The ABINIT Users Mailing List ( CLOSED )

[Cem Sevik <sevik@fen.bilkent.edu.tr>]

Hi
I have compiled the abinit-4.6 with pgf90. (Platform = sun firev20z, it 
has two AMD opteron )
I have observed same error with Paul durin the calculation of tpaw1_1.in 
with tpaw1_x.files.
The calculation of first data set finished and the program gives 
Segmentation fault error and never starts the data set two.

I am testing the PAW implementation in Abinit 4.6.4 (and before that 4.6.3) and I have observed the same error for all the calculations, not just the relaxations. If you make a convergence test with a lot of datasets, the calculation will stop, giving a max memory error after a few datasets. It seems like there is an internal memory that always grows. 





Marc Torrent wrote:

> Hello Eric and Paul,
>
> I'm not afraid by the problems you encounter: the way to a stable 
> version (with new features: PAW, mixing, ...) is a long way, but we 
> only can reach the end by solving problems.
>
> 1) Concerning the problem on forces:
> ====================================
> I think this is not related to PAW.
> I recently have found a way to increase (by a factor of 10-100) the 
> convergence of forces in the case of the MIXING ON DENSITY (default 
> behaviour in PAW, iscf>=10). Present forces are not wrong but they 
> need a better convergence on energy to be converged themselves.
> What I can suggest to you (to verify my theory) is:
> - to send me your input file (I would be able to comment it)
> - to test the same run using a MIXING ON POTENTIAL (i.e. : iscf=4 or 
> iscf=7 ; not iscf=5 in PAW!).
> - to test the same run (mixing on density) using a sharper convergence 
> criterium (toldff, toldfe, etc...) at each ionic step.
> - to report the results of the above tests.
>
> 2) Concerning the memory problem:
> =================================
> We have to solve it.
> I propose that we work together on the problem:
>
> - For you: try to investigate a little bit more the problem ; in 
> particular, try to determine whether the memory increases between:
>   * two SCF cycleS ?
>   * two ionic steps ?
>   * two different datasets ?
>
> - For me: I will use debbuging/profiling tools to investigate the 
> problem. But I need all the files you used to be able to reproduce the 
> behaviour. Send me your input files and give me the exact reference of 
> your atomic data files.
>
> Cheers,
>
> Marc