EPW Forum

Yes, using those options no restart is possible as the main quantities are not stored on disk.

Hi Samuel,
After playing with different versions of intel compilers on Peregrine, I think my build is sound. Everything else I've used in the QE distribution with this build, e.g. ph.x and pw.x, have worked fine. The non-polar EPW also seems to work fine. So, I'm guessing that either impi doesn't play nicely with the polar code, or perhaps there is a bug in the polar code. It could always be something else, but those seem like the two top options to me.

Next week I'll try building with gnu and openmpi and see if that changes anything. If you try to debug it and find anything, please let me know.

best,

Derek

Dear Derek,

Thank you for reporting the issue and your extra testings.

I will try to reproduce the NaN and debug.

From my experience, impi is usually a bit faster than free alternatives like openmpi or mpich but somewhat more unstable. Currently EPW is tested the most with openmpi so you might look at that if you want.

Best,

Samuel

Dear Derek

I have tried running the GaN example again. To reduce slightly the computational time, I changed fsthick to 2 and the k path is only along the MGA directions, so that doesn't really affect the results.

I cannot reproduce the NAN results that you have for the polar case. I tried using both 4x4x4 and 6x6x6 coarse starting grids, as already pointed out with the 6x6x6 grids there's the caveat to use the etf_mem=.false. flag if there is a memory issue. Anyway a 4x4x4 grid is quite fast and almost converged.

We did find a small bug related to the implementation of the inversion symmetry (inv. symm. was not implemented in EPW-3, and has to be included in GaN). We will commit it, in the meanwhile you need to add in elphon_shuffle_wrap.f90, line 580:
CALL loadumat ( nbnd, nbndsub, nks, nkstot, xq, cu, cuq, lwin, lwinq )
!
IF (lpolar) CALL compute_bmn_para3 ( nbnd, nbndsub, nks, cu, cuq, bmat(:,:,:,nqc) )

Let us know if that fixes the problem?

Best
Carla

Hi Carla and Samuel,
Many thanks for your input. I was traveling and hence had to put this issue aside. I am returning to it now and will test the issue of openmpi vs impi and the fix suggested by Carla. I'll report back to you what I find.

best,

Derek