!> @file cpulog_mod.f90 !--------------------------------------------------------------------------------------------------! ! This file is part of the PALM model system. ! ! PALM is free software: you can redistribute it and/or modify it under the terms of the GNU General ! Public License as published by the Free Software Foundation, either version 3 of the License, or ! (at your option) any later version. ! ! PALM is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the ! implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General ! Public License for more details. ! ! You should have received a copy of the GNU General Public License along with PALM. If not, see ! . ! ! Copyright 1997-2020 Leibniz Universitaet Hannover !------------------------------------------------------------------------------! ! ! Current revisions: ! ----------------- ! ! ! Former revisions: ! ----------------- ! $Id: cpulog_mod.f90 4665 2020-09-03 14:04:24Z schwenkel $ ! Format descriptor 102 slightly modified. ! ! 4577 2020-06-25 09:53:58Z raasch ! further re-formatting concerning Fortran parameter variables ! ! 4559 2020-06-11 08:51:48Z raasch ! file re-formatted to follow the PALM coding standard ! ! 4549 2020-05-29 09:27:29Z raasch ! bugfix for r4539: values for min/max/rms stored in separate arrays ! ! 4539 2020-05-18 14:05:17Z raasch ! code re-structured, ! cpu time per grid point and timestep does not included initialization and spinup any more ! ! 4536 2020-05-17 17:24:13Z raasch ! restart I/O transfer speed added ! ! 4429 2020-02-27 15:24:30Z raasch ! bugfix: cpp-directives added for serial mode ! ! 4378 2020-01-16 13:22:48Z Giersch ! Format of rms output changed to allow values >= 100 ! ! 4360 2020-01-07 11:25:50Z suehring ! Corrected "Former revisions" section ! ! 4015 2019-06-05 13:25:35Z raasch ! all reals changed to double precision in order to work with 32-bit working precision, otherwise ! calculated time intervals would mostly give zero ! ! 3885 2019-04-11 11:29:34Z kanani ! Changes related to global restructuring of location messages and introduction of additional debug ! messages ! ! 3655 2019-01-07 16:51:22Z knoop ! output format limited to a maximum line length of 80 ! ! Revision 1.1 1997/07/24 11:12:29 raasch ! Initial revision ! ! ! Description: ! ------------ !> CPU-time measurements for any program part whatever. Results of the measurements are output at !> the end of the run in local file CPU_MEASURES. !> !> To measure the CPU-time (better to say the wallclock time) of a specific code segment, two calls !> of cpu_log have to be used as brackets in front and at the end of the segment: !> !> CALL cpu_log( log_point(n), 'any identifier', 'start' ) !> ... code segment ... !> CALL cpu_log( log_point(n), 'any identifier', 'stop' ) !> !> Parts of the code segment can be excluded from the measurement by additional call of cpu_log: !> !> ... first segment to be measured !> CALL cpu_log( log_point(n), 'any identifier', 'pause' ) !> ... oart of segment to be excluded from measurement !> CALL cpu_log( log_point(n), 'any identifier', 'continue' ) !> ... second segment to be mesasured !> !> n is an INTEGER within the interval [1,100] defining the id of the specific code segment, !> 'any identifier' is a string describing the code segment to be measured. It can be freely chosen !> and results will appear under this name in file CPU_MEASURES. ids can only be used once. If you !> like to do a measurement of a new part of the code, please look for an id which is unused so far. !> !> runtime_parameters-parameter cpu_log_barrierwait can be used to set an MPI barrier at the !> beginning of the measurement (modus 'start' or 'continue'), to avoid that idle times (due to MPI !> calls in the code segment, which are waiting for other processes to be finished) affect the !> measurements. !> If barriers shall not be used at all, a fourth, optional parameter has to be !> given: !> !