Use launch.sh on the login node for work the browser does not cover: more CPU
or memory than the course menu offers, a GPU class the menu does not include,
or a job that keeps running after you disconnect. Start here to launch a
container, watch it while it runs, and make a long run resumable after an
interruption.
| Page | Covers |
|---|---|
launch.sh Reference |
Invocation and wrapper scripts, default resources and resource tiers, resource requests and limits, and the resource, GPU, image, workspace, placement, and job execution flags. |
| Job Modes, Runtime Limits & Configuration | Interactive, background, and batch job modes; running several jobs at once; the runtime limit and raising it; and configuring a launch with environment variables. |
| Watching a Running Job | Monitoring CPU, memory, and GPU use from the browser, from a shell inside the container, and from TensorBoard or another dashboard; memory limits and OOMKilled; checking on a detached job; and listing and stopping sessions. |
| Checkpointing & Logging Long Runs | Causes of an interrupted run, resumable and atomic checkpoints, checkpoint frequency, the termination warning, resuming, and logging for unattended jobs. |
| Direct Kubernetes Use and Session Events | Scope of support, the namespace and kubectl, running a service from a manifest, multi-pod topologies, and the pod and reservation events a session emits. |