[CP2K-user] [CP2K:21192] Re: Hybrid Functionals and GPU usage

[Frederick Stein]

Deare Laurie,
Calculations with hybrid functionals are generally significantly (usually 
at least an order of magnitude) more expensive than non-hybrid 
calculations. This also affects the amount of necessary ressources for 
these kind of calculations. You may try the following things for testing:
- switch the RUN_TYPE to ENERGY_FORCE (default)
- restart your calculation from a non-hybrid calculation (PBE is usually a 
good guess). Add SCREEN_ON_INITIAL_P to enable density matrix screening to 
reduce the memory demands further.
- try a smaller system (to get a feeling for the usage of the methods) see 
https://github.com/cp2k/cp2k/tree/master/benchmarks/QS for some examples. 
Try H2O-32 and H2O-64, but you need to switch to HF/hybrid calculations 
first.
- try a smaller AUX_FIT basis set
- RI-Hybrid methods are not linear scaling for large systems (I do not know 
the crossing-over points as I am not familiar with the RI flavor)
- I do not know the amount of ressources you have available. For H2O-64 
with ADMM but without RI, I need ca. 4 nodes of Piz Daint for 
PBE-calculations but ca. 16 for hybrid calculations(1 P100, 24 cores, 64 GB 
memory per node). For you system, it may be a factor of 4-8 more depending 
on the scaling behavior.
- Check the timings at the end of the output file (ask here if you need 
more advice on that)
- If it still breaks, enabling tracing of routines (add the TRACE and 
TRACE_MASTER keywords to your GLOBAL section) and post the last ca. 100 
lines of the output file as the output file becomes huge (millions of lines 
of output)
- accuracy-wise: check EPS_DEFAULT in the QS section as it determines the 
accuracy of your calculation as well.
Best,
Frederick

Laurie Stevens schrieb am Freitag, 21. Februar 2025 um 16:32:18 UTC+1:

>
> Dear CP2K community, 
>
>
> I'm trying to run a REFTRAJ with the revPBE0 functional and I'm facing 
> issues. I have been working with lower levels of theory before (mostly 
> BLYP-D3, revPVBE and SCAN). When switching to a hybrid functional, I first 
> tried to use only CPUs but it seemed way slower than I would have expected 
> and I am now trying to move to GPUs. 
>
> However, I'm still encountering issues during my tests: either I get a OoM 
> error, or the first SCF step is not finished after 45 minutes. (When I 
> tried again with the CPUs, I had the same issues). I'm unsure if the SCF is 
> even starting or if the GPU is correclty used. 
> The same REFTRAJ calculation was running perfectly fine for BLYP-D3 and 
> revPBE functionals. 
>
> Is that a typical timing that one could expect in this case or am I doing 
> something wrong? 
> Is it recommended to use GPUs when working with revPBE0 or is it 
> reasonable to run it only on CPUs? Is there a way to speed up the 
> calculation? 
>
> I'm really not familiar with hybrid functionals, so I don't know what to 
> expect. Is there any recommendations regarding the setup on a cluster 
> (number of nodes, MPI processes, OMP threads, etc) which are specific to 
> this level of theory? 
>
>
> Thank you in advance for you help, 
> Best regards,
> Laurie Stevens
>
>

-- 
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+unsubscribe at googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/cp2k/978f1f50-cc69-4e9f-918d-62be3677b961n%40googlegroups.com.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20250221/4f95800d/attachment.htm>