abnormally long simulation time for a 2000-gold-atom system using DFTB

[chenhanning]

Hi, all

    I am a new user of QuantumATK, and wanted to test its performance with DFTB.

    For a 576-gold-atom system, it takes 90 seconds on my desktop for each SCF cycle with a maximum memory usage of 1.8GB.

    While for a 2304-gold-atom system, it takes several hours for each SCF cycle with a maximum memory usage of 14.2 GB.

    Because my desktop is equipped with 32GB memory, the memory usage should not be an issue. But this huge difference on simulation time upon a moderate increase of system size is hard to understand.

   Enclosed please find my input files for both the 576-atom and 2304-atom systems.

    Please let me know if any change is needed on my input file to accelerate the 2304-atom simulation using DFTB.

   Thanks.

Hanning

[Anders Blom]

On a general level, the complexity increase for these problems is cubic in the number of atoms (orbitals), so 5x more atoms could quite easily take ~100x longer per step, meaning 2-3 hours in your case.

So the real question is sort of more, how you can reduce the 90 s for the small system! The standard answer to that is to use MPI parallelization, but that is not controlled in the input file but rather from the Job Manager or command line, depending on how you submit the job. I assume your nice machine has something like 14-20 cores so running in parallel would provide a substantial benefit, but I cannot tell if you are already doing that?

[chenhanning]

Yes. You are right. MPI parallelization may reduce the simulation time. But, I also found that by choosing the MultigridSolver, the simulation becomes much more efficient too particularly for large systems even on a single CPU core.

Thanks. 

[Anders Blom]

Interesting, more efficient than FFT?