> CALL cpu_log( ..., ..., 'start', cpu_log_nowait ) !> !> Variable log_point should be used for non-overlapping code segments, and they should sum up to !> the total cpu-time required by the complete run. !> Variable log_point_s can be used for any other special (s) measurements. !--------------------------------------------------------------------------------------------------! MODULE cpulog USE control_parameters, & ONLY: message_string, nr_timesteps_this_run, restart_data_format_output, & restart_file_size, run_description_header, synchronous_exchange, write_binary USE indices, & ONLY: ngp_3d, nx, ny, nz USE kinds USE pegrid IMPLICIT NONE PRIVATE PUBLIC cpu_log, cpu_log_barrierwait, cpu_log_nowait, cpu_statistics, initial_wallclock_time, & log_point, log_point_s INTERFACE cpu_log MODULE PROCEDURE cpu_log END INTERFACE cpu_log INTERFACE cpu_statistics MODULE PROCEDURE cpu_statistics END INTERFACE cpu_statistics INTEGER(iwp), PARAMETER :: cpu_log_continue = 0 !< INTEGER(iwp), PARAMETER :: cpu_log_pause = 1 !< INTEGER(iwp), PARAMETER :: cpu_log_start = 2 !< INTEGER(iwp), PARAMETER :: cpu_log_stop = 3 !< LOGICAL, PARAMETER :: cpu_log_nowait = .FALSE. !< LOGICAL :: cpu_log_barrierwait = .FALSE. !< REAL(dp) :: initial_wallclock_time !< TYPE logpoint REAL(dp) :: isum !< REAL(dp) :: ivect !< REAL(dp) :: mean !< REAL(dp) :: mtime !< REAL(dp) :: mtimevec !< REAL(dp) :: sum !< REAL(dp) :: vector !< INTEGER(iwp) :: counts !< CHARACTER (LEN=25) :: place !< END TYPE logpoint TYPE(logpoint), DIMENSION(100) :: log_point = logpoint( 0.0_dp, 0.0_dp, 0.0_dp, 0.0_dp, & 0.0_dp, 0.0_dp, 0.0_dp, 0, ' ' ) TYPE(logpoint), DIMENSION(100) :: log_point_s = logpoint( 0.0_dp, 0.0_dp, 0.0_dp, 0.0_dp, & 0.0_dp, 0.0_dp, 0.0_dp, 0, ' ' ) SAVE CONTAINS !------------------------------------------------------------------------------! ! Description: ! ------------ !> @todo Missing subroutine description. !------------------------------------------------------------------------------! SUBROUTINE cpu_log( log_event, place, modus, barrierwait ) IMPLICIT NONE CHARACTER (LEN=*) :: modus !< CHARACTER (LEN=*) :: place !< LOGICAL :: wait_allowed !< LOGICAL, OPTIONAL :: barrierwait !< LOGICAL, SAVE :: first = .TRUE. !< REAL(dp) :: mtime = 0.0_dp !< REAL(dp) :: mtimevec = 0.0_dp !< TYPE(logpoint) :: log_event !< INTEGER(idp) :: count !< INTEGER(idp) :: count_rate !< ! !-- Initialize and check, respectively, point of measurement IF ( log_event%place == ' ' ) THEN log_event%place = place ELSEIF ( log_event%place /= place ) THEN WRITE( message_string, * ) 'wrong argument expected: ', & TRIM(log_event%place), ' given: ', TRIM( place ) CALL message( 'cpu_log', 'PA0174', 1, 2, 0, 6, 0 ) ENDIF ! !-- Determine, if barriers are allowed to set IF ( PRESENT( barrierwait ) ) THEN wait_allowed = barrierwait ELSE wait_allowed = .TRUE. ENDIF ! !-- MPI barrier, if requested, in order to avoid measuring wait times !-- caused by MPI routines waiting for other MPI routines of other !-- PEs that have not yet finished #if defined( __parallel ) IF ( cpu_log_barrierwait .AND. wait_allowed .AND. & ( modus == 'start' .OR. modus == 'continue' ) ) THEN CALL MPI_BARRIER( comm2d, ierr ) ENDIF #endif ! !-- Take current time CALL SYSTEM_CLOCK( count, count_rate ) mtime = REAL( count, KIND=dp ) / REAL( count_rate, KIND=dp ) ! !-- Start, stop or pause measurement IF ( modus == 'start' .OR. modus == 'continue' ) THEN log_event%mtime = mtime log_event%mtimevec = mtimevec ELSEIF ( modus == 'pause' ) THEN IF ( ( mtime - log_event%mtime ) < 0.0 .AND. first ) THEN WRITE( message_string, * ) 'negative time interval occured', & '&PE',myid,' L=PAUSE "',TRIM(log_event%place), & '" new=', mtime,' last=',log_event%mtime CALL message( 'cpu_log', 'PA0176', 0, 1, -1, 6, 0 ) first = .FALSE. ENDIF log_event%isum = log_event%isum + mtime - log_event%mtime log_event%ivect = log_event%ivect + mtimevec - log_event%mtimevec ELSEIF ( modus == 'stop' ) THEN IF ( ( mtime - log_event%mtime + log_event%isum ) < 0.0 .AND. first ) THEN WRITE( message_string, * ) 'negative time interval occured', & '&PE',myid,' L=STOP "',TRIM(log_event%place),'" new=', & mtime,' last=',log_event%mtime,' isum=',log_event%isum CALL message( 'cpu_log', 'PA0177', 0, 1, -1, 6, 0 ) first = .FALSE. ENDIF log_event%mtime = mtime - log_event%mtime + log_event%isum log_event%mtimevec = mtimevec - log_event%mtimevec + log_event%ivect log_event%sum = log_event%sum + log_event%mtime IF ( log_event%sum < 0.0 .AND. first ) THEN WRITE( message_string, * ) 'negative time interval occured', & '&PE',myid,' L=STOP "',TRIM(log_event%place),'" sum=', & log_event%sum,' mtime=',log_event%mtime CALL message( 'cpu_log', 'PA0178', 0, 1, -1, 6, 0 ) first = .FALSE. ENDIF log_event%vector = log_event%vector + log_event%mtimevec log_event%counts = log_event%counts + 1 log_event%isum = 0.0_dp log_event%ivect = 0.0_dp ELSE message_string = 'unknown modus of time measurement: ' // TRIM( modus ) CALL message( 'cpu_log', 'PA0179', 0, 1, -1, 6, 0 ) ENDIF END SUBROUTINE cpu_log !------------------------------------------------------------------------------! ! 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. !------------------------------------------------------------------------------! SUBROUTINE cpu_statistics IMPLICIT NONE INTEGER(iwp) :: i !< INTEGER(iwp) :: ii(1) !< #if defined( __parallel ) INTEGER(iwp) :: iii !< INTEGER(iwp) :: sender !< #endif REAL(dp) :: average_cputime !< REAL(dp), SAVE :: norm = 1.0_dp !< REAL(dp), DIMENSION(:), ALLOCATABLE :: pe_max !< REAL(dp), DIMENSION(:), ALLOCATABLE :: pe_max_s !< REAL(dp), DIMENSION(:), ALLOCATABLE :: pe_min !< REAL(dp), DIMENSION(:), ALLOCATABLE :: pe_min_s !< REAL(dp), DIMENSION(:), ALLOCATABLE :: pe_rms !< REAL(dp), DIMENSION(:), ALLOCATABLE :: pe_rms_s !< REAL(dp), DIMENSION(:), ALLOCATABLE :: pe_tmp !< REAL(dp), DIMENSION(:), ALLOCATABLE :: pe_tmp_s !< REAL(dp), DIMENSION(:), ALLOCATABLE :: sum !< REAL(dp), DIMENSION(:), ALLOCATABLE :: sum_s !< REAL(dp), DIMENSION(:,:), ALLOCATABLE :: pe_log_points !< REAL(dp), DIMENSION(:,:), ALLOCATABLE :: pe_log_points_s !< CALL location_message( 'calculating cpu statistics', 'start' ) ! !-- Compute CPU-times in seconds for the global non-overlapping measurements. 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_tmp( 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_dp pe_tmp = 0.0_dp #if defined( __parallel ) ! !-- Receive data from all PEs DO i = 1, numprocs-1 CALL MPI_RECV( pe_tmp(1), SIZE( log_point ), MPI_DOUBLE_PRECISION, i, i, comm2d, & status, ierr ) sender = status(MPI_SOURCE) pe_log_points(:,sender) = pe_tmp 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 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_DOUBLE_PRECISION, 0, myid, comm2d, ierr ) #endif ENDIF ! !-- The same procedure again for the special 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 ! !-- Allocate and initialize temporary arrays needed for statistics ALLOCATE( pe_max_s( SIZE( log_point_s ) ), pe_min_s( SIZE( log_point_s ) ), & pe_rms_s( SIZE( log_point_s ) ), pe_tmp_s( SIZE( log_point_s ) ), & pe_log_points_s( SIZE( log_point_s ), 0:numprocs-1 ) ) pe_min_s = log_point_s%sum pe_max_s = log_point_s%sum ! need to be set in case of 1 PE pe_rms_s = 0.0_dp #if defined( __parallel ) ! !-- Receive data from all PEs DO i = 1, numprocs-1 CALL MPI_RECV( pe_tmp_s(1), SIZE( log_point_s ), MPI_DOUBLE_PRECISION, & MPI_ANY_SOURCE, MPI_ANY_TAG, comm2d, status, ierr ) sender = status(MPI_SOURCE) pe_log_points_s(:,sender) = pe_tmp_s ENDDO pe_log_points_s(:,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_s ) DO i = 1, numprocs-1 log_point_s(iii)%sum = log_point_s(iii)%sum + pe_log_points_s(iii,i) pe_min_s(iii) = MIN( pe_min_s(iii), pe_log_points_s(iii,i) ) pe_max_s(iii) = MAX( pe_max_s(iii), pe_log_points_s(iii,i) ) ENDDO log_point_s(iii)%sum = log_point_s(iii)%sum / numprocs ! !