[CP2K-user] [CP2K:22210] Re: Regarding improvement of speed calculation HSE06

[Johann Pototschnig]

You need to check you bindings / number of processes. 

You request 52 processes. Does mpi use them? To make sure run with:
 mpirun -n 52 ...

If you don't have the CPU's and increase the number of OMP threads then 
they are run on the same processor, which will slow down the calculation as 
you have seen. 

You can also split between MPI and OPENMP:
mpirun -n 26 -c 2 --bind-to none .. 
(The options might have different names depending on the MPI distribution). 


Which splitting works the best depends on your system and on the type of 
computation. 
In general more MPI processes require more memory as they have their 
copies, while openmp does not work over different nodes. 
There is more MPI parallelization in CP2K, so this should work for most 
types of calculation. 



On Tuesday, April 28, 2026 at 4:30:01 PM UTC+2 Lorenzo Lagasco wrote:

> Ok, maybe I can test a bit more carefully the EPS_DEFAULT and EPS_ORB (I 
> have already launched different tests for checking convergence on 
> NGRIDS,CUTOFF and REL_CUTOFF). Meanwhile, this is my slurm submission 
> script:
> #!/bin/bash
> #SBATCH --job-name=HSEO6
> #SBATCH --nodes=1
> #SBATCH --ntasks-per-node=52
> #SBATCH --partition=taras2-6230r
>
> module purge
> module load cp2k/may2025-gnu14.2.0-openmpi4.1.6-psm211.2.230
> export OMP_NUM_THREADS=1
>
> mpirun cp2k.popt -i DYE+SLAB.inp -o DYE+SLAB.out
>
>
> I tried to increase the number of OMP_NUM_THREADS from 1 to 2 or 3 and, as 
> a result, the time required for each scf step is longer and longer. 
> Moreover, I regulated the MAX_MEMORY in relation to the memory free of the 
> node. 
>
> Il giorno mar 28 apr 2026 alle ore 16:00 Frederick Stein <f.s... at hzdr.de> 
> ha scritto:
>
>> Dear Lorenzo,
>> Are you sure that you need these tight cutoffs (EPS_DEFAULT and 
>> especially EPS_PGF_ORB)? Can you check how many integrals are recalculated 
>> (check the output file)? If they are, is it possible for you to increase 
>> MAX_MEMORY? How are you running CP2K (number of MPI ranks, number of OpenMP 
>> threads)?
>> Best,
>> Frederick
>>
>> Lorenzo Lagasco schrieb am Dienstag, 28. April 2026 um 15:43:45 UTC+2:
>>
>>> Good afternoon everyone,
>>>
>>> I am writing to ask for advice on how to improve the computational 
>>> efficiency of an HSE06 calculation used to evaluate inter-state couplings 
>>> between diabatic states in a slab–organic dye system (using the Kondov 
>>> diabatization scheme, system containing 395 atoms). I have attached the 
>>> input file for reference. In a preliminary test on a CPU-only machine 
>>> (single node with 52 processors), a single SCF iteration takes 
>>> approximately 30 minutes.
>>>
>>> Any suggestions would be greatly appreciated.
>>>
>>> Best regards
>>>
>>> Lorenzo Lagasco
>>>
>>>
>>> -- 
>> 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... at googlegroups.com.
>> To view this discussion visit 
>> https://groups.google.com/d/msgid/cp2k/439c2d51-5c6f-4fc9-be1b-1ff2884ab77an%40googlegroups.com 
>> <https://groups.google.com/d/msgid/cp2k/439c2d51-5c6f-4fc9-be1b-1ff2884ab77an%40googlegroups.com?utm_medium=email&utm_source=footer>
>> .
>>
>

-- 
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/173b3016-1da9-4054-bf0c-df730349b6fcn%40googlegroups.com.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20260428/e57a0544/attachment.htm>