[Pw_forum] converged to wrong magnetic state of hcp cobalt in Opteron
Nicola Marzari
marzari at MIT.EDU
Fri Apr 14 10:58:50 CEST 2006
Dear William,
I haven't really followed the discussion, but
1) the amount of smearing you use is always an important
parameter to asses, and it is strictly linked to the
number of k-points you use (i.e. for very small smearings,
in a metal, you need to use a lot of k-points. That's the
whole point - increasing the smearing can get you to a more
accurate result, if you know what you are doing).
2) if you are using the same smearing, and same input,
on two different machines, and you get two different results,
there maybe some good reasons (two different physical minima, reached
through different initizializations of wavefunctions; or
one machine is broken; or...), but the fact that the smearing is
finite is not one of the reasons, strictly speaking. The potential
energy surface of the system, if the machines work properly,
should be the same", with one or many minima. The minimum in
which you end up will depend on where you start, and how you proceed,
that can have an amount of randomness, or numerical noise.
nicola
wlyim at puccini.che.pitt.edu wrote:
> Dear Kostya and Axel,
>
> Thanks for all valuable discussions.
>
> Eventually I figured out that it was not the platform problem, but the
> smearing parameter. Cold smearing was supposed to give reasonable result
> at larger smearing value. It was true for Pt, Pd, Ni those I have tested.
> However, it is not true for cobalt. After I changed smearing parameter to
> 0.007 Ryd, Opteron and Origin gave almost the same results.
> ================================================
> Opteron:
> ! total energy = -148.83154364 ryd
> estimated scf accuracy < 2.6E-12 ryd
>
> band energy sum = 15.10288935 ryd
> one-electron contribution = 5.98708540 ryd
> hartree contribution = 21.83747050 ryd
> xc contribution = -65.80523901 ryd
> ewald contribution = -110.85073891 ryd
> correction for metals = -0.00012162 ryd
>
> total magnetization = 3.15 Bohr mag/cell
> absolute magnetization = 3.56 Bohr mag/cell
> =================================================
> Origin:
> ! total energy = -148.83189344 ryd
> estimated scf accuracy < 1.0E-12 ryd
>
> band energy sum = 15.10215635 ryd
> one-electron contribution = 5.98854047 ryd
> hartree contribution = 21.83686466 ryd
> xc contribution = -65.80577746 ryd
> ewald contribution = -110.85073891 ryd
> correction for metals = -0.00078219 ryd
>
> total magnetization = 3.16 Bohr mag/cell
> absolute magnetization = 3.57 Bohr mag/cell
> =======================================================
>
> It seems to me that the term "correction for metals" is a very important
> parameter to be checked...
>
> Thanks again.
>
> Best regards,
> William
>
> On Thu, 13 Apr 2006, Konstantin Kudin wrote:
>
>
>> Hi all,
>>
>> I'd like to point out that as long as the code converges to a proper
>>*local* minimum, all is OK. One cannot really expect that the code is
>>going to find the *global* minimum in a system where there are many
>>accessible low energy minima. The way things are implemented is to find
>>*a* minimum, not *the* minimum! Note that finding *the* minimum is a
>>quite tricky issue, and not a solved one.
>>
>> Different platforms (compilers/architectures) might have different
>>initial random wavefunctions, leading to different local minima. Is
>>that a sign of a problem? - Not at all!
>>
>> Until things improve, making sure that the electronic state obtained
>>is lowest in energy (i.e. the global minimum) is the operator's
>>responsibility.
>>
>> Kostya
>>
>>
>>--- wlyim at puccini.che.pitt.edu wrote:
>>
>>
>>>Dear Axel,
>>>
>>>Thanks for your advice.
>>>
>>>I have compiled pwscf.2.1.4 using EM64T Version 9.0 with MKL 8.0.1.
>>>Again,
>>>the electronic structure went to the wrong state.
