[CP2K-user] [CP2K:14472] Number of sph. DERIV's calculated on the fly

[hut... at chem.uzh.ch]

Hi

Derivative ERI's are only used once (in the gradient calculation) 
are therefore not stored. Typical cost is 4x for the normal ERIs.
The advantage for the derivative ERI calculation is, that it
can be screened much more efficiently. 
Have a look at the section

CP2K_INPUT / FORCE_EVAL / DFT / XC / HF / SCREENING

EPS_SCHWARZ_FORCES
SCREEN_P_FORCES

Optimal settings of these Keywords can make a huge difference.

In addition, timings are very dependent on the ADMM basis and
the radial cutoff if you are using a truncated operator.

regards

Juerg Hutter
--------------------------------------------------------------
Juerg Hutter                         Phone : ++41 44 635 4491
Institut für Chemie C                FAX   : ++41 44 635 6838
Universität Zürich                   E-mail: hut... at chem.uzh.ch
Winterthurerstrasse 190
CH-8057 Zürich, Switzerland
---------------------------------------------------------------

-----cp... at googlegroups.com wrote: -----
To: "cp2k" <cp... at googlegroups.com>
From: "Sun Geng" 
Sent by: cp... at googlegroups.com
Date: 01/04/2021 09:31PM
Subject: [CP2K:14472] Number of sph. DERIV's calculated on the fly

Dear CP2K users,

I am using CP2K with the ADMM method for structure optimization.
The system is an inorganic bulk with around 100 atoms in a cubic box 12x12x12 angstrom^3.
I am able to set up the calculation resulting smooth SCF convergence. 
The "Number of sph. ERI's calculated on the fly" is 0 and hence the SCF converges quickly. The first SCF takes about 1800 seconds and the following SCF steps take only an impressive 20 seconds. In the end, CP2K can converge the SCF in approximate 2500 seconds.
However, I notice that in between the optimization steps, the derivatives of integrals are still calculated on the fly (see the output below), and this step appears to be very slow. CP2K did not print how long it took to compute the derivatives, but I can tell by checking the updating time in the output. My estimation is that it takes more than 2 hours to calculate the derivatives.

  HFX_MEM_INFO| Number of cart. primitive DERIV's calculated:     11208181891200
  HFX_MEM_INFO| Number of sph. DERIV's calculated:                  140954718240
  HFX_MEM_INFO| Number of sph. DERIV's stored in-core:                         0
  HFX_MEM_INFO| Number of sph. DERIV's calculated on the fly:       140954718240
  HFX_MEM_INFO| Total memory consumption DERIV's RAM [MiB]:                    0
  HFX_MEM_INFO| Whereof max-vals [MiB]:                                        1
  HFX_MEM_INFO| Total compression factor DERIV's RAM:                       0.00


Finally, the overall efficiency of CP2K for optimization is much slower than that of VASP code for this system (both of them used HSE06 functional). In VASP, although an SCF step takes 200 seconds and SCF converged in 15 steps (resulting in 3000 seconds per optimization step), the forces are calculated with negligible cost in VASP.

I wonder are there any tricks in the cp2k inputs that can speed up the derivatives of the integrals? Or any pieces of advice that I can use in using CP2K/ADMM methods?
My personal impression is that CP2K/ADMM is not the optimal choice (in terms of optimization efficiency) for a small system (maybe within 150~200 atoms, I did not test though) compared with VASP code. Maybe, for a larger system, the overhead of calculating forces in CP2K/ADMM is relatively small, and in that case, CP2K/ADMM will be more efficient than VASP? Please let me know if this is a realistic point for choosing VASP and CP2K.

Thank you very much in advance!

Best 
Geng








  
  -- 
 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 cp... at googlegroups.com.
 To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/e8bc30d7-297c-4217-839d-29a86a8f38f3n%40googlegroups.com.