本文主要内容

  1. 4种方式实现计数器功能,对比其性能
  2. 介绍LongAdder
  3. 介绍LongAccumulator

来个需求

一个jvm中实现一个计数器功能,需保证多线程情况下数据正确性。

我们来模拟50个线程,每个线程对计数器递增100万次,最终结果应该是5000万。

我们使用4种方式实现,看一下其性能,然后引出为什么需要使用LongAdderLongAccumulator

方式一:synchronized方式实现

package com.itsoku.chat32;

import java.util.ArrayList;
import java.util.List;
import java.util.concurrent.CompletableFuture;
import java.util.concurrent.CountDownLatch;
import java.util.concurrent.ExecutionException;
import java.util.concurrent.atomic.LongAccumulator;

/**
 * 跟着阿里p7学并发,微信公众号:javacode2018
 */
public class Demo1 {
    static int count = 0;

    public static synchronized void incr() {
        count++;
    }

    public static void main(String[] args) throws ExecutionException, InterruptedException {
        for (int i = 0; i < 10; i++) {
            count = 0;
            m1();
        }
    }

    private static void m1() throws InterruptedException {
        long t1 = System.currentTimeMillis();
        int threadCount = 50;
        CountDownLatch countDownLatch = new CountDownLatch(threadCount);
        for (int i = 0; i < threadCount; i++) {
            new Thread(() -> {
                try {
                    for (int j = 0; j < 1000000; j++) {
                        incr();
                    }
                } finally {
                    countDownLatch.countDown();
                }
            }).start();
        }
        countDownLatch.await();
        long t2 = System.currentTimeMillis();
        System.out.println(String.format("结果:%s,耗时(ms):%s", count, (t2 - t1)));
    }
}

输出:

结果:50000000,耗时(ms):1437
结果:50000000,耗时(ms):1913
结果:50000000,耗时(ms):386
结果:50000000,耗时(ms):383
结果:50000000,耗时(ms):381
结果:50000000,耗时(ms):382
结果:50000000,耗时(ms):379
结果:50000000,耗时(ms):379
结果:50000000,耗时(ms):392
结果:50000000,耗时(ms):384

平均耗时:390毫秒

方式2:AtomicLong实现

package com.itsoku.chat32;

import java.util.concurrent.CountDownLatch;
import java.util.concurrent.ExecutionException;
import java.util.concurrent.atomic.AtomicLong;

/**
 * 跟着阿里p7学并发,微信公众号:javacode2018
 */
public class Demo2 {
    static AtomicLong count = new AtomicLong(0);

    public static void incr() {
        count.incrementAndGet();
    }

    public static void main(String[] args) throws ExecutionException, InterruptedException {
        for (int i = 0; i < 10; i++) {
            count.set(0);
            m1();
        }
    }

    private static void m1() throws InterruptedException {
        long t1 = System.currentTimeMillis();
        int threadCount = 50;
        CountDownLatch countDownLatch = new CountDownLatch(threadCount);
        for (int i = 0; i < threadCount; i++) {
            new Thread(() -> {
                try {
                    for (int j = 0; j < 1000000; j++) {
                        incr();
                    }
                } finally {
                    countDownLatch.countDown();
                }
            }).start();
        }
        countDownLatch.await();
        long t2 = System.currentTimeMillis();
        System.out.println(String.format("结果:%s,耗时(ms):%s", count, (t2 - t1)));
    }
}

输出:

结果:50000000,耗时(ms):971
结果:50000000,耗时(ms):915
结果:50000000,耗时(ms):920
结果:50000000,耗时(ms):923
结果:50000000,耗时(ms):910
结果:50000000,耗时(ms):916
结果:50000000,耗时(ms):923
结果:50000000,耗时(ms):916
结果:50000000,耗时(ms):912
结果:50000000,耗时(ms):908

平均耗时:920毫秒

AtomicLong内部采用CAS的方式实现,并发量大的情况下,CAS失败率比较高,导致性能比synchronized还低一些。并发量不是太大的情况下,CAS性能还是可以的。

AtomicLong属于JUC中的原子类,还不是很熟悉的可以看一下:JUC中原子类,一篇就够了

方式3:LongAdder实现

先介绍一下LongAdder,说到LongAdder,不得不提的就是AtomicLong,AtomicLong是JDK1.5开始出现的,里面主要使用了一个long类型的value作为成员变量,然后使用循环的CAS操作去操作value的值,并发量比较大的情况下,CAS操作失败的概率较高,内部失败了会重试,导致耗时可能会增加。

LongAdder是JDK1.8开始出现的,所提供的API基本上可以替换掉原先的AtomicLong。LongAdder在并发量比较大的情况下,操作数据的时候,相当于把这个数字分成了很多份数字,然后交给多个人去管控,每个管控者负责保证部分数字在多线程情况下操作的正确性。当多线程访问的时,通过hash算法映射到具体管控者去操作数据,最后再汇总所有的管控者的数据,得到最终结果。相当于降低了并发情况下锁的粒度,所以效率比较高,看一下下面的图,方便理解:

