我想运行一个Fortran程序,它调用一个我希望与MPI并行化的子程序。我知道这听起来很复杂,但我希望能够为每个调用指定进程数。我想要使用的是这样的结构:
program my_program
implicit none
!Define variables
nprocs = !formula for calculating number of processes.
call my_subroutine(output,nprocs,other input vars)
end my_program
我想以与此相同的效果运行my_subroutine:
mpirun -n nprocs my_subroutine.o
其中my_subroutine已使用“其他输入变量”进行编译。
这可能吗?
这是一个简单的例子。我尝试编译如下: $ mpif90 -o my_program WAVE_2D_FP_TUNER_mpi.f90 randgen.f SIMPLE_ROUTINE.f90 我尝试像这样运行它: $ mpirun -np(1或2)my_program
PROGRAM WAVE_2D_FP_TUNER_mpi
USE MPI
IMPLICIT NONE
REAL(KIND=8) :: T,PARAM(1:3),Z,ZBQLU01
REAL(KIND=8) :: ERRORS,COSTS,CMAX,CMAX_V(1:1000),THRESHOLD,Z_MIN,Z_MAX
REAL(KIND=8) :: U,S,R(1:6),MATRIX(1:15)
INTEGER :: EN,INC,I,J,M,P
INTEGER :: NPROCS,IERR
!0.8,-0.4,0.4,10,4,4,7 -- [0.003,0.534]
!0.8,-0.2,0.2,10,4,4,7 -- [0.190,0.588]
CALL MPI_INIT(IERR)
CALL MPI_COMM_SIZE(MPI_COMM_WORLD,NPROCS,IERR)
THRESHOLD = 0.D0
EN = 81
INC = 1
Z_MIN = -2.D-1; Z_MAX = 2.D-1
T = 1.D0
PARAM(1) = 10.D0; PARAM(2) = 4.D0; PARAM(3) = 4.D0
CMAX = 7.D0 !Max that wave speed could possibly be.
CALL ZBQLINI(0.D0)
OPEN(UNIT = 1, FILE = "TUNER_F.txt")
WRITE(1,*) 'Grid Size: '
WRITE(1,*) T/(EN-1)
DO P = 1,15
S = 0
Z = Z_MIN + (1.d0/(15-1))*dble((P-1))*(Z_MAX - Z_MIN)
WRITE(1,*) 'Z: ',Z
DO I = 1,1000
DO J = 1,6
R(J) = ZBQLU01(0.D0)
END DO
!CALL PDE_WAVE_F_mpi(T,PARAM,R,Z,CMAX,EN,INC,NPROCS,U)
CALL SIMPLE_ROUTINE(T,PARAM,R,Z,CMAX,EN,INC,NPROCS,U)
IF (U<=threshold) THEN
S = S + 1.D0
ELSE
S = S + 0.D0
END IF
END DO
MATRIX(P) = (1.D0/1000)*S
END DO
DO I = 1,15
WRITE(1,*) MATRIX(I)
END DO
PRINT *,MINVAL(MATRIX)
PRINT *,MAXVAL(MATRIX)
CLOSE(1)
CALL MPI_FINALIZE(IERR)
END PROGRAM WAVE_2D_FP_TUNER_mpi
这是我希望与mpi并行化的子程序。
SUBROUTINE SIMPLE_ROUTINE(T,PARAM,R,Z,CMAX,EN,INC,NPROCS,U)
! Outputs scalar U = T*Z*CMAX*INC*SUM(PARAM)*SUM(R)*SUM(Y)
USE MPI
IMPLICIT NONE
REAL(KIND=8), INTENT(IN) :: T,PARAM(1:3),R(1:6),Z,CMAX
INTEGER, INTENT(IN) :: EN,INC
INTEGER, INTENT(IN) :: NPROCS
REAL(KIND=8), INTENT(OUT) :: U
REAL(KIND=8) :: H,LOCAL_SUM,SUM_OF_X
REAL(KIND=8), DIMENSION(:), ALLOCATABLE :: X
INTEGER :: PX,PX_MAX,NXL,REMX,IX_OFF,P_LEFT,P_RIGHT
INTEGER :: J
INTEGER :: IERR,MYID
! Broadcast nprocs handle to all processes in MPRI_COMM_WORLD
CALL MPI_BCAST(&NPROCS, NPROCS, MPI_INT, 0, MPI_COMM_WORLD,IERR)
! Create subcommunicator SUBCOMM (Do not know how to define WORLD_GROUP?)
