An MPI_BARRIER added in cdnval.F90 to sinchronize RMA read.

cdn/cdnval.F90

@@ -540,8 +540,10 @@ CONTAINS
                noccbd,n_start,n_end,&
                lapw%nmat,lapw%nv,ello,evdu,epar,kveclo,&
                lapw%k1,lapw%k2,lapw%k3,bkpt,wk,nbands,eig,zMat)
-
-
+#ifdef CPP_MPI
+          ! Sinchronizes the RMA operations
+          if (l_evp) CALL MPI_BARRIER(mpi%mpi_comm,ie)
+#endif
           !IF (l_evp.AND.(isize.GT.1)) THEN
           !  eig(1:noccbd) = eig(n_start:n_end)
           !ENDIF

io/cdn_read.F

@@ -158,17 +158,6 @@ c
      <                  zmat=zmat)
 
       ENDIF
-#ifdef CPP_MPI
-!     The barrier here is not algorithmically required.
-!     Nevertheless it improves the performance, presumably because it
-!     synchronizes the processes and prevent the latency accumulating.
-!     U.Alekseeva 2015
-
-!Note: this does not work, two problems:
-! a) MPI_COMM_WORLD should not be used 
-! b) Not all processes might reach this point due to k-point parallelization
-!      call MPI_BARRIER(MPI_COMM_WORLD,mpiierr) 
-#endif
       CALL timestop("cdn_read")
       
 !      IF (nbands>neigd)   CALL juDFT_error("nbands.GT.neigd",calledby