当 corePoolSize = 0 时，ScheduledExecutorService 消耗 100% CPU

Question

我运行在生产中遇到了一个有趣的问题。

我有以下 ScheduledThreadPool 分配代码：

ScheduledExecutorService executorService =
            Executors.newScheduledThreadPool(Runtime.getRuntime().availableProcessors() - 1);

线程池正在定期处理队列中的一些任务。在将服务部署到单核环境之前，一切都运行良好。显然，上面的行转换为：

ScheduledExecutorService executorService = Executors.newScheduledThreadPool(0);

从那时起，JVM 进程 CPU 利用率一直保持在 100% 左右。当我将 Runtime.getRuntime().availableProcessors() - 1 更改为常量 1 时，问题就消失了。

找出根本原因花了一些时间，但我仍然不知道背后的原因。 ScheduledExecutorService JavaDoc 指出：

/**
 * Creates a thread pool that can schedule commands to run after a
 * given delay, or to execute periodically.
 * @param corePoolSize the number of threads to keep in the pool,
 * even if they are idle
 * @return a newly created scheduled thread pool
 * @throws IllegalArgumentException if {@code corePoolSize < 0}
 */
public static ScheduledExecutorService newScheduledThreadPool(int corePoolSize) {
    return new ScheduledThreadPoolExecutor(corePoolSize);
}

基本上，0（零）是线程池实例化的有效参数，但它与此值一起工作时超级奇怪。

有人可以解释一下为什么吗？

简单且可验证的测试用例

import java.util.Queue;
import java.util.concurrent.*;

public class Test {

    public static void main(String[] args) throws InterruptedException {
        MessageTaskExecutor asyncEmailGatewayTaskExecutor = new MessageTaskExecutor();

        // Infinitely add new tasks to the queue every second
        for (int i = 1; ; i++) {
            System.out.println(String.format("Adding message #%s to the queue", i));

            asyncEmailGatewayTaskExecutor.putMessageIntoQueue(i);

            Thread.sleep(1_000);
        }
    }

    static class MessageTaskExecutor {

        static final int INITIAL_DELAY_SECONDS = 1;
        static final int PROCESSING_RATE_MILLISECONDS = 5_000;

        final Queue<Runnable> messageQueue = new ArrayBlockingQueue<>(1_000_000);
        final ScheduledExecutorService executorService = Executors.newScheduledThreadPool(0);

        MessageTaskExecutor() {
            // Scavenging Message Tasks Queue every 'PROCESSING_RATE_MILLISECONDS'. Initial delay is fixed for 'INITIAL_DELAY_SECONDS'
            executorService.schedule(this::processEmailTasks, INITIAL_DELAY_SECONDS, TimeUnit.SECONDS);
        }

        void putMessageIntoQueue(int messageId) {
            Runnable messageTask = () -> System.out.println(String.format("Message #%s is getting processed!", messageId));

            messageQueue.offer(messageTask);
        }

        void processEmailTasks() {
            System.out.println(String.format("There are %s messages in the queue. Processing the messages...", messageQueue.size()));

            // Processing messages queue
            while (!messageQueue.isEmpty()) {
                executorService.submit(messageQueue.poll()); // Submitting task to executor service
            }

            // Re-scheduling processing job
            executorService.schedule(this::processEmailTasks, PROCESSING_RATE_MILLISECONDS, TimeUnit.MILLISECONDS);
        }
    }
}

此代码在单核虚拟机上分配 ~30 MB 并且 JVM 进程消耗 ~100% CPU （在 Win 7/CentOS 7 上测试）。 JDK1.8.0.181.

通过更改字段：

final ScheduledExecutorService executorService = Executors.newScheduledThreadPool(0);

至：

final ScheduledExecutorService executorService = Executors.newScheduledThreadPool(1);

CPU 消耗下降到正常 3-5%。

Answer 1

这是一个已知错误：JDK-8129861。它已在 JDK 9.

中修复

解决方法是将核心池大小至少设置为 1:

int corePoolSize = Math.max(Runtime.getRuntime().availableProcessors() - 1, 1);

当 corePoolSize = 0 时，ScheduledExecutorService 消耗 100% CPU

ScheduledExecutorService consumes 100% CPU when corePoolSize = 0

java

performance

threadpool

threadpoolexecutor

简单且可验证的测试用例