SELFish

SELF image sharing hub

This repository defines a suite of common container images for developing SELF to target different platforms. Each image recipe, defined here as a combination of a spack.yaml and a generated Dockerfile, contains all the instructions necessary to install SELF's dependencies. The intention is to provide container images wherein someone can easily get set up to develop SELF in a common environment.

While SELF does support bare-metal builds and those are regularly tested, the core SELF team at Fluid Numerics is working on a standardized develop, build, and deploy workflow that leverages container environments. We are opting to build Docker container images as these are easily shareable broadly through Dockerhub. Additionally, Docker images can be easily converted for use on shared traditional HPC platforms as singularity/apptainer or enroot images.

The core SELF team at Fluid Numerics has adopted enroot+pyxis with Slurm for our deployment model due to positive experience with this approach.

See Repository Guidelines for contributor expectations, build commands, and review checklists.

Organization

The envs/ subdirectory defines all of the base environments that are aimed at providing base images with all the dependencies required for developing SELF. The subdirectory structure is envs/{cpu_platform}/{gpu_backend}.

Directory	GPU backend	Build args	MPI
envs/x86/rocm/	AMD ROCm	GPU_ARCH (e.g. gfx90a), GPU_BACKEND_VERSION (e.g. 6.4.3)	OpenMPI
envs/x86/cuda/	NVIDIA CUDA	CUDA_ARCH (e.g. 70, 100), CUDA_VERSION (e.g. 12.4, 13.0)	OpenMPI
envs/x86/none/	None (CPU-only)	—	OpenMPI

Each directory contains a spack.yaml manifest, a Dockerfile, and a feq-parse.patch.

Container Images

SELFish provides pre-built container images with all dependencies for spectral element computations. Images are published to Docker Hub under higherordermethods/selfish.

Image Tagging Scheme

Images follow the pattern: higherordermethods/selfish:<version>-<cpu_platform>-<gpu_backend>-<gpu_arch>

• <version>: latest or a commit SHA
• <cpu_platform>: Target CPU architecture (e.g. x86)
• <gpu_backend>: GPU backend with version (e.g. rocm643, cuda124) or none for CPU-only
• <gpu_arch>: Target GPU architecture (e.g. gfx90a, sm70); omitted for CPU-only images

Examples

# AMD MI210/MI250 (gfx90a) with ROCm 6.4.3
docker pull higherordermethods/selfish:latest-x86-rocm643-gfx90a

# NVIDIA V100 (sm70) with CUDA 12.4
docker pull higherordermethods/selfish:latest-x86-cuda124-sm70

# NVIDIA Blackwell (sm100) with CUDA 13.0
docker pull higherordermethods/selfish:latest-x86-cuda130-sm100

# CPU-only
docker pull higherordermethods/selfish:latest-x86-none

Supported Architectures

AMD ROCm

Architecture	GPU Models	Tag
gfx906	MI50, MI60, Radeon VII	latest-x86-rocm643-gfx906
gfx90a	MI210, MI250, MI250X	latest-x86-rocm643-gfx90a
gfx942	MI300A, MI300X	latest-x86-rocm643-gfx942

NVIDIA CUDA

Architecture	GPU Models	Tag
sm70	V100	latest-x86-cuda124-sm70
sm100	B200, B300	latest-x86-cuda130-sm100

CPU-only

Tag
latest-x86-none

Determining Your GPU Architecture

For AMD GPUs:

rocminfo | grep "Name:" | grep "gfx"

For NVIDIA GPUs:

nvidia-smi --query-gpu=compute_cap --format=csv,noheader

Using with Slurm

Specify the architecture-specific image in your job script:

#!/bin/bash
#SBATCH --gpus=1
#SBATCH --container-image=higherordermethods/selfish:latest-x86-rocm643-gfx90a

./run_simulation.sh

Image Metadata

All images include OCI labels for programmatic inspection:

docker inspect higherordermethods/selfish:latest-x86-cuda124-sm70 | grep -A5 Labels

Key labels:

• com.fluidnumerics.rocm.target / com.fluidnumerics.cuda.target: GPU architecture target
• com.fluidnumerics.rocm.version / com.fluidnumerics.cuda.version: Backend version
• org.opencontainers.image.source: Source repository
• org.opencontainers.image.revision: Git commit SHA