When using OMP atom order is not always the same about atomsk HOT 4 CLOSED

Dear jfikar,

Indeed, it is a consequence of parallelization, where each thread (each CPU core) performs a task independently from the others.

The way I see it, this is not a bug. If I modify the code so that atoms appear in a given order, then it means that CPU cores will be allowed to write their atom positions in a specific order, thus loosing all gain of efficiency of the parallelization.

So, you have to choose between efficiency at the cost of reproducibility (atoms should still be at the same positions, but not in the same order), or reproducibility at the cost of efficiency (use only one thread should produce exactly the same ordering of atoms every time). I don't know any method that would keep good efficiency, and produce the same ordering of atoms.

As a side note: to use a single thread you can either set the environment variable "OMP_NUM_THREADS=1", or use the command-line "-nthreads 1" when running Atomsk.

Best regards

from atomsk.

I see, thank you for explaining the problem.

Is there a way to first sort the atoms by x like -sort x, but when there are more atoms with the same x, sort these atoms by y and finally if there is a subset of atoms with the same x and y, sort them by z? It would be sane, as there should not be two atoms with all coordinates equal.

from atomsk.

Sorting by Z, then by Y, and finally by X should do the trick, something like:

-sort Z up -sort Y up -sort X up

Nonetheless, if many atoms share an exactly identical coordinate, the order may differ. Atomsk uses quicksort, which splits the list in several pieces. Depending on the size of the list and the split, atoms that have exactly the same coordinate may appear in a different order.

In the end, using a single thread when constructing a polycrystal seems more reliable if you want to have exactly the same order of atoms.

from atomsk.

In absence of reply or comment for over 1 month I close this issue.

from atomsk.