Migrating from Slurm

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, and sacctmgr (via the symlinks described in Slurm Compatibility).
Job script directives — #SBATCH directives in batch scripts are parsed the same way. #PBS directives are also converted on a best-effort basis.
Environment variables — Spur sets a SLURM_* twin for every SPUR_* variable it injects into a job, so Slurm-aware software (MPI launchers, training frameworks) sees the SLURM_* 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 slurmdbd and no slurmrestd — accounting and the REST API are served by the controller (spurctld). Accounting storage is configured under [accounting] in spur.conf rather than in a separate slurmdbd.conf.
Raft-based state. Controller state is replicated through Raft and persists across restarts, so there is no StateSaveLocation handling 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	Defined in `spur.conf`; there is no runtime `scontrol create/update/delete partition`. Edit the config and reload the controller to change a partition.
`--noconvert`	Accepted for compatibility on `squeue`, `sinfo`, `sacct`, `sstat`, `sshare`, and `sreport`. Spur already reports raw values where applicable (`sinfo`’s memory columns are MB integers; `squeue` has no memory column yet), so there is nothing to suppress. `sstat` still prints `MemAlloc` with an `M` suffix whether or not the flag is passed.
`sacct --jobs`	Accepted for compatibility; the job-id filter is not yet applied server-side.
`sacct` `ReqMem`	The `ReqMem` format field has no value yet and renders empty.
`sstat`	Per-process `Ave` and `Max` metrics (for example `AveRSS`, `MaxVMSize`) show `N/A`.
`sprio`	The `FAIRSHARE` column is a placeholder and does not yet reflect usage.
`sacctmgr` `-p`/`-P`	Delimited output covers `show account`, `show qos`, and `show txn`; for `show user`, `show association`, and `show tres` it errors instead of printing a table no script can parse. `-n` works for every entity.
`scancel` arrays	Array-element syntax (`123_4`) is not yet supported client-side; cancel by plain job id.
`scontrol show`	Output is always the multi-line `Key=Value` block; there is no `--oneliner`. Unset diagnostic fields print as `(null)` rather than being omitted — notably `NodeList=(null)` on a pending job, where earlier Spur releases omitted the line. Note that `ReqNodeList`, `ExcNodeList`, and `SchedNodeList` all contain the literal `NodeList=`, and `ReqNodeList` is printed first: a script matching `NodeList=` without anchoring the field name will now read the requested list instead of the allocated one. Match `^\s*NodeList=` (or an equivalent boundary) instead. `ReqTRES` omits Slurm’s `billing=` component, as Spur computes no TRES billing weights. `MinCPUsNode` is the per-node CPU floor implied by the task layout, which can read higher than Slurm’s `pn_min_cpus` for an uneven split such as `-N2 -n7`. A pending job now reports the scheduler’s projected `StartTime` (and an `EndTime` derived from it) where earlier releases printed `N/A`, so a script treating `StartTime != N/A` as “the job has started” needs to check `JobState` instead.
`scontrol show assoc_mgr`	Caps print as `Limit(Consumed)` with `N` for no limit, as in Slurm, and an `OverLimit` field names caps already exceeded. Filters to one user (`users=<name>` or a bare name); Slurm’s `accounts=`, `qos=` and `flags=` selectors are not accepted (an unknown `<key>=` errors rather than being read as a username). As in Slurm, an unprivileged caller is scoped to their own associations and administrators see every scope. Users are listed on their own lines under a scope rather than nested in a `User Limits=` field, and an account’s group figures are stated once for the account instead of repeated per association row. Fairshare and priority internals Slurm prints (`SharesRaw`, `UsageRaw`, `Lft-Rgt`, `GrpTRESMins`) are not included.
`sshare`/`sreport` usage units	`sshare` `RawUsage` and `CPURawUsage` report cpu-seconds, matching Slurm. Prior Spur releases displayed cpu-hours under the same column names. `sreport` columns are likewise cpu-seconds. Scripts parsing `sshare --parsable` output will see values 3600× larger than before. Upgrade `spurctld` and `spur-cli` together to avoid a mixed-unit window.
QOS `GrpWall`	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 (`grp_wall_window_days` under `[accounting]`) rather than decayed usage, and the cap exists on a QOS only, not on an association. See Group wall-clock budgets (GrpWall).

Accounting and QOS Mapping#

Accounting entities map directly from Slurm:

sacctmgr add account, sacctmgr add user, and sacctmgr add qos work as in Slurm, following the same cluster → account → user → association → QOS ordering.
The [accounting] require_qos setting is the equivalent of Slurm’s AccountingStorageEnforce=qos.
The [accounting] require_association setting is the equivalent of Slurm’s AccountingStorageEnforce=associations.
The default_qos setting is the equivalent of Slurm’s fallback QOS.

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).