我要评论前言
在即将发布的 .net 6 runtime 中，默认的线程池实现从 c++ 代码改为了 c#，更方便我们学习线程池的设计了。
https://github.com/dotnet/runtime/tree/release/6.0/src/libraries/system.threading.threadpool
新的线程池实现位于 portablethreadpool 中，原 threadpool 中的对外公开的接口会直接调用 portablethreadpool 中的实现。
通过设置环境变量 threadpool_useportablethreadpool 为 0 可以设置成使用老的线程池实现。
https://github.com/dotnet/runtime/pull/43841/commits/b0d47b84a6845a70f011d1b0d3ce5adde9a4d7b7
本文以 .net 6 runtime 源码作为学习材料，对线程池的设计进行介绍。从目前的理解上来看，其整体的设计与原来 c++ 的实现并没有特别大的出入。
注意：
- 本文不涉及细节的代码实现，主要为大家介绍其整体设计。所展示的代码并非原封不动的源码，而是为了方便理解的简化版。
threadpool.setmaxthreads(int workerthreads, int completionportthreads)中的completionportthreads所相关的iocp线程池是 .net framework 时代的遗留产物，用于管理 windows 平台专有的 iocp 的回调线程池。目前没看到有什么地方在用它了，completionportthreads 这个参数也已经没有意义，底层io库是自己维护的io等待线程池。本文只涉及 worker thread 池的介绍。- 本文理解并不完整也不一定完全正确，有异议的地方欢迎留言讨论。
- 为了解释问题，一部分代码会运行在 .net 6 之前的环境中。
任务的调度
线程池的待执行任务被存放在一个队列系统中。这个系统包括一个 全局队列，以及绑定在每一个 worker thread 上 的 本地队列 。而线程池中的每一个线程都在执行 while(true) 的循环，从这个队列系统中领取并执行任务。
在 threadpool.queueuserworkitem 的重载方法 threadpool.queueuserworkitem<tstate>(action<tstate> callback, tstate state, bool preferlocal) 里有一个 preferlocal 参数。
调用不带 preferlocal 参数的 threadpool.queueuserworkitem 方法重载，任务会被放到全局队列。
当 preferlocal 为 true 的时候，如果调用 threadpool.queueuserworkitem 代码的线程正好是个线程池里的某个线程，则该任务就会进入该线程的本地队列中。除此之外的情况则会被放到全局队列中等待未来被某个 worker thread 捡走。
在线程池外的线程中调用，不管 preferlocal 传的是什么，任务都会被放到全局队列。
基本调度单元
本地队列和全局队列的元素类型被定义为 object，实际的任务类型分为两类，在从队列系统取到任务之后会判断类型并执行对应的方法。
ithreadpoolworkitem 实现类的实例。
/// <summary>represents a work item that can be executed by the threadpool.</summary>
public interface ithreadpoolworkitem
{
void execute();
}
执行 execute 方法也就代表着任务的执行。
ithreadpoolworkitem 的具体实现有很多，例如通过 threadpool.queueuserworkitem(waitcallback callback) 传入的 callback 委托实例会被包装到一个 queueuserworkitemcallback 实例里。queueuserworkitemcallback 是 ithreadpoolworkitem 的实现类。
task
class task
{
internal void innerinvoke();
}
执行 innerinvoke 会执行 task 所包含的委托。
全局队列
全局队列 是由 threadpoolworkqueue 维护的，同时它也是整个队列系统的入口，直接被 threadpool 所引用。
public static class threadpool
{
internal static readonly threadpoolworkqueue s_workqueue = new threadpoolworkqueue();
public static bool queueuserworkitem(waitcallback callback, object state)
{
object tpcallback = new queueuserworkitemcallback(callback!, state);
s_workqueue.enqueue(tpcallback, forceglobal: true);
return true;
}
}
internal sealed class threadpoolworkqueue
{
// 全局队列
internal readonly concurrentqueue<object> workitems = new concurrentqueue<object>();
// forceglobal 为 true 时，push 到全局队列，否则就放到本地队列
public void enqueue(object callback, bool forceglobal);
}
本地队列
线程池中的每一个线程都会绑定一个 threadpoolworkqueuethreadlocals 实例，在 workstealingqueue 这个字段上保存着本地队列。
internal sealed class threadpoolworkqueuethreadlocals
{
// 绑定在线程池线程上
[threadstatic]
public static threadpoolworkqueuethreadlocals threadlocals;
// 持有全局队列的引用，以便能在需要的时候将任务转移到全局队列上
public readonly threadpoolworkqueue workqueue;
// 本地队列的直接维护者
public readonly threadpoolworkqueue.workstealingqueue workstealingqueue;
public readonly thread currentthread;
public threadpoolworkqueuethreadlocals(threadpoolworkqueue tpq)
{
workqueue = tpq;
workstealingqueue = new threadpoolworkqueue.workstealingqueue();
// workstealingqueuelist 会集中管理 workstealingqueue
threadpoolworkqueue.workstealingqueuelist.add(workstealingqueue);
currentthread = thread.currentthread;
}
// 提供将本地队列中的任务转移到全局队列中去的功能，
// 当 threadpool 通过后文将会介绍的 hillclimbing 算法判断得出当前线程是多余的线程后，
