parallelization for kpm.calc_spatial_ldos

[doronwu]

Hi developer,

I am currently calculating a large bilayer system, and the radius of the system is about 100 nm.

I have checked the API, there's only a parallelization implementation for calc_ldos ( named as deferred_ldos).

If I directly run a calc_spatial_ldos calculation, the ETA from the output file says this calculation may take several days.

I wonder if there's a way to parallel the calc_spatial_ldos calculation.

[doronwu]

oh! I just found that the kpm will do automatically threads.

[BertJorissen]

Hi there
First of all, are you trying to calculate the specific states or the general ldos?
This is important, as a state can be calculated using diagonalization techniques (solver, lapack/arpack) but the 'full' ldos can better be calculated using KPM.
The eigenstate calculation can be parallelized over each k vector independently. The 'calc_spatial_ldos' used the last set k-vector and the available eigenstates and eigenvalues.
In essence, you would make a script that constructs a pb.Model object, extract the Hamiltonian, solve for the eigenvectors and eigenstates for each vector using a parallel technique like joblob/Parallel, and combine it back in a pb.Wavefunction (see 'pip install pybinding-dev ' and https://bertjorissen.github.io/pybinding/_api/pybinding.solver.html ).
As I'm not sure what you're aiming for, I can't really help you right now.
Please provide more details about the calculation.
For the 'parallelization' of these eigenstates, it is best to sample the full Brillouin Zone, which will be quite costly for a spatial large system.
If you have a script that can generate a pb.Model for your system, I'm happy to help you.
Best
Bert