templateHPC.jl

A starting point for a parameter-sweep study on a cluster, wired to ParamIO.jl, DataVault.jl and SweepRunner.jl.

Press Use this template, then:

./setup.sh MyStudy FirstSweep

That renames the module, renames the example project, and — the part that is easy to forget — issues fresh UUIDs. The template ships placeholders, and every copy that keeps them declares itself to be the same package as every other copy; Pkg resolves one into the other's place without reporting anything wrong.

The shape

src/, test/            the shared library — physics, models, anything more than
                       one project needs. No HPC dependency.
projects/<name>/       one study. Copy the directory to start another.
  configs/             the sweep, at three sizes: smoke / debug / production
  scripts/compute.jl   the run. Same command on a laptop and on 1000 cores.
  batch/               SLURM submission — the ONLY place site-specific lines go
  report/              a SEPARATE environment that draws from the finished vault
  out/                 DataVault writes here; gitignored
  note/                working notes
docs/, notes are per-project; docs/ is for the shared library.

Why it is split this way

The compute environment stays lean. projects/<name>/Project.toml has no plotting backend and no Pinax. Those are in report/, which reads the same vault off disk. A figure dependency can then never find its way onto a compute node.

The library is separate from the run. Projects depend on the root by [sources] {path = "../.."}, so the same solve runs in a smoke test and in the production sweep, and neither can drift from the other.

The config is the only thing that changes with size. smoke.toml and production.toml have the same shape; a config that works small is the config that runs big.

Getting started

setup.sh instantiates the library and every project, so after it there is one thing to run (report/ is instantiated the first time you draw — it carries the plotting backend and is slow to install):

julia --project=projects/FirstSweep projects/FirstSweep/scripts/compute.jl \
      projects/FirstSweep/configs/smoke.toml

Two points, seconds on a laptop. Then write MyStudy.solve — it is the one function the sweep calls — and grow configs/ until sbatch batch/run.sh configs/production.toml is worth submitting.

Re-running is cheap and safe: a finished point is skipped after a single manifest read, two processes pointed at the same vault never compute the same point twice, and a run that is killed is continued by the next one rather than restarted.