Direct comparison
Apptainer vs Docker for HPC Research
Apptainer vs Docker compared: privilege model, HPC scheduler support, image format, GPU access, and which to use for reproducible research computing.
Written and maintained by CASRAI Editorial Board
Last updated
Ask CASRAI · free to try
Ask about Apptainer vs Docker for HPC Research
Ask your first 2 questions free below. Subscribers get 150 a day for $29 a month.
An AI assistant specialized in research administration. It cites the sources behind every answer, labels web answers and says when it can't answer.
Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.
Works on this site and inside Claude, Cursor and the AI tools you already use.
Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays free to read, with no account and no card.
How do Apptainer, Docker compare side by side?
The table below compares Apptainer, Docker across 10 procurement-relevant dimensions, from primary design target through best fit.
Side-by-side comparison
| Dimension | Apptainer | Docker |
|---|---|---|
| Primary design target | Shared, multi-tenant HPC clusters | Single-user dev machines, cloud VMs, CI/CD |
| Privilege model | Rootless by default — same user inside and outside the container | Root-owned background daemon (dockerd); container processes commonly run as root |
| Architecture | No persistent daemon; runs as the invoking user's process | Client-server: CLI talks to a long-running daemon |
| Image format | SIF (Singularity Image Format) — single immutable file | Layered image (union filesystem), multiple stored layers |
| Governance | Linux Foundation project since Nov 2021, Technical Steering Committee | Docker, Inc. (commercial); Moby is the open-source upstream |
| HPC scheduler integration | Native integration with Slurm, PBS, LSF | None native — pairs with Kubernetes/Swarm instead |
| GPU access | Direct passthrough (--nv for NVIDIA, --rocm for AMD) | Via NVIDIA Container Toolkit / --gpus flag |
| Runs the other's images? | Yes — pulls/converts Docker/OCI images directly | No — cannot run a SIF file |
| Default filesystem behavior | Read-only by default; host paths bind-mounted in | Writable container filesystem by default |
| Best fit | Reproducible pipelines on shared research computing clusters | App development, microservices, CI/CD, cloud deployment |
Common questions
Common questions about Apptainer vs Docker
Is Apptainer the same thing as Singularity?
+
Effectively yes, with a governance change. Apptainer is the direct continuation of the Singularity project under Linux Foundation governance since November 2021, and remains command-line compatible. SingularityCE is Sylabs’ separate, continuing commercial fork.
Can I just use Docker on our university HPC cluster?
+
Almost certainly not — most shared academic clusters disallow Docker because its daemon requires root, which a multi-tenant cluster cannot safely grant every user. Apptainer exists specifically to avoid that requirement.
Can Apptainer run an existing Docker image without changes?
+
Yes, in most cases. Apptainer pulls Docker/OCI images directly (apptainer pull docker://...) and converts them to a SIF file automatically. Images that assume root-owned processes at runtime may need adjustment.
Which should a new lab standardize on?
+
Build with Docker (or a compatible tool) for the broad base-image ecosystem, and run on shared HPC compute with Apptainer to satisfy cluster security requirements — most research groups end up using both, at different stages.
Does switching from Docker to Apptainer change my results?
+
It shouldn't, if the container's software environment (dependencies, versions, OS base layer) is unchanged — both are running the same binaries in a namespaced environment. The National Academies' distinction between reproducibility (same data, same methods, consistent results) and replicability (new study, same question) is the relevant framing: preserving the exact container image, not the choice of runtime used to execute it, is what protects computational reproducibility.
