Run an executable under Apptainer on GPU

This article describes how to run an applicatiopn that was compiled in an Apptainer environment on a GPU.

Example application

We'll use the fidibench set of benchmarks to illustrate the process. On mahuika,

git clone https://github.com/pletzer/fidibench
cd fidibench
mkdir build-nv
cd build-nv

A container with NVIDIA compilers can be obtained here /nesi/nobackup/pletzera/ngarch_nvhpc.sif. To launch the container:

module purge
module load CUDA/12.6.3 Apptainer/1.2.5
apptainer shell --nv /nesi/nobackup/pletzera/ngarch_nvhpc.sif

Note the --nv option which exposes the GPU hardware to the container.

We can now build an application using the nvc++ compiler inside the container:

Apptainer> CXX=nvc++ cmake -DOPENACC=1 ..
Apptainer> cd upwind/cxx
Apptainer> make upwindAccCxx
Apptainer> exit

The application (upwindAccCxx) contains OpenACC directives to offload to a GPU.

Run the application

srun --gpus-per-node=A100-1g.5gb:1 apptainer exec --nv \
     /nesi/nobackup/pletzera/ngarch_nvhpc.sif ./upwind/cxx/upwindAccCxx

should print

number of cells:  128 128 128
number of time steps: 10
check sum: 1