[Pw_forum] How to estimate the memory requirement of ph.x?

lfhuang lfhuang at theory.issp.ac.cn
Thu Dec 9 09:38:15 CET 2010

Dear QE developers and users:
     I am calculating the vibrational frequencies of a graphene supercell (90 atoms total). I used 24N_p = 6N_pk*4N_pr first, and the pw.x calculation completed perfectly with about 500 MB memory per-processor. However, the ph.x calculation always stopped after several hours at the place as below:
Atomic displacements:
     There are 270 irreducible representations
     Representation     1      1 modes -A To be done
     Representation     2      1 modes -A To be done
      Representation     270      1 modes -A To be done
STOPPED HERE due to the breakdown of some processors.  
If I use   24N_p = 24N_pr,the ph.x can also completed perfectly with about 80 MB (.=500/6) per-processor in the pw.x calculation. From DFPT, it is understandable that ph.x requires more memory than pw.x. But how to estimate it to be qualitatively precise? The architecture of the machine I am using is that:
One node has 8 processors with shared memory of 24GB (3GB per processor).
How can a 3GB processor breaks down when doing a ph.x job, when the requirement of it pw.x job is just 500MB? 
By the way, is it better to assign the N_pr to be m**2 (m=2,3,4...) than other values (e.g. 6, 10, 11, 14...) if I want to divide the R-&G-space in parallelization? 
 I'd like to highly appreciate any help from you!
Best Wishes! 
Yours Sincerely
L. F. Huang 
L.F.Huang(黄良锋) DFT and phonon physics
Add: Research Laboratory for Computational Materials Sciences,
Instutue of Solid State Physics,the Chinese Academy of Sciences,
P.O.Box 1129, Hefei 230031, P.R.China
Tel: 86-551-5591464-326(office)
Fax: 86-551-5591434
Our group: http://theory.issp.ac.cn

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20101209/0d2f2921/attachment.html>

More information about the users mailing list