`makefile.include` for Perlmutter

[chiang-yuan]

Thank you for developing this wonderful package. I saw you have benchmark tests on NERSC Perlmutter CPU and GPU. Could you share the makefile.include you use and the modules loaded for compilation if you still have access to somewhere? I have been trying to moodify the ones NERSC provides but it seems to be incompatible with vasp-interactive. If you could kindly share that, I will really appreciate that. That would definitely save me a lot of time matching right module and flags.

@alchem0x2A @jmusiel @zulissi

[alchem0x2A]

Thx for bringing this up. What we meant for Perlmutter CPU / GPU were just the native builds provided by NERSC. While in principle using the correct makefile.include would work with our patch, it's best contacting the NERSC team to see if they have any hints. (Zhengji should be the best to contact). Another solution is to build your shifter image containing the patched vasp. We've been using NVHPC (20.04) base container image and seems to work fine

[chiang-yuan]

Thanks @alchem0x2A ! I finally sort it out by myself, but I found unstable behavior of passing test example ex00_vasp_test.py on Perlmutter. Sometime it passes all the test including lattice dynamics sometime it doesn't. I played around with OMP_NUM_TRHEADS and other spread and bind settings but it is still unstable. Would it be safe to ignore that because you have set time_limit in the test?

[chiang-yuan]

And now I have compilation script and makefile.include for both NERSC Perlmutter and Bridges2 working. If you are interested in including that, I am happy to open a PR

[alchem0x2A]

have you specified the "slurm -u" option to disable output buffer? We have seen similar cases that the default buffering mechanism may cause unexpected behavior.

yes please open a PR for the working makefile.include, just place in the same folder as vasp-build/makefile_examples should be good

[chiang-yuan]

Yes I added -u otherwise I won't pass any test at all.

[hsulab]

Hi @chiang-yuan ,

I am also doing some tests on Perlmutter. However, I cannot even run a normal vasp calculation successfully. I compiled a vasp5.4.4-cpu with intel but somehow it only runs 2 steps and exits with a segment fault error. I checked the NERSC provided makefile.include for vasp5.4.4-cpu but it seems the compilation flags are for nvidia build. Can you please share your makefile.include with me? Any advice is appreciated.

All the best,
Jiayan