记录测试中发现的sleep_for时间的微小差异，产生的性能的巨大差异

// PerfTest.cpp : This file contains the 'main' function. Program execution begins and ends there.
//

#include <windows.h>
#include <tchar.h>
#include <stdio.h>
#include <thread>
#include <chrono>

//
// Thread pool timer callback function template
//
VOID
CALLBACK
MyTimerCallback(
    PTP_CALLBACK_INSTANCE Instance,
    PVOID                 Parameter,
    PTP_TIMER             Timer
)
{
    // Instance, Parameter, and Timer not used in this example.
    UNREFERENCED_PARAMETER(Instance);
    UNREFERENCED_PARAMETER(Parameter);
    UNREFERENCED_PARAMETER(Timer);

    //
    // Do something when the timer fires.
    //
    _tprintf(_T("MyTimerCallback: timer has fired.\n"));

}


//
// This is the thread pool work callback function.
//
VOID
CALLBACK
MyWorkCallback(
    PTP_CALLBACK_INSTANCE Instance,
    PVOID                 Parameter,
    PTP_WORK              Work
)
{
    // Instance, Parameter, and Work not used in this example.
    UNREFERENCED_PARAMETER(Instance);
    UNREFERENCED_PARAMETER(Parameter);
    UNREFERENCED_PARAMETER(Work);

    BOOL bRet = FALSE;
    thread_local int idx = 0;
    idx++;
    int sum = 0;
    for (int i = 0; i < 1000000; i++)
    {
        sum += i;
    }

    //
    // Do something when the work callback is invoked.
    //
    {
        _tprintf(_T("MyWorkCallback: Task performed. sum: %d, idx: %d\n"), sum, idx);
    }

    return;
}

VOID
DemoCleanupPersistentWorkTimer()
{
    BOOL bRet = FALSE;
    PTP_WORK work = NULL;
    PTP_TIMER timer = NULL;
    PTP_POOL pool = NULL;
    PTP_WORK_CALLBACK workcallback = MyWorkCallback;
    PTP_TIMER_CALLBACK timercallback = MyTimerCallback;
    TP_CALLBACK_ENVIRON CallBackEnviron;
    PTP_CLEANUP_GROUP cleanupgroup = NULL;
    FILETIME FileDueTime;
    ULARGE_INTEGER ulDueTime;
    UINT rollback = 0;

    InitializeThreadpoolEnvironment(&CallBackEnviron);

    //
    // Create a custom, dedicated thread pool.
    //
    pool = CreateThreadpool(NULL);

    if (NULL == pool) {
        _tprintf(_T("CreateThreadpool failed. LastError: %u\n"),
            GetLastError());
        goto main_cleanup;
    }

    rollback = 1; // pool creation succeeded

    //
    // The thread pool is made persistent simply by setting
    // both the minimum and maximum threads to 1.
    //
    SetThreadpoolThreadMaximum(pool, 5);

    bRet = SetThreadpoolThreadMinimum(pool, 1);

    if (FALSE == bRet) {
        _tprintf(_T("SetThreadpoolThreadMinimum failed. LastError: %u\n"),
            GetLastError());
        goto main_cleanup;
    }

    //
    // Create a cleanup group for this thread pool.
    //
    cleanupgroup = CreateThreadpoolCleanupGroup();

    if (NULL == cleanupgroup) {
        _tprintf(_T("CreateThreadpoolCleanupGroup failed. LastError: %u\n"),
            GetLastError());
        goto main_cleanup;
    }

    rollback = 2;  // Cleanup group creation succeeded

    //
    // Associate the callback environment with our thread pool.
    //
    SetThreadpoolCallbackPool(&CallBackEnviron, pool);

    //
    // Associate the cleanup group with our thread pool.
    // Objects created with the same callback environment
    // as the cleanup group become members of the cleanup group.
    //
    SetThreadpoolCallbackCleanupGroup(&CallBackEnviron,
        cleanupgroup,
        NULL);

    //
    // Create work with the callback environment.
    //
    work = CreateThreadpoolWork(workcallback,
        NULL,
        &CallBackEnviron);

    if (NULL == work) {
        _tprintf(_T("CreateThreadpoolWork failed. LastError: %u\n"),
            GetLastError());
        goto main_cleanup;
    }

    rollback = 3;  // Creation of work succeeded

    for (int i = 0; i < 1000000000; i++)
    {
        //
        // Submit the work to the pool. Because this was a pre-allocated
        // work item (using CreateThreadpoolWork), it is guaranteed to execute.
        //
        SubmitThreadpoolWork(work);
        std::this_thread::sleep_for(std::chrono::nanoseconds(100));
    }


    //
    // Create a timer with the same callback environment.
    //
    timer = CreateThreadpoolTimer(timercallback,
        NULL,
        &CallBackEnviron);


    if (NULL == timer) {
        _tprintf(_T("CreateThreadpoolTimer failed. LastError: %u\n"),
            GetLastError());
        goto main_cleanup;
    }

    rollback = 4;  // Timer creation succeeded

    //
    // Set the timer to fire in one second.
    //
    ulDueTime.QuadPart = (ULONGLONG)-(1 * 10 * 1000 * 1000);
    FileDueTime.dwHighDateTime = ulDueTime.HighPart;
    FileDueTime.dwLowDateTime = ulDueTime.LowPart;

    SetThreadpoolTimer(timer,
        &FileDueTime,
        0,
        0);

    //
    // Delay for the timer to be fired
    //
    Sleep(1500);

    //
    // Wait for all callbacks to finish.
    // CloseThreadpoolCleanupGroupMembers also releases objects
    // that are members of the cleanup group, so it is not necessary 
    // to call close functions on individual objects 
    // after calling CloseThreadpoolCleanupGroupMembers.
    //
    CloseThreadpoolCleanupGroupMembers(cleanupgroup,
        FALSE,
        NULL);

    //
    // Already cleaned up the work item with the
    // CloseThreadpoolCleanupGroupMembers, so set rollback to 2.
    //
    rollback = 2;
    goto main_cleanup;

main_cleanup:
    //
    // Clean up any individual pieces manually
    // Notice the fall-through structure of the switch.
    // Clean up in reverse order.
    //

    switch (rollback) {
    case 4:
    case 3:
        // Clean up the cleanup group members.
        CloseThreadpoolCleanupGroupMembers(cleanupgroup,
            FALSE, NULL);
    case 2:
        // Clean up the cleanup group.
        CloseThreadpoolCleanupGroup(cleanupgroup);

    case 1:
        // Clean up the pool.
        CloseThreadpool(pool);

    default:
        break;
    }

    return;
}


int main(void)
{
    DemoCleanupPersistentWorkTimer();
    return 0;
}

　　上面是测试用的代码，我测试了如下两种情况：

（1）std::this_thread::sleep_for(std::chrono::nanoseconds(100));
（2）std::this_thread::sleep_for(std::chrono::nanoseconds(101));
测试使用的IDE是VS2022，如下图所示

 测试的是release编译，使用Start Without Debugging启动。根据测试结果，使用100纳秒的sleep时间，CPU占用大约为14%到22%，而使用101纳秒的sleep时间，CPU占用为0%到4%，其中大多数时间显示为0%（表示低于1%）。暂时没有调查具体原因，先记录下来。如果您对这个原因有想法，欢迎指教。