call to ZHEGV failed for constrained magnetism calculation

[sophie_weber]

Hello,

I am beating my head against a wall regarding this issue. I am doing a constrained magnetism calculation (I_CONSTRAINED_M=1) with a slab geometry with vacuum; I'm constraining the outer few layers of moments to point along the surface normal whereas I constrain the rest in the bulk to point along [001] (the easy axis of the bulk structure...this is Chromia). This has worked fine with a slab which is half the height, with the exact same input parameters (I know my examples have a couple differences, but trust me, I have also tried IDENTICAL calculations except for the POSCARs), but no matter what I do with the doubled slab, I get the following error.

"Error EDDDAV: Call to ZHEGV failed. Returncode= x y z", where x y and z are some integers (they change depending on the number of cores.

I've tried all the suggested tricks I've found. These include
1. checking the geometry (it's fine; the doubled structure comes from the same relaxed primitive slab, and also, self-consistent calculations with this slab work fine).
2. Changing number of cores, number of tasks-per-node, and NPAR
3. Switching the algorithm from ALGO=Normal to ALGO=All (I tried ALGO=Exact as well but couldn't stop this from crashing due to memory overload).

It seems to be something with the constrained magnetism, but this is again weird because it works fine for a slab that's half the size.

I have attached tar files with inputs as well as some outputs for the calculation which fails ('failed_calc') and the calculation for the smaller slab which works ('good_calc'). Note that the ZHEGV fail occurs on two different supercomputers. Any hints would be greatly appreciated.

Thank you so much in advance!::

[ferenc_karsai]

So it took now some time to be able to run your calculation, since the parameters you used make it computationially intractable for me (I could not fit your settings into 256 GB RAM. I used ENCUT=300, a small k-mesh (3x3x1) and ADDGRID = .FALSE. I could reproduce the error and cure it.
The problem was, that in the big calculation you misspelled M_CONSTR. You wrote insead MCONSTR.
Some tips for your calculation:
ADDGRID = .TRUE. is really adding a huge computational overhead.
You have set LAMBDA=10, that is probably too huge, try something like 2 maybe.

[sophie_weber]

Thank you so much, I just caught the MCONSTR error. Thanks also for the tips regarding reducing overhead. I really appreciate it!