[CP2K-user] [CP2K:21032] Re: Performance degregation for DCDFT starting with CP2K 2024.2

[Augustin Bussy]

Hi Ralf,

the density matrix-based screening of the derivatives was only removed in 
the special case of response forces (RPA, MP2, TDDFT, EC). Standard HFX 
calculations still use it by default. Note, however, that the default value 
of EPS_SCHWARZ_FORCES was changed to 100 x EPS_SCHWARZ (vs 1.0E-6 
originally), in order to keep the accuracy of the derivatives consistent 
with that of of the integrals.

Best,

Augustin


On Friday, 10 January 2025 at 13:49:42 UTC+1 Ralf Frischmann wrote:

> Thanks for the quick and comprehensive reply.
>
> Is the density matrix-based derivative integral screening still in place 
> for "standard" HFX calculations (i.e. non-response type), or is the 
> screening also   
> removed here? This should then have an effect on how to properly choose 
> the EPS_SCHWARZ_FORCES threshold in relation to EPS_SCHWARZ?
>  
>
>
> On Wednesday, January 8, 2025 at 10:35:06 AM UTC+1 Augustin Bussy wrote:
>
>> Hi Ralf,
>>
>> I believe this comes from a bug fix in HFX response forces: 
>> https://github.com/cp2k/cp2k/commit/233cc0496608329821996a6a19a46b58c16292f1#diff-e05447976b45d06ff57c2ca772d89269da230e4970b6001c4ee2f25c70584cffR428-R430. 
>> Before that, by default, the density matrix was used to screen the integral 
>> derivative, also for response forces. This is not a safe assumption, and it 
>> was therefore removed.
>>
>> To verify this hypothesis, you can try running CP2K v2024.1 with the 
>> added keyword SCREEN_P_FORCES set to FALSE in the HFX%SCREENING section. 
>> This should make the 2 runs equally slow.
>>
>> Best,
>>
>> Augustin
>>
>>
>> On Wednesday, 8 January 2025 at 10:09:31 UTC+1 Ralf Frischmann wrote:
>>
>>> Dear all,
>>>
>>> when running a DCDFT calculation with pre-compiled CP2K binaries from 
>>> GitHub, I do observe a significant reduction in overall runtime performance 
>>> when moving from CP2K version 2024.1 to 2024.2 (and versions beyond). 
>>>
>>> The increase occurs in ENERGY_FORCE runs only. The crucial spot seems to 
>>> be the execution part AFTER convergence of the gradient response solver 
>>> "The linear solver converged in    7 iterations." 
>>>
>>> From the HFX_MEM_INFO output, it is apparent that version 2024.2 
>>> computes a somewhat larger number of "sph. DERIV's calculated on the fly" 
>>> (see attached outputs for a 20 water PBC cell), but the difference in my 
>>> opinion is too small to explain the increase in runtime of about 50% for 
>>> that system size. 
>>>
>>> For larger system with 32 or 64 H2O molecules, the described runtime 
>>> increase becomes more pronounced, reaching 250% for 64 H2O molecules.
>>>
>>> Is that expected/intended behavior?
>>>
>>> Any explanation, hint or advice would be greatly appreciated.
>>>
>>> Regards,
>>>
>>> Ralf
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>  
>>>
>>>

-- 
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/ceb4f4d9-488d-4491-af84-b54e22135f0an%40googlegroups.com.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20250110/12f03e75/attachment-0001.htm>