unexpected low running speed

[hl375]

Dear all,
I am a new user of ATK. I was testing the DFT calculation performed by ATK in comparison with that by VASP and CASTEP. My model is a nine atoms' simple supercell, sampled by 16 k points. I performed nonpolarized PBE single point calculation with SG15 basis set on 12 cores.  The output of the log file stuck at the line "Calculating Nonlocal Part " for at least  10 min. Similar calculation by VASP or CASTEP on 12 cores takes only tens of seconds. Attachment is the log file for the calculation.
Thanks for providing some suggestions.
Best,
Li

[zh]

Please check the explanation of SG15 basis set from the manual:
http://docs.quantumwise.com/technicalnotes/atk_pseudo_basis/atk_pseudo_basis.html

You need to make a trade-off between the accuracy and the efficiency.

[hl375]

Hi zh,

Thank you. I just tried FHI basis.  The calculation still got stuck at line "Calculating Nonlocal Part" for about 7 min.

[hl375]

When I use 48 cores, the calculation begins to get stuck after the CPU information sector.   

[Jess Wellendorff]

A few comments:
1) It would be very helpful if you attach the ATK python script you have run. This makes it much easier to comment on your post.
2) Please use the newest version of ATK, which is 2016.3. You seem to use the 2016.2 version.
3) You have 16 irreducible k-points, so distributing the calculation over 48 MPI processes is computationally inefficient. Distributing over 12 MPI processes is also not optimum, because 8 cores will be idle while the remaining 4 k-points are computed. Please try to use a more suitable number of MPI processes, e.g. 2, 4, 8, or 16.

[Anders Blom]

This apparent low speed can be caused by a competition between OpenMP and MPI. If you didn't explicitly turn off threading, each of your 12 MPI processes will attempt to thread on all cores of the machine, and in total the calculation will run very slowly. You should set

export OMP_NUM_THREADS=1

in the shell, before running "mpiexec ... atkpython" to avoid this.
In ATK 2017, this will no longer be necessary, but it is for 2016 and any older version.

Also, as the "atklog" file and Jess points out, 12 is not a good number to MPI processes for this system, which again contributes to bad performance.