// 会调用此方法对任务进行转移
public void transferlocalwork()
{
while (workstealingqueue.localpop() is object cb)
{
workqueue.enqueue(cb, forceglobal: true);
}
}
~threadpoolworkqueuethreadlocals()
{
if (null != workstealingqueue)
{
// transferlocalwork 真正的目的并非是为了在这里被调用，这边只是确保任务不会丢的 fallback 逻辑
transferlocalwork();
threadpoolworkqueue.workstealingqueuelist.remove(workstealingqueue);
}
}
}
偷窃机制
这里思考一个问题，为什么本地队列的名字会被叫做 workstealingqueue 呢？
所有 worker thread 的 workstealingqueue 都被集中在 workstealingqueuelist 中。对线程池中其他所有线程可见。
worker thread 的 while(true) 中优先会从自身的 workstealingqueue 中取任务。如果本地队列已经被清空，就会从全局队列中取任务。例如下图的 thread1 取全局队列中领取了一个任务。
同时 thread3 也没活干了，但是全局队列中的任务被 thread1 抢走了。这时候就会去 从 thread2 的本地队列中抢 thread2 的活。
worker thread 的生命周期管理
接下来我们把格局放大，关注点从 worker thread 的打工日常转移到对它们的生命周期管理上来。
为了更方便的解释线程管理的机制，这边使用下面使用一些代码做演示。
代码参考自 https://devblogs.microsoft.com/dotnet/performance-improvements-in-net-6/。
线程注入实验
task.run 会将 task 调度到线程池中执行，下面的示例代码中等效于 threadpool.queueuserworkitem(waitcallback callback)，会把 task 放到队列系统的全局队列中（顺便一提，如果在一个线程池线程中执行 task.run 会将 task 调度到此线程池线程的本地队列中）。
.net 5 实验一 默认线程池配置
static void main(string[] args)
{
var sw = stopwatch.startnew();
var tcs = new taskcompletionsource();
var tasks = new list<task>();
for (int i = 1; i <= environment.processorcount * 2; i++)
{
int id = i;
console.writeline($"loop id: {id:00} | {sw.elapsed.totalseconds:0.000} | busy threads: {getbusythreads()}");
tasks.add(task.run(() =>
{
console.writeline($"task id: {id:00} | {sw.elapsed.totalseconds:0.000} | busy threads: {getbusythreads()}");
tcs.task.wait();
}));
}
tasks.add(task.run(() =>
{
console.writeline($"task setresult | {sw.elapsed.totalseconds:0.000} | busy threads: {getbusythreads()}");
tcs.setresult();
}));
task.waitall(tasks.toarray());
console.writeline($"done: | {sw.elapsed.totalseconds:0.000}");
}
static int getbusythreads()
{
threadpool.getavailablethreads(out var available, out _);
threadpool.getmaxthreads(out var max, out _);
return max - available;
}
首先在代码在 .net 5 环境中运行以下代码，cpu 逻辑核心数 12。
loop id: 01 | 0.000 | busy threads: 0 loop id: 02 | 0.112 | busy threads: 1 loop id: 03 | 0.112 | busy threads: 2 loop id: 04 | 0.113 | busy threads: 4 loop id: 05 | 0.113 | busy threads: 7 loop id: 06 | 0.113 | busy threads: 10 loop id: 07 | 0.113 | busy threads: 10 task id: 01 | 0.113 | busy threads: 11 task id: 02 | 0.113 | busy threads: 12 task id: 03 | 0.113 | busy threads: 12 task id: 07 | 0.113 | busy threads: 12 task id: 04 | 0.113 | busy threads: 12 task id: 05 | 0.113 | busy threads: 12 loop id: 08 | 0.113 | busy threads: 10 task id: 08 | 0.113 | busy threads: 12 loop id: 09 | 0.113 | busy threads: 11 loop id: 10 | 0.113 | busy threads: 12 loop id: 11 | 0.114 | busy threads: 12 loop id: 12 | 0.114 | busy threads: 12 loop id: 13 | 0.114 | busy threads: 12 loop id: 14 | 0.114 | busy threads: 12 loop id: 15 | 0.114 | busy threads: 12 loop id: 16 | 0.114 | busy threads: 12 loop id: 17 | 0.114 | busy threads: 12 loop id: 18 | 0.114 | busy threads: 12 loop id: 19 | 0.114 | busy threads: 12 loop id: 20 | 0.114 | busy threads: 12 loop id: 21 | 0.114 | busy threads: 12 loop id: 22 | 0.114 | busy threads: 12 loop id: 23 | 0.114 | busy threads: 12 loop id: 24 | 0.114 | busy threads: 12 task id: 09 | 0.114 | busy threads: 12 task id: 06 | 0.114 | busy threads: 12 task id: 10 | 0.114 | busy threads: 12 task id: 11 | 0.114 | busy threads: 12 task id: 12 | 0.114 | busy