Java针对多线程下的数值安全计数器设计了一些类,这些类叫做原子类,其中一部分如下:
java.util.concurrent.atomic.AtomicBoolean;java.util.concurrent.atomic.AtomicInteger;java.util.concurrent.atomic.AtomicLong;java.util.concurrent.atomic.AtomicReference;
下面是一个对比 AtomicInteger 与 普通 int 值在多线程下的递增测试,使用的是 junit4;
完整代码:
package test.java;import java.util.concurrent.CountDownLatch;import java.util.concurrent.atomic.AtomicInteger;import org.junit.Assert;import org.junit.Before;import org.junit.Test;/** * 测试AtomicInteger与普通int值在多线程下的递增操作 */public class TestAtomic { // 原子Integer递增对象 public static AtomicInteger counter_integer;// = new AtomicInteger(0); // 一个int类型的变量 public static int count_int = 0; @Before public void setUp() { // 所有测试开始之前执行初始设置工作 counter_integer = new AtomicInteger(0); } @Test public void testAtomic() throws InterruptedException { // 创建的线程数量 int threadCount = 100; // 其他附属线程内部循环多少次 int loopCount = 10000600; // 控制附属线程的辅助对象;(其他await的线程先等着主线程喊开始) CountDownLatch latch_1 = new CountDownLatch(1); // 控制主线程的辅助对象;(主线程等着所有附属线程都运行完毕再继续) CountDownLatch latch_n = new CountDownLatch(threadCount); // 创建并启动其他附属线程 for (int i = 0; i < threadCount; i++) { Thread thread = new AtomicIntegerThread(latch_1, latch_n, loopCount); thread.start(); } long startNano = System.nanoTime(); // 让其他等待的线程统一开始 latch_1.countDown(); // 等待其他线程执行完 latch_n.await(); // long endNano = System.nanoTime(); int sum = counter_integer.get(); // Assert.assertEquals("sum 不等于 threadCount * loopCount,测试失败", sum, threadCount * loopCount); System.out.println("--------testAtomic(); 预期两者相等------------"); System.out.println("耗时: " + ((endNano - startNano) / (1000 * 1000)) + "ms"); System.out.println("threadCount = " + (threadCount) + ";"); System.out.println("loopCount = " + (loopCount) + ";"); System.out.println("sum = " + (sum) + ";"); } @Test public void testIntAdd() throws InterruptedException { // 创建的线程数量 int threadCount = 100; // 其他附属线程内部循环多少次 int loopCount = 10000600; // 控制附属线程的辅助对象;(其他await的线程先等着主线程喊开始) CountDownLatch latch_1 = new CountDownLatch(1); // 控制主线程的辅助对象;(主线程等着所有附属线程都运行完毕再继续) CountDownLatch latch_n = new CountDownLatch(threadCount); // 创建并启动其他附属线程 for (int i = 0; i < threadCount; i++) { Thread thread = new IntegerThread(latch_1, latch_n, loopCount); thread.start(); } long startNano = System.nanoTime(); // 让其他等待的线程统一开始 latch_1.countDown(); // 等待其他线程执行完 latch_n.await(); // long endNano = System.nanoTime(); int sum = count_int; // Assert.assertNotEquals( "sum 等于 threadCount * loopCount,testIntAdd()测试失败", sum, threadCount * loopCount); System.out.println("-------testIntAdd(); 预期两者不相等---------"); System.out.println("耗时: " + ((endNano - startNano) / (1000*1000))+ "ms"); System.out.println("threadCount = " + (threadCount) + ";"); System.out.println("loopCount = " + (loopCount) + ";"); System.out.println("sum = " + (sum) + ";"); } // 线程 class AtomicIntegerThread extends Thread { private CountDownLatch latch = null; private CountDownLatch latchdown = null; private int loopCount; public AtomicIntegerThread(CountDownLatch latch, CountDownLatch latchdown, int loopCount) { this.latch = latch; this.latchdown = latchdown; this.loopCount = loopCount; } @Override public void run() { // 等待信号同步 try { this.latch.await(); } catch (InterruptedException e) { e.printStackTrace(); } // for (int i = 0; i < loopCount; i++) { counter_integer.getAndIncrement(); } // 通知递减1次 latchdown.countDown(); } } // 线程 class IntegerThread extends Thread { private CountDownLatch latch = null; private CountDownLatch latchdown = null; private int loopCount; public IntegerThread(CountDownLatch latch, CountDownLatch latchdown, int loopCount) { this.latch = latch; this.latchdown = latchdown; this.loopCount = loopCount; } @Override public void run() { // 等待信号同步 try { this.latch.await(); } catch (InterruptedException e) { e.printStackTrace(); } // for (int i = 0; i < loopCount; i++) { count_int++; } // 通知递减1次 latchdown.countDown(); } }}普通PC机上的执行结果类似如下:
--------------testAtomic(); 预期两者相等-------------------耗时: 85366msthreadCount = 100;loopCount = 10000600;sum = 1000060000;--------------testIntAdd(); 预期两者不相等-------------------耗时: 1406msthreadCount = 100;loopCount = 10000600;sum = 119428988;
从中可以看出, AtomicInteger操作 与 int操作的效率大致相差在50-80倍上下,当然,int很不消耗时间,这个对比只是提供一个参照。
如果确定是单线程执行,那应该使用 int; 而int在多线程下的操作执行的效率还是蛮高的, 10亿次只花了1.5秒钟；
(假设CPU是 2GHZ，双核4线程,理论最大8GHZ，则每秒理论上有80亿个时钟周期,
10亿次Java的int增加消耗了1.5秒,即 120亿次运算, 算下来每次循环消耗CPU周期 12个;
个人觉得效率不错, C 语言也应该需要4个以上的时钟周期(判断,执行内部代码,自增判断,跳转)
前提是: JVM和CPU没有进行激进优化.
)
而 AtomicInteger 效率其实也不低,10亿次消耗了80秒, 那100万次大约也就是千分之一,80毫秒的样子.