polycrystal crashes when creating really huge sample (8 millions of atoms, 8 grains)

[jfikar]

Dear Pierre,

I think, I'm hitting some limit, when producing very large polycrystal:

atomsk --polycrystal W.xsf tmp.txt poly.cfg -wrap
 ___________________________________________________
|              ___________                          |
|     o---o    A T O M S K                          |
|    o---o|    Version master-2024-05-17            |
|    |   |o    (C) 2010 Pierre Hirel                |
|    o---o     https://atomsk.univ-lille.fr         |
|___________________________________________________|
>>> Constructing a polycrystal using Voronoi method.
>>> Opening the input file: W.xsf
..> Input file was read successfully (2 atoms).
>>> Reading parameters for Voronoi construction from: tmp.txt
..> File was successfully read.
..> Number of grains to be constructed: 8
..> Using a 3-D Voronoi tesselation.

Program received signal SIGSEGV: Segmentation fault - invalid memory reference.

Backtrace for this error:
#0  0x7efe527ba842 in ???
#1  0x7efe527b99d5 in ???
#2  0x7efe524776ef in ???
#3  0x64fb99 in ???
#4  0x664ec1 in ???
#5  0x7a24f4 in ???
#6  0x40256c in ???
#7  0x7efe5246258f in ???
#8  0x7efe5246263f in ???
#9  0x4025a4 in ???
#10  0xffffffffffffffff in ???
Segmentation fault (core dumped)

W.xsf is atomsk --create bcc 3.16519556286392 W xsf and tpm.txt is:

box 214.914 186.121 3200          
node 0.25*box 0.00*box 0.25*box         136.975847 8.366605 -116.116910
node 0.25*box 0.00*box 0.75*box         167.013630 -54.780334 18.100615
node 0.75*box 0.00*box 0.25*box         152.994318 5.012912 147.076081
node 0.75*box 0.00*box 0.75*box         43.115810 -13.310758 141.873019
node 0.50*box 0.50*box 0.25*box         117.138918 -78.712257 65.991134
node 0.50*box 0.50*box 0.75*box         131.501305 -24.487653 109.780401
node 0.00*box 0.50*box 0.25*box         52.894277 -3.522170 125.605655
node 0.00*box 0.50*box 0.75*box         30.053892 -2.480763 131.306464

The half as tall sample worked well, i.e. box 214.914 186.121 1600 on a computer with 128GB RAM and it has 4 millions of atoms. The bigger sample is thus expected to be 8 millions of atoms. I tried the huge sample also on a computer with 256GB RAM, but it seems to allocate at most approx. 124GB and then it crashes.

Might be some internal limitation?

[pierrehirel]

Dear Jan,

Thank you for reporting this. In this mode Atomsk constructs a "template grain", that is afterwards rotated and truncated to produce each grain. To make sure that after rotation, this template fills the volume of all possible grains, it is duplicated according to the largest direction of the box.

In your case that means that Atomsk attempts to duplicate the W unit cell by 3200x3200x3200, which exceeds the limit of integer notation. Although your final polycrystal is expected to host 8 million atoms, the "template grain" would be made of more than 2 billions of atoms.

I can not think of a workaround right now, but I have just pushed a commit to display a warning message instead of the ugly segfault.

Bests,

Pierre

[jfikar]

Oh, I see. I tried two quick and ugly solution to increase the INT4 to INT8 to get around 2^31=2.1e9 limit of INTEGER.

First one is to compile everything with a switch -fdefault-integer-8 . That didn't work. I get invalid memory reference.

The second one is to change just the needed variables in mode_poly.f90 to INT8 by:

use iso_fortran_env
integer(kind=int64):: var

But I'm not sure, which ones need to be INT8. So still getting invalid memory reference. For the testing I also increased your limit of 2.1e9 to some higher number as 8.1e9.