slurmwatch 0.1.1

Live, process-isolated node-local hardware telemetry for active Slurm jobs

slurmwatch

Live, per-process CPU / memory / GPU telemetry for a running Slurm job — with a plain-language efficiency verdict.

Features

• Efficiency verdict — grades CPU, memory, and GPU (GOOD / UNDERUSED / IDLE / WARNING) so you know when you're wasting cores or GPUs.
• Per-process GPU attribution — NVML sees only your PIDs, so a neighbor's job never inflates your numbers.
• Honest memory — working set (RSS minus reclaimable cache) with a configurable OOM guard.
• Works anywhere — full live telemetry on the node; auto-falls back to Slurm accounting (sstat) from a login node.
• Zero config — slurmwatch <jobid> auto-discovers jobs, cgroup v1/v2, and whether it's on the node. No flags to memorize.

Install

pip install slurmwatch

Requires Python 3.10+ and Linux with cgroup v1 or v2. One install works across a mixed cluster: GPU monitoring (NVIDIA, via pynvml) auto-activates on GPU nodes and is silently skipped on CPU-only nodes. Works with pipx / uv too — e.g. uv tool install slurmwatch.

Usage

slurmwatch                       # auto-discover and attach to your running job
slurmwatch 12345                 # attach to a job (array: 12345_3, het: 12345+1)
slurmwatch --demo                # try the live TUI right now — no Slurm needed
slurmwatch 12345 --once --json   # one machine-readable snapshot, then exit
slurmwatch 12345 --log run.jsonl # headless logging (JSON Lines or CSV)

For full live telemetry, run on the node executing the job: srun --jobid 12345 --overlap slurmwatch 12345. From a login node you get an sstat summary (peak memory + CPU time + allocation) instead — GPU utilization isn't available remotely, since Slurm tracks GPU count, not per-device util.

TUI keys: c/m/g/v focus a panel, arrows/PgUp/PgDn scroll, q quits.

Exit codes: 0 success · 1 runtime failure · 2 bad config. Errors go to stderr, so piped --once/--log output stays clean.

See slurmwatch --help for the full flag list. Behavior is also tunable via SLURMWATCH_* environment variables — e.g. SLURMWATCH_OOM_WARN, SLURMWATCH_GPU_IDLE_PCT, SLURMWATCH_POLL_INTERVAL (plus ASCII mode and more).

Library

import asyncio
from slurmwatch import TelemetryCollector, resolve_job_context

async def sample(job_id: str):
    collector = TelemetryCollector(resolve_job_context(job_id))
    await collector.start()
    try:
        print((await collector.next_snapshot()).to_json())
    finally:
        await collector.stop()

asyncio.run(sample("12345"))

Limitations

• NVIDIA-only GPU support (no AMD/ROCm).
• Single-node view — multi-node jobs show data for the node you're on.
• Live GPU utilization and working-set memory require running on the job's node.

License

MIT