04_hpc_gpu.md

Module 4: HPC Integration (GPU & MPI)

This module encapsulates the primary use cases for Apptainer in High Performance Computing.

GPU Support

Apptainer makes it trivial to use GPUs. You do NOT need to install the NVIDIA driver inside the container. You only need the CUDA toolkit (and cuDNN) inside. The host driver is mounted at runtime.

The --nv flag

To enable GPU support, simply add --nv:

apptainer exec --nv pytorch.sif python train.py

MPI Support

Apptainer supports the "Hybrid MPI" model. You have an MPI installed on the host (e.g., OpenMPI) and a compatible MPI installed inside the container.

Execution

You call mpirun from the host, which then launches apptainer containers.

mpirun -n 4 apptainer exec my_mpi_app.sif /app/mpi_binary

Bind Paths

By default, Apptainer binds $HOME, /tmp, and $PWD. On HPC, you often need access to scratch storage or shared datasets.

apptainer exec --bind /scratch/user:/data my_container.sif python script.py

This maps /scratch/user on the host to /data inside the container.

Exercise: GPU Machine Learning

See exercises/04_ml_gpu.def for a PyTorch setup.

To test it (if you have a GPU):

apptainer exec --nv ml_gpu.sif python -c "import torch; print(torch.cuda.is_available())"