Threading options in mumax and oommf

[filvag]

Dear ubermag team,

Recently I have been trying to run simulations with both mumax and oommf in my university cluster. However, I have had some issues regarding efficiency and the % of time used by the program for input/output operations, especially for mumax. From what I have run so far, my impression is that oommf can only utiliize four core at a time as when I try to use more, the efficiency of the program drops significantly. I was wondering if you can verify this, if it is specific to what I'm doing (using mag2exp for hysteresis) and if there is a way to use more cores. In mumax the problem is similar, but worse when on top of low efficiency, regardless of the number of cores, the program spends a large portion of the time reading and writing data to the cluster. I imagine that this is partly due to the use of a GPU; however, I was wondering if you have ideas on how I may improve on that.

I have also attached my python script, the job scripts I use in the cluster and slurm files from jobs I have run in case any of them may be useful to you. Any feedback or insight would be helpful and appreciated.

Kind regards

slurm-19269383.out.txt
slurm-19295547.out.txt

mumax3c_m2e.txt
oommfc_m2e.txt

M2E.py

[lang-m]

Hi @filvag, I can only give you some hints no definitive answer:

• when running with OOMMF you will likely see better performance when you first calculate the full hysteresis loop using the HysteresisDriver and do the post-processing/mag2exp afterwards using micromagneticdata to read the output from the HysteresisDriver. Like that you may benefit from more than 4 cores but to know for sure you will need to run some scaling tests to find the most efficient number of cores.
• using mumax you do not benefit from multiple cores and also cannot control its threading (the threads option is just ignored by Ubermag when it is calling mumax). There is no HysteresisDriver for mumax, so you can't avoid the for-loop here, and therefore probably cannot do much about regarding performance (for your job efficiency to look better, you can ask for fewer CPU cores, as you anyway do not benefit from 8 cores).