Skip to main content

Features

What a TIR Slurm Cluster gives you beyond a scheduler, and where each capability is documented.


Feature Reference

FeatureSummaryDetails
Slurm on SlinkyA real, managed Slurm cluster — sbatch, squeue, sacct, partitions, priorities and accounting all behave as documented. You never install or patch a schedulerHow it works
Container jobsRun any Docker or OCI image inside a job with --container-image, unprivileged, with native GPU access. Two jobs on one cluster can use completely different framework versionsContainers
Editable slurm.confAppend your own directives from the console. The platform keeps only the keys that wire the cluster togetherExtra slurm.conf
Custom partitionsUp to 20 named queues over chosen nodes, each with its own time limits, defaults and node membershipPartitions
Prolog & epilog scriptsFour script slots — worker and controller, before and after every job — for setup, cleanup, health gating and auditingProlog & Epilog
Elastic scalingAdd or remove worker nodes on a live cluster. Scale-up leaves running jobs untouchedScale Cluster
Shared storage everywherePFS, SFS, Weka and read-only datasets mounted at the same path on the login node and every workerStorage
GPU health & DCGM metricsPer-node and per-GPU utilisation, memory, temperature and power, with XID error detection and single-node rebootNodes
Jobs & monitoring without SSHLive squeue in the browser, GPU and scheduler charts over selectable time ranges, controller and worker logsJobs · Monitoring
AlertsAttach alerts to a cluster so problems find you instead of the other way roundAlerts
Per-user loginsEach teammate gets their own SSH identity and a private home directory on shared storage, managed from the consoleLogin User Management
High availabilityNode failures are isolated and the cluster recovers automatically. Job-level resilience is yours, via checkpointingHigh availability
Network controlReserved and VPC IPs, and security groups you manage per clusterNetwork & security
Full API coverageEvery console action maps to a REST call, so provisioning and management can be automated end to endAPI reference

Slurm Cluster vs a Single GPU Node

Slurm ClusterSingle node
Multi-node trainingYes, with an RDMA fabric between nodesNo
Job queue and prioritiesYesNo
Sharing between peopleYes, with per-user identities and partitionsAwkward
Container per jobYesYou manage the container yourself
BillingPer node, per hour; jobs freePer node, per hour
Best forDistributed training, shared team capacity, batch workloadsInteractive development, single-GPU experiments

Best Practices for Slurm Clusters

Write everything you keep to shared storage

A worker or login pod's own filesystem is scratch — it is wiped on restart, scale, image update and node reboot. Checkpoints, logs, datasets and Python environments belong on a PFS or SFS mount, which is identical on the login node and every worker.

Cache container images as squash files

Run enroot import once on the login node into a shared path, then point jobs at the .sqsh file. Otherwise every node and every array task pulls the image again.

Checkpoint frequently on long runs

A node failure, a preemption or a restart ends the job. Your progress is only as good as the last checkpoint written to shared storage.

Never end a worker prolog with a non-zero exit

Slurm drains the node when a prolog fails. Finish every worker prolog with exit 0 unless draining is exactly what you intend.

Check the queue before any disruptive action

No action inspects the Slurm queue. Restart, scale-down and terminate all proceed with jobs in flight and lose them — only Update Image drains nodes first.

Give each person their own login

With everyone as root, job ownership is meaningless and scancel -u root cancels the whole cluster. Enable Login User Management for per-user identities and home directories.

Give short and long work separate partitions

A partition with a one-hour Max time keeps interactive debugging out from behind a 24-hour training run, without anyone having to coordinate.

Remember storage bills after the cluster is gone

Terminating a cluster unmounts your volumes but does not delete them. Review PFS, SFS and datasets separately when you clean up.


Last updated on September 10, 2026.