-- Calculate rms DO i = 0, numprocs-1 pe_rms_s(iii) = pe_rms_s(iii) + ( pe_log_points_s(iii,i) - log_point_s(iii)%sum )**2 ENDDO pe_rms_s(iii) = SQRT( pe_rms_s(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_s( SIZE( log_point_s ) ) ) pe_max_s = log_point_s%sum CALL MPI_SEND( pe_max_s(1), SIZE( log_point_s ), MPI_DOUBLE_PRECISION, 0, 0, comm2d, & ierr ) #endif ENDIF ! !-- Write CPU-times for the non-overlapping measurements. 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_dp ENDWHERE ! !-- Get total time in order to calculate CPU-time per gridpoint and timestep. !-- Time for initialization (2) and spinup (41) are not included here! IF ( nr_timesteps_this_run /= 0 ) THEN average_cputime = log_point_s(10)%sum / REAL( ngp_3d(0), KIND=dp ) / & REAL( nr_timesteps_this_run, KIND=dp ) * 1E6_dp ! in micro-sec ELSE average_cputime = -1.0_dp ENDIF ! !-- Write cpu-times sorted by size CALL check_open( 18 ) #if defined( __parallel ) WRITE ( 18, 100 ) TRIM( run_description_header ), numprocs * threads_per_task, & pdims(1), pdims(2), threads_per_task, nx+1, ny+1, nz, & nr_timesteps_this_run, average_cputime WRITE ( 18, 110 ) #else WRITE ( 18, 100 ) TRIM( run_description_header ), numprocs * threads_per_task, 1, 1, & threads_per_task, nx+1, ny+1, nz, nr_timesteps_this_run, & average_cputime WRITE ( 18, 110 ) #endif DO ii = MAXLOC( sum ) i = ii(1) IF ( sum(i) /= -1.0_dp ) THEN WRITE ( 18, 102 ) log_point(i)%place, log_point(i)%sum, & log_point(i)%sum / log_point(1)%sum * 100.0_dp, & log_point(i)%counts, pe_min(i), pe_max(i), pe_rms(i) sum(i) = -1.0_dp ELSE EXIT ENDIF ENDDO ENDIF ! !-- Write CPU-times for special measurements. IF ( myid == 0 ) THEN ! !-- Re-store sums ALLOCATE( sum_s( SIZE( log_point_s ) ) ) WHERE ( log_point_s%counts /= 0 ) sum_s = log_point_s%sum ELSEWHERE sum_s = -1.0_dp ENDWHERE ! !-- Write cpu-times sorted by size WRITE ( 18, 101 ) DO ii = MAXLOC( sum_s ) i = ii(1) IF ( sum_s(i) /= -1.0_dp ) 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_dp, & log_point_s(i)%counts, pe_min_s(i), pe_max_s(i), pe_rms_s(i) sum_s(i) = -1.0_dp ELSE EXIT ENDIF ENDDO ! !-- Output of handling of MPI operations IF ( collective_wait ) THEN WRITE ( 18, 103 ) ELSE WRITE ( 18, 104 ) ENDIF IF ( cpu_log_barrierwait ) WRITE ( 18, 111 ) IF ( synchronous_exchange ) THEN WRITE ( 18, 105 ) ELSE WRITE ( 18, 106 ) ENDIF ! !-- Output of restart data transfer speed IF ( write_binary .AND. restart_data_format_output(1:3) == 'mpi' ) THEN WRITE ( 18, 107 ) restart_file_size, restart_file_size / log_point(22)%sum ENDIF ! !-- Empty lines in order to create a gap to the results from the next restart run WRITE ( 18, 108 ) ! !-- Unit 18 is not needed anymore CALL close_file( 18 ) ENDIF CALL location_message( 'calculating cpu statistics', 'finished' ) 100 FORMAT (A/11('-')//'CPU measures for ',I5,' PEs (',I5,'(x) * ',I5,'(y', & &') tasks *',I5,' threads):'// & 'gridpoints (x/y/z): ',20X,I5,' * ',I5,' * ',I5/ & 'nr of timesteps: ',22X,I6/ & 'cpu time per grid point and timestep: ',5X,F8.5,' * 10**-6 s') 101 FORMAT (/'special measures:'/ & &'-----------------------------------------------------------', & &'---------------------') 102 FORMAT (A25,2X,F10.3,2X,F7.2,1X,I7,2(1X,F10.3),1X,F6.2) 103 FORMAT (/'Barriers are set in front of collective operations') 104 FORMAT (/'No barriers are set in front of collective operations') 105 FORMAT (/'Exchange of ghostpoints via MPI_SENDRCV') 106 FORMAT (/'Exchange of ghostpoints via MPI_ISEND/MPI_IRECV') 107 FORMAT (/'Restart file size: ',F12.1,' MByte'/ & &'I/O transfer speed: ',F12.1,' MByte / sec') 108 FORMAT (//) 110 FORMAT ('------------------------------------------------------------', & &'----------'// & &'place: mean counts ', & &' min max rms'/ & &' sec. % ', & &'sec. sec. sec.'/ & &'-----------------------------------------------------------', & &'---------------------') 111 FORMAT (/'Barriers are set at beginning (start/continue) of measurements') END SUBROUTINE cpu_statistics END MODULE cpulog