【问题标题】:MPI (Summation)MPI(求和)
【发布时间】:2016-09-24 00:06:00
【问题描述】:

我正在编写一个程序来计算每个数字的总和,直到 1000。例如,1+2+3+4+5....+100。首先,我将求和作业分配给 10 个处理器:处理器 0 得到 1-100,处理器 1 得到 101-200,依此类推。总和存储在一个数组中。

在并行完成所有求和后,处理器将它们的值发送到处理器 0(处理器 0 使用非阻塞发送/接收接收值),处理器 0 对所有值求和并显示结果。

代码如下:

#include <mpi.h>
#include <iostream>

using namespace std;

int summation(int, int);

int main(int argc, char ** argv)
{
    int * array;
    int total_proc;
    int curr_proc;
    int limit = 0;
    int partial_sum = 0;
    int upperlimit = 0, lowerlimit = 0;

    MPI_Init(&argc, &argv);
    MPI_Comm_size(MPI_COMM_WORLD, &total_proc);
    MPI_Comm_rank(MPI_COMM_WORLD, &curr_proc);
    MPI_Request send_request, recv_request;

    /* checking if 1000 is divisible by number of procs, else quit */
    if(1000 % total_proc != 0)
    {
        MPI_Finalize();
        if(curr_proc == 0)
            cout << "**** 1000 is not divisible by " << total_proc << " ...quitting..."<< endl;
        return 0;
    }

    /* number of partial summations */
    limit = 1000/total_proc;

    array = new int [total_proc];

    /* assigning jobs to processors */
    for(int i = 0; i < total_proc; i++)
    {
        if(curr_proc == i)
        {
            upperlimit = upperlimit + limit;
            lowerlimit = (upperlimit - limit) + 1;
            partial_sum = summation(upperlimit, lowerlimit);
            array[i] = partial_sum;
        }
        else
        {
            upperlimit = upperlimit + limit;
            lowerlimit = (upperlimit - limit) + 1;
        }
    }

    cout << "** Partial Sum From Process " << curr_proc << " is " << array[curr_proc] << endl;

    /* send and receive - non blocking */
    for(int i = 1; i < total_proc; i++)
    {
        if(curr_proc == i) /* (i = current processor) */
        {
            MPI_Isend(&array[i], 1, MPI_INT, 0, i, MPI_COMM_WORLD, &send_request);
            cout << "-> Process " << i << " sent " << array[i] << " to Process 0" << endl;

            MPI_Irecv(&array[i], 1, MPI_INT, i, i, MPI_COMM_WORLD, &recv_request);
            //cout << "<- Process 0 received " << array[i] << " from Process " << i << endl;
        }
    }

    MPI_Finalize();

    if(curr_proc == 0)
    {
        for(int i = 1; i < total_proc; i++)
            array[0] = array[0] + array[i];
        cout << "Sum is " << array[0] << endl;
    }

    return 0;
}

int summation(int u, int l)
{
    int result = 0; 
    for(int i = l; i <= u; i++)
        result = result + i;
    return result;
}

输出:

** Partial Sum From Process 0 is 5050
** Partial Sum From Process 3 is 35050
-> Process 3 sent 35050 to Process 0
<- Process 0 received 35050 from Process 3
** Partial Sum From Process 4 is 45050
-> Process 4 sent 45050 to Process 0
<- Process 0 received 45050 from Process 4
** Partial Sum From Process 5 is 55050
-> Process 5 sent 55050 to Process 0
<- Process 0 received 55050 from Process 5
** Partial Sum From Process 6 is 65050
** Partial Sum From Process 8 is 85050
-> Process 8 sent 85050 to Process 0
<- Process 0 received 85050 from Process 8
-> Process 6 sent 65050 to Process 0
** Partial Sum From Process 1 is 15050
** Partial Sum From Process 2 is 25050
-> Process 2 sent 25050 to Process 0
<- Process 0 received 25050 from Process 2
<- Process 0 received 65050 from Process 6
** Partial Sum From Process 7 is 75050
-> Process 1 sent 15050 to Process 0
<- Process 0 received 15050 from Process 1
-> Process 7 sent 75050 to Process 0
<- Process 0 received 75050 from Process 7
** Partial Sum From Process 9 is 95050
-> Process 9 sent 95050 to Process 0
<- Process 0 received 95050 from Process 9
Sum is -1544080023

打印数组的内容:

5050
536870912
-1579286148
-268433415
501219332
32666
501222192
32666
1
0

我想知道是什么原因造成的。

如果我在调用 MPI_Finalize 之前打印数组,它可以正常工作。

【问题讨论】:

    标签: c++ mpi partial


    【解决方案1】:

    你的程序最重要的缺陷是你如何划分工作。在 MPI 中,每个进程都在执行主函数。因此,如果您希望所有进程协作构建结果,则必须确保所有进程都执行您的 summation 函数。

    您不需要 for 循环。每个进程都分别执行 main 。它们只是具有不同的curr_proc 值,您可以根据该值计算它们必须执行的工作部分:

    /* assigning jobs to processors */
    int chunk_size = 1000 / total_proc;
    lowerlimit = curr_proc * chunk_size;
    upperlimit = (curr_proc+1) * chunk_size;
    partial_sum = summation(upperlimit, lowerlimit);
    

    那么,主进程如何接收所有其他进程的部分和是不正确的。

    • MPI 等级值 (curr_proc) 从 0 开始到 MPI_Comm_size 输出值 (total_proc-1)。
    • 只有进程#1 正在发送/接收数据。
    • 您正在使用即时版本的发送和接收:MPI_IsendMPI_recv,但您不会等到这些请求完成。为此,您应该使用 MPI_Waitall

    正确的版本应该是这样的:

    if( curr_proc == 0 ) {
       // master process receives all data
       for( int i = 1; i < total_proc; i++ )
          MPI_Recv( &array[i], MPI_INT, 1, i, 0, MPI_COMM_WORLD );
    } else {
       // other processes send data to the master
       MPI_Send( &partial_sum, MPI_INT, 1, 0, 0, MPI_COMM_WORLD );
    }
    

    这种多对一的通信模式被称为 gather。在 MPI 中有一个函数已经执行了这个功能:MPI_Gather

    最后,您打算执行的操作称为reduction:获取给定数量的数值并通过连续执行单个操作(在您的情况下为总和)生成单个输出值。在 MPI 中也有一个函数可以做到这一点:MPI_Reduce

    我强烈建议您在尝试自己制作之前先做some basic guided exercises。 MPI 一开始很难理解。建立良好的基础对于您以后能够增加复杂性至关重要。 hands on tutorial 也是 MPI 入门的好方法。

    编辑:忘记提及您不需要强制将问题大小(在本例中为 1000)除以资源数量 (total_proc)。根据具体情况,您可以将剩余部分分配给单个进程:

    chunk_size = 1000 / total_proc;
    if( curr_proc == 0 )
        chunk_size += 1000 % total_proc;
    

    或者尽量平衡一下:

    int remainder = curr_proc < ( 1000 % proc )? 1 : 0;
    lowerlimit = curr_proc * chunk_size /* as usual */
               + curr_proc;             /* cumulative remainder */
    upperlimit = (curr_proc + 1) * chunk_size /* as usual */
               + remainder;                   /* curr_proc remainder */
    

    第二种情况,负载不平衡会高达1,而第一种情况最坏情况下负载不平衡可以达到total_proc-1

    【讨论】:

      【解决方案2】:

      您只是在初始化array[i],即与curr_proc id 对应的元素。该数组中的其他元素将未初始化,从而产生随机值。在您的发送/接收打印循环中,您只能访问已初始化的元素。

      我对 MPI 不太熟悉,所以我猜,但您可能希望在调用 MPI_Init 之前分配 array。或者在进程 0 上调用 MPI_Receive,而不是每个单独的进程。

      【讨论】:

        猜你喜欢
        • 2018-05-13
        • 1970-01-01
        • 1970-01-01
        • 2015-10-02
        • 1970-01-01
        • 1970-01-01
        • 2014-02-05
        • 2019-04-29
        • 2011-11-02
        相关资源
        最近更新 更多