threads: 12 task id: 13 | 1.091 | busy threads: 13 task id: 14 | 1.594 | busy threads: 14 task id: 15 | 2.099 | busy threads: 15 task id: 16 | 3.102 | busy threads: 16 task id: 17 | 3.603 | busy threads: 17 task id: 18 | 4.107 | busy threads: 18 task id: 19 | 4.611 | busy threads: 19 task id: 20 | 5.113 | busy threads: 20 task id: 21 | 5.617 | busy threads: 21 task id: 22 | 6.122 | busy threads: 22 task id: 23 | 7.128 | busy threads: 23 task id: 24 | 7.632 | busy threads: 24 task setresult | 8.135 | busy threads: 25 done: | 8.136
task.run 会把 task 调度到线程池上执行，前 24 个 task 都会被阻塞住，直到第 25 个被执行。每次都会打印出当前线程池中正在执行任务的线程数（也就是创建完成的线程数）。
可以观察到以下结果：
- 前几次循环，线程随着 task 数量递增，后面几次循环直到循环结束为止，线程数一直维持在 12 没有发生变化。
- 线程数在达到 12 之前，零间隔时间增加。第 12 到 第 13 线程间隔 1s 不到，往后约 500ms 增加一个线程。
.net 5 实验二 调整 threadpool 设置
在上面的代码最前面加入以下两行代码，继续在 .net 5 环境运行一次。
threadpool.getminthreads(out int defaultminthreads, out int completionportthreads);
console.writeline($"defaultminthreads: {defaultminthreads}");
threadpool.setminthreads(14, completionportthreads);
运行结果如下
defaultminthreads: 12 loop id: 01 | 0.000 | busy threads: 0 loop id: 02 | 0.003 | busy threads: 1 loop id: 03 | 0.003 | busy threads: 2 loop id: 04 | 0.003 | busy threads: 5 loop id: 05 | 0.004 | busy threads: 8 task id: 01 | 0.004 | busy threads: 10 task id: 03 | 0.004 | busy threads: 10 loop id: 06 | 0.004 | busy threads: 10 task id: 02 | 0.004 | busy threads: 10 task id: 04 | 0.004 | busy threads: 10 task id: 05 | 0.004 | busy threads: 12 loop id: 07 | 0.004 | busy threads: 9 loop id: 08 | 0.004 | busy threads: 10 loop id: 09 | 0.004 | busy threads: 11 loop id: 10 | 0.004 | busy threads: 12 task id: 08 | 0.004 | busy threads: 14 task id: 06 | 0.004 | busy threads: 14 task id: 09 | 0.004 | busy threads: 14 task id: 10 | 0.004 | busy threads: 14 loop id: 11 | 0.004 | busy threads: 14 loop id: 12 | 0.004 | busy threads: 14 loop id: 13 | 0.004 | busy threads: 14 loop id: 14 | 0.004 | busy threads: 14 loop id: 15 | 0.004 | busy threads: 14 loop id: 16 | 0.004 | busy threads: 14 loop id: 17 | 0.004 | busy threads: 14 loop id: 18 | 0.004 | busy threads: 14 loop id: 19 | 0.004 | busy threads: 14 loop id: 20 | 0.004 | busy threads: 14 loop id: 21 | 0.004 | busy threads: 14 loop id: 22 | 0.004 | busy threads: 14 task id: 11 | 0.004 | busy threads: 14 loop id: 23 | 0.004 | busy threads: 14 loop id: 24 | 0.005 | busy threads: 14 task id: 07 | 0.005 | busy threads: 14 task id: 12 | 0.005 | busy threads: 14 task id: 13 | 0.005 | busy threads: 14 task id: 14 | 0.005 | busy threads: 14 task id: 15 | 0.982 | busy threads: 15 task id: 16 | 1.486 | busy threads: 16 task id: 17 | 1.991 | busy threads: 17 task id: 18 | 2.997 | busy threads: 18 task id: 19 | 3.501 | busy threads: 19 task id: 20 | 4.004 | busy threads: 20 task id: 21 | 4.509 | busy threads: 21 task id: 22 | 5.014 | busy threads: 22 task id: 23 | 5.517 | busy threads: 23 task id: 24 | 6.021 | busy threads: 24 task setresult | 6.522 | busy threads: 25 done: | 6.523
在调整完线程池的最小线程数量之后，线程注入速度发生转折的时间点从第 12（默认min threads） 个线程换到了第 14（修改后的min threads）个线程。
整体时间也从 8s 缩到 6s。
.net 5 实验三 tcs.task.wait() 改为 thread.sleep
static void main(string[] args)
{
var sw = stopwatch.startnew();
var tasks = new list<task>();
for (int i = 1; i <= environment.processorcount * 2; i++)
{
int id = i;
console.writeline(
$"loop id: {id:00} | {sw.elapsed.totalseconds:0.000} | busy threads: {getbusythreads()}");
