Page 1 of 1

Very strange crash during geometry optimization but not MD (same POSCAR)

Posted: Fri Oct 06, 2023 12:55 pm
by jchapmanbu
Hello,
My name is James Chapman, I'm an assistant professor at Boston university. I have a strange crash with vasp (error files is attached) that only seems to occur during geometry optimizations and not during MD. The root of the problem seems to be the POSCAR, more specifically the interatomic distances. In the vasp files I've attached there are 2 POSCARs, one that works and one that doesn't. For the one that works, the only difference is that I scaled the lattice parameters to be slightly larger and increase the interatomic distances. The INCAR attached is for a geometry optimization, but during MD the POSCAR that fails during the GO works, which is where I'm completely confused. How can the same POSCAR fail during a GO but not during MD? The failure occurs during electronic minimization, and from the error file it looks like it fails during a call to scalapack during diagonalization. Hopefully, the fact that the same POSCAR works during MD but fails during a GO can help elucidate an answer.

Also, the parallelization scheme doesn't seem to matter. I've tried changing the number of nodes to alter the data distribution. I've tried increasing the omp stacksize up to 8G, which is just absurd but also has no effect. I've even tried changing the number of OMP threads which also has no effect. We've gone as far as to create 6 versions of vasp all with different compilers and the problem persists. So, it clearly has something to do with the POSCAR and the interatomic distances and nothing else.

Any help you can provide is greatly appreciated, and please let me know if there's anything I can help with.

Re: Very strange crash during geometry optimization but not MD (same POSCAR)

Posted: Mon Oct 09, 2023 12:31 pm
by alex
Hello James,

the algorithms differ from MD to GO, this might be the reason for your observation.

Two more things I've spotted in your inputs:
You are using std VASP with Gamma-point, change to the gamma-version. Speed-up 2. No other drawbacks.
10% shrinkage is quite a bit, you are far off the harmonic oscillator and hence most likely experience weird stuff, esp. if you shrink the lattice.

Hth,

alex