[mpich-discuss] MPI_REDUCE and complex numbers

[Jeff Hammond]

I wrote a simple program (see below) to test this but i do not see any
error when I run on BGP.

Brian - Can you try this program on your machine and let me know if it
does or does not produce the error?

Thanks,

Jeff

PS Every time I write Fortran, I die a little on the inside.

=====================================
program main
use mpi

    integer rc
    integer rank, worldsize

    integer i, n
    complex*16 x(1000000)
    complex*16 y(1000000)
    complex*16 z(1000000)

    call mpi_init(rc)
    call mpi_comm_rank(MPI_COMM_WORLD,rank,rc)
    call mpi_comm_size(MPI_COMM_WORLD,worldsize,rc)

    n = 1000000

    do i = 1, n
       x(i) = (1.0,1.0)
    end do

    do i = 1, n
       y(i) = (0.0,0.0)
    end do

    call mpi_reduce(x,y,n,MPI_DOUBLE_COMPLEX,MPI_SUM,0,MPI_COMM_WORLD,rc)

    call mpi_allreduce(x,z,n,MPI_DOUBLE_COMPLEX,MPI_SUM,MPI_COMM_WORLD,rc)

    if (rank .eq. 0) then
        print*,'i, x(i), y(i), z(i)'
        do i = 1, n
           print*, i, x(i), y(i), z(i)
        end do
        print*,'the right answer is ',worldsize*(1.0,1.0)
    end if

    call mpi_barrier(MPI_COMM_WORLD,rc)
    call mpi_finalize()

end program main
=====================================

On Fri, Aug 19, 2011 at 10:02 AM, Darius Buntinas <buntinas at mcs.anl.gov> wrote:
>
> There isn't a max buffer size (aside from the running out of memory, but you'd get an error for that).  Can you send us a short sample program that demonstrates this?  How many  processes are you using?
>
> Thanks,
> -d
>
> On Aug 18, 2011, at 11:04 PM, Brian Dushaw wrote:
>
>> I recently ran across what may be a bug in MPI_REDUCE using the latest version of mpich2-1.4., (although I can't be sure...)
>>
>> I have fortran 90 code that calculates a complex array of size
>> about 400X6000.  Each row of the array is calculated independently,
>> so I apportion subsets of rows across a small cluster for calculating.
>> On each node, the rows not calculated are just filled with zeros.
>>
>> when all done, I called:
>>
>> call MPI_REDUCE(psif,psiff,itemp,MPI_DOUBLE_COMPLEX,MPI_SUM,0,MPI_COMM_WORLD,ierr)
>>
>> which just adds up all the arrays from each node, giving me the final
>> filled up array complex psiff on the master node (myid=0). itemp is size
>> ncol*nrow of psiff (or about 400*6000).
>>
>> This worked fine for smaller sized arrays, but for the larger size
>> of 400X6000 (which doesn't seem all that large to me), I would just get
>> almost all zeros in the psiff array.
>>
>> I wondered if the array size was so large as to overflow a maximum buffer
>> size for MPI_REDUCE.
>>
>> In any case, I rewrote the above to send back real and imaginary parts
>> separately, calling instead:
>>
>> call MPI_REDUCE(rpsif,rpsiff,itemp,MPI_DOUBLE_PRECISION,MPI_SUM,0,MPI_COMM_WORLD,ierr)
>> call MPI_REDUCE(ipsif,ipsiff,itemp,MPI_DOUBLE_PRECISION,MPI_SUM,0,MPI_COMM_WORLD,ierr)
>>
>> And this seems to work fine.
>>
>> Does anybody have any comments or wise words to enlighten this situation?
>> The code takes quite a while to run, so it is slow to debug.  I suppose I could write a small test case to check more simply whether this is a bug
>> or not.  Does MPI_REDUCE have a maximum buffer size that one should worry about?
>>
>> Thanks -
>>
>> _______________________________________________
>> mpich-discuss mailing list
>> mpich-discuss at mcs.anl.gov
>> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>
> _______________________________________________
> mpich-discuss mailing list
> mpich-discuss at mcs.anl.gov
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>



-- 
Jeff Hammond
Argonne Leadership Computing Facility
University of Chicago Computation Institute
jhammond at alcf.anl.gov / (630) 252-5381
http://www.linkedin.com/in/jeffhammond
https://wiki.alcf.anl.gov/index.php/User:Jhammond