CALL MPI_COMM_SPLIT(MPI_COMM_WORLD,WORLD_GROUP,SUBCOMM,IERR)
! Assign IDs to processes in SUBCOMM
CALL MPI_COMM_RANK(SUBCOMM,MYID,IERR)
! Give NPROCS - 1 to SUBCOMM
CALL MPI_COMM_SIZE(SUBCOMM,NPROCS-1,IERR)
H = 2.D0/(EN-1)
! LABEL THE PROCESSES FROM 1 TO PX_MAX.
PX = MYID + 1
PX_MAX = NPROCS
! SPLIT UP THE GRID IN THE X-DIRECTION.
NXL = EN/PX_MAX !nxl = 10/3 = 3
REMX = EN-NXL*PX_MAX !remx = 10-3*3 = 1
IF (PX .LE. REMX) THEN !for px = 1,nxl = 3
NXL = NXL+1 !nxl = 4
IX_OFF = (PX-1)*NXL !ix_off = 0
ELSE
IX_OFF = REMX*(NXL+1)+(PX-(REMX+1))*NXL !for px = 2 and px = 3, ix_off = 1*(3+1)+(2-(1+1))*3 = 4, ix_off = 1*(3+1)+(3-(1+1))*3 = 7
END IF
! ALLOCATE MEMORY FOR VARIOUS ARRAYS.
ALLOCATE(X(0:NXL+1))
X(:) = (/(-1.D0+DBLE(J-1+IX_OFF)*H, J=1,EN)/)
LOCAL_SUM = SUM(X(1:NXL))
CALL MPI_REDUCE(LOCAL_SUM,SUM_OF_X,1,&
MPI_DOUBLE_PRECISION,MPI_SUM,&
0,MPI_COMM_WORLD,IERR)
U = T*Z*CMAX*INC*SUM(PARAM)*SUM(R)*SUM_OF_X
DEALLOCATE(X)
CALL MPI_COMM_FREE(SUBCOMM,IERR)
CALL MPI_BARRIER(MPI_COMM_WORLD,IERR)
END SUBROUTINE SIMPLE_ROUTINE
最终,我希望能够更改子程序中使用的处理器数量,我希望从EN的值计算nproc。
答案 0 :(得分:2)
一种简单的方法是使用最大进程数启动MPI应用程序。
然后my_subroutine
将首先MPI_Bcast(&nprocs, ...)
和MPI_COMM_SPLIT(MPI_COMM_WORLD, ..., &subcomm)
,以便与subcomm
创建子通信器nprocs
(您可以使用MPI_UNDEFINED
,以便&#34;其他&#34;传播者将MPI_COMM_NULL
。
然后,属于subcomm
的MPI任务将执行计算。
最后,MPI_Comm_free(&subcomm)
和MPI_Barrier(MPI_COMM_WORLD)
从性能的角度来看,注意子通信器创建可能很昂贵,但希望与计算时间相比并不重要。
如果没有,您宁愿改进算法,以便nprocs
个任务完成工作,其他任务等待。
另一种方法是使用一个MPI任务,MPI_Comm_spawn()
nprocs-1
任务启动您的应用程序,合并交互器,执行计算,并终止生成的任务。
创建任务的开销更为重要,资源管理器可能无法完全支持,因此我不建议使用此选项。