search

m-a-d-n-e-s-s / madness

Multiresolution Adaptive Numerical Environment for Scientific Simulation
GNU General Public License v2.0
175 stars 61 forks source link

MolDFT -- pseudopotentials -- pseudopotential code is slow #90

Open wsttiger opened 10 years ago

wsttiger commented 10 years ago

Hopefully some of this will be alleviated by incorporating the inner product and gaxpy with analytic functions.

wsttiger commented 9 years ago

I've sped this code up a bit. It's still slower than the all-electron version (at least with a Carbon atom). Waiting for Adam to update his gaxpy_ext() function, then we can incorporate with the PP code.

lratcliff commented 9 years ago

The new screening seems to have broken some molecules, e.g. CS2 (input below). Before the most recent PSP related commit I got good agreement with BigDFT, with the added screening the calculation fails to converge but after the limit of the iterations the total energy is at -31.69316384 Ha rather than -26.13383585 Ha so it's clearly going wrong! Switching screening back to false reverts to the correct result.

I could file this as a separate issue but as it's related to the improvements in speed I figured I may as well comment here.

Input for CS2: dft psp_calc xc lda end

geometry units angstrom S .000000 .000000 1.561117 C .000000 .000000 .000000 S .000000 .000000 -1.561117 end

naromero77 commented 9 years ago

Laura,

Please file a separate issue for the screening aspect.

All,

I ran all-electron nacl vs. psp nacl and the later was about four times faster. This experiment was conducted on 1 MPI task. Perhaps the performance is with parallel runs and has to do with load balancing.