segmantation fault with fortran parallel program using openmp

Question

I have a fortran code in which I have a parallel part. It is compiled by gfortran successfully, but I get a segmentation fault by running it. The serial compiled run file does not show any fault. Also I have examined the parallel program with very small input matrix (rho1 & rho2) and tested the step by step parameters. There were no fault. If I understand correctly, when I determine the variables as PRIVATE, there is no need to use $OMP ATOMIC. Here the matrices rho1 and rho2, have a dimension of about 15,000,000. Here is the parallel part of code:

!$OMP PARALLEL DO ORDERED DEFAULT(PRIVATE)
  do ix = 1 , nx
    do iy = 1 , ny
      do iz = 1 , nz
         k = iz + (iy-1) * nz + (ix-1) * ny * nz
         if (rho1(k) .GT. 0.d0) then
           x1 = x0 + ((ix-1) * dx)
           y1 = y0 + ((iy-1) * dy)
           z1 = z0 + ((iz-1) * dz)
           rr = (x1-xa)**2 + (y1-ya)**2 + (z1-za)**2
           r1a = dsqrt (rr)
           rr = (x1-xb)**2 + (y1-yb)**2 + (z1-zb)**2
           r1b = dsqrt (rr)
           if (r1a == 0.d0) Vnuc = (rho1(k) * Znb)/r1b
           if (r1b == 0.d0) Vnuc = (rho1(K) * Zna)/r1a
           if (r1a .GT. 0.d0 .AND. r1b .GT. 0.d0) then
             Vnuc = (rho1(k) * Zna)/r1a + (rho1(K) * Znb)/r1b
           endif
           Ve = 0
           !$OMP ORDERED
           j = 1
           do jx = 1 , nx
            do jy = 1 , ny
             do jz = 1 , nz
               if (rho2(j) .GT. 0.d0) then
                x2 = x0 + ((jx-1) * dx)
                y2 = y0 + ((jy-1) * dy)
                z2 = z0 + ((jz-1) * dz)
                rr= (x1-x2)**2 + (y1-y2)**2 + (z1-z2)**2
                r12 = dsqrt (rr)
                if (r12 .GT. 0.d0) then
                 Ve = Ve + (rho1(k)*rho2(j))/r12
                endif
               endif
               j = j + 1
             enddo
            enddo
           enddo
           !$OMP END ORDERED
           V1 = (Ve * dx * dy * dz * 0.529177d0) - Vnuc
           rr = (x1-xmid)**2 + (y1-ymid)**2 + (z1-zmid)**2
           r = dsqrt (rr)
           zef1(k) = V1 * r            
          endif
       enddo
    enddo
  enddo
  !$OMP END PARALLEL DO

as a start, did you try to compile with e.g. gfortran -fbounds-check -O0 -g -ggdb or something and then run? — steabert
– steabert, Commented Nov 3, 2013 at 14:03
a second remark: shouldn't ny and nz be firstprivate ? — steabert
– steabert, Commented Nov 3, 2013 at 15:23

High Performance Mark · Accepted Answer · 2013-11-03 16:08:27Z

2

It looks as if your arrays rho1 and rho2 are declared private. One consequence of this is that each thread will, on entry to the parallel region, have a private instance of those arrays. If the arrays are large your program may simply be trying to allocate more memory than is available.

It's relatively unusual to write OpenMP programs on large arrays which are not shared; multiple threads operating on different parts of large shared arrays is probably the canonical application for OpenMP.

answered Nov 3, 2013 at 16:08

High Performance Mark

78.6k7 gold badges109 silver badges168 bronze badges

Sign up to request clarification or add additional context in comments.

1 Comment

elham keshavarz Over a year ago

I tried making rho1 and rho2 variables shared, but the fault is the same. Also I have tried leaving all variables shared and the error message was the same, although I traced the program with a simple input and the process went correct.

elham keshavarz · Accepted Answer · 2013-11-06 06:18:01Z

1

I found the solution from some other posts on the website, "Why Segmentation fault is happening in this openmp code". The problem was due to the limitation on stack size. It was solved by the following command: ulimit -s unlimited

answered Nov 6, 2013 at 6:18

elham keshavarz

495 bronze badges

Comments

user3225087 · Accepted Answer · 2014-06-06 17:11:07Z

1

One issue may be also that when variables are declared private, each thread creates an unitialized private replica of the object. You may need to include the copyin openmp statement so that the private variables have the value of the variable before the parallelization region.

answered Jun 6, 2014 at 17:11

user3225087

2253 silver badges10 bronze badges

Collectives™ on Stack Overflow

segmantation fault with fortran parallel program using openmp

3 Answers 3

1 Comment

Comments

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

3 Answers 3

1 Comment

Comments

Comments

Your Answer

Sign up or log in

Post as a guest

Related