Best way to handle large private arrays in openmp parallel region

Question

I am using OpenMP with Fortran.

Within an OpenMP parallel region private variables are allocated on the stack. I have some variables that are 1000x1000 arrays and this is causing some problems. I can get round it but I seem to have to do ulimit -s unlimited as well as setting OMP_STACKSIZE environment variable. This is OK for me, but not great for other users of my code.

At the moment I have arrays passed in the old fashioned way:

integer, intent(in) :: N
real(c_double) :: private_var(N,N)

!$omp parallel do default(shared) private(private_var)
...

Is there a better way to declare these arrays that avoids users having to manually set the stack size on their machine? I don't want to allocate within the loop, but perhaps I can open the parallel region first, have each worker allocate their own private work array, the have the do loop?

Make the arrays allocatable. There are several duplicates about this here. — Vladimir F Героям слава, Aug 20 '15 at 08:34
You might find the second part of [that answer](http://stackoverflow.com/a/13561722/1374437) helpful. — Hristo Iliev, Aug 20 '15 at 10:28
Another one very detailed http://stackoverflow.com/questions/13264274/why-segmentation-fault-is-happening-in-this-openmp-code?lq=1 — Vladimir F Героям слава, Aug 20 '15 at 14:15
Vladimir - thanks. I had seen that one which is how I know about changing the stack size. But the question here was a bit different - how to allocate the arrays so that changing the stack size is not required. The duplicate you marked answers that though (I hadn't found that) — robince, Aug 20 '15 at 14:18
@casey, I don't think `heap-arrays` has any effect on openmp private variables. — robince, Aug 20 '15 at 14:20

Best way to handle large private arrays in openmp parallel region

0 Answers0