Clusters and Slurm

Today: the login node

Add the cluster’s login node as a remote environment. verlet installs verletrc in your home on the shared filesystem, scans the node, and runs there with mpirun -np N. Modules work: list them in a build spec and they are loaded before building and before every run with that executable.

That is enough for checks, short tests and builds, and for clusters that allow small runs on login nodes. It is not how production runs should be done on most clusters.

Caution

Check your centre’s rules. Many limit CPU time or process counts on login nodes. Some prefer builds on a build or compute node.

Planned: Slurm

planned

• Submit: partition, account, nodes, tasks and time limit in the Run sheet. verlet writes the job script, shows it, and submits it with sbatch.
• Queue state in words: “waiting for 64 cores, position 12”, from squeue and sacct.
• Live view unchanged: verletrc on the login node reads the log on the shared filesystem as it grows.
• Time estimate first: a short trial run measures speed, so you know the job fits its limit.
• Builds as jobs: a build spec can ask to be compiled in a job instead of on the login node.

Planned: cloud environments

planned

Start nodes on your cloud account for a run, prepare them with the same build specs, run, fetch the outputs and shut the nodes down.

The roadmap lists the order.