Outcome#
You can distinguish login from compute nodes, submit one tiny CPU job with Slurm, inspect it, cancel it, and find its logs without computing on the login node.
Concept#
Euler login nodes are shared control points. Resource-intensive programs belong in Slurm allocations, where CPU, memory, time, logs, and ownership are explicit.
Worked Example#
The job completes on a compute node with one CPU, 1 GiB per CPU, no GPU, a zero exit code, and bounded logs.
A correct example uses these decisions:
- Where does the actual computation run? On a Slurm-allocated compute node.
- The job is absent from squeue. What next? Use sacct and inspect the job logs; it may already have finished.
Common Trap#
Running work on the login node, resubmitting because squeue is empty, or losing the job ID before checking sacct.
If Blocked#
For an invalid account, stop and check my_share_info. For a pending job,
inspect myjobs -j "$job_id" instead of submitting duplicates. Read the first
meaningful error before changing resources. Use the
first Euler job lab for recovery.
Useful references:
Understand Before Accepting AI Output#
Personally read the script, record the real job ID, know how to cancel it, and inspect output and accounting. An agent must not submit or enlarge the job.