考慮AtomicInteger的作用,AtomicInteger是用來實現多執行緒情況下原子性操作Integer,現在並沒有多個執行緒來併發修改a和b,使用AtomicInteger不能解決問題
- 同事C貌似是看到了問題的所在,但是他也沒理清楚,add()方法僅僅只有一個執行緒在執行為這個方法加上鎖是沒有用的,現在的問題在於add()和compare()的干擾,它們需要序列執行才能確保a和b整體的完整
所以要進行簡單修復這個問題的話就是為add()和compare()都加上synchronized關鍵字,除了這個鎖的方式有沒有其它方式呢?你可以想想。
小結
本文簡單測試了一下java.util.concurrent.atomic包下面的一些常用Atomic操作類,最後分享了一個網友的問題和疑惑,希望文字對你有用。
同樣,程式碼見我的Github,歡迎clone後自己把玩,歡迎點贊。
歡迎關注我的微信公眾號:隨緣主人的園子