Increase of accuracy

[sponce]

Dear Vahid,

Carla will for sure reply to this question better than me but from what I understand in the paper, it means:
- Ground-state properties are computed on a 6x6x6 k-grid. This means the scf cycle.
- Response properties are computed using DFPT using a 4x4x4 q-grid. This means the nscf.in is probably a 4x4x4 k-grid to be commensurate with the q-grid.
Anyway, this is a nscf cycle, therefore it uses the "good" 6x6x6 electronic density computed during the scf cycle.

Best,

Samuel

[carla.verdi]

Dear Vahid,

I can confirm as Samuel said that the 6x6x6 k-grid was used for the scf runs, whereas the wannierization is performed using a 4x4x4 uniform mesh. Sorry for not including this detail!

Best
Carla

[Vahid]

Dear Carla and Samuel,

Thank you for clarifying the coarse kgrid size. I finished the TiO2 run successfully and reproduced the electron lifetimes for the two segments in the BZ as done in the PRL115 paper.

I am curious as to how long it took for you to finish the TiO2 run. My run (which used exactly the same parameter as the PRL115 paper) took 92 hours on 64 cpus with a total of 220GB of memory. I am asking this to see if I should find ways to better optimise while compiling EPW.

Many thanks,

Vahid

[Vahid]

Dear EPW Users,

After several tests, it seems that EPW (and QE) slow down significantly if they are linked against the internal FFT library that accompanies the source code. I linked against an optimised external FFTW3 library and the speed changed significantly (nearly a factor of 3). Therefore my TiO2 run probably took too long for this reason. Hope this helps others.

Best,

Vahid

[sponce]

Dear Vahid,

Thank you for reporting this.

I personally found, for example, a 30% speedup just by going from Intel 12 to 13 (with similar compilation flags).

Bottom line is that, yes, compiler and external libraries do matter.