Performance of Electron-Phonon Calculation for Primitive Cell

[jiong_yang1]

Dear all:
I hope this message finds you well. I am currently performing renormalization calculations for the graphene primitive cell POSCAR (containing 2 atoms), with a focus on obtaining the electronic self-energy. The INCAR settings I am using are as follows:

Code: Select all

PREC = Accurate   
ENCUT = 600  
EDIFF = 1E-8       
LREAL = .FALSE. 
ISMEAR = 0       
SIGMA = 0.01 eV    
LWAVE = F
NCORE = 2

ELPH_RUN = .TRUE. 
ELPH_DRIVER = EL
ELPH_NBANDS = -2    
KPOINTS_OPT_MODE = 1  
ELPH_IGNORE_IMAG_PHONONS = T
ELPH_WF_CACHE_MB = 100000

ELPH_SELFEN_GAPS = F 
ELPH_SELFEN_DELTA = 0.01 
ELPH_SELFEN_FAN = .TRUE.   
ELPH_SELFEN_DW = .TRUE.   
ELPH_SELFEN_ENERGY_WINDOW = 1.500 5.500
ELPH_SELFEN_TEMPS = 0 100 200 300 400 500 

Meanwhile, the KPOINTS mesh is set to 15*15*1, and KPOINTS_ELPH is set to 30*30*1. The calculation is being performed on a 48-core CPU node, and it has started to slow down significantly at the following step:

Code: Select all

 ========== interatomic force constants  ============
 treatment of LR: F
 ====================================================
 =========== electron-phonon potential  =============
 treatment of LR: F
 shmem: F
 spin-unpolarized (ispin=1, nrspinors=1, ncdij=1)
 ====================================================
 ============== electron-phonon wfs =================
 fft_mesh:             20          20         120
 spin-unpolarized (ispin=1, nrspinors=1, ncdij=1)
 nbands_k':          2681
 nbands_k:           2681
 elph_nbands:        2681
 ====================================================
fbz qpoint [       1 -      100 /     900]

I would like to ask that, is this slowdown a normal phenomenon? What specific calculation is being performed at this step that causes the slowdown? In addition, I observed that the memory usage at this time is 406 GB, which is a relatively large value. Is this memory usage normal?

I look forward to your reply. Thank you very much indeed!

Best regards

[christopher_sheldon1]

Dear Jiong,

Thank you for your question. We have two suggestions for the large memory requirements as it stands:

• You use all available bands with

  Code: Select all

  ELPH_NBANDS = -2

  . This will increase the memory requirements to calculate the electron-phonon matrix elements. You could try reducing this to see what effect this has on memory requirements.

• The

  Code: Select all

  ELPH_SELFEN_ENERGY_WINDOW = 1.500 5.500

  will likely cover a lot of states, which will require a lot of time and memory. This is not necessarily an issue, but could partially explain.

Could you provide the POTCAR and OUTCAR files up to this point in the calculation? This will make it easier for us to work out at which step this is occurring and why it slows down.

Best wishes,

Chris

[jiong_yang1]

Dear Jiong,

Thank you for your question. We have two suggestions for the large memory requirements as it stands:

• You use all available bands with

  Code: Select all

  ELPH_NBANDS = -2

  . This will increase the memory requirements to calculate the electron-phonon matrix elements. You could try reducing this to see what effect this has on memory requirements.

• The

  Code: Select all

  ELPH_SELFEN_ENERGY_WINDOW = 1.500 5.500

  will likely cover a lot of states, which will require a lot of time and memory. This is not necessarily an issue, but could partially explain.

Could you provide the POTCAR and OUTCAR files up to this point in the calculation? This will make it easier for us to work out at which step this is occurring and why it slows down.

Best wishes,

Chris

Thanks for your suggestion! To make it easier, the POTCAR and OUTCAR files are uploaded. Thank you very much!

[christopher_sheldon1]

Dear Joing,

Thank you for the files. Your calculation is taking too many nbands_k at this this point:

Code: Select all

 ============== electron-phonon wfs =================
 fft_mesh:             20          20         120
 spin-unpolarized (ispin=1, nrspinors=1, ncdij=1)
 nbands_k':          2681
 nbands_k:           2681
 elph_nbands:        2681

This means that all of the electron-phonon matrix elements are calculated between each of the k- and k'-points, significantly increasing the cost of the calculation. We will try to improve this for the future. For the moment, we recommend reducing the number of nbands_k by using ELPH_SELFEN_BAND_STOP= 8. This will speed up your calculation significantly. ELPH_SELFEN_BAND_STOP will compute the electron-phonon band coupling only for bands up to this number. You should increase this to include the bands that you are interested in.

Let me know if this helps.

Best wishes,

Chris