在 MPI 中发送和接收数组

Question

我是 MPI 新手，正在编写一个简单的 MPI 程序来获取矩阵和向量的点积，即 A*b=c。但是，我的代码不起作用。源代码如下。

如果我将 A、b、c 和缓冲区的声明替换为

double A[16], b[4], c[4], buffer[8];

并注释与分配和释放操作相关的那些行，我的代码有效并且结果是正确的。在这种情况下，我在想问题应该与指针有关，但我不知道如何解决问题。

还有一点，在我的代码中，buffer只有4个元素，但是buffer的大小必须大于8，否则不起作用。

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

using namespace std;

int nx = 4, ny = 4, nxny;
int ix, iy;
double *A = nullptr, *b = nullptr, *c = nullptr, *buffer = nullptr;
double ans;

// info MPI
int myGlobalID, root = 0, numProc;
int numSent;
MPI_Status status;

// functions
void get_ixiy(int);

int main(){

  MPI_Init(NULL, NULL);
  MPI_Comm_size(MPI_COMM_WORLD, &numProc);
  MPI_Comm_rank(MPI_COMM_WORLD, &myGlobalID);

  nxny = nx * ny;

  A = new double(nxny);
  b = new double(ny);
  c = new double(nx);
  buffer = new double(ny);

  if(myGlobalID == root){
    // init A, b
    for(int k = 0; k < nxny; ++k){
      get_ixiy(k);
      b[iy] = 1;
      A[k] = k;
    }
    numSent = 0;

    // send b to each worker processor
    MPI_Bcast(&b, ny, MPI_DOUBLE, root, MPI_COMM_WORLD);

    // send a row of A to each worker processor, tag with row number
    for(ix = 0; ix < min(numProc - 1, nx); ++ix){
      for(iy = 0; iy < ny; ++iy){
        buffer[iy] = A[iy + ix * ny];
      }
      MPI_Send(&buffer, ny, MPI_DOUBLE, ix+1, ix+1, MPI_COMM_WORLD);
      numSent += 1;
    }

    for(ix = 0; ix < nx; ++ix){
      MPI_Recv(&ans, 1, MPI_DOUBLE, MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
      int sender = status.MPI_SOURCE;
      int ansType = status.MPI_TAG;
      c[ansType] = ans;

      // send another row to worker process
      if(numSent < nx){
        for(iy = 0; iy < ny; ++iy){
          buffer[iy] = A[iy + numSent * ny];
        }
        MPI_Send(&buffer, ny, MPI_DOUBLE, sender, numSent+1, 
        MPI_COMM_WORLD);
        numSent += 1;
      }
      else
        MPI_Send(MPI_BOTTOM, 0, MPI_DOUBLE, sender, 0, MPI_COMM_WORLD);
    }

    for(ix = 0; ix < nx; ++ix){
      std::cout << c[ix] << " ";
    }
    std::cout << std::endl;

    delete [] A;
    delete [] b;
    delete [] c;
    delete [] buffer;
  }
  else{
    MPI_Bcast(&b, ny, MPI_DOUBLE, root, MPI_COMM_WORLD);
      if(myGlobalID <= nx){
        while(1){
          MPI_Recv(&buffer, ny, MPI_DOUBLE, root, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
          if(status.MPI_TAG == 0) break;
          int row = status.MPI_TAG - 1;
          ans = 0.0;

          for(iy = 0; iy < ny; ++iy) ans += buffer[iy] * b[iy];

          MPI_Send(&ans, 1, MPI_DOUBLE, root, row, MPI_COMM_WORLD);
      }
    }
  }

  MPI_Finalize();
  return 0;
} // main

void get_ixiy(int k){
  ix = k / ny;
  iy = k % ny;
}

错误信息如下。

=   BAD TERMINATION OF ONE OF YOUR APPLICATION PROCESSES
=   PID 7455 RUNNING AT ***
=   EXIT CODE: 11
=   CLEANING UP REMAINING PROCESSES
=   YOU CAN IGNORE THE BELOW CLEANUP MESSAGES

YOUR APPLICATION TERMINATED WITH THE EXIT STRING: Segmentation fault: 
11 (signal 11)
This typically refers to a problem with your application.
Please see the FAQ page for debugging suggestions

Answer 1

您的代码中有几个问题，您必须先解决这些问题。

首先，你想访问一个不存在的b[]元素，在这个for循环中：

for(int k = 0; k < nxny; ++k){
  get_ixiy(k);
  b[k] = 1;     // WARNING: this is an error
  A[k] = k;
}

其次，您正在删除仅为根进程分配的内存。这会导致内存泄漏：

if(myGlobalID == root){
  // ...
  delete [] A;
  delete [] b;
  delete [] c;
  delete [] buffer;
}

您必须删除为所有进程分配的内存。

第三，你有一个无用的函数void get_ixiy(int);，它改变了全局变量ix，iy。这是无用的，因为在调用此函数后，您永远不会使用 ix, iy 直到您手动更改它们。见这里：

for(ix = 0; ix < min(numProc - 1, nx); ++ix){
    for(iy = 0; iy < ny; ++iy){
        // ...
    }
}

第四，您以完全错误的方式使用MPI_Send() 和MPI_Recv()。你很幸运没有遇到更多错误。