[Pw_forum] testing HSE for BCC Fe
Hello, As a new QE user, I do some tests with the HSE Functional for BCC ferromagnetic Fe with a cell containing 2 atoms (because a want to go to CoFe after that). My input file is : &control title='Fe CC HSE', calculation = 'scf', prefix = 'Fe_CC_HSE', pseudo_dir='./', verbosity = 'high' / &system ibrav = 6, celldm(1) = 5.45, celldm(3) = 1.0, nat = 2, ntyp = 1, nspin = 2, occupations='smearing', smearing='gauss', degauss=0.01, starting_magnetization(1) = 1.0, ecutwfc = 33, input_dft='HSE', report = 5 / &electrons mixing_beta = 0.7 / ATOMIC_SPECIES Fe 55.845 Fe.pbe-sp-van.UPF ATOMIC_POSITIONS (crystal) Fe 0.000 0.000 0.000 Fe 0.500 0.500 0.500 K_POINTS (automatic) 7 7 7 0 0 0 Surprisingly, the computation time is huge as writen into the output file : === iteration # 9 ecut= 33.00 Ry beta=0.70 Davidson diagonalization with overlap ethr = 5.49E-08, avg # of iterations = 2.3 Magnetic moment per site: atom: 1 charge: 14.4350 magn: 2.6346 constr: 0. atom: 2 charge: 14.4351 magn: 2.6343 constr: 0. total cpu time spent up to now is 7.2 secs End of self-consistent calculation convergence has been achieved in 9 iterations EXX: now go back to refine exchange calculation total cpu time spent up to now is 4239.4 secs Self-consistent Calculation iteration # 1 ecut= 33.00 Ry beta=0.70 Davidson diagonalization with overlap ethr = 5.49E-08, avg # of iterations = 4.9 total cpu time spent up to now is 52490.2 secs iteration # 2 ecut= 33.00 Ry beta=0.70 Davidson diagonalization with overlap === So, the calculation takes 7 seconds for converging with PBE, 1h10min for calculating the EXX term and 14h35min for the first HSE iteration. Is this correct ? And if so, please, could you try to explain me how to reduce the computation time. Thanks in advance. Best wishes, Daniel ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum
Re: [Pw_forum] testing HSE for BCC Fe
Please read the literature about hybrid functionals in plane waves; understand why the computational time is, not unexpectedly, huge and start think about it. If you come up with some smart idea on how to reduce the computational time without compromising accuracy (as for the ACE transformation recently proposed by Lin Lin and implemented in the latest QE version) let us know Best, stefano (sent from my phone) > On 09 Sep 2016, at 09:30, Daniel Stoeffler > wrote: > > > > Hello, > > As a new QE user, I do some tests with the HSE Functional for BCC > ferromagnetic Fe with a cell containing 2 atoms (because a want to go to CoFe > after that). > > My input file is : > > &control > title='Fe CC HSE', > calculation = 'scf', > prefix = 'Fe_CC_HSE', > pseudo_dir='./', > verbosity = 'high' > / > &system > ibrav = 6, > celldm(1) = 5.45, > celldm(3) = 1.0, > nat = 2, > ntyp = 1, > nspin = 2, > occupations='smearing', smearing='gauss', degauss=0.01, > starting_magnetization(1) = 1.0, > ecutwfc = 33, > input_dft='HSE', > report = 5 > / > &electrons > mixing_beta = 0.7 > / > > ATOMIC_SPECIES > Fe 55.845 Fe.pbe-sp-van.UPF > > ATOMIC_POSITIONS (crystal) > Fe 0.000 0.000 0.000 > Fe 0.500 0.500 0.500 > > K_POINTS (automatic) > 7 7 7 0 0 0 > > > Surprisingly, the computation time is huge as writen into the output file : > > === > > iteration # 9 ecut=33.00 Ry beta=0.70 > Davidson diagonalization with overlap > ethr = 5.49E-08, avg # of iterations = 2.3 > > Magnetic moment per site: > atom:1charge: 14.4350magn:2.6346constr:0. > atom:2charge: 14.4351magn:2.6343constr:0. > > total cpu time spent up to now is7.2 secs > > End of self-consistent calculation > > > convergence has been achieved in 9 iterations > > EXX: now go back to refine exchange calculation > > total