Revert slurm.sim to the pre-2.0 template (fixes -l local launches)

[JRPan]

Problem

The Accel-Sim 2.0 merge (#553) rewrote util/job_launching/slurm.sim to stage the working directory in /tmp and rsync it back to NFS. That template is now hard-bound to a real Slurm + NFS cluster, but it is also the template used by the local launcher:

elif options.launcher == "local":
    job_submit_call = os.path.join(this_directory, "procman.py")
    job_template = "slurm.sim"

procman.py provides no $SLURM_JOB_ID, and a CI container has neither squeue nor (necessarily) rsync. Three things then go wrong at once:

1. #SBATCH --output=/dev/null / --error=/dev/null — procman.py parses exactly these lines to decide where to write a job's stdout/stderr:

   outFMatch = re.match(r"#SBATCH --output=(.*)", line.strip())
   if outFMatch: job.outF = outFMatch.group(1)

   So every byte the simulator prints is now discarded, leaving no diagnostics at all.

2. The watchdog self-destructs. while squeue -j $SLURM_JOB_ID &>/dev/null; do ... done returns 127 immediately when squeue is absent, so the subshell falls straight through to sync_to_nfs && rm -rf "$TMP_DIR" at t≈0 — deleting the staging directory while the main script is still racing to rsync into it and cd there. With set -eE and trap copy_output EXIT, the job dies before the simulator produces anything.

3. ${SLURM_JOB_ID} is not substituted. procman.py does re.sub(r"\$SLURM_JOB_ID", str(job.id), line), which matches the bare form on the exec > and squeue lines but not the braced form on TMP_DIR=/tmp/${SLURM_JOB_ID}_REPLACE_NAME. The staging directory and the log file inside it end up disagreeing about the job id.

Net effect: no <name>.o<jobid> ever lands in the run directory, so job_status.py reports NOT_RUNNING_NO_OUTPUT for every job, within seconds.

Impact

This is what is currently breaking gpgpu-sim CI, which clones this repo unpinned and launches with -l local. Every matrix entry — QV100, TITANV, TITANV-LOCALXBAR, RTX2060, RTX3070, H100 — fails identically on all 10 rodinia_2.0-ft apps, with zero output to diagnose from. Last green run was before #553 landed; everything after it fails.

Fix

Revert slurm.sim to the pre-2.0 template. It only needs coreutils mv, so it works under both sbatch and procman.

Follow-up (not in this PR): the tmp-staging behaviour is genuinely useful on the Slurm runners and can be reintroduced guarded on [ -n "$SLURM_JOB_ID" ] plus command -v squeue rsync, keeping #SBATCH --output=/--error= pointed at real paths so procman.py keeps working.

🤖 Generated with Claude Code