Is there any way to split calculations involving denser k-grid to make it more efficient?
I was doing a phonon calculation for CaO with PBE pseudopotentials, ecut=35 Ha and k-grid: 16x16x40 and was getting OUT_OF_MEMORY errors.
I was wondering if there is a way to split the calculation—perhaps by distributing the k-points or using another method—to reduce memory usage while maintaining accuracy.
It is not possible to calculate different pieces of the wavefunction in different runs since you need to update that wavefunction after each SCF step.
There are several things to check to ensure you make the best use of all techniques ABINIT has available:
ensure a proper balance of resource distribution: for this are you using paral_kgb?
target accuracy by property: against what property are you trying to achieve convergence? phonon frequency? for this >10000 k-points for the full Brillouin zone seems excessive. For example you can see in this benchmark article that usually you reach a great accuracy for 2000 kpra (which is number of k-points in BZ multiplied by the number of atom in the cell). In your case this would be the equivalent of 80000 kpra (40 times more) if you are using the cubic 8 atom CaO cell.
enforcing symmetry: ABINIT will identify the symmetry equivalent k-points and will reduce the full BZ k-point number (for example if your system presents time-reversal symmetry, this number of kpts will already be reduced to half).