cpu time spent up to now is 4239.4 secs > > Self-consistent Calculation > > iteration # 1 ecut=33.00 Ry beta=0.70 > Davidson diagonalization with overlap > ethr = 5.49E-08, avg # of iterations = 4.9 > > total cpu time spent up to now is52490.2 secs > > iteration # 2 ecut=33.00 Ry beta=0.70 > Davidson diagonalization with overlap > > > === > > So, the calculation takes 7 seconds for converging with PBE, 1h10min for > calculating the EXX term and 14h35min for the first HSE iteration. Is this > correct ? And if so, please, could you try to explain me how to reduce the > computation time. Thanks in advance. > > Best wishes, > > Daniel > > > > > ___ > Pw_forum mailing list > Pw_forum@pwscf.org > http://pwscf.org/mailman/listinfo/pw_forum ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum
Re: [Pw_forum] testing HSE for BCC Fe
On Fri, Sep 9, 2016 at 9:30 AM, Daniel Stoeffler < daniel.stoeff...@ipcms.unistra.fr> wrote: Surprisingly, the computation time is huge no sir: "unsurprisingly" the computation time is huge. The computational complexity of a calculation with hybrid functionals goes as (Nk*Nb)^2*Npw log(Npw) or so, where Nk=number of k-points, Nb=number of occupied bands, Npw=number of plane waves; with GGA functionals, as Nk*Nb*Npw log(Npw) or Nk*Nb*Npw^2 could you try to explain me how to reduce the computation time > first of all you may try to reduce one of the sums over k-points using options nq1,nq2,nq3, as explained in the example for hybrid. For a metal, not sure it works well, though. Since you are using UltraSoft PseudoPotentials, try option "tqr=.true." (real-space treatment of augmentation functions): it should give a significant speedup. Another significant improvement is contained in the forthcoming 6.0 version of QE, but it doesn't work yet for USPP Paolo -- Paolo Giannozzi, Dip. Scienze Matematiche Informatiche e Fisiche, Univ. Udine, via delle Scienze 208, 33100 Udine, Italy Phone +39-0432-558216, fax +39-0432-558222 ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum
Re: [Pw_forum] testing HSE for BCC Fe
> The computational complexity of a calculation with hybrid functionals goes > as (Nk*Nb)^2*Npw log(Npw) or so, where Nk=number of k-points, Nb=number of > occupied bands, Npw=number of plane waves; with GGA functionals, as > Nk*Nb*Npw log(Npw) or Nk*Nb*Npw^2 > well, no, as Nk*Nb*Npw log(Npw) or Nk*Nb^2*Npw Paolo > > > could you try to explain me how to reduce the computation time >> > > first of all you may try to reduce one of the sums over k-points using > options nq1,nq2,nq3, as explained in the example for hybrid. For a metal, > not sure it works well, though. Since you are using UltraSoft > PseudoPotentials, try option "tqr=.true." (real-space treatment of > augmentation functions): it should give a significant speedup. Another > significant improvement is contained in the forthcoming 6.0 version of QE, > but it doesn't work yet for USPP > > Paolo > -- > Paolo Giannozzi, Dip. Scienze Matematiche Informatiche e Fisiche, > Univ. Udine, via delle Scienze 208, 33100 Udine, Italy > Phone +39-0432-558216, fax +39-0432-558222 > > -- Paolo Giannozzi, Dip. Scienze Matematiche Informatiche e Fisiche, Univ. Udine, via delle Scienze 208, 33100 Udine, Italy Phone +39-0432-558216, fax +39-0432-558222 ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum
Re: [Pw_forum] testing HSE for BCC Fe
