cluster-run

Cluster Run

Use this when a workload needs software from Spack and has to run somewhere other than the current shell. It spans two servers, and the handoff between them is the part that goes wrong.

The Spack server deliberately refuses to load anything into an environment, because a load performed inside a tool call dies with that call. JARVIS owns the runtime environment instead — spack://capabilities says so directly: stateful_load_exposed: false, runtime_owner: jarvis_run.

Pick the route first

There are two ways work reaches a cluster, and they are alternatives, not steps:

• Package-backed workload — a Spack package, run through a JARVIS pipeline. clio-jarvis:jarvis_run takes execution intent (local, cluster, hostfile) and resolves the scheduler itself. This is the path below.
• A batch script you wrote yourself — clio-slurm:slurm_submit, which returns a scheduler-native job ID.

Do not submit a JARVIS pipeline with clio-slurm:slurm_submit. Pass the cluster intent to clio-jarvis:jarvis_run and let it own the scheduler.

Steps

1. Size the request against the machine.

Use clio-slurm:slurm_cluster to inspect target partitions and bounded node records. clio-node-hardware:get_cpu_info and get_gpu_info describe the MCP host, which may be a login node rather than a compute node. Verify scheduler limits and the workload's rank/thread requirements before setting resources.

2. Find out whether the package is already installed.

Call clio-spack:spack_find with the constraint. No matches is a successful result with count=0, not an error — do not retry it as if the call failed.

If it returns nothing, clio-spack:spack_search answers a different question: whether a recipe exists at all, in which repo, and whether it is already installed. Reach for it before concluding the software is unavailable.

3. Install only if it is missing.

Call clio-spack:spack_install with an explicit concretization choice. Skip this entirely when step 2 already found a match — installing is slow and runs synchronously, and reinstalling a package that is present wastes the wait.