If you scanned this from the poster board: the board argues that moving geospatial AI from a laptop to a shared cluster is a contract design problem. This site is that argument with the mechanics filled in.
Download the poster (PDF, 36 x 42 in)
The argument
Geospatial AI workflows often break when teams move from cloud development to HPC production. Containers package dependencies, but execution assumptions change across schedulers, GPU allocation models, and storage layouts. Geospatial outputs add a second source of drift, because they are spatial assets that have to stay discoverable, tileable, and traceable back to their inputs and run settings. Standardize three contracts and heterogeneous backends can sit behind them.
Three contracts
Runtime contract. One job schema across local and HPC execution. Local workers run containerized inference directly. HPC workers execute equivalent jobs under scheduler allocation with a rootless runtime. The API contract, input bindings, output paths, and status schema do not change; only the runtime layer beneath them does.
This is the same boundary Module 1 calls the scheduler as orchestrator.
Data contract. Outputs are catalog objects, not loose files. Register them as STAC items with asset metadata and store rasters in COG-compatible form so tile services can stream them directly.
Execution contract. Queue-backed GPU dispatch with explicit lifecycle state and persisted run metadata. Status transitions, errors, and artifact URIs belong to the contract, not to optional logs.
Read further
- Module 1: Scheduler as Orchestrator is where the runtime contract comes from. The scheduler, not a container daemon, owns the orchestration boundary.
- Module 2: Baseline Single-Node Pattern is the pattern the poster’s local-worker path generalizes.
- Get started runs the baseline single-node workflow end to end.
- Troubleshooting covers the failures that show up first: readiness timeouts, Apptainer, Slurm, and storage.
The poster’s central move, mapping Compose services to scheduler jobs, depends_on to
readiness checks, volumes to job-scoped scratch, and ports to loopback plus a tunnel, is
summarized on the home page.
Citation and contact
Sinha, P. (2026). Geospatial AI on HPC: Infrastructure Contracts for Reproducible Containerized Workflows. Poster #23, I-GUIDE Forum 2026, co-located with the NSF HDR Ecosystem Conference, University of Illinois Chicago, August 2026.
Parmanand Sinha, Computational Scientist, Research Computing Center, University of Chicago.
2026 Better Scientific Software Fellow. [email protected]