Hello Stefano, As said, I am a new user of QE but not of the DFT ;). I use the FLEUR, Abinit and the VASP codes and I want to test if it would be interesting for my studies to use also QE. I said "surprisingly" huge because I compared to the VASP code with the same parameters (Ecutoff, NKpoints, Nq) and with the HSE functional. Indeed, with the VASP code during one night on my computer I got a few tens of iterations compared to one with QE :(. >From my point of view, I compare 2 pseudopotential-based methods for the same parameters which I can control via the input file : my question would then more preferably be "which default parameter should I modify to make the QE calculation time similar to the VASP one ?" Best wishes, Daniel Le 2016-09-09 09:47, Stefano de Gironcoli a écrit : > Please read the literature about hybrid functionals in plane waves; understand why the computational time is, not unexpectedly, huge and start think about it. > If you come up with some smart idea on how to reduce the computational time without compromising accuracy (as for the ACE transformation recently proposed by Lin Lin and implemented in the latest QE version) let us know > Best, > stefano > (sent from my phone) > > On 09 Sep 2016, at 09:30, Daniel Stoeffler wrote: > >> Hello, >> >> As a new QE user, I do some tests with the HSE Functional for BCC ferromagnetic Fe with a cell containing 2 atoms (because a want to go to CoFe after that). >> >> My input file is : >> >> &control >> title='Fe CC HSE', >> calculation = 'scf', >> prefix = 'Fe_CC_HSE', >> pseudo_dir='./', >> verbosity = 'high' >> / >> &system >> ibrav = 6, >> celldm(1) = 5.45, >> celldm(3) = 1.0, >> nat = 2, >> ntyp = 1, >> nspin = 2, >> occupations='smearing', smearing='gauss', degauss=0.01, >> starting_magnetization(1) = 1.0, >> ecutwfc = 33, >> input_dft='HSE', >> report = 5 >> / >> &electrons >> mixing_beta = 0.7 >> / >> >> ATOMIC_SPECIES >> Fe 55.845 Fe.pbe-sp-van.UPF >> >> ATOMIC_POSITIONS (crystal) >> Fe 0.000 0.000 0.000 >> Fe 0.500 0.500 0.500 >> >> K_POINTS (automatic) >> 7 7 7 0 0 0 >> >> Surprisingly, the computation time is huge as writen into the output file : >> >> === >> >> iteration # 9 ecut= 33.00 Ry beta=0.70 >> Davidson diagonalization with overlap >> ethr = 5.49E-08, avg # of iterations = 2.3 >> >> Magnetic moment per site: >> atom: 1 charge: 14.4350 magn: 2.6346 constr: 0. >> atom: 2 charge: 14.4351 magn: 2.6343 constr: 0. >> >> total cpu time spent up to now is 7.2 secs >> >> End of self-consistent calculation >> >> convergence has been achieved in 9 iterations >> >> EXX: now go back to refine exchange calculation >> >> total cpu time spent up to now is 4239.4 secs >> >> Self-consistent Calculation >> >> iteration # 1 ecut= 33.00 Ry beta=0.70 >> Davidson diagonalization with overlap >> ethr = 5.49E-08, avg # of iterations = 4.9 >> >> total cpu time spent up to now is 52490.2 secs >> >> iteration # 2 ecut= 33.00 Ry beta=0.70 >> Davidson diagonalization with overlap >> >> === >> >> So, the calculation takes 7 seconds for converging with PBE, 1h10min for calculating the EXX term and 14h35min for the first HSE iteration. Is this correct ? And if so, please, could you try to explain me how to reduce the computation time. Thanks in advance. >> >> Best wishes, >> >> Daniel > >> ___ >> Pw_forum mailing list >> Pw_forum@pwscf.org >> http://pwscf.org/mailman/listinfo/pw_forum [1] > > ___ > Pw_forum mailing list > Pw_forum@pwscf.org > http://pwscf.org/mailman/listinfo/pw_forum [1] Links: -- [1] http://pwscf.org/mailman/listinfo/pw_forum ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum
Re: [Pw_forum] testing HSE for BCC Fe
