Use MPI_Allreduce values instead of each processor's values
(cherry picked from commit f30232b41ed2ad0db8fdb90155da01dadd9e5574)