Hey Oualid Chinoune

Before switching to QE or VASP, perhaps you first should try reducing grid size or using a coarser k-point mesh to lower memory requirements.

Best,

Gabriel Vinicius

Dear Gabriel Vinicius

The same problems persist even when I use the lowest parameters.

Hello Oualid Chinoune

I think you should check if your model is reliable, because a system with only 32 atoms in DFT calculation wouldn't require 140 GB, even using HSE06. Changing from CASTEP to Quantum ESPRESSO wouldn't make the calculation faster, but I think there must something making this calculation so heavy. If it's possible, I can help you with this, but I would have to check your input.

Best regards,

Ricardo Tadeu

There are many factors that can determine the memory requirements for a DFT calculations. These for a code in plane waves are

1) energy cut-off;

2) number of k-points in the BZ;

3) number of electrons or better number of bands to be considered;

4) algorithm used to perform the different SCF steps.

How heavy is the same calculation with a standard pseudopotential (PBE, PBESOL)? Hybrid pseudopotentials are particularly heave for plane waves codes.

So try to evaluate the memory need for PBE then look at HSE06.

This seems strange, I suspect that your convergence parameters are either very large or you have over-parallelised the calculation.

CASTEP replicates some data over the MPI processes when using nonlocal XC functionals, so if these processes are on the same node this can cause issues - I recommend you reduce the number of MPI processes and use CASTEP's threading to parallelise instead. This may not be the best way to run this calculation, but hopefully it will get you started and we can worry about optimising it later.

You control CASTEP's number of threads per process with the environment variable CASTEP_NUM_THREADS. For example, if you're using a bash shell you would set:

export CASTEP_NUM_THREADS=8

to tell CASTEP to use 8 threads per MPI process. This also means you should reduce the number of MPI processes by a factor of 8, otherwise you will overload your computer!

This should reduce the replicated data by a factor of 8, which should allow your calculation to proceed if that was the bottleneck. Either way, at least it will help us to understand what's going on!

(You're welcome to send me your cell and param files if you'd like me to investigate the issues myself.)

Hope that helps,

Phil Hasnip

(CASTEP developer)

Philip James Hasnip

I appreciate your explanation and suggested solution, sir. I'll give it a go and see if it works. I'll forward the .param and .cell file to you to inspect

thanks,

Oualid CHINOUNE

PhD student

Ferhat Abbas University Setif 1