the use of MPI_Init()

Question

I encountered a question about the use of MPI_Init().

I want to initialize random number "randv" only on the root processor with the code in the context below. To see if my goal is fulfilled, I have the program print out the array "randv" by placing a do loop immediately after the line "call RANDOM_NUMBER(randv)."

However, what is shown on the outcome screen is the repetition of the random number array by 8 times (given the number of processors is 8). My question is why the processors other than the root one are initialized before call MPI_Init(). If all the processors are awaken and have the same random number array before evoking MPI_Init, why bother to place call MPI_Init() for initialization? Thanks.

Lee

Here is the example I use:

program main
  include 'mpif.h'
  integer :: i
  integer :: ierr
  integer :: irank
  integer :: nrow, ncol
  real, dimension(:,:), allocatable :: randv
  nrow = 4
  ncol = 2
  allocate(randv(nrow,ncol))   
  call RANDOM_SEED
  call RANDOM_NUMBER(randv)
  do i = 1, nrow
    write(*,'(2(f5.2,x))') randv(i,:)    
  enddo   
  call MPI_Init ( ierr )
  allocate(row_list(ncol), col_list(nrow))  
  call MPI_Comm_rank ( MPI_COMM_WORLD, irank, ierr )
  if( irank == 0 )then
    do i = 1, nrow
      write(*,'(2(f5.2,x))') randv(i,:)    
    enddo          
  endif
  call MPI_Finalize ( ierr ) 
  deallocate( randv )
end program

Answer 1

I think you misunderstand how MPI works. The program you wrote is executed by every process. MPI_Init initializes the MPI environment s.t. those processes can interact. After initialization every process is uniquely identified by its rank. You have to make sure that, based on these ranks, each process works on different portions of your data, or performs different tasks.

Typically, you should run MPI_Init before anything else in your program.

Using MPI_Comm_rank you can obtain the ID of the current process (its rank). The first process always has the rank 0. Therefore, if you want to run parts of the code on the "master" process only, you can test for irank == 0:

program main
  include 'mpif.h'
  integer :: i
  integer :: ierr
  integer :: irank
  integer :: nrow, ncol
  real, dimension(:,:), allocatable :: randv

  ! Initialize MPI
  call MPI_Init ( ierr )
  ! Get process ID
  call MPI_Comm_rank ( MPI_COMM_WORLD, irank, ierr )

  ! Executed on all processes
  nrow = 4
  ncol = 2
  allocate(randv(nrow,ncol))   

  ! Only exectued on the master process
  if ( irank == 0 ) then
    call RANDOM_SEED
    call RANDOM_NUMBER(randv)
    do i = 1, nrow
      write(*,'(2(f5.2,x))') randv(i,:)    
    enddo   
  endif

  ! Executed on all threads
  allocate(row_list(ncol), col_list(nrow))  

  ! Only exectued on the master process
  if ( irank == 0 ) then
    do i = 1, nrow
      write(*,'(2(f5.2,x))') randv(i,:)    
    enddo          
  endif

  deallocate( randv )

  ! Finalize MPI, should always be executed last
  call MPI_Finalize ( ierr ) 
end program