pthread 在 Windows 中比串行快，但在 Linux 中比串行慢

Question

我正在尝试在 Windows 和 Linux 上以相同数量的线程（4 个线程和 4 个 CPU）运行相同的 C++ 并行代码以使用蒙特卡洛算法计算 pi。虽然并行代码在 Windows 上比串行实现要快，但在 Linux 上要慢得多。

这是程序：

#include <iostream>
#include <cstdlib>
#include <ctime>
#include <cmath>
#include <pthread.h>
#include <chrono>

using namespace std;
using ns = chrono::nanoseconds;
using get_time = chrono::steady_clock;

static int thread_count = 4;
pthread_mutex_t myMutex;

struct args{
    int id;
    int random_count;
    double *pi;
};

double compute_pi(long n)
{
     double pi = 0;
     double x, y;
     for(long i=0; i<n; i++){
         x = -1 + 2 * double(rand())/RAND_MAX;
         y = -1 + 2 * double(rand())/RAND_MAX;
         if (sqrt(x*x + y*y) <= 1.0) pi++;
    }
    return 4*pi/n;
}

void* threadFunc(void *argin){
    args *inputs = (args*) argin;
    double my_sum = 0;
    double x, y;
    for(int i=0; i<inputs->random_count; i++){
        x = -1 + 2 * double(rand())/RAND_MAX;
        y = -1 + 2 * double(rand())/RAND_MAX;
        if (sqrt(x*x + y*y) <= 1.0) my_sum++;
    }
    pthread_mutex_lock(&myMutex);
    *(inputs->pi) += my_sum;
    pthread_mutex_unlock(&myMutex);
    return nullptr;
}

double compute_pi_parallel(long n)
{
    double pi = 0;
    int count_per_thread = n/thread_count;
    pthread_t *threads = new pthread_t[thread_count];
    args *funcInputs = new args[thread_count];
    pthread_mutex_init(&myMutex, nullptr);

    for(int i=0; i<thread_count; i++){
        funcInputs[i].id = i;
        funcInputs[i].random_count = i<n%thread_count ? count_per_thread+1 : 
    count_per_thread;
        funcInputs[i].pi = &pi;
        int rc = pthread_create(&threads[i], nullptr, threadFunc, (void *) 
                                &funcInputs[i]);
        if(rc) cerr << "error in thread creation!\n";
    }

    for(int i=0; i<thread_count; i++){
        int rc = pthread_join(threads[i], nullptr);
        if(rc) cerr << "Error in thread join!\n";
    }

    pthread_mutex_destroy(&myMutex);
    delete [] funcInputs;
    delete [] threads;
    return 4*pi/n;
}

int main(int argc, char* argv[])
{
    srand(time(nullptr));
    long n = 100000000;
    auto start = get_time::now();
    if (argc > 1){
        n = atol(argv[1]);
        if (argc == 3){
            thread_count = atoi(argv[2]);
            cout << "pi(parallel) = " << compute_pi_parallel(n) << endl;
            auto stop = get_time::now();
            auto diff = stop - start;
            cout<<"Elapsed time is :  "<< chrono::duration_cast<ns>
    (diff).count()/1e9<<" s "<<endl;
            return 0;
        }
    }

    cout << "pi = " << compute_pi(n) << endl;
    auto stop_s = get_time::now();
    auto diff_s = stop_s - start;
    cout << "pi(parallel) = " << compute_pi_parallel(n) << endl;
    auto stop_p = get_time::now();
    auto diff = stop_p - stop_s;
    cout<<"Elapsed time for serial is :  "<< chrono::duration_cast<ns>
    (diff_s).count()/1e9<<" s "<<endl;
    cout<<"Number of threads: "<< thread_count<< endl;
    cout<<"Elapsed time for parallel is :  "<< chrono::duration_cast<ns>
    (diff).count()/1e9<<" s "<<endl;
    return 0;
}

Windows 上的输出：

pi = 3.14146
pi(parallel) = 3.14087
Elapsed time for serial is :  6.16426 s
Number of threads: 4
Elapsed time for parallel is :  1.0659 s

在 Linux 上：$g++ -std=c++11 -g -Wall -o mc mc.cpp -lpthread

输出：

pi = 3.14138
pi(parallel) = 3.14166
Elapsed time for serial is :  3.10837 s 
Number of threads: 4
Elapsed time for parallel is :  19.8226 s

我用 $lscpu 检查了 Linux 上的 CPU 数量，并用 $top 监控了 CPU 使用情况，似乎 Linux 正在使用所有可用的内核，但仍然它比串行代码慢。我正在 Windows 虚拟机上的 Ubuntu 16.04 LTS 上运行该程序。

我想知道我在 Linux 上是否做错了什么。

Answer 1

您正在使用 rand。如果 rand 是线程安全的，则它是实现定义的。它可以简单地调用互斥锁。使用没有全局状态的现代 C++ 随机数生成器。