[ase-users] speeding up thermochemistry calculations

[Claudio A. Perottoni]

Dear Eric,

The performance issue I referred to may be observed even in the example 
for gold bulk provided at 
https://wiki.fysik.dtu.dk/ase/ase/thermochemistry/thermochemistry.html 
and listed below:

from  ase.spacegroup  import  crystal
from  ase.calculators.emt  import  EMT
from  ase.optimize  import  QuasiNewton
from  ase.phonons  import  Phonons
from  ase.thermochemistry  import  CrystalThermo

# Set up gold bulk and attach EMT calculator
a  =  4.078
atoms  =  crystal('Au',  (0.,  0.,  0.),
                 spacegroup=225,
                 cellpar=[a,  a,  a,  90,  90,  90],
                 pbc=(1,  1,  1))
calc  =  EMT()
atoms.set_calculator(calc)
qn  =  QuasiNewton(atoms)
qn.run(fmax=0.05)
potentialenergy  =  atoms.get_potential_energy()

# Phonon analysis
N  =  5
ph  =  Phonons(atoms,  calc,  supercell=(N,  N,  N),  delta=0.05)
ph.run()
ph.read(acoustic=True)
phonon_energies,  phonon_DOS  =  ph.dos(kpts=(40,  40,  40),  npts=3000,
                                      delta=5e-4)

# Calculate the Helmholtz free energy
thermo  =  CrystalThermo(phonon_energies=phonon_energies,
                        phonon_DOS=phonon_DOS,
                        potentialenergy=potentialenergy,
                        formula_units=4)
F  =  thermo.get_helmholtz_energy(temperature=298.15)

This example takes about one minute on a 12-core shared-memory machine. 
However fast, it seems to take longer than expected to perform the 
(mostly) linear algebra operations after calculation of the numerical 
derivatives and creation of the phonon.*.pckl files. I observed the same 
on my calculations: creation of phonon.*.pckl files (in my case 
calculations are performed in parallel using LAMMPS/OpenMP) is 
relatively fast while the last portion of the code (which, I believe, 
uses numpy) is taking longer than expected (in fact, calculations using 
numpy compiled against OpenBLAS are taking longer when using 
OMP_NUM_THREADS=12 than OMP_NUM_THREADS=1). I know that sometimes 
running in parallel will take longer than running on a single core. 
However, I would expect that after generation of the dynamical matrix 
(which is relatively fast) the remaining calculations should be much 
faster...

Thank you very much for your interest in this issue!

Best regards,

Claudio


Em 20-Mar-17 11:05, Eric Hermes via ase-users escreveu:
> On Mon, 2017-03-20 at 09:28 -0300, Claudio A. Perottoni via ase-users
> wrote:
>> Dear ASE Users,
>>
>> I am using ase.thermochemistry.CrystalThermo class to calculate the
>> Helmholtz free energy of crystals and I wonder is there any way to
>> speed
>> up these calculations. It is taking over a day to calculate the
>> Helmholtz free energy of a 64 atom unit cell, replicated to a 5x5x5
>> supercell, in a 20-core shared-memory machine using numpy compiled
>> against multithreading OpenBLAS.  Comments on this execution time
>> and
>> how to improve this performance  would be very much appreciated!
>>
>> Best regards,
>>
>> Claudio
> Claudio,
>
> Can you please provide an example script that exhibits this problem? If
> possible, a smaller example (one that still takes longer to run than
> expected) would be preferable.
>
> Eric
>
> _______________________________________________
> ase-users mailing list
> ase-users at listserv.fysik.dtu.dk
> https://listserv.fysik.dtu.dk/mailman/listinfo/ase-users

-- 
*********************************************************************
Cláudio A. Perottoni

Universidade de Caxias do Sul
Programa de Pós-Graduação em Engenharia e Ciência dos Materiais
Rua Francisco Getúlio Vargas, 1130
95070-560 Caxias do Sul - RS - Brazil
Phone: +55 54 3218-2607

Universidade de Caxias do Sul
IMC - Instituto de Materiais Cerâmicos
Rua Irmão Moretto, 75
95765-000  Bom Princípio - RS - Brazil
Phone: +55 51 3634-1100

http://www.ucs.br/ccet/defq/caperott/
*********************************************************************



---
This email has been checked for viruses by Avast antivirus software.
https://www.avast.com/antivirus
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listserv.fysik.dtu.dk/pipermail/ase-users/attachments/20170320/b8b05ffc/attachment-0001.html>