Hello Paolo, Thank you for your answer. Up to now, I used the FLEUR, Abinit and the VASP codes and I am doing some tests to see if it would be interesting (or not) for my studies to use also QE. I found the computation "surprisingly" huge because when I compare to the VASP code with the same parameters (Ecutoff, NKpoints, Nq) and with the HSE06 functional, after onr night of run, I got a few tens of iterations whereas I got only one with QE :(. For these tests, I compare 2 pseudopotential-based methods for the same parameters which I can control via the input file (Ecutoff and Kpoints). I have compared the (defaut) q-points and they are identical for the VASP and the QE codes. I will try the "tqr=.true." option and let you know. Concerning the Pseudopotential, I decided to use Fe.pbe-sp-van.UPF in order to keep the Ecutoff small : I tried HGH but the Ecutoff needed for magnetic systems is very large and I have avoided to use NLCC pseudopotentials which are not compatible with hybrid functional (is this correct ?). The final system I would investigate is the Fe(FeCo)O4 spinel with a 28-atoms into cell using hybrid functionals : from your experience and knowlege, it this hopeless with QE or not ? Best wishes, Daniel Le 2016-09-09 10:21, Paolo Giannozzi a écrit : >> The computational complexity of a calculation with hybrid functionals goes as (Nk*Nb)^2*Npw log(Npw) or so, where Nk=number of k-points, Nb=number of occupied bands, Npw=number of plane waves; with GGA functionals, as Nk*Nb*Npw log(Npw) or Nk*Nb*Npw^2 > > well, no, as Nk*Nb*Npw log(Npw) or Nk*Nb^2*Npw > > Paolo > >>> could you try to explain me how to reduce the computation time >> >> first of all you may try to reduce one of the sums over k-points using options nq1,nq2,nq3, as explained in the example for hybrid. For a metal, not sure it works well, though. Since you are using UltraSoft PseudoPotentials, try option "tqr=.true." (real-space treatment of augmentation functions): it should give a significant speedup. Another significant improvement is contained in the forthcoming 6.0 version of QE, but it doesn't work yet for USPP >> >> Paolo -- >> >> Paolo Giannozzi, Dip. Scienze Matematiche Informatiche e Fisiche, >> Univ. Udine, via delle Scienze 208, 33100 Udine, Italy >> Phone +39-0432-558216 [1], fax +39-0432-558222 [2] > > -- > > Paolo Giannozzi, Dip. Scienze Matematiche Informatiche e Fisiche, > Univ. Udine, via delle Scienze 208, 33100 Udine, Italy > Phone +39-0432-558216, fax +39-0432-558222 > > ___ > Pw_forum mailing list > Pw_forum@pwscf.org > http://pwscf.org/mailman/listinfo/pw_forum [3] Links: -- [1] tel:%2B39-0432-558216 [2] tel:%2B39-0432-558222 [3] http://pwscf.org/mailman/listinfo/pw_forum ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum
Re: [Pw_forum] testing HSE for BCC Fe
Dear Daniel I said "surprisingly" huge because I compared to the VASP code with the > same parameters (Ecutoff, NKpoints, Nq) and with the HSE functional. > Indeed, with the VASP code during one night on my computer I got a few tens > of iterations compared to one with QE :(. > With exactly the same parameters the computational time (on one processor) should be comparable. There are two important points you should consider to make a comparison: -Different type of methodology: Using EXX with Ultrasoft pseudopotentials a very large number of augmentation charges are computed. As already suggested setting tqr=.true. should solve this problem. -If you changed the default values of some variables such as Precfock or Nkred in Vasp the computational parameters are not anymore the same in Vasp and QE. Best, Dario > From my point of view, I compare 2 pseudopotential-based methods for the > same parameters which I can control via the input file : my question would > then more preferably be "which default parameter should I modify to make > the QE calculation time similar to the VASP one ?" > > Best wishes, > > Daniel > > > > > > > Le 2016-09-09 09:47, Stefano de Gironcoli a écrit : > > Please read