speedup problem using openMP in intel fortran

bohluly — Mon, 12 Jan 2015 20:03:59 GMT

Dear all,

I have developed a program and unfortunately I have speedup problem in it. My program is so big so I have tried to write a sample similar to my program, fortunately this simple program has a same problem with my program.

I need other experiences and your help if it is possible.

Thanks,

I am using VS2010 and Intel FORTRAN XE 2011

Program:

    TYPE var
        REAL(8),POINTER :: A, B, C
   END TYPE var
     REAL(8),POINTER :: A(:), B(:), C(:)
     TYPE(var),POINTER :: vars(:)
     TYPE(var),POINTER :: varOMP

     REAL*8 t1,t2 ,ai,bi,ci,di,ei,fi
     INTEGER(4) c1,c2
     INTEGER N, CHUNKSIZE, I, id, f , l
     PARAMETER (N=200)
     PARAMETER (CHUNKSIZE=10)

     Allocate (A(N), B(N), C(N),vars(N))

!     initializations

      DO I = 1, N
         A(N)      =   I * 1.0
         B(N)      =   A(N)
         vars(I)%A => A(N)
         vars(I)%B => B(N)
         vars(I)%C => C(N)
         vars(I)%A = 0.51
         vars(I)%B = 0.45
      ENDDO

      CALL SYSTEM_CLOCK(c1)
      Do Itration=1,1000000
!z$OMP PARALLEL PRIVATE(I,varOMP,ai,bi,ci,ei ,di,fi )
!z$OMP DO SCHEDULE(STATIC,CHUNKSIZE)
      DO I = 1, N
        varOMP => vars(I)
       ai = varOMP%A
       bi = varOMP%B
       di = ai*2.2 + bi * 2.0 + ai*2.1
       ei = ai*2.1 + bi * 2.3 + ai*2.15
       fi = di * ( ai + bi )*2.1 + ei * ( di + bi )*2.1
       ci = bi*2.1 + ai*2.0 + di*2.0 + ei*2.0 + fi*2.0
        varOMP%C = ci
      ENDDO
!z$OMP END DO
!z$OMP END PARALLEL
      ENDDO

      CALL SYSTEM_CLOCK(c2)
      WRITE(*,*) c2-c1

      STOP

END

Sorry, in last compile I have

bohluly — Mon, 12 Jan 2015 20:34:49 GMT

Sorry, in last compile I have changed !$OMP to !z$OMP for testing the program using openMP please remove z from !z$OMP in sample.

To see an advantage for

TimP — Tue, 13 Jan 2015 00:13:00 GMT

To see an advantage for threading you will need an outer parallel loop with a count such as 1000 with an inner vectorizable loop of at least the size shown. System_clock works better with 64 bit arguments.mm

threading can't compensate for inefficient data access. If multiple threads write to the same cache line (false sharing) it will perform poorly.

topic Sorry, in last compile I have in Intel® Moderncode for Parallel Architectures

speedup problem using openMP in intel fortran

Sorry, in last compile I have

To see an advantage for