如何在 Linux 中创建高分辨率计时器来测量程序性能？

问题描述：

我正在尝试比较 GPU 与 CPU 的性能。对于 NVIDIA GPU，我一直使用这些cudaEvent_t类型来获得非常精确的时间。

对于 CPU，我一直使用以下代码：

// Timers
clock_t start, stop;
float elapsedTime = 0;

// Capture the start time

start = clock();

// Do something here
.......

// Capture the stop time
stop = clock();
// Retrieve time elapsed in milliseconds
elapsedTime = (float)(stop - start) / (float)CLOCKS_PER_SEC * 1000.0f;

显然，这段代码只适用于以秒为单位计算的情况。而且，结果有时会很奇怪。

有人知道在 Linux 中创建高分辨率计时器的方法吗？

解决方案 1：

查看clock_gettime，它是高分辨率计时器的 POSIX 接口。

CLOCK_REALTIME如果阅读了手册页后，您仍对和之间的区别感到疑惑CLOCK_MONOTONIC，请参阅CLOCK_REALTIME 和 CLOCK_MONOTONIC 之间的区别？

请参阅以下页面以获取完整示例：http://www.guyrutenberg.com/2007/09/22/profiling-code-using-clock_gettime/

#include <iostream>
#include <time.h>
using namespace std;

timespec diff(timespec start, timespec end);

int main()
{
    timespec time1, time2;
    int temp;
    clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &time1);
    for (int i = 0; i< 242000000; i++)
        temp+=temp;
    clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &time2);
    cout<<diff(time1,time2).tv_sec<<":"<<diff(time1,time2).tv_nsec<<endl;
    return 0;
}

timespec diff(timespec start, timespec end)
{
    timespec temp;
    if ((end.tv_nsec-start.tv_nsec)<0) {
        temp.tv_sec = end.tv_sec-start.tv_sec-1;
        temp.tv_nsec = 1000000000+end.tv_nsec-start.tv_nsec;
    } else {
        temp.tv_sec = end.tv_sec-start.tv_sec;
        temp.tv_nsec = end.tv_nsec-start.tv_nsec;
    }
    return temp;
}

解决方案 2：

总结目前提供的信息，这是典型应用所需的两个功能。

#include <time.h>

// call this function to start a nanosecond-resolution timer
struct timespec timer_start(){
    struct timespec start_time;
    clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &start_time);
    return start_time;
}

// call this function to end a timer, returning nanoseconds elapsed as a long
long timer_end(struct timespec start_time){
    struct timespec end_time;
    clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &end_time);
    long diffInNanos = (end_time.tv_sec - start_time.tv_sec) * (long)1e9 + (end_time.tv_nsec - start_time.tv_nsec);
    return diffInNanos;
}

下面是一个示例，说明如何使用它们来计算输入列表的方差所需的时间。

struct timespec vartime = timer_start();  // begin a timer called 'vartime'
double variance = var(input, MAXLEN);  // perform the task we want to time
long time_elapsed_nanos = timer_end(vartime);
printf("Variance = %f, Time taken (nanoseconds): %ld
", variance, time_elapsed_nanos);

解决方案 3：

struct timespec t;
clock_gettime(CLOCK_REALTIME, &t);

还有 CLOCK_REALTIME_HR，但我不确定它是否有任何区别。

解决方案 4：

您对挂钟时间 (实际经过了多少时间) 或循环计数 (有多少个循环) 感兴趣吗？在第一种情况下，您应该使用类似 的内容gettimeofday。

最高分辨率计时器使用RDTSCx86 汇编指令。但是，这会测量时钟滴答，因此您应该确保已禁用省电模式。

TSC 的 wiki 页面给出了一些示例：http://en.wikipedia.org/wiki/Time_Stamp_Counter

解决方案 5：

阅读完该帖子后，我开始测试 clock_gettime 的代码与 c++11 的 chrono 的代码，但它们似乎不匹配。

他们之间的差距实在太大了！

std ::chrono::seconds(1)似乎相当于clock_gettime的~70,000

#include <ctime>
#include <cstdlib>
#include <cstring>
#include <iostream>
#include <thread>
#include <chrono>
#include <iomanip>
#include <vector>
#include <mutex>

timespec diff(timespec start, timespec end);
timespec get_cpu_now_time();
std::vector<timespec> get_start_end_pairs();
std::vector<timespec> get_start_end_pairs2();
void output_deltas(const std::vector<timespec> &start_end_pairs);