the literature about hybrid functionals in plane waves; > understand why the computational time is, not unexpectedly, huge and start > think about it. > If you come up with some smart idea on how to reduce the computational > time without compromising accuracy (as for the ACE transformation recently > proposed by Lin Lin and implemented in the latest QE version) let us know > Best, > stefano > (sent from my phone) > > On 09 Sep 2016, at 09:30, Daniel Stoeffler unistra.fr> wrote: > > > > Hello, > > As a new QE user, I do some tests with the HSE Functional for BCC > ferromagnetic Fe with a cell containing 2 atoms (because a want to go to > CoFe after that). > > My input file is : > > &control > title='Fe CC HSE', > calculation = 'scf', > prefix = 'Fe_CC_HSE', > pseudo_dir='./', > verbosity = 'high' > / > &system > ibrav = 6, > celldm(1) = 5.45, > celldm(3) = 1.0, > nat = 2, > ntyp = 1, > nspin = 2, > occupations='smearing', smearing='gauss', degauss=0.01, > starting_magnetization(1) = 1.0, > ecutwfc = 33, > input_dft='HSE', > report = 5 > / > &electrons > mixing_beta = 0.7 > / > > ATOMIC_SPECIES > Fe 55.845 Fe.pbe-sp-van.UPF > > ATOMIC_POSITIONS (crystal) > Fe 0.000 0.000 0.000 > Fe 0.500 0.500 0.500 > > K_POINTS (automatic) > 7 7 7 0 0 0 > > Surprisingly, the computation time is huge as writen into the output file : > > === > > iteration # 9 ecut=33.00 Ry beta=0.70 > Davidson diagonalization with overlap > ethr = 5.49E-08, avg # of iterations = 2.3 > > Magnetic moment per site: > atom:1charge: 14.4350magn:2.6346constr: > 0. > atom:2charge: 14.4351magn:2.6343constr: > 0. > > total cpu time spent up to now is7.2 secs > > End of self-consistent calculation > > convergence has been achieved in 9 iterations > > EXX: now go back to refine exchange calculation > > total cpu time spent up to now is 4239.4 secs > > Self-consistent Calculation > > iteration # 1 ecut=33.00 Ry beta=0.70 > Davidson diagonalization with overlap > ethr = 5.49E-08, avg # of iterations = 4.9 > > total cpu time spent up to now is52490.2 secs > > iteration # 2 ecut=33.00 Ry beta=0.70 > Davidson diagonalization with overlap > > === > > So, the calculation takes 7 seconds for converging with PBE, 1h10min for > calculating the EXX term and 14h35min for the first HSE iteration. Is this > correct ? And if so, please, could you try to explain me how to reduce the > computation time. Thanks in advance. > > Best wishes, > > Daniel > > > > > ___ > Pw_forum mailing list > Pw_forum@pwscf.org > http://pwscf.org/mailman/listinfo/pw_forum > > > ___ > Pw_forum mailing > listPw_forum@pwscf.orghttp://pwscf.org/mailman/listinfo/pw_forum > > > ___ > Pw_forum mailing list > Pw_forum@pwscf.org > http://pwscf.org/mailman/listinfo/pw_forum > ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum
Re: [Pw_forum] testing HSE for BCC Fe
Dear Daniel, Stefano and Paolo. My added penny: in my hands US/EXX has never worked in a useful way...:-( Moreover there is a nasty constraint that force ecutrho=4*ecutwfc, that is not exactly the best for US PPs. It might turn out that a NC/EXX calculation (always very satisfying in my hands) is significantly faster end quite reliable, even if you must use 80/320 Ry cutoffs (or even higher; first row transition metals with 2s,2p semicore electrons are usually the hardest NC PPs). You can try to speedup the calculation by using a) adaptive_thr=.true. In the case of smooth convergence this makes inner scf cycles significantly faster. b) ecutfock~[1/2-->1/4]*ecutrho Fantastic with isolated systems; handle with care with metals because this might change occupations at k-points. You should check convergence on this, but in the case of "almost