tasks.add(task.run(() =>
{
console.writeline(
$"task id: {id:00} | {sw.elapsed.totalseconds:0.000} | busy threads: {getbusythreads()}");
thread.sleep(environment.processorcount * 1000);
}));
}
task.whenall(tasks.toarray()).continuewith(_ =>
{
console.writeline($"done: | {sw.elapsed.totalseconds:0.000}");
});
console.readline();
}
loop id: 01 | 0.000 | busy threads: 0 loop id: 02 | 0.027 | busy threads: 1 loop id: 03 | 0.027 | busy threads: 2 loop id: 04 | 0.027 | busy threads: 3 loop id: 05 | 0.028 | busy threads: 4 loop id: 06 | 0.028 | busy threads: 10 loop id: 07 | 0.028 | busy threads: 9 loop id: 08 | 0.028 | busy threads: 9 loop id: 09 | 0.028 | busy threads: 10 loop id: 10 | 0.028 | busy threads: 12 loop id: 11 | 0.028 | busy threads: 12 loop id: 12 | 0.028 | busy threads: 12 loop id: 13 | 0.028 | busy threads: 12 loop id: 14 | 0.028 | busy threads: 12 loop id: 15 | 0.028 | busy threads: 12 loop id: 16 | 0.028 | busy threads: 12 loop id: 17 | 0.028 | busy threads: 12 loop id: 18 | 0.028 | busy threads: 12 loop id: 19 | 0.028 | busy threads: 12 loop id: 20 | 0.028 | busy threads: 12 loop id: 21 | 0.028 | busy threads: 12 loop id: 22 | 0.028 | busy threads: 12 loop id: 23 | 0.028 | busy threads: 12 loop id: 24 | 0.028 | busy threads: 12 task id: 01 | 0.029 | busy threads: 12 task id: 05 | 0.029 | busy threads: 12 task id: 03 | 0.029 | busy threads: 12 task id: 08 | 0.029 | busy threads: 12 task id: 09 | 0.029 | busy threads: 12 task id: 10 | 0.029 | busy threads: 12 task id: 06 | 0.029 | busy threads: 12 task id: 11 | 0.029 | busy threads: 12 task id: 12 | 0.029 | busy threads: 12 task id: 04 | 0.029 | busy threads: 12 task id: 02 | 0.029 | busy threads: 12 task id: 07 | 0.029 | busy threads: 12 task id: 13 | 1.018 | busy threads: 13 task id: 14 | 1.522 | busy threads: 14 task id: 15 | 2.025 | busy threads: 15 task id: 16 | 2.530 | busy threads: 16 task id: 17 | 3.530 | busy threads: 17 task id: 18 | 4.035 | busy threads: 18 task id: 19 | 4.537 | busy threads: 19 task id: 20 | 5.040 | busy threads: 20 task id: 21 | 5.545 | busy threads: 21 task id: 22 | 6.048 | busy threads: 22 task id: 23 | 7.049 | busy threads: 23 task id: 24 | 8.056 | busy threads: 24 done: | 20.060
达到 min threads （默认12）之后，线程注入速度明显变慢，最快间隔 500ms。
.net 6 实验一 默认 threadpool 设置
将 .net 5 实验一的代码在 .net 6 执行一次
loop id: 01 | 0.001 | busy threads: 0 loop id: 02 | 0.018 | busy threads: 1 loop id: 03 | 0.018 | busy threads: 3 loop id: 04 | 0.018 | busy threads: 6 loop id: 05 | 0.018 | busy threads: 4 loop id: 06 | 0.018 | busy threads: 5 loop id: 07 | 0.018 | busy threads: 6 loop id: 08 | 0.018 | busy threads: 8 task id: 01 | 0.018 | busy threads: 11 task id: 04 | 0.018 | busy threads: 11 task id: 03 | 0.018 | busy threads: 11 task id: 02 | 0.018 | busy threads: 11 task id: 05 | 0.018 | busy threads: 11 loop id: 09 | 0.018 | busy threads: 12 loop id: 10 | 0.018 | busy threads: 12 loop id: 11 | 0.018 | busy threads: 12 loop id: 12 | 0.018 | busy threads: 12 loop id: 13 | 0.018 | busy threads: 12 task id: 09 | 0.018 | busy threads: 12 loop id: 14 | 0.018 | busy threads: 12 loop id: 15 | 0.018 | busy threads: 12 loop id: 16 | 0.018 | busy threads: 12 loop id: 17 | 0.018 | busy threads: 12 task id: 06 | 0.018 | busy threads: 12 loop id: 18 | 0.018 | busy threads: 12 loop id: 19 | 0.018 | busy threads: 12 loop id: 20 | 0.018 | busy threads: 12 loop id: 21 | 0.018 | busy threads: 