A desktop-only model now runs as batch jobs on an HPC cluster

Starting point: a desktop model needed on a cluster

The customer’s simulation ran only on desktops. The customer needed wide parameter sweeps of it, a lot of runs each, on its HPC cluster, where a job scheduler shares the machines among many users. The cluster sits on an air-gapped network, so everything for it arrives by disc, and nothing can be fetched once it is there.

• Runs had to start on the cluster without compiling anything there.
• Every run needed its own scheduler job and its own logs.
• Analysis of the results had to work with no network.

Approach: a surrogate first, then runs from a spreadsheet

1. Stood up a surrogate cluster. It runs the same scheduler as the customer’s cluster, and every package and job script ran there first.
2. Packaged releases that need no compiling. The model arrives on the cluster ready to run, with nothing to build there.
3. Turned a spreadsheet into a batch of runs. Each row of a design-of-experiments spreadsheet becomes a run with its own scheduler job and its own logs.
4. Shipped post-run analysis in the release. A local dashboard and notebooks read the results with no network, and they run in the release’s own Python, so they behave the same on the cluster’s network.
5. Carried the kit in and worked on site. The kit went in by disc, and we worked on site with the customer’s staff, as needed and for as long as the work needed.

Results: the model runs as batch jobs on a cluster

A desktop-only model now runs its parameter sweeps as batch jobs on the customer’s HPC cluster, many jobs per sweep, and the release carries the offline analysis for its results.

Engineering write-up: Parameterized simulation runs on a PBS cluster from one spreadsheet

Buy from us or team with us: Contracting