AI Potluck
Back to Gap Map Infrastructure / ML orchestration

Slurm

SchedMD
open source / Overall score: 3.2

Workload manager for Linux clusters, and a long-standing scheduler for HPC and GPU training. Jobs request generic resources with --gres or --gpus, and Slurm's controller allocates whole nodes or specific GPUs to a job, enforces the allocation, and runs the job's steps inside it. Multi-node allocations are granted together, which is what all-or-nothing training jobs need.

Maintained by SchedMD. The GitHub API reports NOASSERTION because the COPYING file opens with a SLURM LICENSE AGREEMENT preamble before naming the GNU General Public License; the body is GPL. Verified 2026-09-15 via the COPYING body, the repository tree and the Slurm GRES documentation.

Openness

5 high confidence
5.0
license
GPL-2.0(OSI
source
public(SchedMD/slurm is the scheduler you run)
core-gated
ungated(a 3,889-entry recursive tree carries no path matching enterprise, ee, commercial or proprietary)

GPL throughout, read from the COPYING body rather than from the API label: GitHub reports NOASSERTION because the file opens with a SLURM LICENSE AGREEMENT preamble, and the text then states that all Slurm code and documentation is available under the GNU General Public License. SchedMD sells support and services beside it rather than a fuller build. GPL-2.0 is named in the shared osi tier, so this pair is computed by the ladder rather than authored; the tier entry records the measurement that no other software-type product on the map declares that license, so adding the name moved this record alone.

  • https://raw.githubusercontent.com/SchedMD/slurm/master/COPYING recorded 2026-09-15

    Slurm's COPYING file: 'SLURM LICENSE AGREEMENT. All Slurm code and documentation is available under the GNU General Public License. Some tools in the contribs directory have other licenses.' Establishes the license from the body rather than from the API label.

  • https://github.com/SchedMD/slurm recorded 2026-09-15

    Repository page for SchedMD/slurm: public and unarchived, described as 'Slurm: A Highly Scalable Workload Manager'. Establishes that the source is public.

  • https://api.github.com/repos/SchedMD/slurm/git/trees/master?recursive=1 recorded 2026-09-15

    Full untruncated recursive tree of the default branch, 3,889 entries. Licenses are the root COPYING plus contribs/ components. No path matches enterprise, ee, commercial or proprietary. A tree lists paths; it is cited for that and for nothing about the vendor's offerings.

  • https://www.schedmd.com/slurm-support/support/ recorded 2026-09-16

    SchedMD's support page: the commercial offering around Slurm is support, training and development services. Establishes that what is sold is service rather than a fuller build.

Adoption

2 low confidence
2.0

4,357 GitHub stars. Slurm is distributed as source tarballs and distribution packages rather than through a package registry this map routes, so no usage-volume channel exists for it and the band rests on stars. Stars measure attention rather than use, so the level says what the available channel shows and not how widely Slurm is deployed - which nothing cited here can establish in either direction.

Capability

5 medium confidence
5.0

Admitted on the accelerator limb: --gres and --gpus are Slurm's own submit-language fields and its controller allocates against them. Band 5 because it picks the machines and grants a multi-node allocation as one unit. General-purpose origins do not lower the band - the test is what the scheduler does, and what it does is exactly what the newer AI-specific schedulers here do.

  • https://slurm.schedmd.com/gres.html recorded 2026-09-15

    Slurm's generic resource (GRES) scheduling documentation, describing --gres and GPU allocation by the controller. Establishes that the accelerator request is consumed by Slurm's own scheduler.

  • https://slurm.schedmd.com/quickstart.html recorded 2026-09-15

    Slurm's quickstart: as a cluster workload manager it 'allocates exclusive and/or non-exclusive access to resources (compute nodes) to users for some duration of time', and then 'provides a framework for starting, executing, and monitoring work (normally a parallel job) on the set of allocated nodes'. Establishes that the allocation is a set of nodes held for one job.

Verified 2026-09-15