12 loop id: 22 | 0.018 | busy threads: 12 loop id: 23 | 0.018 | busy threads: 12 loop id: 24 | 0.018 | busy threads: 12 task id: 10 | 0.018 | busy threads: 12 task id: 07 | 0.019 | busy threads: 12 task id: 11 | 0.019 | busy threads: 12 task id: 08 | 0.019 | busy threads: 12 task id: 12 | 0.019 | busy threads: 12 task id: 13 | 0.020 | busy threads: 16 task id: 14 | 0.020 | busy threads: 17 task id: 15 | 0.020 | busy threads: 18 task id: 16 | 0.020 | busy threads: 19 task id: 17 | 0.020 | busy threads: 20 task id: 18 | 0.020 | busy threads: 21 task id: 19 | 0.020 | busy threads: 22 task id: 20 | 0.020 | busy threads: 23 task id: 21 | 0.020 | busy threads: 24 task id: 23 | 0.020 | busy threads: 24 task id: 22 | 0.020 | busy threads: 24 task id: 24 | 0.020 | busy threads: 24 task setresult | 0.045 | busy threads: 25 done: | 0.046
与实验一相比，虽然线程数仍然停留在 12 了一段时间，但随后线程就立即增长了，后文会介绍 .net 6 在这方面做出的改进。
.net 6 实验二 调整 threadpool 设置
将 .net 5 实验二的代码在 .net 6 中执行一次
defaultminthreads: 12 loop id: 01 | 0.001 | busy threads: 0 loop id: 02 | 0.014 | busy threads: 1 loop id: 03 | 0.014 | busy threads: 2 loop id: 04 | 0.015 | busy threads: 5 loop id: 05 | 0.015 | busy threads: 4 loop id: 06 | 0.015 | busy threads: 5 loop id: 07 | 0.015 | busy threads: 7 loop id: 08 | 0.015 | busy threads: 8 loop id: 09 | 0.015 | busy threads: 11 task id: 06 | 0.015 | busy threads: 9 task id: 01 | 0.015 | busy threads: 9 task id: 02 | 0.015 | busy threads: 9 task id: 05 | 0.015 | busy threads: 9 task id: 03 | 0.015 | busy threads: 9 task id: 04 | 0.015 | busy threads: 9 task id: 07 | 0.015 | busy threads: 9 task id: 08 | 0.016 | busy threads: 9 task id: 09 | 0.016 | busy threads: 9 loop id: 10 | 0.016 | busy threads: 9 loop id: 11 | 0.016 | busy threads: 10 loop id: 12 | 0.016 | busy threads: 11 loop id: 13 | 0.016 | busy threads: 13 task id: 10 | 0.016 | busy threads: 14 loop id: 14 | 0.016 | busy threads: 14 loop id: 15 | 0.016 | busy threads: 14 loop id: 16 | 0.016 | busy threads: 14 task id: 11 | 0.016 | busy threads: 14 loop id: 17 | 0.016 | busy threads: 14 loop id: 18 | 0.016 | busy threads: 14 loop id: 19 | 0.016 | busy threads: 14 loop id: 20 | 0.016 | busy threads: 14 loop id: 21 | 0.016 | busy threads: 14 loop id: 22 | 0.016 | busy threads: 14 loop id: 23 | 0.016 | busy threads: 14 loop id: 24 | 0.016 | busy threads: 14 task id: 12 | 0.016 | busy threads: 14 task id: 13 | 0.016 | busy threads: 14 task id: 14 | 0.016 | busy threads: 14 task id: 15 | 0.017 | busy threads: 18 task id: 16 | 0.017 | busy threads: 19 task id: 17 | 0.017 | busy threads: 20 task id: 18 | 0.017 | busy threads: 21 task id: 19 | 0.017 | busy threads: 22 task id: 20 | 0.018 | busy threads: 23 task id: 21 | 0.018 | busy threads: 24 task id: 22 | 0.018 | busy threads: 25 task id: 23 | 0.018 | busy threads: 26 task id: 24 | 0.018 | busy threads: 26 task setresult | 0.018 | busy threads: 25 done: | 0.019
前半部分有部分日志乱序，可以看到，与实验三一样，维持在最大线程数一小段时间之后，立即就开始了线程增长。
.net 6 实验三 tcs.task.wait() 改为 thread.sleep
将 .net 5 实验三的代码在 .net 6 中执行一次
