!************************************************************************ ! ! ! Module : MPI_Times (Version : 3.4) ! ! ! ! Goal : Measure and print on stdout CPU and Elapsed user times ! ! and the ratio CPU/Elapsed of MPI programs. ! ! ! ! Usage : Insert a "USE MPI_TIMES" instruction inside each MPI ! ! Fortran program unit to instrument, then make calls to ! ! the MPI_TIME subroutine as shown in the example below : ! ! ! ! PROGRAM foo ! ! USE MPI_TIMES ! ! ... ! ! CALL MPI_INIT(ierr) ! ! ! ! !... Set elapsed and CPU user times ! ! CALL MPI_TIME(0) ! ! ... ! ! ... Instruction block to instrument ... ! ! ... ! ! !... Measure and print elapsed and CPU user times ! ! CALL MPI_TIME(1) ! ! ! ! CALL MPI_FINALIZE(ierr) ! ! END PROGRAM foo ! ! ! ! Notes : 1) Standard Fortran 95 compiler has to be used to compile ! ! MPI_TIMES module. ! ! ! ! 2) MPI_TIME subroutine is collective over all processes ! ! of MPI_COMM_WORLD communicator. ! ! ! ! 3) On some machines, default CPU user time may also ! ! include MPI wait times on communication to complete. ! ! ! ! 4) If Te and Tc respectively denotes the elapsed and CPU ! ! user times, then the ratio R=Tc/Te > 0 may lead to ! ! different interpretations depending on R<1 or R=1 ! ! or R>1. ! ! ! ! a) If R<1, then wait time on communications and/or ! ! system load could be the reason of such performance. ! ! ! ! b) If R is close to 1 and no hybrid parallelization ! ! (e.g. MPI + OpenMP) is implemented, then wait time ! ! on communications and/or system load are far to be ! ! considered unless point 3) and then, one can assume ! ! that 99% of the time, processes are busy performing ! ! useful computations on dedicated processors. ! ! ! ! c) If R>1, then process might has been multi-threaded ! ! during execution as what would happen in hybrid ! ! parallelization (e.g. MPI + OpenMP) on cluster ! ! of SMP nodes. In such case, R may reflect the speed ! ! up of the process. ! ! ! ! 5) On IBM SP machine, do not compile MPI_TIMES module ! ! using "-qrealsize=8" switch. This will transform ! ! MPI_WTIME function type from 8 to 16 bytes floating ! ! point precision. ! ! ! ! 6) No special switch is needed to compile this file. ! ! The following should be sufficient on many platforms: ! ! f90 -c -Ipath_to_MPI_header_file MPI_Time.f90 ! ! ! ! Output : At normal termination of the MPI program, process of ! ! rank 0 prints on stdout elapsed time, cpu time and ratio ! ! cpu/elapsed of all MPI_COMM_WORLD processes. ! ! The following is an output example from an execution ! ! with 4 processes: ! ! ! !............ ! ! (C) January 2003, CNRS/IDRIS, FRANCE. ! ! MPI_Time (release 3.4) summary report: ! ! ! ! Process Rank | Elapsed Time (s) | CPU Time (s) | Ratio CPU/Elapsed ! ! -------------|------------------|--------------|------------------ ! ! 0 | 427.098 | 270.393 | 0.633 ! ! 1 | 427.099 | 279.818 | 0.655 ! ! 2 | 427.099 | 276.064 | 0.646 ! ! 