scimate-ai/scicli/hpc-orchestration

HPC Orchestration

Use this skill as the repository-level execution layer above solver-specific HPC skills.

Start

1. Read references/cluster-operations-manual.md first for any end-to-end cluster workflow.
2. Read references/lifecycle-manual.md when coordinating solver skills with orchestration stages.
3. Read references/scheduler-and-parallelism.md when choosing scheduler directives or MPI sizing.
4. Read references/slurm-launch-patterns.md when the main decision is how to launch work inside a Slurm allocation.
5. Read references/environment-and-storage-hygiene.md when environment modules, scratch policy, or filesystem behavior may decide workflow reliability.
6. Read references/data-transfer-and-staging.md when files must be synchronized, staged, archived, or verified.
7. Read references/software-build-and-reproducibility.md when compilation, package stacks, or rebuildability are in scope.
8. Read references/interactive-debugging-and-profiling.md when the task needs live diagnosis, scaling studies, or profiler evidence.
9. Read references/remote-development-and-notebooks.md when VS Code Remote SSH, Jupyter, or port forwarding is involved.
10. Read references/container-workflows.md when Apptainer or Singularity-style execution is in scope.
11. Read references/public-protocol.md when aligning solver skills with the repository lifecycle contract.
12. Read references/error-pattern-dictionary.md when scheduler, monitoring, or log-tracking failures occur.

Additional References

Load these on demand:

• references/tools-and-scripts.md for shared execution tool roles
• references/ecosystem-roadmap.md for repository-wide coverage goals
• references/legacy-template.md when adapting an older solver skill draft into the current format

Reusable Templates

Use assets/templates/ when a concrete scheduler scaffold is needed:

• slurm-basic.sh
• slurm-array.sh
• slurm-packed-single-node.sh
• slurm-apptainer.sh
• slurm-perf-report.sh
• pbs-basic.sh
• lsf-basic.sh
• rsync-stage-in.sh
• jupyter-lab-compute.sh
• ssh-config-compute-proxy.example

Shared Scripts

Use scripts/ for deterministic orchestration tasks:

• hpc_job_submitter.py
• hpc_job_monitor.py
• hpc_log_tracker.py
• hpc_slurm_deploy.py

Guardrails

• Do not run heavy MPI workloads on login nodes.
• Do not separate queue submission from runtime monitoring in a production workflow.
• Do not resubmit unchanged failing jobs when a solver-specific error dictionary exists.
• Do not choose core counts without a scale-based heuristic.