🕵️ SLURM Job Detective

Are you wasting GPU memory? Underusing CPUs? Thrashing disk?
sjdet shows you instantly — one command, no environment to activate.

---

Why sjdet?

HPC clusters are expensive. Running a GPU job that uses 2% of VRAM, or a CPU job at 5% utilization, wastes both your compute quota and valuable cluster resources. Normally you'd need to ssh into nodes, run nvidia-smi, parse sstat output, and do mental math — every time.

sjdet does all of that and fits it in one table:

What you see	What it tells you
CPU eff bar	Are your cores actually doing work, or are they idle?
RAM Max Util %	Peak RAM observed vs. requested RAM.
GPU VRAM %	GPU memory used vs. detected GPU memory.
MaxPages / MaxDisk	Page fault and disk IO rates between fresh samples.
Node column	Exactly which node your job landed on, so you can ssh or srun --overlap instantly

All of this from two SLURM calls (squeue + sstat), batched, with a local cache to avoid hammering the scheduler.

---

Install

No root. No sudo. Works on any cluster where you have a home directory.

Requires uv (recommended) or pipx — both install sjdet to ~/.local/bin so you can type sjdet from anywhere, forever, without activating anything.

uv (recommended)

curl -LsSf https://astral.sh/uv/install.sh | sh  # install uv itself (once)
uv tool install git+https://github.com/e-candeloro/slurm_job_detective

Upgrade later:

uv tool upgrade slurm-job-detective

pipx

python3 -m pip install --user pipx
pipx install git+https://github.com/e-candeloro/slurm_job_detective

Upgrade later:

pipx upgrade slurm-job-detective

---

Usage

sjdet                             # auto-detects $USER, shows all your jobs
sjdet --user alice                # inspect another user's jobs
sjdet --max-jobs 20               # show up to 20 jobs (default: 10)
sjdet --interval 120              # minimum seconds between sstat polls (default: 60)
sjdet --headroom 0.30             # memory suggestion headroom: ceil(MaxRSS × 1.30) (default: 20%)
sjdet --force-update-nodes        # forces an update to the node info cache
sjdet --clear-cache               # clears the local user cache completely and exits
sjdet --version                   # print installed version and exit
sjdet --update                    # upgrades sjdet and exits

Update checks

• Normal runs perform a lightweight upstream update check and print a notice when a newer version is detected.
• Notices are cached to avoid spam (cooldown: 7 days).
• sjdet does not prompt interactively for updates during normal runs.
• Update notices show current_version -> target_version.
• sjdet --update upgrades directly to the latest published GitHub release version.

Update-source strategy:

1. GitHub latest release only

Update-command strategy for --update:

1. uv tool install --upgrade git+https://github.com/e-candeloro/slurm_job_detective.git@v<latest_version>
2. pipx install --force git+https://github.com/e-candeloro/slurm_job_detective.git@v<latest_version>
3. python -m pip install --upgrade git+https://github.com/e-candeloro/slurm_job_detective.git@v<latest_version>

Reading the GPU column:

• Green/yellow/red follow the same utilization brackets as CPU and RAM.
• If GPU memory is unavailable, sjdet falls back to GPU utilization.

Reading MaxPages / MaxDisk columns:

• These columns show rate changes between fresh sstat samples.
• baseline means this run established the first comparison point.
• Re-run after the cache interval to see measured page and disk rates.

To jump to your job's node:

ssh <node>               # SSH directly (node shown in table)
srun --overlap --jobid=<JOBID> --pty nvidia-smi   # run nvidia-smi inside your allocation

---

How it works (no scheduler spam)

Each invocation makes exactly two SLURM calls:

1. squeue — one call for all your jobs (state, node, GRES)
2. sstat — one batched call for all running job IDs

GPU hardware info (scontrol show node) is queried once per node, then cached locally forever — GPU models don't change between runs.

The sstat result is cached for --interval seconds (default 60s) to respect cluster etiquette. Re-running sjdet within that window is instant.

---

Development

git clone https://github.com/e-candeloro/slurm_job_detective
cd slurm_job_detective
uv sync --dev        # creates .venv, installs deps + torch (for GPU load testing)
uv run sjdet         # run without activating the venv

GPU load test (to see real VRAM numbers in the table):

# inside a GPU srun session:
uv run scripts/gpu_load_test.py --gb 8 --seconds 300
# in another terminal:
uv run sjdet

---

Project layout

src/sjdet/
├── cli.py      ← argument parsing + main() orchestration
├── slurm.py    ← squeue/sstat/scontrol calls, data model, parsing
├── display.py  ← rich table, progress bars, color logic
└── cache.py    ← local JSON cache (throttles sstat + persists node info)
scripts/
├── gpu_load_test.py  ← dev tool to burn VRAM and verify GPU column
└── mock_cli.py ← dev tool to test the CLI without the access of a SLURM scheduler