The ABINIT Users Mailing List ( CLOSED )

[<6671011@163.com>]

Dear Matthieu,

I did some energy-cut-off convergence tests, but nothing changed. After
comparing with the files which give consistent results, I got the cause:  I
used nonscf in the Phonon+Efield calculation for the two-atom system. After
correcting it, everthing works fine.

Then, my next question is: do you have any plan for implementing this
parallelization strategy---"dividing perturbation-atoms into groups" ---into
the future abinit?  (In my opinion, this is a both simple and efficient method
to speed up the RF calculation)


Sincerely,
Gunagfu Luo


----------------------------------------------------------------------------------------------------
Hello Guangfu Luo,

First of all, I don't understand why it would be different to send an
attachment to the forum. Several other people have said the same, but it is
just an email address...

Anyhow, having looked at your (very systematic and clear) outputs, and
comparing the tnlo_2.out files, you can see at the end the block  ==> Compute
Derivative Database <==. This contains the 2dTE for all the perturbations 
which
were actually calculated, _plus_ those which come from non-stationary
expressions (NSE) of the 2dTE (see Xavier's prb_55_10337 and following paper
with Changyol Lee). 

The NSE allow you to calculate the derivative of E_tot wrt R_alpha and R_beta,
if you are given just the derivatives of the potentials wrt R_alpha only 
(which
abinit just calculated). In this way, from just the perturbations wrt 1 atom,
you can complete whole rows and columns of the full interatomic force constant
(IFC) matrix.

However, the non-stationary expressions are less precise, and in your case I
believe that they are completely off (sometimes by a factor of 2), probably
because of the very low ecut (I can't see any other unconverged parameters). 
As
a result, all of the calculations you did (Usual, and merged) are roughly at
the same level of convergence, but the way the IFC matrix is completed was
different, because of different errors in the NSE, and they give different
results at the end (during the mergeddb the 2dTE are now overcomplete, and I
don't know which values mrgddb will choose to prefer - those coming from Part1
or Part2, or an average...).

Could you try an ecut-converged calculation to see if this is indeed the case?

Matthieu



---------------------------------------------------------------------------------------------------
From: Guangfu Luo <6671011@163.com>
Date: 2009/3/7
Subject: Re: [abinit-forum] Possible method to speed RF calculation?
To: matthieu.jean.verstraete@gmail.com



Dear Matthieu,

 

Because I do not know how to attach enclosure to the Forum, I send the 
two-atom
test directly to you. The files come from the built-in tests after a little
revision. 

 

 The structure of the tar file is,

(1) UsualMethod means "rfatpol 1 2"

 

(2) Part1 and Part2 mean "rfatpol 1 1" and "rfatpol 2 2", respectively.

(3) MergePart1-2_andthen_AnalyseIR means merge the DDBs from Part1 and Part2
and then analyze the IR.

 

I compared their spectra of "Oscillator strengths", and found (1) show two
distinct peaks, while the (2)-(3) method shows only one peak (or two nearly
degenerate peaks). 

 

Thanks

 
 
Sincerely,
Guangfu Luo


 
---------------------------------------------------------------------------------------------------
2009-03-08，"matthieu verstraete" <matthieu.jean.verstraete@gmail.com>
wrote：


In principle yes, 2 atoms should be the same. Send in your in/output for the
test cases.

Matthieu


----------------------------------------------------------------------------------------------------
On Sat, Mar 7, 2009 at 6:33 PM, <6671011@163.com> wrote:

Dear friends,

During a RF calculation, if one wants to do the perturbations of n atoms,
usually he/she calculates them in a single file by using rfatpol 1 n.


I wonder if we can speed the RF calculation by splitting it into several
separate files, such as
    file1: rfatpol 1  i
    file2: rfatpol i+1 j
    file3: rfatpol j+1 n
Then the files can be submitted at the same time; when they finish, one can
merge their DDBs.


I tested the forementioned methods in several systems, and most of them give
almost the same results, except a two-atom system.  My question is: are the 
two
ways equivalent in principle?


Thanks.

Sincerely,
Gunagfu Luo