JAVA accept方法 java accumulate_JAVA accept方法

代码:

package com.itsoku.chat32;

import java.util.concurrent.CountDownLatch;
import java.util.concurrent.ExecutionException;
import java.util.concurrent.atomic.AtomicLong;
import java.util.concurrent.atomic.LongAdder;

/**
 * 跟着阿里p7学并发,微信公众号:javacode2018
 */
public class Demo3 {
    static LongAdder count = new LongAdder();

    public static void incr() {
        count.increment();
    }

    public static void main(String[] args) throws ExecutionException, InterruptedException {
        for (int i = 0; i < 10; i++) {
            count.reset();
            m1();
        }
    }

    private static void m1() throws ExecutionException, InterruptedException {
        long t1 = System.currentTimeMillis();
        int threadCount = 50;
        CountDownLatch countDownLatch = new CountDownLatch(threadCount);
        for (int i = 0; i < threadCount; i++) {
            new Thread(() -> {
                try {
                    for (int j = 0; j < 1000000; j++) {
                        incr();
                    }
                } finally {
                    countDownLatch.countDown();
                }
            }).start();
        }
        countDownLatch.await();
        long t2 = System.currentTimeMillis();
        System.out.println(String.format("结果:%s,耗时(ms):%s", count.sum(), (t2 - t1)));
    }
}

输出:

结果:50000000,耗时(ms):206
结果:50000000,耗时(ms):105
结果:50000000,耗时(ms):107
结果:50000000,耗时(ms):107
结果:50000000,耗时(ms):105
结果:50000000,耗时(ms):99
结果:50000000,耗时(ms):106
结果:50000000,耗时(ms):102
结果:50000000,耗时(ms):106
结果:50000000,耗时(ms):102

平均耗时:100毫秒

代码中new LongAdder()创建一个LongAdder对象,内部数字初始值是0,调用increment()方法可以对LongAdder内部的值原子递增1。reset()方法可以重置LongAdder的值,使其归0。

方式4:LongAccumulator实现

LongAccumulator介绍

LongAccumulator是LongAdder的功能增强版。LongAdder的API只有对数值的加减,而LongAccumulator提供了自定义的函数操作,其构造函数如下:

/**
  * accumulatorFunction:需要执行的二元函数(接收2个long作为形参,并返回1个long)
  * identity:初始值
 **/
public LongAccumulator(LongBinaryOperator accumulatorFunction, long identity) {
    this.function = accumulatorFunction;
    base = this.identity = identity;
}

示例代码:

package com.itsoku.chat32;

import java.util.concurrent.CountDownLatch;
import java.util.concurrent.ExecutionException;
import java.util.concurrent.atomic.LongAccumulator;
import java.util.concurrent.atomic.LongAdder;

/**
 * 跟着阿里p7学并发,微信公众号:javacode2018
 */
public class Demo4 {
    static LongAccumulator count = new LongAccumulator((x, y) -> x + y, 0L);

    public static void incr() {
        count.accumulate(1);
    }

    public static void main(String[] args) throws ExecutionException, InterruptedException {
        for (int i = 0; i < 10; i++) {
            count.reset();
            m1();
        }
    }

    private static void m1() throws ExecutionException, InterruptedException {
        long t1 = System.currentTimeMillis();
        int threadCount = 50;
        CountDownLatch countDownLatch = new CountDownLatch(threadCount);
        for (int i = 0; i < threadCount; i++) {
            new Thread(() -> {
                try {
                    for (int j = 0; j < 1000000; j++) {
                        incr();
                    }
                } finally {
                    countDownLatch.countDown();
                }
            }).start();
        }
        countDownLatch.await();
        long t2 = System.currentTimeMillis();
        System.out.println(String.format("结果:%s,耗时(ms):%s", count.longValue(), (t2 - t1)));
    }
}

输出:

结果:50000000,耗时(ms):138
结果:50000000,耗时(ms):111
结果:50000000,耗时(ms):111
结果:50000000,耗时(ms):103
结果:50000000,耗时(ms):103
结果:50000000,耗时(ms):105
结果:50000000,耗时(ms):101
结果:50000000,耗时(ms):106
结果:50000000,耗时(ms):102
结果:50000000,耗时(ms):103

平均耗时:100毫秒

LongAccumulator的效率和LongAdder差不多,不过更灵活一些。

调用new LongAdder()等价于new LongAccumulator((x, y) -> x + y, 0L)

从上面4个示例的结果来看,LongAdder、LongAccumulator全面超越同步锁及AtomicLong的方式,建议在使用AtomicLong的地方可以直接替换为LongAdder、LongAccumulator,吞吐量更高一些。