[Pw_forum] Graphene phonon calculations too slow
Lorenzo Paulatto
lorenzo.paulatto at impmc.upmc.fr
Thu May 26 09:44:53 CEST 2016
For an el-ph calculation you need the same grid for the (electronic) k and
(vibrational) q points, but the q points can be safely obtained from a coarse
grid (4x4x2 or 8x8x2) using Fourier interpolation (q2r+matdyn).
On the other hand, the phonons on this coarse grid require 32x32x4 electronic
k-points to be well converged.
So: compute a coarse grid of phonon q-points with a fine grid of k-points, then
interpolate the coarse grid to the fine grid and do el-ph with that.
On Wednesday, 25 May 2016 19:47:21 CEST Paolo Giannozzi wrote:
> Are you sure you need that many k-points and such a big cell in the z
> direction?
> Attached a quick-and-dirty test. Parameters are likely far from yielding
> converged values, but it takes 50' on a single processor of my 2007 vintage
> PC. Before starting a full-fledged calculation at convergence, you should
> experiment a bit. Also, tuning the parallelism may make a significant
> difference.
>
> Paolo
>
> On Wed, May 25, 2016 at 5:13 PM, Abualnaja, Faris <
>
> f.abualnaja15 at imperial.ac.uk> wrote:
> > Dear all,
> >
> >
> > I am running phonon calculations on graphene, and the calculations are
> > taking days to complete! The calculations are run on a supercomputer and
> > are running in parallel. I am using QE version 4.3.1, the amount of
> > k-points that are calculated are 12x12x1 for an MP-grid which reduces down
> > to 19 irreducible symmetry points. My phonon input is below:
> >
> >
> > Phonon Dispersion of graphene
> >
> > &inputph
> >
> > recover=.true.
> >
> > tr2_ph=1.0d-16
> >
> > ldisp=.true.
> >
> > nq1=12 nq2=12 nq3=1
> >
> > prefix='graphene'
> >
> > amass(1)=12.01
> >
> > fildyn='graphene.dyn'
> >
> > /
> >
> > and part of the output is shown here:
> >
> >
> > Parallel version (MPI), running on 48 processors
> >
> > K-points division: npool = 3
> >
> > R & G space division: proc/pool = 16
> >
> > Computing dynamical matrix for
> >
> > q = ( 0.0000000 0.0000000 0.0000000 )
> >
> > 25 Sym.Ops. (with q -> -q+G )
> >
> > G cutoff = 245.0079 ( 16214 G-vectors) FFT grid: ( 32, 32,600)
> >
> > number of k points= 217 Methfessel-Paxton smearing, width (Ry)=
> >
> > 0.0100
> >
> >
> >
> > There are 4 irreducible representations
> >
> > Representation 1 1 modes -B_2g To be done
> >
> >
> > Representation 2 2 modes -E_2g To be done
> >
> >
> > Representation 3 1 modes -A_2u To be done
> >
> >
> > Representation 4 2 modes -E_1u To be done
> >
> > G cutoff = 245.0079 ( 259459 G-vectors) FFT grid: ( 32, 32,600)
> >
> > Largest allocated arrays est. size (Mb) dimensions
> >
> > Kohn-Sham Wavefunctions 0.34 Mb ( 2220, 10)
> >
> > NL pseudopotentials 0.07 Mb ( 2220, 2)
> >
> > Each V/rho on FFT grid 0.59 Mb ( 38912)
> >
> > Each G-vector array 0.12 Mb ( 16224)
> >
> > G-vector shells 0.05 Mb ( 6856)
> >
> > Largest temporary arrays est. size (Mb) dimensions
> >
> > Auxiliary wavefunctions 1.35 Mb ( 2220, 40)
> >
> > Each subspace H/S matrix 0.02 Mb ( 40, 40)
> >
> > Each <psi_i|beta_j> matrix 0.00 Mb ( 2, 10)
> >
> > The potential is recalculated from file :
> >
> > ./_ph0/graphene.save/charge-density.dat
> >
> >
> > Starting wfc are 8 atomic + 2 random wfc
> >
> >
> > total cpu time spent up to now is 2.64 secs
> >
> >
> > per-process dynamical memory: 15.0 Mb
> >
> >
> > Band Structure Calculation
> >
> > Davidson diagonalization with overlap
> >
> > c_bands: 1 eigenvalues not converged
> >
> > c_bands: 1 eigenvalues not converged
> >
> > c_bands: 1 eigenvalues not converged
> >
> > c_bands: 1 eigenvalues not converged
> >
> > c_bands: 1 eigenvalues not converged
> >
> >
> > ethr = 1.25E-10, avg # of iterations = 63.7
> >
> >
> > total cpu time spent up to now is 4392.92 secs
> >
> >
> > End of band structure calculation
> >
> > Any suggestions on how to speed up the process and if I am missing
> > something?! Thank you in advance for all your kind help and time.
> >
> > Regards,
> >
> > Faris
> > Imperial
> >
> >
> >
> > _______________________________________________
> > Pw_forum mailing list
> > Pw_forum at pwscf.org
> > http://pwscf.org/mailman/listinfo/pw_forum
--
Dr. Lorenzo Paulatto
IdR @ IMPMC -- CNRS & Université Paris 6
+33 (0)1 44 275 084 / skype: paulatz
http://www.impmc.upmc.fr/~paulatto/
23-24/4é16 Boîte courrier 115,
4 place Jussieu 75252 Paris Cédex 05
More information about the users
mailing list