3 | 427.182 | 271.001 | 0.634 ! ! -------------|------------------|--------------|------------------ ! ! Total | 1708.477 | 1097.275 | 2.569 ! ! -------------|------------------|--------------|------------------ ! ! Minimum | 427.098 | 270.393 | 0.633 ! ! -------------|------------------|--------------|------------------ ! ! Maximum | 427.182 | 279.818 | 0.655 ! ! -------------|------------------|--------------|------------------ ! ! Average | 427.119 | 274.319 | 0.642 ! ! -------------|------------------|--------------|------------------ ! ! ! ! MPI_Time started on 13/11/2002 at 16:54:59 MET +01:00 from GMT ! ! MPI_Time ended on 13/11/2002 at 17:02:06 MET +01:00 from GMT ! !............ ! ! ! ! Author : Jalel Chergui ; E-mail: Jalel.Chergui@idris.fr ! ! ! ! Permission is garanted to copy and distribute this file or modified ! ! versions of this file for no fee, provided the copyright notice and ! ! this permission notice are preserved on all copies. ! ! ! ! Copyright (C) January 2003, CNRS/IDRIS, FRANCE. ! !************************************************************************ MODULE MPI_TIMES IMPLICIT NONE PRIVATE !... Shared variables INTEGER, PARAMETER :: p = SELECTED_REAL_KIND(12) REAL(kind=p) :: Eoverhead, Coverhead REAL(kind=p), DIMENSION(2) :: Etime, Ctime INTEGER, DIMENSION(8) :: values CHARACTER(LEN=8), DIMENSION(2) :: date CHARACTER(LEN=10), DIMENSION(2) :: time CHARACTER(LEN=5) :: zone PUBLIC :: MPI_Time CONTAINS SUBROUTINE MPI_Time(flag) IMPLICIT NONE !... MPI Header files INCLUDE "mpif.h" !... Input dummy parameter INTEGER, INTENT(IN) :: flag !... Local variables INTEGER :: rank, nb_procs, i, code INTEGER, ALLOCATABLE, DIMENSION(:) :: All_Rank REAL(KIND=p), ALLOCATABLE, DIMENSION(:) :: All_Etime, All_Ctime, All_Ratio REAL(KIND=p) :: Total_Etime,Total_Ctime,Total_Ratio,& Max_Etime, Max_Ctime, Max_Ratio, & Min_Etime, Min_Ctime, Min_Ratio, & Avg_Etime, Avg_Ctime, Avg_Ratio, & dummy CHARACTER(LEN=128), dimension(8) :: lignes CHARACTER(LEN=128) :: hline, start_date, final_date CHARACTER(LEN=2048) :: fmt SELECT CASE(flag) CASE(0) !... Compute clock overhead Eoverhead = MPI_WTIME() Eoverhead = MPI_WTIME() - Eoverhead CALL CPU_TIME(dummy) CALL CPU_TIME(Coverhead) if (dummy < 0.0_p) & PRINT *,"Warning, MPI_TIME: CPU user time is not available on this machine." Coverhead = Coverhead - dummy CALL MPI_COMM_RANK(MPI_COMM_WORLD, rank, code) !... Start of timings on "date & time" IF ( rank == 0 ) & CALL DATE_AND_TIME(date(1),time(1),zone,values) !... Start elapsed and CPU time counters Etime(1) = MPI_WTIME() CALL CPU_TIME(Ctime(1)) CASE(1) !... Final CPU and elapsed times CALL CPU_TIME(Ctime(2)) Etime(2) = MPI_WTIME() - Etime(1) - Eoverhead - Coverhead Ctime(2) = Ctime(2) - Ctime(1) - Coverhead !... Gather all times CALL MPI_COMM_RANK(MPI_COMM_WORLD, rank, code) CALL MPI_COMM_SIZE(MPI_COMM_WORLD, nb_procs, code) IF ( rank == 0) ALLOCATE(All_Etime(nb_procs), & All_Ctime(nb_procs), & All_Ratio(nb_procs), & All_Rank(nb_procs) ) CALL MPI_GATHER(Etime(2), 1, MPI_DOUBLE_PRECISION, & All_Etime, 1, MPI_DOUBLE_PRECISION, & 0, MPI_COMM_WORLD, code) CALL MPI_GATHER(Ctime(2), 