[Question] Polaris Allocation Estimation

[ZoeRichter]

Question or discussion topic.

LAMMPS and OpenMC both have GPU capabilities, and LAMMPS simulations can be very intense. Delta is a GPU machine on campus that could be used to run intense LAMMPS simulations. However, in order to use Delta, you need to make a request and estimate the allocation needed.

Provide details.

In order to estimate the needed allocation, we need an understanding of the run times and other simulation information needed for the sort of LAMMPS sims being run, and a list of the simulations that we would be using Delta for.

External context or references.

info on Delta
Output text from end of recent/common sim:

Full LAMMPS log (includes above):
lammps-output.txt

How can this issue be closed?

This issue can be closed once there is an estimation of the Delta allocation needed.

[katyhuff]

Assigning to myself to remember to submit the request.

[ZoeRichter]

I've added two sub-issues for myself - the first is to send updated performance information, now that I've tested LAMMPS on a newer machine. I will close the issue with a csv or similar file summarizing results, but the short version is that the best performance is when I run 1 OpenMP thread per MPI task, and as many tasks as I have physical cores available. So on Quokka, which has 20 cores, I get ~190 sim steps per CPU second when I use 20 MPI tasks and 1 thread per task, and this is the best performance out of the combinations I tried. Performance is hardware and problem specific, however.

The second issue is to generate some short inputs to test on Delta, to figure out what will work on that hardware in particular. The only barrier for this is figuring out the maximum cores a single user should use at a time - the guidance on the Delta webpage only discusses CPU hours.

[ZoeRichter]

@katyhuff - #34 has been updated with a brief summary of how both LAMMPS models performed on our hardware in the best-case scenario. I didn't include the full log files, but I can if you think they would be helpful (they are quite long).