c++ - 多线程执行时间与随机数之和

Question

我正在尝试创建一个多线程程序，该程序将 N 个随机数 [-100,100] 的数组与 K 个工作线程相加，这些工作线程由程序员实现的自旋锁(忙等待)序列化。在尝试使用随机数之前，出于测试目的，我将整个数组初始化为 1，正如您将在我的代码中看到的那样。由于我不知道问题出在哪里，所以我将发布完整代码:

#include <iostream>
#include <string.h>
#include <pthread.h>
#include <cstdlib>
#include <time.h>
#include <atomic>
#include <chrono>

using namespace std;
using namespace chrono;

struct lock {

    long double sum = 0;
    atomic_flag m_flag = ATOMIC_FLAG_INIT; // Inicializa com m_flag = 0

    void acquire() {
        while(m_flag.test_and_set());
    }
    void release() {
        m_flag.clear();
    }
};

struct t_data{
    int t_id;
    char* sumArray;
    struct lock* spinlock;
};

void* sum(void* thread_data) {

    struct t_data *my_data;
    long double m_sum=0;
    my_data = (struct t_data *) thread_data; 

    for (int i=0;i<strlen(my_data->sumArray);i++) {
        m_sum += my_data->sumArray[i];
    }

    my_data->spinlock->acquire();
    cout << "THREAD ID: " << my_data->t_id << endl;
    cout << "Acquired lock." << endl;
    my_data->spinlock->sum += m_sum;
    cout << "Releasing lock..." << endl << endl;
    my_data->spinlock->release();

}

int main(int argc, char** argv) {

    // Inicializar cronômetro, arrays, spinlock,etc.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            , spinlock, etc. 
    system_clock::time_point starting_time = system_clock::now();
    int K = atoi(argv[1]);
    int N = atoi(argv[2]);
    int temp;
    double expected_sum = 0;
    pthread_t threads[K];
    struct t_data threads_data[K];
    struct lock spinlock;
    const long int numElements = (long int) N/K;  //Divisão inteira de N/K para dividir array em parcelas

    // Criar array[K] de arrays para delegar cada sub-lista a uma thread
    char** numArrays = new char*[K]; 
    for(int i=0;i<K;i++)
        numArrays[i] = new char[numElements];  //Char utilizado para que seja alocado apenas 1 byte por número

    // Inicializar seed aleatória para preenchimento de arrays
    srand(time(NULL)); 

    //Preencher arrays que serão passados às threads criadas
    for (int i=0;i<K;i++) {
        for(int j=0;j<numElements;j++) {
            temp = 1;//rand() % 201 - 100; (CHANGING THIS GIVES UNEXPECTED RESULTS)
            numArrays[i][j] = temp;
            expected_sum+=temp;
        }
        //Criar threads e passando argumentos(id,spinlock,array)
        threads_data[i].t_id = i;
        threads_data[i].spinlock = &spinlock;
        threads_data[i].sumArray = numArrays[i];
        pthread_create(&threads[i],NULL,sum,(void*)&threads_data[i]);
    }

    // Parar o programa até que todas as threads terminem para imprimir soma correta
    for (int i=0;i<K;i++){
        if(pthread_join(threads[i],NULL)) cout << "Error waiting for threads." << endl;
    }

    // Somando últimos valores restantes no caso de N%K != 0 (esta parcela torna-se irrelevante à medida que N >> K)
    for(int i=0;i<(int)N%K;i++) {
        temp = 1;//rand() % 201 - 100; (CHANGING THIS GIVES UNEXPECTED RESULTS)
        spinlock.sum+=temp;
        expected_sum+=temp;
    }

    // Printar resultado esperado, o calculado e tempo de execução
    cout << "EXPECTED SUM = " << expected_sum << endl;
    cout << "CALCULATED SUM = " << spinlock.sum << endl;

    // Liberar memória alocada
    for(int i=0;i<K;i++)
        delete[] numArrays[i];

    delete[] numArrays;

    auto start_ms = time_point_cast<milliseconds>(starting_time);
    auto now = system_clock::now();
    auto now_ms = time_point_cast<milliseconds>(now);
    auto value = now_ms - start_ms;
    long execution_time = value.count();
    cout << "-----------------------" << endl;
    cout << "Execution time: " << execution_time << "ms" << endl;
    return 0;
}

这在计算总和时效果很好，但会带来执行时间问题:它应该与 (N/K) 成线性比例，但测试 K=10，N=10⁶:

EXPECTED SUM = 1e+06
CALCULATED SUM = 1e+06
-----------------------
Execution time: 1310ms

并且 K=10，N=2*10⁶:

EXPECTED SUM = 2e+06
CALCULATED SUM = 2e+06
-----------------------
Execution time: 7144ms

我不知道为什么会这样。应该加倍。更改 K 正常工作。另外，如果我使用 rand() % 201-100 而不是 1，事情就会变得一团糟。对于 K=10,N=10⁶:

EXPECTED SUM = -16307
CALCULATED SUM = 1695
-----------------------
Execution time: 95ms

关于执行时间的变化，N 是固定的(线性缩放)但 K 不再有任何区别。这些对我来说都没有意义。

提前致谢!

Answer 1

strlen(my_data->sumArray)将停在第一个 0在字符数组/c 字符串中继续总结 temp expected_sum 的值.使用 vector对于非 ascii 数据(毕竟这是 C++):

// use a vector in t_data
struct t_data{
    int t_id;
    std::vector<char> sumArray;
    lock* spinlock;
};

// adjust summing up in sum(void* thread_data)
for (char value : my_data->sumArray) {
    m_sum += value;
}

// initialise like this
threads_data[i].sumArray.resize(numElements);
for(size_t j = 0; j < threads_data[i].sumArray.size(); ++j) {
    char temp = 1; //or (char)(rand() % 201 - 100);
    threads_data[i].sumArray[j] = temp;
    expected_sum += temp;
}

现在考虑一下你的时间安排:移动 threads_data[i] 的初始化和 expected_sum在计时区域之外，否则数百万rand电话肯定会主宰一切。在任何情况下，您都在测量顺序版本和并行版本，因此您不能指望 K在时间上有所不同:您总是至少测量顺序版本 + 最后一个并行版本(加入时)。