//=============================================================
int main()
{
    std::cout << "Hello waiter" << std::endl; // flush is intentional
    std::vector<timespec> start_end_pairs = get_start_end_pairs2();
    output_deltas(start_end_pairs);

    return EXIT_SUCCESS;
}

//=============================================================
std::vector<timespec> get_start_end_pairs()
{
    std::vector<timespec> start_end_pairs;
    for (int i = 0; i < 20; ++i)
    {
        start_end_pairs.push_back(get_cpu_now_time());
        std::this_thread::sleep_for(std::chrono::seconds(1));
        start_end_pairs.push_back(get_cpu_now_time());
    }

    return start_end_pairs;
}


//=============================================================
std::vector<timespec> get_start_end_pairs2()
{
    std::mutex mu;
    std::vector<std::thread> workers;
    std::vector<timespec> start_end_pairs;
    for (int i = 0; i < 20; ++i) {
        workers.emplace_back([&]()->void {
            auto start_time = get_cpu_now_time();
            std::this_thread::sleep_for(std::chrono::seconds(1));
            auto end_time = get_cpu_now_time();
            std::lock_guard<std::mutex> locker(mu);
            start_end_pairs.emplace_back(start_time);
            start_end_pairs.emplace_back(end_time);
        });
    }

    for (auto &worker: workers) {
        if (worker.joinable()) {
            worker.join();
        }
    }

    return start_end_pairs;
}

//=============================================================
void output_deltas(const std::vector<timespec> &start_end_pairs)
{
    std::cout << "size: " << start_end_pairs.size() << std::endl;
    for (auto it_start = start_end_pairs.begin(); it_start < start_end_pairs.end(); it_start += 2)
    {
        auto it_end = it_start + 1;
        auto delta = diff(*it_start, *it_end);

        std::cout
                << std::setw(2)
                << std::setfill(' ')
                << std::distance(start_end_pairs.begin(), it_start) / 2
                << " Waited ("
                << delta.tv_sec
                << "    s    "
                << std::setw(9)
                << std::setfill('0')
                << delta.tv_nsec
                << "    ns)"
                << std::endl;
    }
}

//=============================================================
timespec diff(timespec start, timespec end)
{
    timespec temp;
    temp.tv_sec = end.tv_sec-start.tv_sec;
    temp.tv_nsec = end.tv_nsec-start.tv_nsec;

    if (temp.tv_nsec < 0) {
        --temp.tv_sec;
        temp.tv_nsec += 1000000000;
    }
    return temp;
}

//=============================================================
timespec get_cpu_now_time()
{
    timespec now_time;
    memset(&now_time, 0, sizeof(timespec));
    clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &now_time);

    return now_time;
}

输出：

Hello waiter
 0 Waited (0    s       000843254       ns)
 1 Waited (0    s       000681141       ns)
 2 Waited (0    s       000685119       ns)
 3 Waited (0    s       000674252       ns)
 4 Waited (0    s       000714877       ns)
 5 Waited (0    s       000624202       ns)
 6 Waited (0    s       000746091       ns)
 7 Waited (0    s       000575267       ns)
 8 Waited (0    s       000860157       ns)
 9 Waited (0    s       000827479       ns)
10 Waited (0    s       000612959       ns)
11 Waited (0    s       000534818       ns)
12 Waited (0    s       000553728       ns)
13 Waited (0    s       000586501       ns)
14 Waited (0    s       000627116       ns)
15 Waited (0    s       000616725       ns)
16 Waited (0    s       000616507       ns)
17 Waited (0    s       000641251       ns)
18 Waited (0    s       000683380       ns)
19 Waited (0    s       000850205       ns)

解决方案 6：

clock_gettime(2)

解决方案 7：

epoll 实现：
https: //github.com/ielife/simple-timer-for-c-language

像这样使用：

timer_server_handle_t *timer_handle = timer_server_init(1024);
if (NULL == timer_handle) {
    fprintf(stderr, "timer_server_init failed
");
    return -1;
}
ctimer timer1;
    timer1.count_ = 3;
    timer1.timer_internal_ = 0.5;
    timer1.timer_cb_ = timer_cb1;
    int *user_data1 = (int *)malloc(sizeof(int));
    *user_data1 = 100;
    timer1.user_data_ = user_data1;
    timer_server_addtimer(timer_handle, &timer1);

    ctimer timer2;
    timer2.count_ = -1;
    timer2.timer_internal_ = 0.5;
    timer2.timer_cb_ = timer_cb2;
    int *user_data2 = (int *)malloc(sizeof(int));
    *user_data2 = 10;
    timer2.user_data_ = user_data2;
    timer_server_addtimer(timer_handle, &timer2);

    sleep(10);

    timer_server_deltimer(timer_handle, timer1.fd);
    timer_server_deltimer(timer_handle, timer2.fd);
    timer_server_uninit(timer_handle);