为什么每个人都说SpinLock更快？

Question

我在网上读了很多文档、文章和帖子。几乎所有的人和任何地方的人都认为SpinLock对于短时间运行的代码来说更快，但是我做了一个测试，在我看来简单的Monitor.Enter比SpinLock.Enter工作得更快(测试是针对.NET 4.5编译的)

using System;
using System.Collections.Concurrent;
using System.Collections.Generic;
using System.Diagnostics;
using System.Threading.Tasks;
using System.Linq;
using System.Globalization;
using System.ComponentModel;
using System.Threading;
using System.Net.Sockets;
using System.Net;

class Program
{
    static int _loopsCount = 1000000;
    static int _threadsCount = -1;

    static ProcessPriorityClass _processPriority = ProcessPriorityClass.RealTime;
    static ThreadPriority _threadPriority = ThreadPriority.Highest;

    static long _testingVar = 0;


    static void Main(string[] args)
    {
        _threadsCount = Environment.ProcessorCount;

        Console.WriteLine("Cores/processors count: {0}", Environment.ProcessorCount);

        Process.GetCurrentProcess().PriorityClass = _processPriority;

        TimeSpan tsInterlocked = ExecuteInterlocked();
        TimeSpan tsSpinLock = ExecuteSpinLock();
        TimeSpan tsMonitor = ExecuteMonitor();

        Console.WriteLine("Test with interlocked: {0} ms\r\nTest with SpinLock: {1} ms\r\nTest with Monitor: {2} ms",
            tsInterlocked.TotalMilliseconds,
            tsSpinLock.TotalMilliseconds,
            tsMonitor.TotalMilliseconds);

        Console.ReadLine();
    }

    static TimeSpan ExecuteInterlocked()
    {
        _testingVar = 0;

        ManualResetEvent _startEvent = new ManualResetEvent(false);
        CountdownEvent _endCountdown = new CountdownEvent(_threadsCount);

        Thread[] threads = new Thread[_threadsCount];

        for (int i = 0; i < threads.Length; i++)
        {
            threads[i] = new Thread(() =>
                {
                    _startEvent.WaitOne();

                    for (int j = 0; j < _loopsCount; j++)
                    {
                        Interlocked.Increment(ref _testingVar);
                    }

                    _endCountdown.Signal();
                });

            threads[i].Priority = _threadPriority;
            threads[i].Start();
        }

        Stopwatch sw = Stopwatch.StartNew();

        _startEvent.Set();
        _endCountdown.Wait();

        return sw.Elapsed;
    }

    static SpinLock _spinLock = new SpinLock();

    static TimeSpan ExecuteSpinLock()
    {
        _testingVar = 0;

        ManualResetEvent _startEvent = new ManualResetEvent(false);
        CountdownEvent _endCountdown = new CountdownEvent(_threadsCount);

        Thread[] threads = new Thread[_threadsCount];

        for (int i = 0; i < threads.Length; i++)
        {
            threads[i] = new Thread(() =>
            {
                _startEvent.WaitOne();

                bool lockTaken;

                for (int j = 0; j < _loopsCount; j++)
                {
                    lockTaken = false;

                    try
                    {
                        _spinLock.Enter(ref lockTaken);

                        _testingVar++;
                    }
                    finally
                    {
                        if (lockTaken)
                        {
                            _spinLock.Exit();
                        }
                    }
                }

                _endCountdown.Signal();
            });

            threads[i].Priority = _threadPriority;
            threads[i].Start();
        }

        Stopwatch sw = Stopwatch.StartNew();

        _startEvent.Set();
        _endCountdown.Wait();

        return sw.Elapsed;
    }

    static object _locker = new object();

    static TimeSpan ExecuteMonitor()
    {
        _testingVar = 0;

        ManualResetEvent _startEvent = new ManualResetEvent(false);
        CountdownEvent _endCountdown = new CountdownEvent(_threadsCount);

        Thread[] threads = new Thread[_threadsCount];

        for (int i = 0; i < threads.Length; i++)
        {
            threads[i] = new Thread(() =>
            {
                _startEvent.WaitOne();

                bool lockTaken;

                for (int j = 0; j < _loopsCount; j++)
                {
                    lockTaken = false;

                    try
                    {
                        Monitor.Enter(_locker, ref lockTaken);

                        _testingVar++;
                    }
                    finally
                    {
                        if (lockTaken)
                        {
                            Monitor.Exit(_locker);
                        }
                    }
                }

                _endCountdown.Signal();
            });

            threads[i].Priority = _threadPriority;
            threads[i].Start();
        }

        Stopwatch sw = Stopwatch.StartNew();

        _startEvent.Set();
        _endCountdown.Wait();

        return sw.Elapsed;
    }
}

在24核为2.5 GHz的服务器上，使用x64编译的应用程序产生了以下结果：

Cores/processors count: 24
Test with interlocked: 1373.0829 ms
Test with SpinLock: 10894.6283 ms
Test with Monitor: 1171.1591 ms

Answer 1

您只是没有测试SpinLock可以改进线程处理的场景。自旋锁背后的核心思想是线程上下文切换是非常昂贵的操作，花费在2000到10,000个cpu周期之间。而且，如果线程很可能通过等待一点点(旋转)获得锁，那么额外的等待周期就可以通过避免线程上下文切换来获得回报。

因此，基本的要求是锁的时间非常短，这在您的情况下是正确的。而且有合理的机会可以获得锁。在您的情况下不是这样的，锁是由不少于24个线程激烈竞争的。所有旋转和燃烧的核心没有机会获得锁。

在本测试中，Monitor将工作得最好，因为它会排队等待线程获取锁。它们被挂起，直到其中一个有机会获得锁，在释放锁时从等待队列中释放。给他们一个公平的机会轮流，从而最大化的机会，他们都将在同一时间完成。Interlocked.Increment也不错，但不能提供公平保证。

要判断Spinlock是否是最前面的正确方法是相当困难的，你必须测量。并发分析器是正确的工具。