[mpich-discuss] Better alternative to MPI_Allreduce() and avoiding deadlock with MPI_Neighbor_alltoallw().