FWIW, this function is described as "Cartesian Convenience Function" and "helps the user select a balanced distribution of processes per coordinate direction", which means that it really isn't required if the user knows math. Obviously, your case is trivially solved because you already know sqrt(nnodes), but I'm sure you are not using it this way in the real application. However, there is nothing stopping you from writing a superior prime factorization routine than MPICH, or whatever is necessary to get behavior you want. Given the number of bugs in this function I've seen reported over the years, I am inclined to ask the MPI Forum to deprecate it and encourage users to do the math themselves, since it seems more often than not that they can do better. Best, Jeff On Tue, Mar 31, 2015 at 12:10 AM, Valery <[email protected]> wrote:
Dear All
I noticed that MPI_DIMS_CREATE cannot split eg 361 into a 19 * 19 processor grid. Is that a feature?
The code follows.
I used MPICH-3.1.4 gcc-4.9.2
valery
cat mpi_dims.f90 program test use mpi implicit none integer :: ierr, nnodes, ndims, dims(2), i call MPI_INIT( ierr ) ndims = 2 do i = 1, 200 nnodes = i**2 dims(:) = 0 call MPI_DIMS_CREATE( nnodes, ndims, dims, ierr ) if( dims(1) /= i ) write(*,*) i, dims enddo call MPI_FINALIZE(ierr) end program test
mpif90 mpi_dims.f90
mpiexec -n 1 ./a.out 19 361 1 41 1681 1 43 1849 1 71 5041 1 73 5329 1 79 6241 1 83 6889 1 89 7921 1 137 18769 1 139 19321 1 149 22201 1 151 22801 1 157 24649 1 163 26569 1 167 27889 1 173 29929 1 179 32041 1 181 32761 1
_______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss
-- Jeff Hammond [email protected] http://jeffhammond.github.io/ _______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss