My Community

Hi,

I have been trying to run an optical job on a 28 atoms 2D system with a vacuum slab before running the GW job, However, it keeps giving me an "out of memory" error.

I have tried to vary the NCORE from 1 to 64 (the number of cores per node).
I have tried to vary the KPAR from 1 to 64 as well.
I have tried a lot of mixed variations between the NCORE and KPAR
I have tried to use up to 30 nodes each have 64 cores each core has 3.7 GB of memory.

I am using a machine of 3.7 GB/core and 64 cores/node.

The lower level is inaccurate, so I should use this high level of the calculations or the work will be trash.

Please, let me know if I am doing something wrong. Attached are the input files and the OUTCAR file.

Thank you

Dear ahmed_biby,

This seems a supercell calculation. Is that the case? If so, could you try to reduce the number of k-points?

You could you try ALGO=Normal. ALGO=Exact might be too demanding for large systems, since it will try to diagonalise the full Hamiltonian.

Kind regards,
Pedro

Thank you Pedro. I am using such a high KPOINTS since this is the converged one and I need to make a GW calculation based on the WAVECAR from this calculation.

For Algo= Exact, this is the recommended algo if I want to make a GW calculation from this calculation. Do you know if is accepted to use Algo= All instead?

Thank you

Dear ahmed_biby,

The ALGO=EXACT selects a full diagonalisation of your DFT hamiltonian. This can be very expensive for large systems. I would suggest ALGO=NORMAl, which employs an iterative scheme to compute the non-occupied states. The consequence is that the last empty states are not well converged compared to others. You can avoid using those states by selecting a number of bands in the GW step, NBANDSGW, that is lower than the value you set for NBANDS.

Kind regards,
Pedro

Dear Pedro,

Thank you so much for your informative reply. I will try this.

Thanks,
Ahmed