I have an MD simulation from which I extracted 50000 steps; the system is made by 27 Zr atoms and it’s a 3x3x3 supercell of the primitive Zr unit cell with one atom only. I am using the script 2_construct_fcp.py as described in
At the moment, I am interested only to the second order, so I set
cutoffs = [4.5]
for the cluster construction. I am running the script on a 8 core intel i7-4770 CPU @ 3.40GHz with 32GB RAM; I see that the training part runs on 8 cores, so I believe it is parallelized with openMP (is it lapack?). However, after 15h 30’ (wallclock time), it didn’t finish yet. To obtain the force constants with 10000 steps, it took instead 170’; by assuming a linear scaling (which apparently is wrong - how does it scale?) for 50k steps I expect that it should have been done in about 14 hours.
I then have the following questions:
- Do you have some hint to speed up the calculation significantly? Is there any option of the optimizer that can affect the maximum amount of RAM to use, hence the computation time?
- Considering that the symmetries have been correctly recognized, apart the cutoff, is there any parameter that can be set before calling the optimizer so that the train time is reduced?
- Which optimizer do you recommend in the case of a standard MD simulation? I read the manual, but I don’t see any hint on this. What quantities should I consider to choose the proper optimizer?
Thanks a lot in advance for your kind reply.
All the best