problems with parallelization in v 23.1 (compiled through conda-forge channel)

[gekbuccella]

Describe the bug
Molecular Dynamics calculations with conda-forge version of dftbplus 23.1 are slow or get stuck when using "mpirun".

-with openmpi version, the problem might be avoided by diminishing the number of requested cores, but this behaviour seems to be random;
-with mpich version, the calculations carry on but very slowly;
-by launching in serial (i.e. without "mpirun"), the calculations seem to go fine.

To Reproduce
Installation steps:

#conda create --name dftbplus23.1
#conda activate dftbplus23.1
#conda install --name dftbplus23.1 mamba
#mamba install --name dftbplus23.1 -c conda-forge dftbplus=23.1=mpi_*_*
#mamba install --name dftbplus23.1 -c conda-forge dftbplus-tools=23.1
#mamba install --name dftbplus23.1 -c conda-forge dftbplus-python=23.1

job script used to run calculations:

#!/bin/bash
#SBATCH --account ttd
#SBATCH --partition ttdFast
#SBATCH --qos ttd
#
#SBATCH -J ER-400
#SBATCH -o ./out.%j.log
#SBATCH -e ./err.%j.log
#SBATCH -D ./
#SBATCH --mail-type=NONE
#SBATCH --mail-user=giacomo.buccella@rse-web.it
#
#SBATCH --get-user-env
#SBATCH --nodes=1
#SBATCH --ntasks=6
#SBATCH --ntasks-per-core=1

cd /home/TTD/giacomo/MD-ER/temperature-tests/400K

module purge

export OMPI_MCA_opal_cuda_support=true

echo "Start time: `date`"

mpirun dftb+ -i dftb_in.hsd > dftb_out.log

echo "End time: `date`"

Expected behaviour
parallel calculations should run well, and converge faster than the serial ones.

[bhourahine]

There is a memory bug in the 23.1 release, which is possibly the cause for your problem. This was fixed by #1281 #1284 and #1294 and will be included in the next release. In the meanwhile, would it be possible for you to compile the latest unreleased version?