盘点Java线程池配置的常见误区，你中了几个？

{"type":"doc","content":[{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"前言","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"由于线程的创建和销毁对操作系统来说都是比较重量级的操作，所以线程的池化在各种语言内都有实践，当然在 Java 语言中线程池是也非常重要的一部分，有 Doug Lea 大神对线程池的封装，我们使用的时候是非常方便，但也可能会因为不了解其具体实现，对线程池的配置参数存在误解。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我们经常在一些技术书籍或博客上看到，向线程池提交任务时，线程池的执行逻辑如下：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"1、当一个任务被提交后，线程池首先检查正在运行的线程数是否达到核心线程数，如果未达到则创建一个线程。","attrs":{}}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"2、如果线程池内正在运行的线程数已经达到了核心线程数，任务将会被放到 BlockingQueue 内。","attrs":{}}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"3、如果 BlockingQueue 已满，线程池将会尝试将线程数扩充到最大线程池容量。","attrs":{}}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"4、如果当前线程池内线程数量已经达到最大线程池容量，则会执行拒绝策略拒绝任务提交。流程如图（摘自美团技术博客）：","attrs":{}}]}]}],"attrs":{}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/70/705649a21d5143fb030d4806fc8d6a09.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"流程描述没有问题，但如果某些点未经过推敲，容易导致误解，而且描述中的情境太理想化，如果配置时不考虑运行时环境，也会出现一些非常诡异的问题。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"核心池","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"线程池内线程数量小于等于 coreSize 的部分我称为核心池，核心池是线程池的常驻部分，内部的线程一般不会被销毁，我们提交的任务也应该绝大部分都由核心池内的线程来执行。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"线程创建时机的误解","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"有关核心池最常见的一个误区是没搞清楚核心池内线程的创建时机，这个问题，我觉得甩 10% 的锅给 Doug Lea 大神应该不算过分，因为他在文档里写道 “If fewer than corePoolSize threads are running, try to start a new thread with the given command as its first task”，其中 \"running\" 这个词就比较有歧义，因为在我们理解里 running 是指当前线程已被操作系统调度，拥有操作系统时间分片，或者被理解为正在执行某个任务。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"基于以上的理解，我们很容易就认为如果任务的 QPS 非常低，线程池内线程数量永远也达不到 coreSize。即如果我们配置了 coreSize 为 1000，实际上 QPS 只有 1，单个任务耗时 1s，那么核心池大小就会一直是 1，即使有流量抖动，核心池也只会被扩容到 3。因为一个线程每秒执行执行一个任务，刚好不用创建新线程就足以应对 1QPS。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"创建过程","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"但如果简单设计一个测试，使用 jstack 打印出线程栈并数一下线程池内线程数量，会发现线程池内的线程数会随着任务的提交而逐渐增大，直到达到 coreSize。因为核心池的设计初衷是想它能作为常驻池，承载日常流量，所以它应该被尽快初始化，于是线程池的逻辑是在没有达到 coreSize 之前，每一个任务都会创建一个新的线程，对应的源码为：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"public void execute(Runnable command) {\\\n ...\\\n int c = ctl.get();\\\n if (workerCountOf(c) < corePoolSize) { // workerCountOf() 方法是获取线程池内线程数量\\\n if (addWorker(command, true))\\\n return;\\\n c = ctl.get();\\\n }\\\n ...\\\n }\n","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"而文档里的 running 状态也指的是线程已经被创建，我们也知道线程被创建后，会在一个 while 循环里尝试从 BlockingQueue 里获取并执行任务，说它正在 running 也不为过。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"基于此，我们对一些高并发服务进行的预热，其实并不是期望 JVM 能对热点代码做 JIT 等优化，对线程池、连接池和本地缓存的预热才是重点。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"BlockingQueue","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"BlockingQueue 是线程池内的另一个重要组件，首先它是线程池”生产者-消费者”模型的中间媒介，另外它也可以为大量突发的流量做缓冲，但理解和配置它也经常会出错。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"运行模型","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"最常见的错误是不理解线程池的运行模型。首先要明确的一点是线程池并没有准确的调度功能，即它无法感知有哪些线程是处于空闲状态的，并把提交的任务派发给空闲线程。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"线程池采用的是”生产者-消费者”模式，除了触发线程创建的任务（线程的 firstTask）不会入 BlockingQueue 外，其他任务都要进入到 BlockingQueue，等待线程池内的线程消费，而任务会被哪个线程消费到完全取决于操作系统的调度。对应的生产者源码如下：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"public void execute(Runnable command) {\\\n ...