metallic" systems (e.g., shallow defects in semiconductors) ecutfock=1/2*ecutrho is often well converged. c) remember that (at least in principle) you should check convergence with respect to the projection of density in real space: you are (implicitly) using one q-point for each k-point, having not assigned any value to nqx1, nqx2, nqx3. More q points increase the computational effort. You might try to reduce the k grid and to increase nqx(i) values to find a correct balance: However a 7 7 7 0 0 0 grid is actually large. You will not be able to perform 'nscf' or 'bands' calculations to have a proper band structure, and densities/forces/stresses/eigenvalues/energies should converge at an acceptable level with less k points. HTH Giuseppe On Friday, September 09, 2016 10:42:49 AM Daniel Stoeffler wrote: > Hello Stefano, > > As said, I am a new user of QE but not of the DFT > ;). I use the FLEUR, Abinit and the VASP codes and I want to test if it > would be interesting for my studies to use also QE. > > I said > "surprisingly" huge because I compared to the VASP code with the same > parameters (Ecutoff, NKpoints, Nq) and with the HSE functional. Indeed, > with the VASP code during one night on my computer I got a few tens of > iterations compared to one with QE :(. > > >From my point of view, I > > compare 2 pseudopotential-based methods for the same parameters which I > can control via the input file : my question would then more preferably > be "which default parameter should I modify to make the QE calculation > time similar to the VASP one ?" > > Best wishes, > > Daniel > > Le 2016-09-09 > > 09:47, Stefano de Gironcoli a écrit : > > Please read the literature > > about hybrid functionals in plane waves; understand why the > computational time is, not unexpectedly, huge and start think about it. > > > If you come up with some smart idea on how to reduce the > > computational time without compromising accuracy (as for the ACE > transformation recently proposed by Lin Lin and implemented in the > latest QE version) let us know > > > Best, > > stefano > > (sent from my phone) > > > > > > On 09 Sep 2016, at 09:30, Daniel Stoeffler > > wrote: > >> Hello, > >> > >> As a new > > QE user, I do some tests with the HSE Functional for BCC ferromagnetic > Fe with a cell containing 2 atoms (because a want to go to CoFe after > that). > > >> My input file is : > >> > >> &control > >> title='Fe CC > > HSE', > > >> calculation = 'scf', > >> prefix = 'Fe_CC_HSE', > > pseudo_dir='./', > > >> verbosity = 'high' > >> / > >> &system > >> ibrav = 6, > > celldm(1) = 5.45, > > >> celldm(3) = 1.0, > >> nat = 2, > >> ntyp = 1, > > nspin = 2, > > >> occupations='smearing', smearing='gauss', degauss=0.01, > > starting_magnetization(1) = 1.0, > > >> ecutwfc = 33, > >> input_dft='HSE', > >> > >> report = 5 > >> / > >> &electrons > >> mixing_beta = 0.7 > >> / > > ATOMIC_SPECIES > > >> Fe 55.845 Fe.pbe-sp-van.UPF > >> > >> ATOMIC_POSITIONS > > (crystal) > > >> Fe 0.000 0.000 0.000 > >> Fe 0.500 0.500 0.500 > > K_POINTS (automatic) > > >> 7 7 7 0 0 0 > >> > >> Surprisingly, the computation > > time is huge as writen into the output file : > > > === > > >> iteration # 9 ecut= 33.00 Ry > > beta=0.70 > > >> Davidson diagonalization with overlap > >> ethr = 5.49E-08, > > avg # of iterations = 2.3 > > >> Magnetic moment per site: > >> atom: 1 > > charge: 14.4350 magn: 2.6346 constr: 0. > > >> atom: 2 charge: 14.4351 > > magn: 2.6343 constr: 0. > > >> total cpu time spent up to now is 7.2 > > secs > > >> End of self-consistent calculation > >> > >> convergence has > > been achieved in 9 iterations > > >> EXX: now go back to refine exchange > > calculation > > >> total cpu time spent up to now is 4239.4 secs > > Self-consistent Calculation > > >> iteration # 1 ecut= 33.00 Ry > > beta=0.70 > > >> Davidson diagonalization with overlap > >> ethr = 5.49E-08, > > avg # of iterations = 4.9 > > >> total cpu time spent up to now is > > 52490.2 secs > > >> iteration # 2 ecut= 33.00 Ry beta=0.7
