TLS 线程局部变量, 网上一堆文章说比用锁要好, 为什么我测试了后和 CRITICAL_SECTION 差不多呢

[Sandrer]

之前开发的程序一直在用 TLS, 各线程玩各自的东西
但这段时间要优化一个服务端程序, 就想着能不能再优化一下 TLS 的东西
然后就写了下面的代码, 但运行结果看来, TLS 和 CRITICAL_SECTION 差不多的时间
反而很多时候要比 CS 慢.........................
难道是我打开的方式不对???

#include "stdafx.h"
#include <process.h>

static DWORD m_dwTLSIndex = 0;

static CRITICAL_SECTION m_lock;

#define MAX_ALLOCATE    1000000u
static volatile DWORD m_dwAllocCount = 0;

struct BUFFER
{
    BUFFER *next;
    BYTE data[16];
};

struct BUFFER_LIST
{
    BUFFER_LIST *next;
    BUFFER *pBuffers;
};

static BUFFER_LIST *m_pBufferList = NULL;
static BUFFER *m_pBuffers = NULL;

uintptr_t __stdcall _thread1(void *)
{
    BUFFER *p;
    BUFFER_LIST *pList = (BUFFER_LIST *)malloc(sizeof(BUFFER_LIST));
    ZeroMemory(pList, sizeof(BUFFER_LIST));

    ::EnterCriticalSection(&m_lock);
    pList->next = m_pBufferList;
    m_pBufferList = pList;
    ::LeaveCriticalSection(&m_lock);

    ::TlsSetValue(m_dwTLSIndex, pList);

    for (;;)
    {
        p = (BUFFER *)malloc(sizeof(BUFFER));
        p->data[0] = 0;

        pList = (BUFFER_LIST *)::TlsGetValue(m_dwTLSIndex);
        p->next = pList->pBuffers;
        pList->pBuffers = p;

        if (InterlockedIncrement(&m_dwAllocCount) >= MAX_ALLOCATE)
            break;
    }

    return 0;
}

uintptr_t __stdcall _thread2(void *)
{
    BUFFER *p;

    for (;;)
    {
        p = (BUFFER *)malloc(sizeof(BUFFER));
        p->data[0] = 0;

        ::EnterCriticalSection(&m_lock);
        p->next = m_pBuffers;
        m_pBuffers = p;
        ::LeaveCriticalSection(&m_lock);

        if (InterlockedIncrement(&m_dwAllocCount) >= MAX_ALLOCATE)
            break;
    }

    return 0;
}

void TestTLSData()
{
    ::InitializeCriticalSection(&m_lock);

    ULONGLONG ullBegin, ullEnd;
    SYSTEM_INFO si;
    ::GetSystemInfo(&si);

    DWORD dwThread = 0;
    DWORD dwCount = si.dwNumberOfProcessors;
    HANDLE *phThreads = new HANDLE[dwCount + 1];
    ZeroMemory(phThreads, sizeof(HANDLE) * (dwCount + 1));

    m_dwTLSIndex = ::TlsAlloc();

    ::GetSystemTimeAsFileTime((LPFILETIME)&ullBegin);

    dwThread = 0;
    while (dwCount--)
        phThreads[dwThread++] = (HANDLE)_beginthreadex(NULL, 0, _thread1, NULL, 0, NULL);

    ::WaitForMultipleObjects(dwThread, phThreads, TRUE, INFINITE);

    ::GetSystemTimeAsFileTime((LPFILETIME)&ullEnd);
    printf("Used TLS: %u.%04u ms\n", (DWORD)(ullEnd - ullBegin) / 10000, (DWORD)(ullEnd - ullBegin) % 10000);
    while (dwThread--)
        ::CloseHandle(phThreads[dwThread]);

    ::TlsFree(m_dwTLSIndex);

    BUFFER *n, *p;
    //BUFFER_LIST *pNext;
    //BUFFER_LIST *pList = m_pBufferList;
    //while (pList != NULL)
    //{
    //    pNext = pList->next;

    //    p = pList->pBuffers;
    //    while (p)
    //    {
    //        n = p->next;
    //        free(p);
    //        p = n;
    //    }

    //    free(pList);
    //    pList = pNext;
    //}

    m_dwAllocCount = 0;
    dwCount = si.dwNumberOfProcessors;
    ZeroMemory(phThreads, sizeof(HANDLE) * (dwCount + 1));

    ::GetSystemTimeAsFileTime((LPFILETIME)&ullBegin);

    dwThread = 0;
    while (dwCount--)
        phThreads[dwThread++] = (HANDLE)_beginthreadex(NULL, 0, _thread2, NULL, 0, NULL);

    ::WaitForMultipleObjects(dwThread, phThreads, TRUE, INFINITE);

    ::GetSystemTimeAsFileTime((LPFILETIME)&ullEnd);
    printf("Not used TLS: %u.%04u ms\n", (DWORD)(ullEnd - ullBegin) / 10000, (DWORD)(ullEnd - ullBegin) % 10000);
    while (dwThread--)
        ::CloseHandle(phThreads[dwThread]);

    //p = m_pBuffers;
    //while (p)
    //{
    //    n = p->next;
    //    free(p);
    //    p = n;
    //}

    ::DeleteCriticalSection(&m_lock);

    system("pause");
}

[赵4老师]

无profiler不要谈效率！！尤其在这个云计算、虚拟机、模拟器、CUDA、多核 、多级cache、指令流水线、多种存储介质、……满天飞的时代！

[风影-取名字真烦]

tls本来就应该比cs方式要慢，因为它本身是为了安全而准备，而不是为效率，如果安全性要求高你可以用，如果安全性要求不高，那还是cs的形式最为方便快捷，你可以详细了解下tls的协议就知道了，本协议为了安全性付出了很多时间上的代价，比如加密，比如额外的协议层数据