loop id: 01 | 0.003 | busy threads: 0 loop id: 02 | 0.024 | busy threads: 1 loop id: 03 | 0.025 | busy threads: 2 loop id: 04 | 0.025 | busy threads: 3 loop id: 05 | 0.025 | busy threads: 7 loop id: 06 | 0.025 | busy threads: 5 loop id: 07 | 0.025 | busy threads: 6 loop id: 08 | 0.025 | busy threads: 7 loop id: 09 | 0.025 | busy threads: 9 loop id: 10 | 0.025 | busy threads: 10 loop id: 11 | 0.026 | busy threads: 10 loop id: 12 | 0.026 | busy threads: 11 loop id: 13 | 0.026 | busy threads: 12 loop id: 14 | 0.026 | busy threads: 12 loop id: 15 | 0.026 | busy threads: 12 loop id: 16 | 0.026 | busy threads: 12 loop id: 17 | 0.026 | busy threads: 12 loop id: 18 | 0.026 | busy threads: 12 loop id: 19 | 0.026 | busy threads: 12 loop id: 20 | 0.026 | busy threads: 12 loop id: 21 | 0.026 | busy threads: 12 loop id: 22 | 0.026 | busy threads: 12 loop id: 23 | 0.026 | busy threads: 12 loop id: 24 | 0.026 | busy threads: 12 task id: 01 | 0.026 | busy threads: 12 task id: 02 | 0.026 | busy threads: 12 task id: 05 | 0.026 | busy threads: 12 task id: 04 | 0.026 | busy threads: 12 task id: 06 | 0.026 | busy threads: 12 task id: 08 | 0.026 | busy threads: 12 task id: 09 | 0.026 | busy threads: 12 task id: 03 | 0.026 | busy threads: 12 task id: 11 | 0.026 | busy threads: 12 task id: 10 | 0.026 | busy threads: 12 task id: 07 | 0.026 | busy threads: 12 task id: 12 | 0.026 | busy threads: 12 task id: 13 | 1.026 | busy threads: 13 task id: 14 | 2.027 | busy threads: 14 task id: 15 | 3.028 | busy threads: 15 task id: 16 | 4.030 | busy threads: 16 task id: 17 | 5.031 | busy threads: 17 task id: 18 | 6.032 | busy threads: 18 task id: 19 | 6.533 | busy threads: 19 task id: 20 | 7.035 | busy threads: 20 task id: 21 | 8.036 | busy threads: 21 task id: 22 | 8.537 | busy threads: 22 task id: 23 | 9.538 | busy threads: 23 task id: 24 | 10.039 | busy threads: 24 done: | 22.041
结果与 .net 5 的实验三相差不大。
线程注入
对照上述的几组实验结果，接下来以 .net 6 中 c# 实现的 threadpool 作为资料来理解一下线程注入的几个阶段（按个人理解进行的划分，仅供参考）。
1. 第一个线程的出现
随着任务被调度到队列上，第一个线程被创建出来。
下面是线程池在执行第一个任务的时候的代码摘要，涉及到计数的并执行相关处理的地方，代码都使用了 while(xxx) + interlocked 的方式来进行并发控制，可以理解成乐观锁。这一阶段，实际上我们只需要关注到 threadpoolworkqueue.ensurethreadrequested 方法就行了。
可利用 rider 的反编译 debug 功能帮助我们学习。
下面是第一个 task.run 的代码执行路径
注意：执行环节是 main thread
public static class threadpool
{
internal static readonly threadpoolworkqueue s_workqueue = new threadpoolworkqueue();
public static bool queueuserworkitem(waitcallback callback, object state)
{
object tpcallback = new queueuserworkitemcallback(callback!, state);
s_workqueue.enqueue(tpcallback, forceglobal: true);
return true;
}
}
internal sealed class threadpoolworkqueue
{
[structlayout(layoutkind.sequential)]
private struct cachelineseparated
{
private readonly internal.paddingfor32 pad1;
public volatile int numoutstandingthreadrequests;
private readonly internal.paddingfor32 pad2;
}
private cachelineseparated _separated;
public void enqueue(object callback, bool forceglobal)
{
// 线程池中执行的任务有两种：ithreadpoolworkitem、task
debug.assert((callback is ithreadpoolworkitem) ^ (callback is task));
if (loggingenabled && frameworkeventsource.log.isenabled())
frameworkeventsource.log.threadpoolenqueueworkobject(callback);
threadpoolworkqueuethreadlocals? tl = null;
if (!forceglobal)
// 获取本地队列，如果执行改代码的线程不是线程池线程，
// 那这边是获取不到的，就算 forceglobal 是 false，
// 也会把任务放到全局队列
tl = threadpoolworkqueuethreadlocals.threadlocals;
if (null != tl)
{
// 放到本地队列
tl.workstealingqueue.localpush(callback);
}
else
{
// 当道全局队列
workitems.enqueue(callback);
}
ensurethreadrequested();
}
internal void ensurethreadrequested()
{
//
// if we have not yet requested #procs threads, then request a new thread.
//
// coreclr: note that there is a separate count in the vm which has already been incremented
// by the vm by the time we reach this point.