\\\n if (isRunning(c) && workQueue.offer(command)) { isRunning() 是判断线程池处理戚状态\\\n int recheck = ctl.get();\\\n if (! isRunning(recheck) && remove(command))\\\n reject(command);\\\n else if (workerCountOf(recheck) == 0)\\\n addWorker(null, false);\\\n }\\\n ...\\\n }\n","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"对应的消费者源码如下：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"private Runnable getTask() {\\\n for (;;) {\\\n ...\\\n Runnable r = timed ?\\\n workQueue.poll(keepAliveTime, TimeUnit.NANOSECONDS) :\\\n workQueue.take();\\\n if (r != null)\\\n return r;\\\n ...\\\n }\\\n }\n","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"BlockingQueue 的缓冲作用","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"基于”生产者-消费者”模型，我们可能会认为如果配置了足够的消费者，线程池就不会有任何问题。其实不然，我们还必须考虑并发量这一因素。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"设想以下情况：有 1000 个任务要同时提交到线程池内并发执行，在线程池被初始化完成的情况下，它们都要被放到 BlockingQueue 内等待被消费，在极限情况下，消费线程一个任务也没有执行完成，那么这 1000 个请求需要同时存在于 BlockingQueue 内，如果配置的 BlockingQueue Size 小于 1000，多余的请求就会被拒绝。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"那么这种极限情况发生的概率有多大呢？答案是非常大，因为操作系统对 I/O 线程的调度优先级是非常高的，一般我们的任务都是由 I/O 的准备或完成（如 tomcat 受理了 http 请求）开始的，所以很有可能被调度到的都是 tomcat 线程，它们在一直往线程池内提交请求，而消费者线程却调度不到，导致请求堆积。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我负责的服务就发生过这种请求被异常拒绝的情况，压测时 QPS 2000，平均响应时间为 20ms，正常情况下，40 个线程就可以平衡生产速度，不会堆积。但在 BlockingQueue Size 为 50 时，即使线程池 coreSize 为 1000，还会出现请求被线程池拒绝的情况。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这种情况下，BlockingQueue 的重要的意义就是它是一个能长时间存储任务的容器，能以很小的代价为线程池提供缓冲。根据上文可知，线程池能支持BlockingQueue Size个任务同时提交，我们把最大同时提交的任务个数，称为并发量，配置线程池时，了解并发量异常重要。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"并发量的计算","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我们常用 QPS 来衡量服务压力，所以配置线程池参数时也经常参考这个值，但有时候 QPS 和并发量有时候相关性并没有那么高，QPS 还要搭配任务执行时间来推算峰值并发量。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"比如请求间隔严格相同的接口，平均 QPS 为 1000，它的并发量峰值是多少呢？我们并没有办法估算，因为如果任务执行时间为 1ms，那么它的并发量只有 1；而如果任务执行时间为 1s，那么并发量峰值为 1000。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"可是知道了任务执行时间，就能算出并发量了吗？也不能，因为如果请求的间隔不同，可能 1min 内的请求都在一秒内发过来，那这个并发量还要乘以 60，所以上面才说知道了 QPS 和任务执行时间，并发量也只能靠推算。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"计算并发量，我一般的经验值是 QPS * 平均响应时间，再留上一倍的冗余，但如果业务重要的话，BlockingQueue Size 设置大一些也无妨（1000 或以上），毕竟每个任务占用的内存量很有限。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"考虑运行时","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"GC","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"除了上面提到的各种情况下，GC 也是一个很重要的影响因素。我们都知道 GC 是 Stop the World 的，但这里的 World 指的是 JVM，而一个请求 I/O 的准备和完成是操作系统在进行的，JVM 停止了，但操作系统还是会正常受理请求，在 JVM 恢复后执行，所以 GC 是会堆积请求的。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"上文中提到的并发量计算一定要考虑到 GC 时间内堆积的请求同时被受理的情况，堆积的请求数可以通过 QPS * GC时间 来简单得出，还有一定要记得留出冗余。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"业务峰值","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"除此之外，配置线程池参数时，一定要考虑业务场景。假如接口的流量大部分来自于一个定时程序，那么平均 QPS 就没有了任何意义，线程池设计时就要考虑给 BlockingQueue 的 Size 设置一个大一些的值；而如果流量非常不平均，一天内只有某一小段时间才有高流量的话，而且线程资源紧张的情况下，就要考虑给线程池的 maxSize 留下较大的冗余；在流量尖刺明显而响应时间不那么敏感时，也可以设置较大的 BlockingQueue，允许任务进行一定程度的堆积。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"当然除了经验和计算外，对服务做定时的压测无疑更能帮助掌握服务真实的情况。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"小结","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"总结线程池的配置时，我最大的感受是一定要读源码！读源码！读源码！只看一些书和文章的总结是无法吃透一些重要概念的，即使搞懂了大部分也很容易会在一些角落踩坑。深入理解原理后，面对复杂情况，才有灵活配置的能力。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"关注公众号：北游学Java，回复【721】即可领取我整理好的线程池学习资料以及精选面试题。","attrs":{}}]}]}