BUG: Illegal memory access error when using coarsening in X2 direction on GPU #226
Assignees
Labels
bug
Something isn't working
Comments
|
Hi @Anto6453 , thanks for the report. In order to reproduce this, could you send me the full setup you are using? (i.e. setup.cpp+idefix.ini) (can be by email if your don't want your setup to appear publicly). |
|
Problem is that gridCoarsening tries to reconstruct the normal component of B from the divergence of the two tangential components of B stored in Vs. When DIMENSIONS<3, one of these components is not defined (because it's not stored on cell faces), resulting in a segfault (as it turns out, also on CPUs). The fix #230 will be part of Idefix v2.0.05 release. |
|
v2.0.05 released with this bug fixed. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Describe the issue:
When using coarsening on X2 direction (corresponding to$\theta$ in spherical coordinates), I obtain an illegal memory access error when I try to run idefix on GPU (seen on A100 and RTX 2080 SUPER). The same configuration works on CPU.
In my case, I use a 2D-axisym spherical grid, and I divide my radial grid in one uniform grid in the inner domain, and one log grid in the outer domain. To avoid that my uniform grid in$r$ becomes too thin in the $\theta$ direction when it comes close to the center of the domain, in need to coars the grid in the $\theta$ direction.
My coarsening function :
where
radiusCoarsis the radius that delimits the uniform and the log grids.The error seems to occure when the number of points become too big.
For example :
- with one process on one RTX 2080 SUPER, the error seems to occure when the total number of point comes close to 2^22 (works for 2^21 points but not for 2^22 points).
- with two mpi processes (on two RTX 2080 SUPER) the limit increase also by a factor 2.
- increasing the number of point only in the log grid (which is not coarsed) seems also to be a trigger.
- if I remove the coarsening, I don't have this error whatever is the number of point.
It also seems to be triggered more easily when I increase the maximum coarsening level (via
coarsIncrementin the coarsening function).By looking at the log in debug mode, the error seems to appear when idefix is trying to coars the magnetic field while keeping divB equal to zero (in the second FLUID_CoarsenFlow_BXsn idefix_for loop in the src/fluid/coarsenFlow.hpp file).
Thanks in advance for your help.
Error message:
runtime information:
Idefix version 2.0.04-00e3db93
Master branch
Kokkos 30500
Seen on GPU A100 and RTX 2080 SUPER
The text was updated successfully, but these errors were encountered: