MPI_BCAST() applied on only a fraction of the base group - parallel-processing

Question

I grouped 8 processors into two groups, each of which contains evenly four processors. I ask the root of each subgroup to do some communication with their subordinates using the subroutine "MPI_BCAST."
I came across a question: to indicate the root of a subgroup, should I use the original rank which the subgroup root corresponds to with the MPI_COMM_WORLD communicator, or the new rank it represents with the new communicator?
Take the code snippet below for example, I want to require P:0 to send data to its subordinates P:1, P:2, and P:3, and similarly, I ask P:4 to send out its data to P:5, P:6, P:7. To reach this goal, I am wondering if I should specify the fourth argument in line 36 as 1, or specify them as 0 and 4 respectively conditional on which head of subgroup I am referring to?
Thanks.
Lee
1 program main
2 include 'mpif.h'
3 integer :: ierr, irank, num_procs, base_group
4 integer :: nrow, ncol, irow, icol
5 integer :: dummy_group, dummy_comm, new_comm, new_rank
6 integer :: i, j, roster(4), data(4)
7
8 call MPI_Init ( ierr )
9 call MPI_COMM_RANK( MPI_comm_world, irank, ierr )
10 call MPI_COMM_SIZE( MPI_comm_world, num_procs, ierr)
11 call MPI_COMM_GROUP( MPI_comm_world, base_group, ierr)
12 nrow = 4
13 ncol = 2
14 irow = mod( irank, nrow ) + 1
15 icol = irank/nrow + 1
16
17 roster(1) = 0
18 do i = 2, nrow
19 roster(i) = roster(i-1) + 1
20 enddo
21
22 do i = 1, ncol
23 call MPI_GROUP_INCL( base_group, nrow, roster, dummy_group, ierr )
24 call MPI_COMM_CREATE( MPI_COMM_WORLD, dummy_group, dummy_comm, ierr )
25 if( icol == i ) new_comm = dummy_comm
26 forall( j=1:nrow ) roster(j) = roster(j) + nrow
27 enddo
28
29 ! Here I want to initialize data for processors P:0 and P:4
30 if( irank == 0 ) data = 0
31 if( irank == 4 ) data = 4
32
33 ! In the code below I want to require P:0 to send data to
34 ! its subordinates P:1, P:2, and P:3. Similarly, I ask P:4
35 ! to send out its data to P:5, P:6, P:7.
36 call MPI_BCAST( data, 4, MPI_INTEGER, 0, new_comm, ierr)
37
38 call MPI_Finalize ( ierr )
39 end program

Answer 1

All rank-type arguments (origin, target, etc.) in MPI must be ranks in the same communicator as that given by the communicator argument. In practice, what this means is that after creating a new communicator, each process in that communicator must call MPI_Comm_rank and MPI_Comm_size to retrieve it's rank and the total size in that communicator (unless you can deduce the new rank and size by other means in your code, of course).
As an aside, as what you're doing is splitting the original communicator into two disjoint communicators, I think an easier way to accomplish that is to use MPI_Comm_split rather than setting up groups manually as you have done.