source: palm/trunk/SOURCE/cpu_statistics.f90 @ 1

 SUBROUTINE cpu_statistics

!-------------------------------------------------------------------------------!
! Actual revisions:
! -----------------
! 
!
! Former revisions:
! -----------------
! $Log: cpu_statistics.f90,v $
! Revision 1.13  2006/04/26 12:10:51  raasch
! Output of number of threads per task, max = min in case of 1 PE
!
! Revision 1.12  2003/03/16 09:30:22  raasch
! Two underscores (_) are placed in front of all define-strings
!
! Revision 1.11  2002/12/19 14:12:08  raasch
! PE results are collected on PE0 in order to calculate mean cpu statistics.
! Output format changed.
!
! Revision 1.10  2001/03/30 07:01:02  raasch
! Translation of remaining German identifiers (variables, subroutines, etc.)
!
! Revision 1.9  2001/01/22 05:49:49  raasch
! Module test_variables removed
!
! Revision 1.8  2001/01/02 17:25:23  raasch
! Unit 18 closed at the end of the subroutine
!
! Revision 1.7  2000/12/20 10:10:26  letzel
! All comments translated into English.
!
! Revision 1.1  1997/07/24 11:11:11  raasch
! Initial revision
!
!
! Description:
! ------------
! Analysis and output of the cpu-times measured. All PE results are collected
! on PE0 in order to calculate the mean cpu-time over all PEs and other
! statistics. The output is sorted according to the amount of cpu-time consumed
! and output on PE0.
!-------------------------------------------------------------------------------!

    USE cpulog
    USE pegrid
    USE control_parameters

    IMPLICIT NONE

    INTEGER    ::  i, ii(1), iii, lp, sender
    REAL, SAVE ::  norm = 1.0
    REAL, DIMENSION(:),   ALLOCATABLE ::  pe_max, pe_min, pe_rms, sum
    REAL, DIMENSION(:,:), ALLOCATABLE ::  pe_log_points


!#if defined( __hpmuk )  &&  ! defined( __chinook )
!    norm = clock_ticks_per_second()
!#endif

!
!-- Compute cpu-times in seconds
    log_point%mtime  = log_point%mtime  / norm
    log_point%sum    = log_point%sum    / norm
    log_point%vector = log_point%vector / norm
    WHERE ( log_point%counts /= 0 )
       log_point%mean = log_point%sum / log_point%counts
    END WHERE


!
!-- Collect cpu-times from all PEs and calculate statistics
    IF ( myid == 0 )  THEN
!
!--    Allocate and initialize temporary arrays needed for statistics
       ALLOCATE( pe_max( SIZE( log_point ) ), pe_min( SIZE( log_point ) ), &
                 pe_rms( SIZE( log_point ) ),                              &
                 pe_log_points( SIZE( log_point ), 0:numprocs-1 ) )
       pe_min = log_point%sum
       pe_max = log_point%sum    ! need to be set in case of 1 PE
       pe_rms = 0.0

#if defined( __parallel )
!
!--    Receive data from all PEs
       DO  i = 1, numprocs-1
          CALL MPI_RECV( pe_max(1), SIZE( log_point ), MPI_REAL, &
                         MPI_ANY_SOURCE, MPI_ANY_TAG, comm2d, status, ierr )
          sender = status(MPI_SOURCE)
          pe_log_points(:,sender) = pe_max
       ENDDO
       pe_log_points(:,0) = log_point%sum   ! Results from PE0
!
!--    Calculate mean of all PEs, store it on log_point%sum
!--    and find minimum and maximum
       DO  iii = 1, SIZE( log_point )
          DO  i = 1, numprocs-1
             log_point(iii)%sum = log_point(iii)%sum + pe_log_points(iii,i)
             pe_min(iii) = MIN( pe_min(iii), pe_log_points(iii,i) )
             pe_max(iii) = MAX( pe_max(iii), pe_log_points(iii,i) )
          ENDDO
          log_point(iii)%sum = log_point(iii)%sum / numprocs
!
!--       Calculate rms
          DO  i = 0, numprocs-1
!             IF ( log_point(iii)%place == 'run_control' )  THEN
!                PRINT*, 'pe_rms=',pe_rms(iii),' plp=',pe_log_points(iii,i), &
!                        ' lps=',log_point(iii)%sum
!             ENDIF
             pe_rms(iii) = pe_rms(iii) + ( &
                                 pe_log_points(iii,i) - log_point(iii)%sum &
                                         )**2
          ENDDO
          pe_rms(iii) = SQRT( pe_rms(iii) / numprocs )
       ENDDO
    ELSE
!
!--    Send data to PE0 (pe_max is used as temporary storage to send
!--    the data in order to avoid sending the data type log)
       ALLOCATE( pe_max( SIZE( log_point ) ) )
       pe_max = log_point%sum
       CALL MPI_SEND( pe_max(1), SIZE( log_point ), MPI_REAL, 0, 0, comm2d, &
                      ierr )
#endif

    ENDIF

!
!-- Write cpu-times
    IF ( myid == 0 )  THEN
!
!--    Re-store sums
       ALLOCATE( sum( SIZE( log_point ) ) )
       WHERE ( log_point%counts /= 0 )
          sum = log_point%sum
       ELSEWHERE
          sum = -1.0
       ENDWHERE

