Why is my fortran code slower with omp parallel do than without

Question

I have coded a matrix * vector multiplication with a matrix stored in the crs format. Here is my code:

call cpu_time(time3)
!$OMP PARALLEL DO
do o = 1, b2
    do p = row(o), row(o+1)-1
        ergcrs(o) = ergcrs(o) + val(p)*v(col(p))
    end do
end do
!$OMP END PARALLEL DO
call cpu_time(time4)

it works with and without openmp but if it is enabled my code runs slower than it did without openmp.

2 questions: 1. why is my code slower if i try to run it parallel und 2. can i run the second do loop in a parallel do loop too?

the same thing happens if i run this test programm:

program test

    use omp_lib

    implicit none
    integer :: i, n = 1000000000, counter = 0   
    real    :: time1, time2

    call cpu_time(time1)
    !$OMP PARALLEL DO
    do i = 1, N
        counter = counter + 1
    end do
    !$OMP END PARALLEL DO
    call cpu_time(time2)

    print *,time2-time1

end program test

is my problem that i store the results in a shared varaible (ergcrs(o) in the first code piece and counter in the second)?

if so, is there a way to fix it?

Thank you very much for your help!

Don't use `CPU_TIME()` with parallel programs, it will result in wrong timings. Use [`SYSTEM_CLOCK`](https://gcc.gnu.org/onlinedocs/gfortran/SYSTEM_005fCLOCK.html) instead. See [here](https://stackoverflow.com/questions/6878246/fortran-intrinsic-timing-routines-which-is-better-cpu-time-or-system-clock) for the differences between the two. — Alexander Vogt, Jun 25 '15 at 16:43

Why is my fortran code slower with omp parallel do than without

0 Answers0