• This version of the model is compatible with NEURON versions newer than 8.1 (#19).
• If you would like to use older NEURON (<= 8.1) then you can use tagged release v1.0.0.

First we need to build NEURON with CoreNEURON support. In order to do so, you can use following script:

module load profile/base
module load profile/advanced
module load hpc-sdk/2021--binary gnu/8.4.0 cuda/11.0
module load spectrum_mpi/10.3.1--binary boost/1.72.0--spectrum_mpi--10.3.1--binary
module load cmake/3.20.0 python/3.8.2

export OMPI_CXX=nvc++
export OMPI_CC=nvc
export CC=mpicc
export CXX=mpicxx

git clone https://github.com/neuronsimulator/nrn.git

mkdir -p nrn/build && cd nrn/build

# cmake, make and make install
cmake .. -DCMAKE_INSTALL_PREFIX=$HOME/install \
	-DNRN_ENABLE_INTERVIEWS=OFF \
	-DNRN_ENABLE_RX3D=OFF \
	-DNRN_ENABLE_CORENEURON=ON \
	-DPYTHON_EXECUTABLE=`which python` \
	-DCORENRN_ENABLE_MPI=ON \
	-DCORENRN_ENABLE_OPENMP=OFF \
	-DCORENRN_ENABLE_GPU=ON
make -j8
make install

Note that above mentioned modules (from May 2022) might change in the future. So make sure to change them if necessary.

Let's clone olfactory bulb model repository from GitHub. For now, we have to use 2to3 branch which has Python 3 support:

git clone https://github.com/HumanBrainProject/olfactory-bulb-3d.git
cd olfactory-bulb-3d/sim
git checkout 2to3

For using NEURON release older than v8.1, checkout tag v1.0.0:

git checkout v1.0.0

We can now build the model nrnivmodl -coreneuron command. Make sure to change path of nrnivmodl if you have installed NEURON from source.

module load profile/base
module load profile/advanced
module load hpc-sdk/2021--binary gnu/8.4.0 cuda/11.0
module load spectrum_mpi/10.3.1--binary boost/1.72.0--spectrum_mpi--10.3.1--binary
module load cmake/3.20.0 python/3.8.2

export OMPI_CXX=nvc++
export OMPI_CC=nvc
export CC=mpicc
export CXX=mpicxx

$HOME/install/bin/nrnivmodl -coreneuron .

If there are no errors, special with CoreNEURON support will be built.

We are now ready to run model. To enable GPU support via CoreNEURON, you can pass the following command line argument to bulb3dtest.py:

--coreneuron --gpu

Also, if you are benchmarking model then make sure to select larger test case i.e. changing runsim.build_part_model in in bulb3dtest.py.

We can now submit the job to the cluster using following job script:

#!/bin/bash
#SBATCH -A Pra18_4575_0
#SBATCH -p m100_usr_prod
#SBATCH --time 00:60:00
#SBATCH --job-name=coreneuron_gpu_test
#SBATCH --nodes=2
#SBATCH --ntasks-per-node=8
#SBATCH --gres=gpu:4

module load profile/base
module load profile/advanced
module load hpc-sdk/2021--binary gnu/8.4.0 cuda/11.0
module load spectrum_mpi/10.3.1--binary boost/1.72.0--spectrum_mpi--10.3.1--binary
module load cmake/3.20.0 python/3.8.2

export PYTHONPATH=$HOME/install/lib/python/:$PYTHONPATH

# gpu run with coreneuron
mpirun ./special_wrapper.ppc64 -python -mpi bulb3dtest.py --tstop=1050 --coreneuron --gpu

Note that we are using special_wrapper.ppc64 instead of ppc64le/special as a binary to launch. This is because MPS service is currently not enabled on Marconi100. Using wrapper script, we start the MPS service before launching an application.