[mpich-devel] mpich3 error

[Raffenetti, Kenneth J.]

On 1/14/21, 11:02 PM, "Brent Morgan via devel" <devel at mpich.org> wrote:

    Hi Hui Zhou,
    Robert and I managed to incorporate multiple communicators- we use MPI_Gather() of Send/Receives.  However the issue remains- for a small (N<50) # of threads, the calculations work and seem fine.  For large (N>=50) # of threads, the issue persists.  We will try compiling for ch4 tonight...  But I wonder if we're doing something wrong?  Should we do Gather of Gathers?   
    
    Our MPI implementation is merely finding the sum of the results of the N processes, where N is large.  Is MPI_Reduce going to be faster?  

It should *never be slower* to use MPI_Reduce, and it is likely to be faster on most machines you run it on.

Ken