optimal performance for frequency calculations in parallel

Ada Sedova ada.a.... at gmail.com
Wed Oct 11 23:28:39 UTC 2017


Hi,

I was wondering if there are best practice methods for running a frequency 
calculation in parallel on multiple nodes of an HPC system. I see 
benchmarks for MD and single-point energies, but not frequencies. I ran a 
60-atom unit cell as a test, and using 32 nodes on a system with 16 intel 
cores/ 32 logical cores I was finished in under three hours, so it seems 
that good performance is possible. However, I did not configure 
OMP_NUM_THREADS, or set the number of processes per node. I would like to 
run these types of calculations on supercells of about 200-500 atoms. Could 
you provide some guidance about possible optimal configurations for the 
parallelization for this? Is there any information about the scaling 
behavior for frequency calculations? I am using TZVP-MOLOPT-GTH basis sets 
and PBE or B3LYP, but would also like to add D3-BJ dispersion corrections.

Thanks,

Ada
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20171011/83c07622/attachment.htm>


More information about the CP2K-user mailing list