[CP2K-user] A single point calculation fails in a vibrational analysis

[Matt W]

Yes, we should add this to the FAQ. There is NPROC_REPLICA 
<https://manual.cp2k.org/cp2k-6_1-branch/CP2K_INPUT/VIBRATIONAL_ANALYSIS.html#list_NPROC_REP> 
keyword which says how many cores run on each single point job. If you run 
with number of cores equal to 6*Natoms*NPROC_REPLICA it should run in one 
shot (excepting time to diagonalise the Hessian which won't fully 
parallelise but shouldn't be significant).

Matt

On Saturday, February 15, 2020 at 4:05:35 PM UTC, Stephen Vicchio wrote:
>
> Is there a keyword to set the parallelization for these vibrational jobs? 
> I ran a small vibrational calculation (19 single point calculations) the 
> other day and I had to select 360 GB on 20 cores because I only had one 
> group. Does CP2K set the number of groups or the group size? I checked your 
> .inp file, but I didn't see any input parameter that controls that setup. 
>
> Also, sorry to diverge from your initial question! This was a question I 
> had a few days ago. 
>
> Stephen 
>  
>
> On Saturday, February 15, 2020 at 10:56:38 AM UTC-5, Torstein Fjermestad 
> wrote:
>>
>> They are parallelized in six groups. 
>> each group run on 144 cores (NPROC_REP  144)
>>  
>>
>> lørdag 15. februar 2020 16.29.26 UTC+1 skrev Stephen Vicchio følgende:
>>>
>>> Thanks for sharing some of your insights/problems! I am curious as well. 
>>> Also, do the 1735 single point calculations run simultaneously 
>>> (parallelized)? Or are they performed sequentially (in series)?
>>>
>>> Stephen 
>>>
>>> On Saturday, February 15, 2020 at 10:27:10 AM UTC-5, Matt W wrote:
>>>>
>>>> Maybe printing forces leads to you running out of disk space or a 
>>>> similar file system problem? It won't be affecting the actual calculation.
>>>>
>>>> Matt
>>>>
>>>> On Saturday, February 15, 2020 at 3:16:13 PM UTC, Torstein Fjermestad 
>>>> wrote:
>>>>>
>>>>> Just a small update:
>>>>>
>>>>> The vibrational analysis calculation completed successfully when I 
>>>>> changed the section parameter of &FORCES in the &PRINT section of 
>>>>> &FORCE_EVAL from ON to SILENT. This was the only thing that changed. Here 
>>>>> is the output of the "diff" command on the two input files:
>>>>> <      &FORCES  ON
>>>>> ---
>>>>> >      &FORCES  SILENT
>>>>>
>>>>> This was surprising to me because I thought the &PRINT section 
>>>>> determined what is printed and had no influence on the outcome of the 
>>>>> calculation. Could it be a bug? I have been using CP2K version 7.0 
>>>>> the output of the command cp2k.popt -v:
>>>>>
>>>>>  CP2K version 7.0 (Development Version)
>>>>>  SVN source code revision
>>>>>  cp2kflags: fftw3 parallel mpi3 scalapack libderiv_max_am1=4 
>>>>> libint_max_am=5 mkl
>>>>>
>>>>> Regards,
>>>>> Torstein
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> fredag 14. februar 2020 11.45.34 UTC+1 skrev Torstein Fjermestad 
>>>>> følgende:
>>>>>>
>>>>>> Dear all, 
>>>>>>
>>>>>> I started a vibrational analysis calculation on a system that 
>>>>>> consists of 289 atoms, and therefore 2*3*289+1 = 1735 single point 
>>>>>> calculations are necessary to complete the vibrational analysis (The input 
>>>>>> file is attached). After having completed 708 single point calculations, 
>>>>>> the vibrational analysis stopped because one single point calculation 
>>>>>> failed.
>>>>>>
>>>>>> The error message I got was the following:
>>>>>>
>>>>>> "Cholesky decomposition failed. Matrix ill conditioned ?"
>>>>>>
>>>>>> While this isolated problem seems to have a straight-forward solution 
>>>>>> (for instance by following the advice provided here: 
>>>>>> https://www.cp2k.org/faq:cholesky_decomp_failed), it seems to be a 
>>>>>> bit more complicated in the context of a vibrational analysis. 
>>>>>>
>>>>>> There are two reasons why I think it would be more complicated. These 
>>>>>> are the following:
>>>>>>
>>>>>>    1. The geometry for which the energy calculation failed is not 
>>>>>>    printed (at least I cannot see it). Only the geometries for which the 
>>>>>>    energy converged are printed. This makes it difficult (impossible) to 
>>>>>>    troubleshoot the problem of why the energy calculation of that structure 
>>>>>>    failed. Is it possible to change the print settings such that the structure 
>>>>>>    is printed before the energy calculation starts?
>>>>>>    2. Suppose that I will be able to compute the energy for the 
>>>>>>    particular geometry. Is it then possible to restart the vibrational 
>>>>>>    analysis from the point where it failed, or do I have to start all over 
>>>>>>    again from the beginning? 
>>>>>>    
>>>>>> The input file and the output file with the error message are 
>>>>>> attached. 
>>>>>>
>>>>>> Thanks a lot for your help. 
>>>>>>
>>>>>> Regards,
>>>>>> Torstein Fjermestad
>>>>>>  
>>>>>>
>>>>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20200215/2d378909/attachment.htm>