Need help to increase compute speed

[Dev Rana]

Hello,

I'm having some trouble conducting an MD simulation. Each SCF iteration takes about 100-200 seconds taking about 15-30 iterations per MD step. Could someone experienced take a look at my input file to see if there is any input which could be added/removed to increase speed? I've attached my restart file containing all information. 

I'm using a computing node with openMPI OMP threads set to 2, and 64 cores on a single exclusive node (no multi-node delays). 

Best Regards,
Devyesh Rana

[Nicholas Winner]

Looks like you can improve your MGRID section. Try a relative cutoff of 50, which should be sufficient for most calculations. You can even go down to 40 if you need.

Then, set CUTOFF by taking the largest exponent in your basis set and multiplying it by REL_CUTOFF. In your case it is ~6 for both Cu and C. 6.13 * 50 = 306.5 ~= 310 

&MGRID

  CUTOFF 310

  REL_CUTOFF 50

&END

Give that a try.

--
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/285a3ca8-132d-479a-b0e6-edcd98c899f5n%40googlegroups.com.

[Lucas Lodeiro]

Hello Rev Rana,

From your input, I see 2 thing that could improve your calculation. First you use a CUTOFF of 600 and a REL_CUTOFF of 250. This is a little weird to me... usually REL_CUTOFF is around 60.... using higher values will slow down your calculation without a real improvement over the calculation results. Do you do a convergence study of these parameters?

Second, you are adding 200 MOs to your calculation... Obviously you need some virtual orbitals to use the smearing, but the greater the number of these, the slower the calculation is. So, 200 is in my opinion too much. Your system has 62 Cu and 1 C atoms, with 686 electrons in total, this is 343 occupied orbitals (only as a mental exercise). Usually adding 20% more orbitals is enough. so adding just 70 orbitals would be sufficient, and make your calculations faster.

Regards - Lucas

--

[Xavier Bidault]

I quote BASIS_MOLOPT file:

# Usage hint : 'NGRIDS 5' (section &MGRID) will deal more efficiently (2x speedup) with the diffuse nature of the basis.

Due to that, don't bother too much with REL_CUTOFF. Just comment or delete the line and let CP2K use the default value.

You indeed have to study the convergence of the CUTOFF. I'm not bothered by 600 Ry. But you should assess this value in the range 300-1500 Ry for example, with a variable-cell optimisation of a CuC crystal at ambient conditions. The energy and volume should converge for a certain value of CUTOFF.

I'm not sure "saving" on EPS_SCF is a good idea. Did you assess 1e-5 was okay? I would say at least 1e-6.

On the other hand, ALPHA = 0.1 is small. How does 0.2 or 0.3 behave?

My 2 cents.

[Dev Rana]

Hi lucas,

Thank you for this. This definitely speeds things up.

I feel like I'm missing something in your explanation. You say that 62 Cu + 1 C = 686 electrons total. Cu has 29 electrons * 62 Cu atoms = 1798. 1 C = 6 electrons. So This should be 1802. With roughly half occupied = 901 and 20% of that is still roughly 200. I agree that 200 is too much - but how you obtained 70 is a bit confusing to me.

Could you explain how you reached your conclusion?

Thanks!

[Lucas Lodeiro]

Hi Dev Rana,

You are usign the GPW method, so you need and your are using pseudopotentials for the atoms. In the KIND section of each atom class, you select the POTENTIAL i.e. the pseudo potential. The number after the "q" means the number of valence electrons explicitly computed. So for Cu is q-11 and for C is q-4: 62*11 + 4 = 686... this information also is printed in the output, before the SCF cycles, where you can check this,

Regards - Lucas

To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/a6c9cc14-476f-4e41-89c6-b849814bf96an%40googlegroups.com.