Hybrids and double hybrids on LUMI

[Sigbjorn Bore]

Hi, first, thank you for this awesome software!

I am planning to run a lot of water single point calculations (forces and energies) for water on LUMI-G. I am currently using this installation https://lumi-supercomputer.github.io/LUMI-EasyBuild-docs/c/CP2K/#user-documentation. This recommends a slightly different setup (6 openMP vs 3 in https://pubs.aip.org/aip/jcp/article/158/16/164109/2886896, which I understand was done during pilot phase). Is this installation up to date, and what is generally the best jobsetup currently?

Then, to get best performance for hybrids and double hybrids on LUMI, do you have any specific recommendations? From https://pubs.aip.org/aip/jcp/article/160/6/064116/3263749,  https://pubs.aip.org/aip/jcp/article/160/6/064116/3263749, https://pubs.aip.org/aip/jcp/article/158/16/164109/2886896, and rapid developments I am little confused on what I should be running to get reasonable numerical accuracy (at each level of theory) for lowest amount of cost. So far it seems standard input file (4c) for hybrid functionals and use the SOS of https://pubs.aip.org/aip/jcp/article/158/16/164109/2886896 for RI-MP2.

All the best,

Sigbjørn

 

[Frederick Stein]

Dear Sigbjørn,

From the manual, I can tell that the available version is not up-to-date (The latest release is 2024.1). If I were you, I would just test which setup is better. The k-point implementation is not compatible with post-Hartree-Fock methods. It will depend on your systems whether the low-scaling or the high-scaling implementation of RPA or SOS-MP2 is faster and whether ADMM and/or RI-HFX is more suitable for your system (You may combine ADMM with RI-HFX if seems reasonable to you). I can give you recommendations regarding the high-scaling methods, in case of the low-scaling methods and RI-HFX, you should consult Augustin Bussy.

Best,

Frederick

[Sigbjorn Bore]

Okay! Will try to reach out to lumi about the version, seems like gpu part is changing rapidly, should not be too hard to update it.

Definitely interested, I will try any specific suggestion and report back on performance! I was thinking 64 water molecules for a start.

--
You received this message because you are subscribed to a topic in the Google Groups "cp2k" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/cp2k/CHsAz_QXXkE/unsubscribe.
To unsubscribe from this group and all its topics, send an email to cp2k+unsubscribe@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/d5734839-2288-4c48-b3b9-50dbcfceee22n%40googlegroups.com.

[Frederick Stein]

From out last tests, the high-scaling algorithm has a better performance up to 128 water molecules. For slightly larger systems, you should try both, for significantly larger systems (512 water molecules), the low-scaling algorithm is preferrable.

PS: Soon, there will be an announcement for a workshop on these methods in a few weeks.

To unsubscribe from this group and all its topics, send an email to cp2k+uns...@googlegroups.com.

[Augustin Bussy]

Hi Sigbjørn,

I will add to Frederick's answer concerning the HFX part of your calculations. RI-HFX is only more performant for molecules and medium size solid state systems. For water, my experience has shown that using the initial 4-center implementation is always faster, especially if ADMM is also used.

Best,

Augustin

You received this message because you are subscribed to the Google Groups "cp2k" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cp2k+uns...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/02fb4998-47e5-4267-ac99-9f1ba0ad2f34n%40googlegroups.com.

[Sigbjorn Bore]

Thank both of you! I saw the 4c did better in one of the papers indeed. Would you happen to have specific inputs I can try out? Are the ones in benchmark (GitHub) and supplementary materials up to date?

PS. I have initiated a ticket to update cp2k LUMI. 

[Augustin Bussy]

Yes, they should be up to date. You can also find ADMM-PBE0 bulk water inputs for 4-center and RI-HFX in https://archive.materialscloud.org/record/2023.50, path: RI_HFX/method_scaling

To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/3531d9fa-9875-4071-81d9-d2c03ea37869n%40googlegroups.com.

[Sigbjorn Bore]

I have the new CP2K installed now on LUMI (it will be made available once tested), and can finally try everything out

I also saw your article J. Chem. Phys. 160, 064116 (2024); doi: 10.1063/5.0189659. Is it worth testing it for hybrids (no test for water in the paper)?