Error when try to use Mpi recv : signal segmentation fault

Question

:)

I got a disturbing question concerning a mpi program. The idea is : each process(slave) send data to master in order to compute the mandelbrot fractal.

Firstly, each slave sent point and it worker. Then they sent line and it worked !

But now, i try to make them sending a block of line(let's suppose 5 lines, so a submatrix).

My idea is to make these five lines into a single line. The master receive the first "new" line but doesn't for the others O_o. I'm disturbed.

I receive for the others(>1) : signal segmentation fault signal code : adress not mapped failing at adress

Please Help me ! Because it's a long time , i'have been looking for :(

Ps : i'm french (so that's why my english is bad )

  //the whole table to be used in a master

  //int table[NX*NY];
  //int count =0;

  if (rank == 0) {

    int res;
    int line[MAXY+MAXY+1];
    int block[5*(MAXY+MAXY+1)];
    int count = 0;    

    /* Begin User Program  - the master */

   //MPI_Recv(&line, MAXY+MAXY+1, MPI_INT,MPI_ANY_SOURCE, DATATAG, MPI_COMM_WORLD, &status);
   MPI_Recv(&block, 5*(MAXY+MAXY+1), MPI_INT,MPI_ANY_SOURCE, DATATAG, MPI_COMM_WORLD, &status);
   printf("sizeof of datablock received is = %d \n",sizeof(block)/sizeof(block[0]));
   recvd = status.MPI_SOURCE;
   printf("i have received blockdata from %d \n",recvd);
   /* remplissage du case */

   for(i = -MAXX; i <= MAXX; i++) { 
      for(j = -MAXY; j <= MAXY; j++) { 
        cases[i + MAXX][j + MAXY] = block[count%(MAXY+MAXY+1)];
        //printf("j'ai fait un bloc[count], pas credible\n"); 
        count++;
        }
    } 

    dump_ppm("mandel.ppm", cases);
    printf("Fini.\n");
  }

  else {

    /* On est l'un des fils */
    /* for the block;let's suppose each son send 5 rows*/ 
    double x, y;
    int i, j, res, rc, rank,count;
    //int line[MAXY + MAXY + 1];
    int block[5*(MAXY+MAXY+1)];
    count = 0;
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    for(i = -MAXX; i <= MAXX; i++) {
      for(j = -MAXY; j <= MAXY; j++) {
        x = 2 * i / (double)MAXX;
        y = 1.5 * j / (double)MAXY;
        res = mandel(x, y);
        //line[j+MAXY] = res;
        block[count] =res;
        if (count % (5*(MAXY+MAXY+1)) == 0){ //we send each five rows
            MPI_Send(&block,5*(MAXY+MAXY+1), MPI_INT, 0, DATATAG, MPI_COMM_WORLD);
            printf("me slave %d, have sent datablock to master\n",rank);
            printf("sizeof of datablock sent is = %d\n",sizeof(block)/sizeof(block[0]));
            }
        count++;
        }
    //MPI_Send(&line, MAXY+MAXY+1 , MPI_INT, 0, DATATAG, MPI_COMM_WORLD);
    }
  }

  MPI_Finalize();
  return 0;
}

Answer 1

The function MPI_Recv() needs the address of the buffer where the data will be received. It is the same thing for MPI_Send(). Since int block[5*(MAXY+MAXY+1)] is an array, block points to the first item of the array block[0]: this is the address that is required. On the other hand, &block points to block: it's similar to a pointer to a pointer to int. But the value of &block is not the address of the first item of the array!

Hence, could you try:

int block[5*(MAXY+MAXY+1)]
...
MPI_Send(block,5*(MAXY+MAXY+1), MPI_INT, 0, DATATAG, MPI_COMM_WORLD);
...
MPI_Recv(block, 5*(MAXY+MAXY+1), MPI_INT,MPI_ANY_SOURCE, DATATAG, MPI_COMM_WORLD, &status);

Which is equivalent to:

int block[5*(MAXY+MAXY+1)]
...
MPI_Send(&block[0],5*(MAXY+MAXY+1), MPI_INT, 0, DATATAG, MPI_COMM_WORLD);
...
MPI_Recv(&block[0], 5*(MAXY+MAXY+1), MPI_INT,MPI_ANY_SOURCE, DATATAG, MPI_COMM_WORLD, &status);

What if you what to send a single integer int a? The address of a (&a) can be provided to MPI_Send(), as performed in many examples devoted to MPI_Send():

int a=42;
MPI_Send(&a,1, MPI_INT, 0, DATATAG, MPI_COMM_WORLD);

Lastly, make sure that MPI_Send() is called as many times as MPI_Recv(). Indeed, in the code you posted, MPI_Recv() is called only once by the root process, while each non-root process will send a message to the root. Consequently, the program will work for 2 processes and it is likely to fail if more processes are used or if a single process is used.