stress tensor calculations hanging

[martin....@gmail.com]

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

[Matt Watkins]

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

[martin....@gmail.com]

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

[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