rt-tddft - benchmark times

[Carles]

Dear all,

I am doing several benchmarks at Theta (ALCF) with the aim of calculating some nonlinear effects with rt-tddft. I find however that I am far from being able of studying those effects with rt-tdfft. I would like to share with you some of the results to confirm/compare with yours, and see whether you find that my results are expected or not.

For the relatively small molecule Zn(H2O)4 (13 atoms), I find the fastest calculation is using 1 single node (of 64 cpu). It takes 1978 seconds to do 600 rt-tddft time steps. It improves a little bit with 2 threads (1706 s). Any attempt to increase of the number of nodes gives slower results.

I use the basis 6-31G** except for Zn, with B2_basis_set_for_Zn, and xc B3LYP.

For a typical nonlinear calculation, using 3.5 fs laser pulses at about 30 keV (hard X-rays), I would need to calculate about 350.000 steps, so that it is far unfeasible.

From the NWChem page, the benchmark shown at https://nwchemgit.github.io/Benchmarks.html , working at cascade and tahoma, gives a scaling optimization to about 9 nodes. Although there the molecule has 533 atoms, and the basis set 7108 functions. The single round is about 524 seconds.

Is that if the molecule is small, the scaling to many nodes through MPI is not efficient at all?

Thanks a lot!

Carles 

[Edoardo Aprà]

I am not sure that the NWChem installation is the most appropriate for multi-node runs.

According to https://www.alcf.anl.gov/support-center/nwchem, the value for ARMCI_NETWORK set at compile time was MPI-TS.

Unfortunately this is not the best choice for performing multi-node runs as discussed in the following section of the NWChem website https://nwchemgit.github.io/ARMCI.html#choosing-the-armci-library

[Carles]

Dear Edoardo,

thanks a lot for your reply. I will inform Theta administration, to see whether something can be done.

Best regards,

Carles

El dia dijous, 14 d’octubre de 2021 a les 20:03:09 UTC+2, Edoardo Aprà va escriure: