Fwd: [Pw_forum] CP breakdown for long simulations?

Nicola Marzari marzari at MIT.EDU
Fri Oct 20 00:34:26 CEST 2006


Thanks to Carlo Cavazzoni for the explanation !

The reason is that the lambda of the ortho constraints
is calculated independently on each processor, and that
can induce a numerical drift in different processors.

The lambda is broadcast to all other processors every time there
is a wavefunction write, but if one has a low frequency of restarts
the drift can raise its ugly head. (Not to mention that our
frequency of restarts was 100, that is quite frequent).

Suggestion: since *especially* on massively-parallel run
(like we do here - 4 cpus, and sometimes even 8 cpus) you
do not want to save the wfcs very often, and since the ortho
is very important, can't it be broadcast at every iteration,
from the root processor to all others ?

				nicola

>>>> Hi all,
>>>>
>>>> I'm running the CVS version of CP, and I'm wondering if anyone else  
>>>> has come across this problem.  After ~10,000-15,000 timesteps of  
>>>> good data for a well-behaved, insulating system, the total  
>>>> integrated electronic charge density is suddenly no longer  
>>>> conserved and begins to fluctuate.  As a result, the total energy  
>>>> drops rapidly and the simulation goes haywire.
>>>>
>>>> It seems like it is due to compounding of small numerical errors.   
>>>> Could there be a problem with the orthonormalization scheme that  
>>>> carries over from timestep to timestep but only manifests itself  
>>>> after sufficient time, for instance?
>>>>
>>>> Thanks,
>>>> Brandon

> 


-- 
---------------------------------------------------------------------
Prof Nicola Marzari   Department of Materials Science and Engineering
13-5066   MIT   77 Massachusetts Avenue   Cambridge MA 02139-4307 USA
tel 617.4522758 fax 2586534 marzari at mit.edu http://quasiamore.mit.edu



More information about the Pw_forum mailing list