[DFTB-Plus-User] How fast is DFTB+ in MD simulation?

[Ben Hourahine]

Hello Moyassar,


you are probably using too many cores to get efficient performance on
such a small job (the matrix dimension is only ~2500).


Are you comparing equivalent tolerances between the CP2K and DFTB+
calculations? There is probably some difference in the self-consistency
requirements between the codes and you may also be using different time
steps for the MD. Also what is your criteria for reliable results?


For standard Born-Oppenheimer MD, the accuracy of the energy is
quadratic in the self-consistency tollerance (for DFTB+ the energy error
in Hartree is approximately equal to the square of the SCC tolerance in
electron charges). The error in forces is closer to linearly proportional.


If you can still get good results (test carefully), you can accelerate
MD calculations by:


1) Change the SCCTolerance to a larger value (the default is 1.0E-5 you
might be OK with 1 - 2 orders of magnitude looser values depending on
what you are doing).


2) Try increasing the accuracy of the evaluated forces to compensate for
the poorer convergence by using  ForceEvaluation = Dynamics (or DynamicsT0).


3) You didn't mention whether you use a finite temperature electron
(this can speed up self-consistency) or if you are using a periodic
calculation (if its a water box, probably a single k-point at 0,0,0 will
work).


4) Test number of processors against performance. You are choosing to
trade off between wall clock time to complete against computation
resources. But if you request too many processes, this can actually
increase wall clock time to finish (look up 'spin locks' for an
explanation).


Unfortunately the XL scheme as implemented currently does not support
thermostats (this might change in the future), but you could thermalize
your calculation with conventional NVT (thermostat) and then restart an
NVE calculation if that can give you the properties you need.


Regards


Ben


On 19/06/2020 21:09, Moyassar Meshhal wrote:
> Hello DFTB+ community,
>
> I've noticed that DFTB+ is super fast as well as efficient in geometry
> optimization.
> On contrary, in MD it's relatively slow compared to DFTB in CP2K. 
>
> For a system consists of about 250 atoms (C, H and O) in a box of
> water (250-300 water molecules), DFTB+ produces about 800 steps per
> day using 16 cpus, and increasing the number of cpus up to 32 does not
> show much improvement in fastening the simulation. On the other
> hand, CP2K is about 8-10 times faster. But DFTB+'s results are much
> reliable than cp2k.
>
> Does this sound normal? As if it is normal, one would expect that it
> may take about 2 months to get a long MD trajectory for each of my
> systems!
> Is there any way to make the calculations faster?
>
> On the "dftbplus-recipes" website, SCC-MD could be speeded up by
> using the extended Lagrangian (XL) scheme, however, according to the
> manual: "The extended Lagrangian implementation only works for the
> (N,V,E) ensemble so far, so neither thermostats nor barostats are
> allowed."
> So, it does not apply in my case as I do NVT MD.
>
> Best Regards,
> Moyassar
>
>
>
> _______________________________________________
> DFTB-Plus-User mailing list
> DFTB-Plus-User at mailman.zfn.uni-bremen.de
> https://eur02.safelinks.protection.outlook.com/?url=https%3A%2F%2Fmailman.zfn.uni-bremen.de%2Fcgi-bin%2Fmailman%2Flistinfo%2Fdftb-plus-user&data=02%7C01%7Cbenjamin.hourahine%40strath.ac.uk%7C483fdc0fcbc340811cb808d8148ca7a6%7C631e0763153347eba5cd0457bee5944e%7C0%7C0%7C637281941903342295&sdata=jELumBQUePJczhi7i32rPpcEmEJJQRdTjPAF41kwyaQ%3D&reserved=0

-- 
        Dr.  B.  Hourahine,  Senior  lecturer
            SUPA, Department  of  Physics,
             University  of  Strathclyde,
              John  Anderson  Building,
          107 Rottenrow, Glasgow G4 0NG, UK.                    
  +44 141 548 2325, benjamin.hourahine at strath.ac.uk

The  Department is  a partner  in SUPA,  the Scottish
            Universities Physics Alliance

The University  of Strathclyde  is a  charitable body,
       registered in Scotland, number SC015263

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.zfn.uni-bremen.de/pipermail/dftb-plus-user/attachments/20200622/b41ccd38/attachment-0001.htm>