My Community

I'm in the process of running DFPT calculations (LEPSILON = .TRUE.) on a wide range of unit cell sizes (from <10 to >40 atoms). About 50 % of my calculations time out since my jobs have a 36 h max runtime cap. To my knowledge, VASP offers no way to checkpoint perturbative calculations?

So I wanted to ask what's the recommended way to speed up DFPT calculations? Is there some combination of settings/flags combined with more cores or more memory that might make DFPT calculations run faster?

Dear Janosh,

The combination of INCAR tags, e.g., NPAR, KPAR , NCORE etc., depends on the architecture of your compute engine. To get the most performance out of VASP, you can contact your system administrator, refer to the documentation of the above-mentioned INCAR tags and run some performance tests.

An additional route you might want to persued is to take advantage of checkpointing tools. Perhaps try MPI-agnostic network-agnostic (MANA) transparent checkpointing. It is implemented as a plugin in DMTCP: Distributed MultiThreaded CheckPointing.

Please consider sharing your experience here afterward!

Best regards,
Marie-Therese