Different scf behavior between serial and parallel ABINIT

Total energy, geometry optimization, DFT+U, spin....

Moderator: bguster

Locked
htakenaka
Posts: 4
Joined: Thu Jun 18, 2015 4:59 pm

Different scf behavior between serial and parallel ABINIT

Post by htakenaka » Wed Oct 07, 2015 6:00 pm

Dear all,

I am relaxing structures for barium titanate using ABINIT-7.10.4.

When I use serial and parallel ABINIT with npband=4, npfft=2, npkpt=4, and para_kgb=1, the number of scf steps with serial ABINIT is larger than that with parallel ABINIT.
I expected that scf convergence would be identical.

I am wondering that this could be a problem when I built ABINIT or I need to use some input variables.

Could someone please give me instruction?

Best regards,

Hiroyuki Takenaka

ecutsm 0.75
dilatmx 1.1
natom 5
ntypat 3
znucl 56 22 8
typat 1*1 1*2 3*3
nband 28


acell 7.5207394431E+00 7.5207394431E+00 7.7203843581E+00 Bohr
xred 0.0000000000E+00 0.0000000000E+00 -3.9146615751E-03
5.0000100000E-01 5.0000000000E-01 4.7857472093E-01
5.0000000000E-01 5.0000000000E-01 2.7891397150E-02
5.0000000000E-01 0.0000000000E+00 5.1272427177E-01
0.0000000000E+00 5.0000000000E-01 5.1272427177E-01

symmorphi 0
shiftk 0.5 0.5 0.5


ngkpt 4 4 4
occopt 1

pawecutdg 56
ecut 40
ixc 23
nstep 100
kptopt 1
iscf 17

#optimization
ionmov 2
optcell 2
ntime 200

toldff 1.0d-9
tolmxf 1.0d-8

Jordan
Posts: 282
Joined: Tue May 07, 2013 9:47 am

Re: Different scf behavior between serial and parallel ABINI

Post by Jordan » Mon Oct 12, 2015 8:13 am

in sequential, the algorithm is the conjugate gradient whereas with parallelization over fft nband and kpt you use lobpcg which use an other scheme.
Therefore you can not expect the same convergence behavior. Indeed, lobpcg might need less SCP steps but each step may be slower.
At convergence, both must give the same ground same.

Jordan

htakenaka
Posts: 4
Joined: Thu Jun 18, 2015 4:59 pm

Re: Different scf behavior between serial and parallel ABINI

Post by htakenaka » Mon Oct 12, 2015 9:19 pm

Dear Jordan,

Thank you very much.

Just to make sure, I have a question.

The jobs were done with PAWs.
According to log files, wfoptalg = 10, which is standard state by state conjugate gradient algorithm, was used for MPI and serial ABINIT calculations.
While I used "wfoptalg 4" in an input file, ABINIT automatically set wflptalg = 10 in the log file.

Even for wfoptalg = 10, will LOBPCG algorithm be used ?

Best regards,

Hiroyuki Takenaka

htakenaka
Posts: 4
Joined: Thu Jun 18, 2015 4:59 pm

Re: Different scf behavior between serial and parallel ABINI

Post by htakenaka » Mon Oct 12, 2015 9:25 pm

Dear Jordan,

Sorry, I misunderstood my log file in.

LOBPCG algorithm,wfoptalg = 14, was used.

Best regards,

Hiroyuki Takenaka

htakenaka
Posts: 4
Joined: Thu Jun 18, 2015 4:59 pm

Re: Different scf behavior between serial and parallel ABINI

Post by htakenaka » Tue Oct 13, 2015 7:47 pm

Dear Jordan,

I carefully looked into log files of serial and MPI runs.

All of their structures are identical at the end of relaxations.

Thank you very much!

Best regards,

Hiroyuki Takenaka

Locked