Derived datatypes describe scattered data (matrix columns, alternating elements, struct fields) so MPI can send them directly without manual packing.

double tmp[N_ROWS];
for (int i = 0; i < N_ROWS; i++)
    tmp[i] = matrix[i][col];
MPI_Send(tmp, N_ROWS, MPI_DOUBLE, dest, 0, MPI_COMM_WORLD);

This adds an allocation and a copy on every send. Derived datatypes describe the layout of the original data so MPI can gather it directly without an intermediate buffer. MPI_Type_vector describes a strided pattern: count blocks of blocklength elements separated by a stride.

MPI_Datatype col_t;
// one column of a row-major matrix: N_ROWS blocks of 1 double, stride = N_COLS
MPI_Type_vector(N_ROWS, 1, N_COLS, MPI_DOUBLE, &col_t);
MPI_Type_commit(&col_t);
MPI_Send(&matrix[0][col], 1, col_t, dest, 0, MPI_COMM_WORLD);
MPI_Type_free(&col_t);

MPI_Type_create_struct describes a struct with mixed types, matching a C struct layout. Every derived type must be committed with MPI_Type_commit before use and freed with MPI_Type_free when done.