1, MPI_DOUBLE_PRECISION, & All_Ctime, 1, MPI_DOUBLE_PRECISION, & 0, MPI_COMM_WORLD, code) IF ( rank == 0) THEN All_Rank(:) = (/ (i,i=0,nb_procs-1) /) !... Compute elapse user time Total_Etime = SUM(All_Etime(:)) Avg_Etime = Total_Etime/REAL(nb_procs,KIND=p) Max_Etime = MAXVAL(All_Etime(:)) Min_Etime = MINVAL(All_Etime(:)) IF( Min_Etime <= 0.0_p ) THEN PRINT *,"Warning, MPI_TIME: Measured elapsed user time seems to be too short" PRINT *,"compared to the clock precision. Timings could be erroneous." END IF !... Compute CPU user time Total_Ctime = SUM(All_Ctime(:)) Avg_Ctime = Total_Ctime/REAL(nb_procs,KIND=p) Max_Ctime = MAXVAL(All_Ctime(:)) Min_Ctime = MINVAL(All_Ctime(:)) IF( Min_Ctime <= 0.0_p ) THEN PRINT *,"Warning, MPI_TIME: Measured CPU user time seems to be too short" PRINT *,"compared to the clock precision. Timings could be erroneous." END IF !... Compute cpu/elapsed ratio All_Ratio(:) = All_Ctime(:) / All_Etime(:) Total_Ratio = SUM(All_Ratio(:)) Avg_Ratio = Total_Ratio/REAL(nb_procs,KIND=p) Max_Ratio = MAXVAL(All_Ratio(:)) Min_Ratio = MINVAL(All_Ratio(:)) !... End of timings on "date & time" CALL DATE_AND_TIME(date(2),time(2),zone,values) !... Output Format hline ='10X,13("-"),"|",18("-"),"|",14("-"),"|",18("-"),/,' lignes(1)='(//,10X,"(C) January 2003, CNRS/IDRIS, FRANCE.",/,' lignes(2)='10X,"MPI_Time (release 3.4) summary report:",//,' lignes(3)='10X,"Process Rank |"," Elapsed Time (s) |"," CPU Time (s) |"," Ratio CPU/Elapsed",/,' lignes(4)=' (10X,I4,9(" "),"|",F12.3,6(" "),"|",F12.3,2(" "),"|",4(" "),F7.3,/),' WRITE(lignes(4)(1:4),'(I4)') nb_procs lignes(5)='10X,"Total |",F12.3,6(" "),"|",F12.3,2(" "),"|",4(" "),F7.3,/,' lignes(6)='10X,"Minimum |",F12.3,6(" "),"|",F12.3,2(" "),"|",4(" "),F7.3,/,' lignes(7)='10X,"Maximum |",F12.3,6(" "),"|",F12.3,2(" "),"|",4(" "),F7.3,/,' lignes(8)='10X,"Average |",F12.3,6(" "),"|",F12.3,2(" "),"|",4(" "),F7.3,/,' start_date='/,10X,"MPI_Time started on ",2(A2,"/"),A4," at ",2(A2,":"),A2," MET ",A3,":",A2," from GMT",/,' final_date='10X, "MPI_Time ended on ",2(A2,"/"),A4," at ",2(A2,":"),A2," MET ",A3,":",A2," from GMT",//)' fmt=TRIM(lignes(1))//TRIM(lignes(2))//TRIM(lignes(3))// & & TRIM(hline)//TRIM(lignes(4))//TRIM(hline)//TRIM(lignes(5))// & & TRIM(hline)//TRIM(lignes(6))//TRIM(hline)//TRIM(lignes(7))// & & TRIM(hline)//TRIM(lignes(8))//TRIM(hline)//TRIM(start_date)// & & TRIM(final_date) WRITE(*, TRIM(fmt)) & (All_rank(i),All_Etime(i),All_Ctime(i),All_Ratio(i),i=1, nb_procs), & Total_Etime, Total_Ctime, Total_Ratio, & Min_Etime, Min_Ctime, Min_Ratio, & Max_Etime, Max_Ctime, Max_Ratio, & Avg_Etime, Avg_Ctime, Avg_Ratio, & date(1)(7:8), date(1)(5:6), date(1)(1:4), & time(1)(1:2), time(1)(3:4), time(1)(5:6), & zone(1:3), zone(4:5), & date(2)(7:8), date(2)(5:6), date(2)(1:4), & time(2)(1:2), time(2)(3:4), time(2)(5:6), & zone(1:3), zone(4:5) DEALLOCATE(All_Etime, All_Ctime, All_Ratio, All_rank) END IF CASE DEFAULT PRINT *,"Error, MPI_TIME: Invalid input parameter" END SELECT END SUBROUTINE MPI_Time END MODULE MPI_TIMES