Speed up pair_style loading time with QUIP/GAP

Hi all,

I’m using pair_style quip to load a GAP interatomic potential:

pair_style quip
pair_coeff * * potentials/gp_iter6_sparse9k.xml “Potential xml_label=GAP_2017_6_17_60_4_3_56_165 14”

The GAP data files are from https://github.com/libAtoms/silicon-testing-framework/tree/master/models/GAP.

Good news: it works!
Bad news: actually loading the GAP data takes forever. This pre-simulation step is often the slowest part of my process and it’s holding me back from prompt results.

I have some sample timings produced by loading the GAP data using pair_coeff and then immediately terminating the programme. The numbers below are both from the regular GAP data set (224MB) and a smaller, non-production sample set (751kB):Sample data set - 1 cores - 1.456s

Sample data set - 1 core - 1.456s
Sample data set - 8 cores - 5.754s
Sample data set - 16 cores - 10.964s
Sample data set - 32 cores - 21.379s

Regular data set - 1 core - 2m 24.885s
Regular data set - 8 cores - 3m 0.430s
Regular data set - 16 cores - 4m 54.074s
Regular data set - 32 cores - 11m 16.834s

A few minutes may not seem like a lot of time to lose, but this is a simulation that may need to be run up to 200 times in succession - at which point an eleven minute delay is a loss of a day and a half. No matter what it seems that I’m losing any time gained by using more processors right at the starting block.

I’d love to be able to speed this step up by using the restart mechanism instead of starting a completely new run each time, but the documentation for pair_style quip clearly states that its information is not written to restart files: “you need to re-specify the pair_style and pair_coeff commands”.

Equally, I’d love to be able to refactor my wrapper script to run 32 of these simulations in parallel on one core each, but each requires data from the previous one.

Is there some way to expedite this initial data loading step?

Many thanks!

Hi all,

[…]

Is there some way to expedite this initial data loading step?

Not on the LAMMPS side. You would have to talk to the QUIP/GAP developers to discuss options for faster loading.

Axel.