MPI_Scatter - not working as expected

问题:

I am writing my first program using MPI and I am having hard time trying to properly send data to other processes using MPI_Scatter, modify them and receive the values using MPI_Gather. The code is as follows:

int** matrix;
int m = 2, n = 2;
int status;

// could have been int matrix[2][2];

matrix = malloc(m*sizeof(int*));

for(i = 0; i < m; i++) {
  matrix[i] = malloc(n*sizeof(int));
}

matrix[0][0] = 1;
matrix[0][1] = 2;
matrix[1][0] = 2;
matrix[1][1] = 3;

MPI_Init( &argc, &argv );
MPI_Comm_rank( MPI_COMM_WORLD, &rank ); 
MPI_Comm_size( MPI_COMM_WORLD, &size );

printf("My name is %d out of %d\n", rank, size);

MPI_Scatter(&matrix, 1, MPI_INT, &status, 1, MPI_INT, 0, MPI_COMM_WORLD);

printf("My name is %d and my status is %d\n", rank, status);

status += 1;

MPI_Gather(&status,1,MPI_INT,&matrix,1,MPI_INT,0,MPI_COMM_WORLD);

The results is as follows:

My name is 0 out of 4
My name is 1 out of 4
My name is 1 and my status is 0
My name is 2 out of 4
My name is 2 and my status is 120
My name is 0 and my status is 17773264
My name is 3 out of 4
My name is 3 and my status is 0

After MPI_Gather when I try to print the content of matrix I get segmentation fault...

Which is not what I expect... I suspect that the problem is I have a 2d array and want to send single values. Is this right? How to correct this?

回答1:

The reason for the segmentation fault is, that you're allocating a 2D array but there is no guarantee that it uses contiguous memory. So, allocate a 1D array and scatter/gather it. This also explains why some nodes get the wrong data.

You named the receiving variable status which I find misleading, since it'll hold the received value(s). So you'll have to allocate it correctly for the case when a node receives more then one value. Also I'd suggest to rename it to e.g. local_matrix or similar.

Side node: You allocate matrix on all nodes, but you should only allocate on a single node. This can be done by checking the current rank and placing a barrier after the check. Furthermore, I assume that you also print on all nodes, but this code isn't visible. As with the allocation, print should done as well only on one node.

Also have a look at How are MPI_Scatter and MPI_Gather used from C? which explains it pretty well.