>>>==================================================
>>>INTEL EM64T result:
>>>! total energy = -148.81668356 ryd
>>> estimated scf accuracy < 2.0E-12 ryd
>>>
>>> band energy sum = 15.05020564 ryd
>>> one-electron contribution = 6.02718174 ryd
>>> hartree contribution = 21.83280277 ryd
>>> xc contribution = -65.85608487 ryd
>>> ewald contribution = -110.85073891 ryd
>>> correction for metals = 0.03015572 ryd
>>>
>>> total magnetization = 3.75 Bohr mag/cell
>>> absolute magnetization = 4.14 Bohr mag/cell
>>>=======================================================
>>>
>>>It seemed to me that in this chemical system, the energy is
>>>constantly
>>>shifted by 0.06 Ry when compared to Origin result. I pick up the
>>>first
>>>iteration step for comparison:
>>>================================================
>>>Origin:
>>> iteration # 1 ecut= 30.00 ryd beta=0.70
>>> Davidson diagonalization (with overlap)
>>> ethr = 1.00E-02, avg # of iterations = 4.2
>>>
>>> npt with rhoup < 0: 2, npt tot 20736, 0.01 %
>>>
>>> npt with rhoup < 0: 2, npt tot 20736, 0.01 %
>>>
>>> total cpu time spent up to now is 15.75 secs
>>>
>>> total energy = -147.96340174 ryd
>>> estimated scf accuracy < 2.11694003 ryd
>>>
>>> total magnetization = 5.92 Bohr mag/cell
>>> absolute magnetization = 5.97 Bohr mag/cell
>>>=================================================
>>>Opteron(PGI) or INTEL(EM64T) result:
>>> iteration # 1 ecut= 30.00 ryd beta=0.70
>>> Davidson diagonalization (with overlap)
>>> ethr = 1.00E-02, avg # of iterations = 4.1
>>>
>>> npt with rhoup < 0: 2, npt tot 20736, 0.01 %
>>>
>>> npt with rhoup < 0: 2, npt tot 20736, 0.01 %
>>>
>>> total cpu time spent up to now is 30.39 secs
>>>
>>> total energy = -147.90474844 ryd
>>> estimated scf accuracy < 2.13792299 ryd
>>>
>>> total magnetization = 5.95 Bohr mag/cell
>>> absolute magnetization = 6.00 Bohr mag/cell
>>>======================================================
>>>
>>>At the initial stage, I would expect that the "total energy" are
>>>quite
>>>similar calculated in two different platforms. However this is not
>>>the
>>>case. Origin's energy is about 0.06 Ry lower than that in Opteron.
>>>And
>>>such a difference was kept till the end. I can't imagine what
>>>reasoned for
>>>this. On the other hand, platinum is free of this problem.
>>>
>>>Is it due to different fft treatments in Origin and Linux machines?
>>>Or
>>>code/library problem?
>>>
>>>Many thanks in advance.
>>>
>>>Best regards,
>>>William
>>>
>>>On Thu, 13 Apr 2006, Axel Kohlmeyer wrote:
>>>
>>>
>>>>On Thu, 13 Apr 2006 wlyim at puccini.che.pitt.edu wrote:
>>>>
>>>>WL> Dear all,
>>>>
>>>>dear william,
>>>>
>>>>please check the mailing list archives. the PGI compilers version
>>>
>>>5.x
>>>
>>>>are _known_ to miscompile several parts of quantum espresso (and
>>>
>>>not
>>>
>>>>only that code, i know of several other DFT codes as well).
>>>>the same is true for quite few SGI compilers...
>>>>if you can, please try with the intel (EM64t) compilers on opteron
>>>>to check, whether it is still going to the wrong state. it
>>>
>>>additionally
>>>
>>>>may depend on the kind of BLAS/LAPACK library you are using. there
>>>
>>>are
>>>
>>>>a few bugs in several of them as well.
>>>>
>>>>best regards,
>>>> axel.
>>>>
>>>>WL>
>>>>WL> I did a test on crystalline hcp cobalt by pwscf.2.1.4, both on
>>>
>>>Opteron
>>>
>>>>WL> (pgf90 5.2-4, lam 7.1.1) and Origin. I found that the
>>>
>>>calculation
>>>
>>>>WL> converged to a wrong magnetic state in Opteron, while the
>>>
>>>calculation in
>>>
>>>>WL> Origin was okay.
>>>>
>>>>[...]
>>>>
>>>>WL> I used the PBE uspp as provided in pwscf website. On the other
>>>
>>>hand, I
>>>
>>>>WL> have also tried the calculations on platinum, which is a
>>>
>>>closed-shell
>>>
>>>>WL> system, and I got essentially the same results from Opteron and
>>>
>>>Origin.
>>>
>>>>WL>
>>>>WL> Does anyone experience a similar thing? I will be appreciated
>>>
>>>if someone
>>>
>>>>WL> can give a help hand.
>>>>WL>
>>>>WL> Many thanks in advance!
>>>>WL>
>>>>WL> I am also attaching the input file at the bottom of this
>>>
>>>meesage.
>>>
>>>>WL>
>>>>WL> regards,
>>>>WL> William
>>>>WL>
>>>>
>>>>[...]
>>>>
>>>>
>>>
>>>
>>>_______________________________________________
>>>Pw_forum mailing list
>>>Pw_forum at pwscf.org
>>>http://www.democritos.it/mailman/listinfo/pw_forum
>>>
>>
>>
>>__________________________________________________
>>Do You Yahoo!?
>>Tired of spam? Yahoo! Mail has the best spam protection around
>>http://mail.yahoo.com
>>_______________________________________________
>>Pw_forum mailing list
>>Pw_forum at pwscf.org
>>http://www.democritos.it/mailman/listinfo/pw_forum
>>
>
>
>
> _______________________________________________
> Pw_forum mailing list
> Pw_forum at pwscf.org
> http://www.democritos.it/mailman/listinfo/pw_forum
--
---------------------------------------------------------------------
Prof Nicola Marzari Department of Materials Science and Engineering
13-5066 MIT 77 Massachusetts Avenue Cambridge MA 02139-4307 USA
tel 617.4522758 fax 2586534 marzari at mit.edu http://quasiamore.mit.edu
More information about the users
mailing list