Migrating from Slurm#
Spur is drop-in compatible with Slurm’s command-line interface, REST API, and C FFI. Most Slurm workloads move over unchanged. This page covers what works as-is, where the configuration model differs, and the behavioral differences a Slurm user should know about before switching.
What Works Unchanged#
The following work without modification:
Commands —
sbatch,srun,salloc,squeue,sinfo,sacct,scancel,scontrol, andsacctmgr(via the symlinks described in Slurm Compatibility).Job script directives —
#SBATCHdirectives in batch scripts are parsed the same way.#PBSdirectives are also converted on a best-effort basis.Environment variables — Spur sets a
SLURM_*twin for everySPUR_*variable it injects into a job, so Slurm-aware software (MPI launchers, training frameworks) sees theSLURM_*names it expects.REST API and C FFI — the REST surface and the FFI library remain Slurm-compatible.
Configuration Differences#
Slurm’s slurm.conf and slurmdbd.conf are replaced by a single TOML file,
spur.conf (default location /etc/spur/spur.conf). Key differences:
One config file, one set of daemons. There is no
slurmdbdand noslurmrestd— accounting and the REST API are served by the controller (spurctld). Accounting storage is configured under[accounting]inspur.confrather than in a separateslurmdbd.conf.Raft-based state. Controller state is replicated through Raft and persists across restarts, so there is no
StateSaveLocationhandling to manage separately.Built-in high availability. HA is provided by Raft (openraft) and is always on — even a single controller runs a one-member cluster. There is no separate backup-controller configuration; add controllers to the Raft cluster instead.
See Configuration Reference (spur.conf) for the full spur.conf reference.
Behavioral Differences and Current Limitations#
The commands below accept Slurm syntax, but the following differences apply. Flags noted as “accepted for compatibility” parse without error but have no effect yet.
Area |
Difference |
|---|---|
Partitions |
|
Node deletion |
|
|
Accepted for compatibility on |
|
The |
|
Per-process |
|
The |
|
Delimited output covers |
|
Array-element syntax ( |
|
Output is always the multi-line |
|
Not supported on |
|
Supported as |
|
Inverted default. Slurm runs its daemons on the specialized cores unless |
|
Supported as local transforms (no controller call). |
|
Caps print as |
|
|
QOS |
Enforced at scheduling admission only: an over-budget QOS stops admitting jobs, and running jobs are never killed where Slurm cancels them. Consumption is a trailing window ( |
|
Supported as |
|
Supported per QOS ( |
Accounting and QOS Mapping#
Accounting entities map directly from Slurm:
sacctmgr add account,sacctmgr add user, andsacctmgr add qoswork as in Slurm, following the samecluster → account → user → association → QOSordering.The
[accounting] require_qossetting is the equivalent of Slurm’sAccountingStorageEnforce=qos.The
[accounting] require_associationsetting is the equivalent of Slurm’sAccountingStorageEnforce=associations.The
default_qossetting is the equivalent of Slurm’s fallback QOS.The
[auth] private_datasetting is the equivalent of theslurm.confside of Slurm’sPrivateData(see Job and usage visibility (private_data)). Like Slurm, it is unset by default, so every user sees every job."jobs","usage","users","accounts", and"reservations"are supported, written as a TOML list (private_data = ["jobs", "usage"]), not a comma-separated string.sacctandsreport jobsizes follow"jobs".sshareandsreportutilization follow"usage".spriofollows"jobs".sreport preemptionsfollows both"jobs"and"usage". User and account listings follow"users"and"accounts".sshareis not filtered by"accounts". A coordinator sees the accounts they coordinate, including sub-accounts, and thesaccthistory,sreport jobsizes, usage, and associations in those accounts.sacctmgr show userstays the caller.show user withcoordandshow account withcoordlist coordinator assignments, andshow coordinatoris the same listing.squeueandspriostay the caller’s own jobs. Unlike Slurm, root has no implicit exemption: grant it a role (cluster_admins,admin_groups, an accounting admin level, orallow_uid_zero_administrator). Coordinators do not gain Slurm’s write powers (add users, change limits, cancel jobs).
See Accounting, Accounts, Users, and QOS for the accounting concept guide.
Tip
Before migrating a production workload, test the golden path on a small cluster
first: submit a job (sbatch), check the queue (squeue), cancel a job
(scancel), and review accounting history (sacct).