Submit Hpc4 Jobs

ssh sqwu@hpc4 bash -l -c 'cd /remote/workdir && sbatch --test-only job.batch'

• If the dry run fails, fix the script or ask the user rather than submitting anyway.
• Ask the user before a real sbatch or scancel action.

Submit Correctly

• For non-interactive SSH, always run SLURM commands inside bash -l -c.
• Use the remote working directory explicitly instead of assuming shell startup files will cd correctly.
• Submit with the smallest command that makes the environment boundary obvious:

ssh sqwu@hpc4 bash -l -c 'cd /remote/workdir && sbatch job.batch'

• If the job script already embeds the environment setup, do not re-encode it in the SSH wrapper.
• If the batch script is derived from a generic template, inspect it for stale engine settings before submission.

Record Metadata

• Capture the returned job id from sbatch.
• Record at least the job id, remote workdir, script path, submission command, and the reason for submission in the project artifacts or notes before moving on.
• If the repo has a durable server note, update it when new facts about login, account, environment, or submission behavior are confirmed.

Monitor Safely

• Prefer light queue checks while the job is pending:

ssh sqwu@hpc4 bash -l -c 'squeue -j JOBID'

• Once the job starts, back off polling unless the user asked for close monitoring.
• When the job leaves the queue, inspect the log and final scheduler state before declaring success or failure.
• Distinguish wrapper noise from real scientific failure by checking the produced outputs, not just the SLURM exit code.

Reference

• Read references/current-hpc4-pattern.md when the repo-local note is absent, stale, or too thin to drive the submission safely.