bands parallelization for KSS computation

Moderators: gmatteo, bruneval, rigna

Post Reply
Zeila
Posts: 3
Joined: Fri Feb 19, 2010 12:34 am

bands parallelization for KSS computation

Post by Zeila » Tue May 18, 2010 2:15 am

Dear all,

I'm running KSS calculations for a system with a huge number of plane waves (from the log: mpw = 453787) and Number of Grid points for Fast Fourier Transform (nfft = 7873200, still from the log).
The calculation proceedes in serial mode, but it takes forever. For instance, for nbandkss 400, it does 4 non-selfconsistent iterations in 3 days.
Hence, I've tried to perform the calculations using the parallelization over bands (I have only 1 kpoint).
However, all the attempts failed: the code crashed without any error message :(((
I'm attaching here my input and the log file, for more details on the calculation.

Can anyone suggest how to use properly the parallelization over bands for the generation of KSS files?
On passing, I've performed the "KGB tests" and chosen the parallelization values from that list.

Thanks!
Zeila

NB: I'm running abinit 6.0.3 on altamira (one of the Barcelona Supercomputing Center machines).
Abinit has been compiled with xlf.

david.waroquiers
Posts: 138
Joined: Sat Aug 15, 2009 12:45 am

Re: bands parallelization for KSS computation

Post by david.waroquiers » Wed May 19, 2010 1:11 pm

Hi Zeila,

The parallelisation over bands is not (yet?) implemented for the generation of the KSS file. What you could do (and what I usually do) is to generate your KSS file in a two step procedure :

1) generate the wavefunction with parallelisation over bands with the number of processors you want
2) generate the KSS file by reading the wavefunction and setting nstep to 0 in sequential

David

User avatar
gmatteo
Posts: 269
Joined: Sun Aug 16, 2009 5:40 pm

Re: bands parallelization for KSS computation

Post by gmatteo » Thu May 20, 2010 1:30 pm

Hi Zeila,

see also

viewtopic.php?f=9&t=85

Note that the KSS generation starts only once the SCF cycle is converged and it only requires
a simple conversion of the final wavefunctions and the writing of some additional matrix elements.
Therefore I would say that the problem you are encountering is not due to the KSS generation but to
the paral_kgb algorithm!

Zeila
Posts: 3
Joined: Fri Feb 19, 2010 12:34 am

Re: bands parallelization for KSS computation

Post by Zeila » Sat May 22, 2010 5:18 pm

Dear David and Matteo,

thanks for the hints. I've then tried to run the computation of the WFK in parallel (kgb parallelization over bands) with the same input as for the KSS without the KSS options. However, the code crashed once more.
In the "out" file I have the following error message:

getcut: ERROR -
Choice of acell, ngfft, and ecut
===> basis sphere extends BEYOND fft box !
Recall that boxcut=Gcut(box)/Gcut(sphere) must be > 1.
Action : try larger ngfft or smaller ecut.
Note that ecut=effcut/boxcut**2 and effcut= 0.000000

I've tried a 'test' with a very small ecut, but I've got the same error.
I'm starting the calculations by reading a WFK generated in serial mode with less nbands: this shouldn't be a problem, isn't it?

Thanks in advance for the help!
Zeila

Post Reply