Re: [Pw_forum] testing HSE for BCC Fe
On 09/09/2016 11:37, dario rocca wrote: > -If you changed the default values of some variables such as Precfock or > Nkred in Vasp the computational parameters are not anymore the same in > Vasp and QE. Ciao Dario, all true, but at the end what counts for a expert in DFT but not in computational implementations is "time-to-accurate-solution". This latter requires time in itself to establish (i.e. what are the parameters that give me a converged result - say, for non symmetric configurations, with forces that are converged to .XXX eV/A), but for a DFT scientist that is a user of codes, rather than a developer of codes, the question above (let's call it TASK "time-to-accurate-solution knowing own CPU budget") is a very important one. nic -- Prof Nicola Marzari, Chair of Theory and Simulation of Materials, EPFL Director, National Centre for Competence in Research NCCR MARVEL, EPFL http://theossrv1.epfl.ch/Main/Contact http://nccr-marvel.ch/en/project ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum
Re: [Pw_forum] testing HSE for BCC Fe
Ciao Nicola, I agree that 'time-to-accurate-solution' is the goal. However, when you have a variable as in Vasp PRECFOCK= Low | Medium | Fast | Normal | Accurate that allows you to do a HF calculation in 10 hours if PREFOCK=Low and in an infinite time if PREFOCK=Accurate, I see the time saving part but the accuracy can be hardly tested. And it makes also hard to do a comparison of performance and results with other codes. Best, Dario On Fri, Sep 9, 2016 at 11:55 AM, Nicola Marzari wrote: > On 09/09/2016 11:37, dario rocca wrote: > > -If you changed the default values of some variables such as Precfock or > > Nkred in Vasp the computational parameters are not anymore the same in > > Vasp and QE. > > > Ciao Dario, > > > all true, but at the end what counts for a expert in DFT but not in > computational implementations is "time-to-accurate-solution". > > This latter requires time in itself to establish (i.e. what are the > parameters > that give me a converged result - say, for non symmetric configurations, > with forces that are converged to .XXX eV/A), but for a DFT > scientist that is a user of codes, rather than a developer of codes, > the question above (let's call it TASK "time-to-accurate-solution > knowing own CPU budget") is a very important one. > > nic > > > -- > Prof Nicola Marzari, Chair of Theory and Simulation of Materials, EPFL > Director, National Centre for Competence in Research NCCR MARVEL, EPFL > http://theossrv1.epfl.ch/Main/Contact http://nccr-marvel.ch/en/project > ___ > Pw_forum mailing list > Pw_forum@pwscf.org > http://pwscf.org/mailman/listinfo/pw_forum > ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum
Re: [Pw_forum] testing HSE for BCC Fe
