[Pw_forum] efficient parallelization of phonon jobs without manual distribution (GRID)

Christoph Wolf(신소재공학과) chwolf at postech.ac.kr
Wed Aug 17 09:43:35 CEST 2016


Dear All!

I am currently looking into phonon calculations and I am a bit confused by the available methods of parallelization and their applicability to my system.

I read chapter 5 (Parallelism) http://www.quantum-espresso.org/wp-content/uploads/Doc/ph_user_guide/node10.html and I briefly checked the paper on the GRID example.

The machine I want to run the job on is a local “supercomputer” (25,408 Intel Xeon CPUs on 3,176 nodes) so I guess resources should not be a problem - unfortunately I am the only person in my group doing this and I am pretty clueless…

I am specifically trying to find out how to choose -np -ni and -nk, so that the calculation doesn’t complain about some processor not having a wavefunction or k-point.

My unit cell is cubic and has 5 atoms of 3 types. I calculate scf on a fine (12 12 12 0 0 0) and coarse (6) k-grid for la2f calculations. Then I use ph.x on a (6 6 6) q-grid (I am following a tutorial I found on this mailing list).

During the ph.x test- run on a single cpu for testing I find that

     There are   5 irreducible representations
     There are  10 irreducible representations
     There are  10 irreducible representations
     There are  10 irreducible representations
     There are  15 irreducible representations

And

     number of k points=    20
     number of k points=   120
     number of k points=   120
     number of k points=   120

irreps (corresponding to images in parallelization) and k-points (npools) constantly change, so I do not understand how to efficiently parallelize this kind of job when submitting, especially since I do not have any a-priori knowledge of the change of these quantities.

Further http://training.uhem.itu.edu.tr/docs/18hazirannano/PW-III-para.pdf notes that npool must be a divider of Nk. So the lowest common denominator in above calculation would be: -npool 20 and -nimage 5 ?

Any comment is greatly appreciated, I am really lost on this!

Yours,
Chris

Christoph Wolf
Postech University, Department of Materials Science and Engineering
Pohang, South Korea
[http://home.postech.ac.kr/wp-content/themes/postech/images/mail_logo_postech30.png]
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20160817/4b732d26/attachment.html>


More information about the users mailing list