//
int count = _separated.numoutstandingthreadrequests;
while (count < environment.processorcount)
{
int prev = interlocked.compareexchange(ref _separated.numoutstandingthreadrequests, count + 1, count);
if (prev == count)
{
threadpool.requestworkerthread();
break;
}
count = prev;
}
}
public static class threadpool
{
/// <summary>
/// this method is called to request a new thread pool worker to handle pending work.
/// </summary>
internal static void requestworkerthread() => portablethreadpool.threadpoolinstance.requestworker();
}
internal sealed class portablethreadpool
{
public static readonly portablethreadpool threadpoolinstance = new portablethreadpool();
internal void requestworker()
{
// the order of operations here is important. maybeaddworkingworker() and ensurerunning() use speculative checks to
// do their work and the memory barrier from the interlocked operation is necessary in this case for correctness.
interlocked.increment(ref _separated.numrequestedworkers);
workerthread.maybeaddworkingworker(this);
// 初始化 gatethread
gatethread.ensurerunning(this);
}
/// <summary>
/// the worker thread infastructure for the clr thread pool.
/// </summary>
private static class workerthread
{
internal static void maybeaddworkingworker(portablethreadpool threadpoolinstance)
{
threadcounts counts = threadpoolinstance._separated.counts;
short numexistingthreads, numprocessingwork, newnumexistingthreads, newnumprocessingwork;
// 这个 while (true) 是确保计算出正确的待创建线程数
while (true)
{
numprocessingwork = counts.numprocessingwork;
if (numprocessingwork >= counts.numthreadsgoal)
{
return;
}
newnumprocessingwork = (short)(numprocessingwork + 1);
numexistingthreads = counts.numexistingthreads;
newnumexistingthreads = math.max(numexistingthreads, newnumprocessingwork);
threadcounts newcounts = counts;
newcounts.numprocessingwork = newnumprocessingwork;
newcounts.numexistingthreads = newnumexistingthreads;
threadcounts oldcounts = threadpoolinstance._separated.counts.interlockedcompareexchange(newcounts, counts);
if (oldcounts == counts)
{
break;
}
counts = oldcounts;
}
int tocreate = newnumexistingthreads - numexistingthreads;
int torelease = newnumprocessingwork - numprocessingwork;
if (torelease > 0)
{
s_semaphore.release(torelease);
}
while (tocreate > 0)
{
if (trycreateworkerthread())
{
tocreate--;
continue;
}
counts = threadpoolinstance._separated.counts;
while (true)
{
threadcounts newcounts = counts;
newcounts.subtractnumprocessingwork((short)tocreate);
newcounts.subtractnumexistingthreads((short)tocreate);
threadcounts oldcounts = threadpoolinstance._separated.counts.interlockedcompareexchange(newcounts, counts);
if (oldcounts == counts)
{
break;
}
counts = oldcounts;
}
break;
}
}
private static bool trycreateworkerthread()
{
try
{
// thread pool threads must start in the default execution context without transferring the context, so
// using unsafestart() instead of start()
thread workerthread = new thread(s_workerthreadstart);
workerthread.isthreadpoolthread = true;
workerthread.isbackground = true;
// thread name will be set in thread proc
workerthread.unsafestart();
}
catch (threadstartexception)
{
return false;
}
catch (outofmemoryexception)
{
return false;
}
return true;
}
}
}
}
2. 达到 min threads 之前的线程数增长
细心的朋友会发现上面代码里 ensurethreadrequested 方法有一个终止条件，_separated.numoutstandingthreadrequests == environment.processorcount，每次新增一个 threadrequested，这个数就会 +1，似乎允许创建的最大 worker thread 是 environment.processorcount？
其实 threadpoolworkqueue 维护的 numoutstandingthreadrequests 这个值会在线程池线程真正跑起来之后，会在 threadpoolworkqueue.dispatch方法中 -1。也就是说，只要有一个线程真正运行起来了，就能创建第 environment.processorcount + 1 个thread。当然，在向 threadpoolworkqueue 加入第13个任务的时候，第13个 worker thread 就算不允许创建也没关系，因为任务已经入队了，会被运行起来的 worker thread 取走。
min threads 初始值为 运行环境 cpu 核心数，可通过 threadpool.setminthreads 进行设置，参数有效范围是 [1, max threads]。
portablethreadpool里维护了一个计数器 portablethreadpool.threadpoolinstance._separated.counts，记录了 worker thread 相关的三个数值：
- numprocessingwork：当前正在执行任务的 worker thread。
- numexistingthreads：当前线程池中实际有的 worker thread。
- numthreadsgoal：当前允许创建的最大 worker thread，初始值为 min threads。
internal class portablethreadpool
{
public static readonly portablethreadpool threadpoolinstance = new portablethreadpool();
private cachelineseparated _separated;
private struct cachelineseparated
{
public threadcounts counts;
}
/// <summary>
/// tracks information on the number of threads we want/have in different states in our thread pool.
