multi-node runs

[Carles]

Dear All,

I would appreciate to know whether you would expect a shorter running time using more than one node (rpn=64) for the following dft code, or you wound't expect any scaling at all?

-------------

title "Zn[H2O]4"

echo

start test

memory 40 gb 

geometry "system" units angstrom nocenter noautoz noautosym

Zn 0.0000000000 0.0000000000 0.0000000000

O 0.0647884400 1.5966752200 1.1383416900

O -1.6031944800 0.2080833800 -1.1242480200

O -0.0894662700 -1.5520184000 1.2184004600

O 1.6120898700 -0.3808227500 -1.0795566600

H -0.3479873200 2.4234365000 0.8571073200

H 1.5230527100 -0.7106965900 -1.9834614900

H -2.3187212300 -0.4402578700 -1.0967684600

H 0.3203851700 -2.3919633900 0.9741995200

H 2.3633840000 0.2272110700 -1.0726091400

H -1.5510030800 0.5457124400 -2.0275830900

H 0.8316793600 1.8231842600 1.6798628400

H -0.8619627100 -1.7521863000 1.7624995400

end

set geometry "system"

basis spherical

* library 6-31G** except Zn

Zn library B2_basis_set_for_Zn

end

dft

xc B3LYP

end

task dft energy

---------

Thanks a lot!

Carles

[Edoardo Aprà]

Not sure how we can look at this case given that it uses a non standard basis set for Zinc

[Carles]

Dear Edoardo, 

thank you for your answer.

I am also trying aug-cc-pVQZ, and it seems it takes a lot of time. I am not sure which basis can be the optimal. I would like to use the largest available basis set, something like QZ4P, which I understand is a Slater-type basis set and not available in NWChem. I would like to get a correct absorption spectrum around the Zn K-edge. 

The ultimate goal is to study nonlinear effects in the core resonances of Zn, using high photon energy ultrashort laser pulses with rt-tddft calculations. Very small time-step and therefore large total N is needed for e.g. 5 fs pulses. Do you have any suggestion for the basis to start with? Which one would be more appropriate for multi-node scaling?

Thank you, 

Regards,

Carles

 

El dia divendres, 5 de novembre de 2021 a les 19:45:48 UTC+1, Edoardo Aprà va escriure:

[Edoardo Aprà]

My suggestion is to use the def2 family of basis sets. You can start with def2-tzvpd and go up to def2-qzvpd.

I am guessing the energy+gradient calculation would be pretty quick and it should not scale beyond a couple of recent 64-core nodes.

[Carles]

Thanks a lot, I will start trying these basis.

Carles

El dia dimarts, 9 de novembre de 2021 a les 20:24:01 UTC+1, Edoardo Aprà va escriure:

[Carles]

I therefore understand that multi-node runs are not useful at all for rt-tddft simulations, and that the best performance will be obtained with a single node of 64 cores, is that correct?

Carles

El dia dimecres, 10 de novembre de 2021 a les 9:16:36 UTC+1, Carles va escriure:

[Edoardo Aprà]

No, that is not what I meant.

I was referring to the specific case of the input file you posted. Since this input file corresponds to just a few hundred basis functions when using a def2-qzvp basis, scaling is going to be limited. Larger cases will scale, instead.

--
You received this message because you are subscribed to a topic in the Google Groups "NWChem Forum" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/nwchem-forum/cY6jeUKLUNQ/unsubscribe.
To unsubscribe from this group and all its topics, send an email to nwchem-forum...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/nwchem-forum/e6511f5f-b700-4bf3-9f48-69fd6a3bd511n%40googlegroups.com.

[Carles]

Ah, ups, I see the point, thanks a lot. But in any case, doing rt-tddft with ultrashort high photon energy laser pulses (e.g. pulses of 5 fs duration and 20 keV central photon energy - i.e. hard x-rays), seems to be something far from what can be done using the present computational tools, since a large basis is needed at those photon energies, with the electron traveling far in the continuum, and although it would then scale with the node number, it will always be slower than the simple def2-qzvp basis with the single node. 

El dia dimecres, 10 de novembre de 2021 a les 16:02:17 UTC+1, Edoardo Aprà va escriure: