【OpenMP】多线程计算过程中任务调度问题

对于OpenMP的任务调度主要针对于并行的for循环，当每一次循环过程中的计算时间复杂度不一致的时候，简单的给每一个线程分配相同次数的迭代，会导致线程计算负载不均衡。不仅如此，对于实时计算的计算机，每一个核心的占用率是不一样的。针对该问题，OpenMP中给出多种线程调度的方式。

1. 基本使用

#include <stdio.h>
#include <omp.h>

const int NUMS = 20;

int main() {
    omp_set_num_threads(4);

#pragma omp parallel for
    for(int i=0; i< NUMS; ++i) {
        printf("id is %3d   thread is %d\n",i, omp_get_thread_num());
    }
    return 0;
}

输出结果：

id is   0   thread is 0
id is   1   thread is 0
id is   2   thread is 0
id is   3   thread is 0
id is  15   thread is 3
id is  16   thread is 3
id is  17   thread is 3
id is  18   thread is 3
id is  19   thread is 3
id is   4   thread is 0
id is  10   thread is 2
id is  11   thread is 2
id is  12   thread is 2
id is  13   thread is 2
id is  14   thread is 2
id is   5   thread is 1
id is   6   thread is 1
id is   7   thread is 1
id is   8   thread is 1
id is   9   thread is 1

可以看到对于循环20次的时候，使用四个线程，每一个线程会平均分配5个迭代运算，分配方式是按顺序分配。

2. schedule子句

语法：schedule(type[, size])
（1）type：表示调度的类型，分别有static、dynamic、guided、runtime。在实际使用的过程中，一般不会使用runtime，当类型为runtime的时候，不能设置size。
（2）size：表示一次性分配给一个线程的循环次数。

3. 静态调度static

编译器在执行没有使用schedule的并行指令时，默认是static子句。static在编译的过程中，就已经确定了线程的分配方案。如果不使用size制定，默认每一个线程会分配N/num_threads个迭代。也就是说，先给第一个线程直接分配N/num_threads个，依次遍历每一个线程。但是如果设置了size，首先给第一个线程分配size个迭代，然后分配到最后一个线程，再给第一个线程分配。（实际的分配过程可能不是这样子的，但是实际看到）

#include <stdio.h>
#include <omp.h>

const int NUMS = 20;

int main() {
    omp_set_num_threads(4);

#pragma omp parallel for schedule(static)
    for(int i=0; i< NUMS; ++i) {
        printf("id is %3d   thread is %d\n",i, omp_get_thread_num());
    }
    return 0;
}

输出结果：

id is   0   thread is 0
id is   1   thread is 0
id is   2   thread is 0
id is   3   thread is 0
id is   4   thread is 0
id is   5   thread is 1
id is   6   thread is 1
id is   7   thread is 1
id is   8   thread is 1
id is   9   thread is 1
id is  10   thread is 2
id is  11   thread is 2
id is  12   thread is 2
id is  13   thread is 2
id is  14   thread is 2
id is  15   thread is 3
id is  16   thread is 3
id is  17   thread is 3
id is  18   thread is 3
id is  19   thread is 3

使用size

#include <stdio.h>
#include <omp.h>

const int NUMS = 20;

int main() {
    omp_set_num_threads(4);

#pragma omp parallel for schedule(static， 2)
    for(int i=0; i< NUMS; ++i) {
        printf("id is %3d   thread is %d\n",i, omp_get_thread_num());
    }
    return 0;
}

输出结果：

id is   0   thread is 0
id is   1   thread is 0
id is   8   thread is 0
id is   9   thread is 0
id is  16   thread is 0
id is  17   thread is 0
id is   4   thread is 2
id is   5   thread is 2
id is  12   thread is 2
id is  13   thread is 2
id is   2   thread is 1
id is   3   thread is 1
id is  10   thread is 1
id is  11   thread is 1
id is  18   thread is 1
id is  19   thread is 1
id is   6   thread is 3
id is   7   thread is 3
id is  14   thread is 3
id is  15   thread is 3

4.动态调度dynamic

对于在实际运行的过程，有的线程执行速度较快，因此在执行完之后会去领取新的任务。如果不设置size，是会迭代每一个循环分配给各个线程，如果使用size，会每次分配size个给一个线程。

#include <stdio.h>
#include <omp.h>

const int NUMS = 20;

int main() {
    omp_set_num_threads(4);

#pragma omp parallel for schedule(dynamic)
    for(int i=0; i< NUMS; ++i) {
        printf("id is %3d   thread is %d\n",i, omp_get_thread_num());
    }
    return 0;
}

输出结果：

id is   0   thread is 0
id is   4   thread is 0
id is   5   thread is 0
id is   6   thread is 0
id is   7   thread is 0
id is   8   thread is 0
id is   3   thread is 3
id is  10   thread is 3
id is  11   thread is 3
id is  12   thread is 3
id is  13   thread is 3
id is   1   thread is 2
id is  15   thread is 2
id is  16   thread is 2
id is  17   thread is 2
id is  18   thread is 2
id is  19   thread is 2
id is   9   thread is 0
id is   2   thread is 1
id is  14   thread is 3

使用size：

#include <stdio.h>
#include <omp.h>

const int NUMS = 20;

int main() {
    omp_set_num_threads(4);

#pragma omp parallel for schedule(dynamic, 3)
    for(int i=0; i< NUMS; ++i) {
        printf("id is %3d   thread is %d\n",i, omp_get_thread_num());
    }
    return 0;
}

输出结果：

id is   0   thread is 0
id is   1   thread is 0
id is   2   thread is 0
id is   3   thread is 3
id is   4   thread is 3
id is   5   thread is 3
id is  15   thread is 3
id is  16   thread is 3
id is  17   thread is 3
id is  18   thread is 3
id is  19   thread is 3
id is  12   thread is 0
id is  13   thread is 0
id is  14   thread is 0
id is   6   thread is 1
id is   7   thread is 1
id is   8   thread is 1
id is   9   thread is 2
id is  10   thread is 2
id is  11   thread is 2

5. 启发式调度guided

采用启发式调度方法进行调度，每次分配给线程迭代次数不同，开始比较大，以后逐渐减小。size表示每次分配的迭代次数的最小值，由于每次分配的迭代次数会逐渐减少，少到size时，将不再减少。如果不知道size的大小，那么默认size为1，即一直减少到1。
每个任务分配的任务是先大后小，指数下降。当有大量任务需要循环时，刚开始为线程分配大量任务，最后任务不多时，给每个线程少量任务，可以达到线程任务均衡。

参考链接：https://www.cnblogs.com/xudong-bupt/p/3622101.html