MPI_Allgather can be thought of as an MPI_Gather where all processes, not just the root, receive the result. The jth block of the receive buffer is reserved for data sent from the jth rank; all the blocks are the same size. MPI_Allgatherv works similarly, except the block size can depend on rank j, in direct analogy to MPI_Gatherv. It's not surprising, then, that the syntax of MPI_Allgather and MPI_Allgatherv turns out to be close to MPI_Gather and MPI_Gatherv, respectively. The main difference is that the argument root is dropped from the argument lists because there is no root process for the "all" type commands.

Allgather[v] syntax

MPI_ALLGATHER(sbuf, scount, stype, rbuf, rcount, 
    rtype, comm, ierr)

MPI_ALLGATHERV(sbuf, scounts, stype, rbuf, rcounts, displs,
     rtype, comm, ierr)

Allgather[v] parameters:

sbuf

is the starting address of the send buffer

scount[s]

is the [array of the] number of elements to send to each process

stype

is the data type of send buffer elements

rbuf

is the address of the receive buffer

rcount[s]

is [an array containing] the number of elements to be received from each process

[displs]

is an array specifying the displacement of data relative to rbuf

rtype

is the data type of the receive buffer elements

comm

is the group communicator

Example: matrix-vector multiplication using MPI_Allgather

• This example is similar to previous ones, but it uses 7 MPI processes.
• The matrix is distributed by rows, so the output vector will be too.
• The output vector is needed in its entirety by ALL processes.

MPI_Allgather Example Code

/* This allgather example assumes 7 total ranks */
float apart[25,100], b[100], cpart[25], ctotal[175];
int i, k;
⋮
for(i=0; i<25; i++)
{
   cpart[i]=0;
   for(k=0; k<100; k++)
   {
      cpart[i]=cpart[i]+apart[i,k]*b[k];
   }
}
MPI_Allgather(cpart, 25, MPI_FLOAT, ctotal, 25, MPI_FLOAT, 
    MPI_COMM_WORLD);

! This allgather example assumes 7 total ranks 
REAL apart(25,100), b(100), cpart(25), ctotal(175)
INTEGER I, K
⋮
! compiler will likely reorder these loops due to Fortran storage order
DO I=1,25
    cpart(I) = 0.
    DO K=1,100
        cpart(I) = cpart(I) + apart(I,K) * b(K)
    END DO
END DO
CALL MPI_ALLGATHER(cpart, 25, MPI_REAL, ctotal, 25, MPI_REAL, &
    MPI_COMM_WORLD, ierr)