EPW Forum

Dear all,

I am trying to understand if it is possible to reduce computational time for electron-phonon scattering rate calculations in larger super cells. I started with computation of el-ph scattering rate for silicon 2 atom u.c. using parameters from Computer Physics Commun, 209, 116, 2016. (Coarse grid: 6x6x6/6x6x6; fine grid(q/k): 150,000/30,000). I perfectly matched the scattering rate plot with one in the paper. Then, I increased the cell to 8 atoms (conventional cell). For this system I used the following parameters: coarse grid: 4x4x4/4x4x4; fine grid(q/k):30,000/10,000; for sp3 hybridization I get 32 wannier functions. Since I use 64 k-points for the coarse grid I can only use 64 cores for calculations (one k-point per processor). Thus, such calculations took about 3 days on HPC. If I keep increasing the size of the cell, the BZ is decreasing and, therefore, the number of k-points can be decreased as well. This means I would have to reduce the number of cores for calculations. I also tried to perform calculations for supercell containing 2 conventional cells (16 atoms) and it took almost 2 weeks on 32 cores. At this point I am not sure if I can further increase the cell size. I would really appreciate if anyone could suggest me how to reduce the computational time.

1. Is there any way to use more cores than number of k-points on coarse grid?
2. For sp3 hybridization in Si I can only specify 4 Wannier function per atom. Therefore, 16 atom u.c. would have 64 w.f. Can I reduce the number of wannier functions somehow?
3. I noticed that the frozen window size doesn't affect the results significantly. Can I use larger window size then?

Thank you very much!

Dear chelsy,

Happy to hear that you could reproduce the result of the paper.
First of all, why do you want to do supercells ? It should give you the same answer I guess (except if you do defects or something else ?)

The code works in two parts (i) the part before the writing of the el-ph on real space and (ii) the interpolated part. For the first part, you can indeed parallelize on k-points, which mean you will be able to parallelize on less if you reduce the coarse k-point grid.
However I implemented a band parallelism for that reason. It is not well tested (I used it for 1 system I was looking at) but it is inside the code but not well documented.
The way it works is by using image parallelization. This only works for part (i).
There is a test in q-e/test-suite/epw* about it.
Also, you should look into the code to really understand it.

For part (ii) it should not be a problem if you are using those fine grids.

Note however that the code is not really design to do supercell calculations so it is not really optimized for this.

PS: I'm going on holiday tomorrow morning for a week so I cannot help you more until then unfortunately.

can you please elaborate on supercell method and DFPT methods. i mean the difference between them? are calculations using supercell method more heavier than DFPT method? lastly, EPW have a feature for DFPT calculations and not supercell method.

topgun