Understanding the Elapsed Time in Machine Learned Force Fields

[burakgurlek]

Hi,

I am running ML force field simulation and confused a bit of the elapsed time info. Here the result of an ionic step


MLFF: cpu time 1.0896: real time 1.0953
LOOP+: cpu time 1.3737: real time 1.3848
RANDOM_SEED = 450668051 55296 0
IONSTEP: cpu time 0.2797: real time 0.2801

Is the total time per ionice step=MLFF+LOOP+ +IONSTEP? How can I calculate the time per force field evaluation?

Best Regards,
Burak

[ferenc_karsai]

No the total time at each step per core is LOOP+.
MLFF shows the time spent in all machine learning parts for a given step.
In force-field only mode it should be easier to understand since there are only two entries LOOP+ and MLFF.
You will see that LOOP+ is always higher since it also includes IO.

[burakgurlek]

Thanks for the answer. Then, what would be the IONSTEP? time to execute the force field calculation?

also I realized that LOOP+!=MLFF+IONSTEP. Do you know why?

Regards,
Burak

[ferenc_karsai]

IONSTEP is the timing for the movement of the ions and related stuff (solving the differential equations, calculating temperature, ...).

LOOP+ will never be the total sum of the other timings because IO (input-output - but mainly writing to different outputs) is not timed separately and is inside LOOP+.