Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Prometheus exporter for performance metrics from Slurm.
| Date | Stars |
|---|---|
| 2026-07-30 | 286 |
| 2026-08-08 | 287 |
| 2026-08-19 | 286 |
| 2026-08-20 | 286 |
| 2026-09-11 | 286 |
| 2026-09-16 | 287 |
| 2026-10-01 | 286 |
| 2026-10-08 | 286 |
| 2026-10-09 | 285 |
| 2026-10-13 | 285 |
| 2026-10-28 | 286 |
| 2026-11-04 | 286 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
35.0
growth rate 0.00%/day
# Prometheus Slurm Exporter
Prometheus collector and exporter for metrics extracted from the [Slurm](https://slurm.schedmd.com/overview.html) resource scheduling system.
**Minimum Slurm version: 23.02.** As of this version the exporter parses `sinfo --json`, `squeue --json` and `sacct --json` output rather than hand-parsed positional text fields, since Slurm's text layouts are not a stable API and have changed between releases (see the former "State of the GPUs" caveat about issue #38 below, now fixed by this change). The `sdiag`/`sshare`-based collectors (scheduler and fair-share metrics) still parse legacy text output - see the comments at the top of `scheduler.go`/`sshare.go`.
## Exported Metrics
### State of the CPUs
* **Allocated**: CPUs which have been allocated to a job.
* **Idle**: CPUs not allocated to a job and thus available for use.
* **Other**: CPUs which are unavailable for use at the moment.
* **Total**: total number of CPUs.
- Information extracted from the SLURM [**sinfo**](https://slurm.schedmd.com/sinfo.html) command.
- [Slurm CPU Management User and Administrator Guide](https://slurm.schedmd.com/cpu_management.html)
### State of the GPUs
* **Allocated**: GPUs which have been allocated to a job.
* **Idle**: GPUs not currently allocated to a job.
* **Total**: total number of GPUs.
* **Utilization**: allocation ratio (alloc/total) per GPU type.
All four metrics carry a **`gpu_type`** label (e.g. `mi250`, `mi300x`, `a100`, `h100`, or an empty string for untyped/legacy GRES configurations), extracted from Slurm's own GRES descriptor strings (`gpu:<type>:<count>`). This works identically for AMD/ROCm and NVIDIA GPUs - and any other vendor - since Slurm's GRES abstraction is vendor-agnostic; the exporter never talks to vendor tooling for these metrics, only to Slurm itself.
- Information extracted from the SLURM [**sinfo**](https://slurm.schedmd.com/sinfo.html) and [**sacct**](https://slurm.schedmd.com/sacct.html) command.
- [Slurm GRES scheduling](https://slurm.schedmd.com/gres.html)
**NOTE**: since version **0.19**, GPU accounting has to be **explicitly** enabled adding the _-gpus-acct_ option to the command line otherwise it will not be activated.
Be aware that:
* Issue #38 (newer Slurm releases produced slightly different `sinfo`/`sacct` output than the old text parser expected, especially for typed GRES like `gpu:mi250:2`) is fixed as of the JSON migration described above - the parser now reads `gres`/`gres_used`/`tres.allocated` from `--json` output and correctly splits `kind:type:count`, instead of assuming every GRES entry had exactly two colon-delimited fields.
* Users who do not have GPUs and/or do not have accounting activated may want to keep GPUs accounting **off** (see issue #45).
### AMD ROCm GPU Telemetry
In addition to the GRES-based scheduling-occupancy metrics above, the exporter can optionally report **real device-level telemetry** for AMD/ROCm GPUs, read directly from `amd-smi` (or `rocm-smi`) on the node it runs on:
* `slurm_rocm_gpu_utilization_percent{node, gpu}`
* `slurm_rocm_gpu_memory_used_bytes{node, gpu}` / `slurm_rocm_gpu_memory_total_bytes{node, gpu}`
* `slurm_rocm_gpu_temperature_celsius{node, gpu, sensor}` (`sensor` is `edge`, `junction`, or `memory`)
* `slurm_rocm_gpu_power_watts{node, gpu}`
Enable it with `-rocm-acct` (default off); `-rocm-smi-cmd` (default `amd-smi`) selects the CLI to invoke, e.g. `-rocm-smi-cmd=rocm-smi` on older fleets. **These metrics are node-local hardware telemetry, not job-scheduling data** - there is no `job_id` label and no cross-reference to Slurm's GRES allocation, so run this collector as one exporter instance per GPU node rather than expecting it from the cluster-wide instance that serves the metrics above. This is AMD/ROCm-specific by design; NVIDIA GPUs remain covered only by the vendor-agnostic GRES metrics above.
### State of the Nodes
* **Allocated**: nodes which has been allocated to one or more jobs.
* **Completing**: all Excerpt of 14,891 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:ed22b64284f3dfa4, topic:slurm