[CP2K-user] [CP2K:16219] Re: stress tensor calculations hanging

[martin....@gmail.com]

Hi all,

Just posting an update in case anyone runs into similar issues:
I've followed Matt's suggestion and re-tested my calculations with 
STRESS_TENSOR ANALYTICAL, but ran into a similar hanging issue. This time, 
the TRACE showed a different MPI call as the last one before hanging 
(mp_bcast_b). Using higher cutoffs (as Matt suggested i should probably be 
doing anyways) didnt change this.

However, I have re-run these calculations on a different cluster using the 
Cray flavour of mpi (cray-mpich-8.0.16), this time without issue. Hence i 
believe the problems were related to the mpi builds (all openmpi-3.X) i was 
relying on.

Cheers,
Martin

On Friday, November 5, 2021 at 2:42:43 PM UTC martin.... at gmail.com wrote:

> Hi Matt,
>
> Good catch, i thought i was running with analytical for all my 
> calculations. I've been testing 3 different parametrisations of SCAN: 
> D3-BJ, D3, and rVV10; the D3 run with analytical stress tensor however does 
> seem to run into the same issue as the other 2 using numerical. I'll retry 
> the other 2 with analytical and see whether that leads to a difference.
>
> Regarding the cutoff, i've followed the approach on the cp2k website (
> https://www.cp2k.org/howto:converging_cutoff) while also keeping an eye 
> on the charge density on the r and g grids. The energy does not tend 
> towards a converged value while the charge density on grids is constant at 
> ~1.0E-08 across cutoffs in the range 500-1200Ry. I'll take a look at forces 
> to see whether i get a smoother convergence.
>
> Thanks for your advice,
> Martin
>
> On Friday, November 5, 2021 at 2:18:17 PM UTC mattwa... at gmail.com wrote:
>
>> Hello,
>> I'd suggest the problem is likely 
>>     STRESS_TENSOR NUMERICAL
>> this is a finite difference approximation to the stress tensor and is 
>> massively expensive. Use ANALYTICAL if possible. 
>>
>> SCAN might need a very high cutoff for sensible stress tensor calcs too.
>> Matt
>>
>> On Thursday, 4 November 2021 at 12:24:42 UTC martin.... at gmail.com wrote:
>>
>>> Hello all,
>>>
>>> I am continuing some work a masters student previously did on a layered 
>>> MOF material. Calculations run smoothly with PBE-D3 for cell optimisations 
>>> and NVT MD, but hang (until walltime runs out, without erroring out) after 
>>> a small number of steps for NPT. Running TRACE TRUE on these jobs 
>>> consistently shows the last line as an mpi communication (17    905 
>>> mp_alltoall_z11v       start Hostmem: 818 MB GPUmem: 0 MB)
>>>
>>> I have tested the same calculation on different HPCs and versions of 
>>> cp2k (5.1 through to 8.1, albeit all using central installs of mpi) and run 
>>> into the same issue; is this just an mpi issue or is there anything i can 
>>> try on CP2K's end?
>>>
>>> Interestingly, trying cell optimisations using SCAN (CP2K 8.2) also 
>>> leads to hanging (same endpoint for TRACE) after the first SCF cycle runs 
>>> to completion. I have attached inputs for the pbe-d3 NPT run and SCAN 
>>> geo-opt.
>>>
>>> Best regards,
>>> Martin
>>>
>>

-- 
You received this message because you are subscribed to the Google Groups "cp2k" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cp2k+unsubscribe at googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/123f54bc-7fec-406a-8e2f-a54a0eb8f51bn%40googlegroups.com.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20211112/cdf28118/attachment.htm>