!
!--    Write cpu-times sorted by size
       CALL check_open( 18 )
       WRITE ( 18, 100 )  TRIM( run_description_header ),        &
                          numprocs * threads_per_task, numprocs, &
                          threads_per_task
       DO
          ii = MAXLOC( sum )
          i = ii(1)
          IF ( sum(i) /= -1.0 )  THEN
             WRITE ( 18, 102 ) &
                log_point(i)%place, log_point(i)%sum,                &
                log_point(i)%sum / log_point(1)%sum * 100.0,         &
                log_point(i)%counts, pe_min(i), pe_max(i), pe_rms(i)
             sum(i) = -1.0
          ELSE
             EXIT 
          ENDIF
       ENDDO
    ENDIF


!
!-- The same procedure again for the individual measurements.
!
!-- Compute cpu-times in seconds
    log_point_s%mtime  = log_point_s%mtime  / norm
    log_point_s%sum    = log_point_s%sum    / norm
    log_point_s%vector = log_point_s%vector / norm
    WHERE ( log_point_s%counts /= 0 )
       log_point_s%mean = log_point_s%sum / log_point_s%counts
    END WHERE

!
!-- Collect cpu-times from all PEs and calculate statistics
#if defined( __parallel )
!
!-- Set barrier in order to avoid that PE0 receives log_point_s-data
!-- while still busy with receiving log_point-data (see above)
    CALL MPI_BARRIER( comm2d, ierr )    
#endif
    IF ( myid == 0 )  THEN
!
!--    Initialize temporary arrays needed for statistics
       pe_min = log_point_s%sum
       pe_max = log_point_s%sum    ! need to be set in case of 1 PE
       pe_rms = 0.0

#if defined( __parallel )
!
!--    Receive data from all PEs
       DO  i = 1, numprocs-1
          CALL MPI_RECV( pe_max(1), SIZE( log_point ), MPI_REAL, &
                         MPI_ANY_SOURCE, MPI_ANY_TAG, comm2d, status, ierr )
          sender = status(MPI_SOURCE)
          pe_log_points(:,sender) = pe_max
       ENDDO
       pe_log_points(:,0) = log_point_s%sum   ! Results from PE0
!
!--    Calculate mean of all PEs, store it on log_point_s%sum
!--    and find minimum and maximum
       DO  iii = 1, SIZE( log_point )
          DO  i = 1, numprocs-1
             log_point_s(iii)%sum = log_point_s(iii)%sum + pe_log_points(iii,i)
             pe_min(iii) = MIN( pe_min(iii), pe_log_points(iii,i) )
             pe_max(iii) = MAX( pe_max(iii), pe_log_points(iii,i) )
          ENDDO
          log_point_s(iii)%sum = log_point_s(iii)%sum / numprocs
!
!--       Calculate rms
          DO  i = 0, numprocs-1
             pe_rms(iii) = pe_rms(iii) + ( &
                                 pe_log_points(iii,i) - log_point_s(iii)%sum &
                                         )**2
          ENDDO
          pe_rms(iii) = SQRT( pe_rms(iii) / numprocs )
       ENDDO
    ELSE
!
!--    Send data to PE0 (pe_max is used as temporary storage to send
!--    the data in order to avoid sending the data type log)
       pe_max = log_point_s%sum
       CALL MPI_SEND( pe_max(1), SIZE( log_point ), MPI_REAL, 0, 0, comm2d, &
                      ierr )
#endif

    ENDIF

!
!-- Write cpu-times
    IF ( myid == 0 )  THEN
!
!--    Re-store sums
       WHERE ( log_point_s%counts /= 0 )
          sum = log_point_s%sum
       ELSEWHERE
          sum = -1.0
       ENDWHERE

!
!--    Write cpu-times sorted by size
       WRITE ( 18, 101 )
       DO
          ii = MAXLOC( sum )
          i = ii(1)
          IF ( sum(i) /= -1.0 )  THEN
             WRITE ( 18, 102 ) &
                log_point_s(i)%place, log_point_s(i)%sum, &
                log_point_s(i)%sum / log_point(1)%sum * 100.0, &
                log_point_s(i)%counts, pe_min(i), pe_max(i), pe_rms(i)
             sum(i) = -1.0
          ELSE
             EXIT 
          ENDIF
       ENDDO

!
!--    Empty lines in order to create a gap to the results of the model 
!--    continuation runs
       WRITE ( 18, 103 )

!
!--    Unit 18 is not needed anymore
       CALL close_file( 18 )

    ENDIF


100 FORMAT (A/11('-')//'CPU measures for ',I3,' PEs (',I3,' tasks *',I3,    &
            &' threads):'/ &
             &'--------------------------------------------------'//        &
            &'place:                        mean        counts      min  ', &
             &'     max       rms'/ &
            &'                           sec.      %                sec. ', &
             &'     sec.      sec.'/  &
            &'-----------------------------------------------------------', &
             &'-------------------')

101 FORMAT (/'special measures:'/ &
            &'-----------------------------------------------------------', &
            &'--------------------')

102 FORMAT (A20,2X,F9.3,2X,F7.2,1X,I7,3(1X,F9.3))
103 FORMAT (//)

 END SUBROUTINE cpu_statistics