/// </summary>
private struct threadcounts
{
/// <summary>
/// number of threads processing work items.
/// </summary>
public short numprocessingwork { get; set; }
/// <summary>
/// number of thread pool threads that currently exist.
/// </summary>
public short numexistingthreads { get; set; }
// <summary>
/// max possible thread pool threads we want to have.
/// </summary>
public short numthreadsgoal { get; set; }
}
}
3. 避免饥饿机制（starvation avoidance）
上面讲到，随着任务进入队列系统，worker thread 将随之增长，直到达到 numthreadsgoal。
numthreadsgoal 是12，前 12 个线程都被堵住了，加入到队列系统的第 13 个任务没办法被这前 12 个线程领走执行。
在这种情况下，线程池的 starvation avoidance 机制就起到作用了。
在上述所说的第一个阶段，除了线程池中的第一个线程会被创建之外，gatethread 也会随之被初始化。在第一阶段的代码摘录中，可以看到 gatethread 的初始化。
internal sealed class portablethreadpool
{
public static readonly portablethreadpool threadpoolinstance = new portablethreadpool();
internal void requestworker()
{
interlocked.increment(ref _separated.numrequestedworkers);
workerthread.maybeaddworkingworker(this);
// 初始化 gatethread
gatethread.ensurerunning(this);
}
}
在 gatethread 是一个独立的线程，每隔 500ms 进行检查一下，如果 numprocessingwork >= numthreadsgoal（workerthread.maybeaddworkingworker 不添加 worker thread 的判断条件），就设置新的 numthreadsgoal = numprocessingwork + 1，并调用 workerthread.maybeaddworkingworker，这样新的 worker thread 就可以被 workerthread.maybeaddworkingworker 创建。
这就解释了，为什么 .net 5 实验一、二在线程数达到min threads（numthreadsgoal 的默认值）之后，后面 worker thread 的增长是每 500ms 一个。
由于在第三阶段中，线程的增长会比较缓慢，有经验的开发会在应用启动的时候设置一个较大的 min threads，使其较晚或不进入第三阶段。
线程注入在 .net 6 中的改进
.net 6 与 .net 5 的实验二相比，达到 min threads 之后，线程的增长速度有明显的差异，而两者的实验三却相差不大。
.net 6 对于 task.wait 导致线程池线程阻塞的场景进行了优化，但如果并非此原因导致的线程数不够用，依旧是 starvation avoidance 的策略。
新的 threadpool 提供了一个 threadpool.notifythreadblocked 的内部接口，里面会调用 gatethread.wake 去唤醒 gatethread 本来 500ms 执行一次的逻辑，这 500ms 的间隔时间是通过 autoresetevent 实现的，所以 gatethread.wake 也很简单。
关键代码示意，非真实代码：
internal class portablethreadpool
{
public bool notifythreadblocked()
{
// ...
gatethread.wake(this);
return true;
}
private static class gatethread
{
private static readonly autoresetevent delayevent = new autoresetevent(initialstate: false);
// gatethread 入口方法
private static void gatethreadstart()
{
while(true)
{
delayevent.waitone(500);
// ...
}
}
public static void wake(portablethreadpool threadpoolinstance)
{
delayevent.set();
ensurerunning(threadpoolinstance);
}
}
爬山算法（hill climbing）
除了上述介绍的线程注入机制外，从clr 4.0开始，线程池内实现了一个根据采集到线程池吞吐率数据（每次任务完成时记录数据），推导出该算法认为最优的线程池线程数量。
算法实现位于 hillclimbing.threadpoolhillclimber.update，有兴趣的朋友可以去看一下。
public (int newthreadcount, int newsamplems) update(int currentthreadcount, double sampledurationseconds, int numcompletions)
- currentthreadcount：当前线程数
- sampledurationseconds：采样间隔
- numcompletions：这段采样时间间隔内完成的任务数
- newthreadcount：新的线程数
- newsample：新的采样间隔时间
不必要线程的销毁
如果线程需要被移除的时候，本地队列还存在待执行任务，则会将这些任务转移到全局队列中。
在以下几个场景中，线程池将会销毁掉不需要的线程，并不一定全面，只限于笔者当前认知。
- 在无法从队列系统领取到任务时。
- 通过爬山算法认定当前线程属于多余线程时。
参考资料
https://www.codeproject.com/articles/3813/net-s-threadpool-class-behind-the-scenes
https://devblogs.microsoft.com/dotnet/performance-improvements-in-net-6/
https://mattwarren.org/2017/04/13/the-clr-thread-pool-thread-injection-algorithm/
https://docs.microsoft.com/zh-cn/previous-versions/msp-n-p/ff963549(v=pandp.10)?redirectedfrom=msdn
到此这篇关于.net 6线程池threadpool实现方法的文章就介绍到这了。希望对大家的学习有所帮助，也希望大家多多支持代码网。
发表评论