Hello Dario, This is what I guessed also and this is why I was surprised by the computation time of QE ;). I am trying tqr=.true. Because BCC Fe is metallic and magnetic, I did not modify the NKRED and PRECFOCK default values as suggested into the VASP manual. So, it is nearly the most computation time case for VASP (the time would be larger only by changing PRECFOCK to Accurate). Best wishes, Daniel Le 2016-09-09 11:37, dario rocca a écrit : > Dear Daniel > >> I said "surprisingly" huge because I compared to the VASP code with the same parameters (Ecutoff, NKpoints, Nq) and with the HSE functional. Indeed, with the VASP code during one night on my computer I got a few tens of iterations compared to one with QE :(. > > With exactly the same parameters the computational time (on one processor) should be comparable. There are two important points you should consider to make a comparison: > -Different type of methodology: Using EXX with Ultrasoft pseudopotentials a very large number of augmentation charges are computed. As already suggested setting tqr=.true. should solve this problem. > -If you changed the default values of some variables such as Precfock or Nkred in Vasp the computational parameters are not anymore the same in Vasp and QE. > Best, > Dario > >> From my point of view, I compare 2 pseudopotential-based methods for the same parameters which I can control via the input file : my question would then more preferably be "which default parameter should I modify to make the QE calculation time similar to the VASP one ?" >> >> Best wishes, >> >> Daniel >> >> Le 2016-09-09 09:47, Stefano de Gironcoli a écrit : >> >>> Please read the literature about hybrid functionals in plane waves; understand why the computational time is, not unexpectedly, huge and start think about it. >>> If you come up with some smart idea on how to reduce the computational time without compromising accuracy (as for the ACE transformation recently proposed by Lin Lin and implemented in the latest QE version) let us know >>> Best, >>> stefano >>> (sent from my phone) >>> >>> On 09 Sep 2016, at 09:30, Daniel Stoeffler wrote: >>> Hello, As a new QE user, I do some tests with the HSE Functional for BCC ferromagnetic Fe with a cell containing 2 atoms (because a want to go to CoFe after that). My input file is : &control title='Fe CC HSE', calculation = 'scf', prefix = 'Fe_CC_HSE', pseudo_dir='./', verbosity = 'high' / &system ibrav = 6, celldm(1) = 5.45, celldm(3) = 1.0, nat = 2, ntyp = 1, nspin = 2, occupations='smearing', smearing='gauss', degauss=0.01, starting_magnetization(1) = 1.0, ecutwfc = 33, input_dft='HSE', report = 5 / &electrons mixing_beta = 0.7 / ATOMIC_SPECIES Fe 55.845 Fe.pbe-sp-van.UPF ATOMIC_POSITIONS (crystal) Fe 0.000 0.000 0.000 Fe 0.500 0.500 0.500 K_POINTS (automatic) 7 7 7 0 0 0 Surprisingly, the computation time is huge as writen into the output file : === iteration # 9 ecut= 33.00 Ry beta=0.70 Davidson diagonalization with overlap ethr = 5.49E-08, avg # of iterations = 2.3 Magnetic moment per site: atom: 1 charge: 14.4350 magn: 2.6346 constr: 0. atom: 2 charge: 14.4351 magn: 2.6343 constr: 0. total cpu time spent up to now is 7.2 secs End of self-consistent calculation convergence has been achieved in 9 iterations EXX: now go back to refine exchange calculation total cpu time spent up to now is 4239.4 secs Self-consistent Calculation iteration # 1 ecut= 33.00 Ry beta=0.70 Davidson diagonalization with overlap ethr = 5.49E-08, avg # of iterations = 4.9 total cpu time spent up to now is 52490.2 secs iteration # 2 ecut= 33.00 Ry beta=0.70 Davidson diagonalization with overlap === So, the calculation takes 7 seconds for converging with PBE, 1h10min for calculating the EXX term and 14h35min for the first HSE iteration. Is this correct ? And if so, please, could you try to explain me how to reduce the computation time. Thanks in advance. Best wishes, Daniel >>> ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum [1] >>> >>> ___ >>> Pw_forum mailing list >>> Pw_forum@pwscf.org >>> http://pwscf.org/mailman/listinfo/pw_forum [1] >> >> ___ >> Pw_forum mailing list >> Pw_forum@pwscf.org >> http://pwscf.org/mailman/listinfo/pw_forum [1] > > ___ > Pw_forum mail
