Skip to main content

Cluster Actions

Every lifecycle operation on a Slurm Cluster runs from the Actions menu. This page documents each one — what it does, what it asks you, what it restarts, and when it is available.


Where the Actions Live

Entry pointHow to reach it
Cluster listTraining Cluster in the sidebar → the ⋮ icon on a cluster row
Cluster detailsOpen the cluster → Actions button in the top-right area
Quick actionsThe details header also carries Connect, Restart Workers and a refresh icon as buttons

Both menus offer the same actions in the destructive-last order below. Actions you do not have update permission for are not shown.

ActionAvailable when
Update SSH KeysAny status
Update ImageRunning
Slurm ConfigurationRunning
Scale ClusterRunning, and the cluster is on an hourly plan or a private cluster
Convert to CommittedHourly, non-private-cluster plans, and not Failed / Terminating / Terminated
Restart All WorkersRunning
Restart ClusterRunning
Restart Login ServiceRunning
Reconfigure ClusterRunning or Creating
Clone ClusterAny status
Terminate ClusterAny status
Delete Training ClusterAny status

For the full picture of what hides or blocks an action, see Action availability and blockers.


Update SSH Keys

Replaces the set of SSH keys that can log in to the cluster.

  1. ActionsUpdate SSH Keys.
  2. The dialog reads "Manage SSH keys for cluster <name>".
  3. Use the Select SSH Keys multi-select to choose the keys. The count shows as e.g. 2 selected.
  4. To add a brand-new key, use the Click here link — it opens Add SSH Key ("Attach a public key to authenticate with your server") without losing your place.
  5. Click Update SSH Keys.
DetailValue
Restart requiredNo
Available atAny cluster status
Confirm blocked whenNothing selected, or the selection is unchanged
This replaces the whole set, it does not add to it

Keys you leave out of the selection lose access. The dialog pre-selects the cluster's current keys — deselecting one removes it. If nothing is selected you see "No SSH keys selected. You need at least one SSH key to access the cluster." and cannot continue.

Prefer per-user identities for a team

Rotating a shared root key affects everyone at once. For team access, use Login User Management — each person gets their own key, their own home directory, and can be revoked individually.


Update Image

Moves the cluster to a different version of its Slurm image — the software stack the controller, login node and workers run.

  1. ActionsUpdate Image. The dialog is titled Update Image: <name>.
  2. Current Image shows what is running, e.g. ubuntu-slurm:v2-slurm25.11-ubuntu24.04.
  3. Under SELECT IMAGE, pick the family — Ubuntu Slurm or NeMo Framework.
  4. Under SELECT VERSION, pick a version. Rows carry badges: Latest for the newest, Current for the one running now.
  5. Click Update Image, then read the confirmation.

The confirmation dialog is Confirm Image Update and is explicit about the cost:

Rolling Restart Required

  • Nodes will restart one at a time to apply the new image
  • Running jobs will be drained before each node restarts
  • Pending jobs will queue until nodes come back online
  • Estimated downtime per node: 3-8 minutes

Confirm with Apply & Rolling Restart. Success toast: "Image update applied successfully. Rolling restart initiated."

This is not the job container image

Update Image changes the cluster's own software — Slurm, the drivers, the base OS. It has nothing to do with the container your jobs run in. To change that, pass --container-image to srun/sbatch; see Containers with Enroot and Pyxis. You do not need to update the cluster image to run a different framework version.

Drain your queue first on a busy cluster

Running jobs are drained node by node. On a cluster with long jobs, an image update can take a long time to work through and your jobs are interrupted. Schedule it, and make sure your jobs checkpoint to shared storage.


Slurm Configuration

Edits slurm.conf extras, partitions, and prolog/epilog scripts. Available only while the cluster is Running.

Covered in full in Slurm Configuration:

Running jobs are never interrupted; queued jobs pick up the new settings; prolog and epilog changes apply from the next job.


Scale Cluster

Adds or removes worker nodes on a live cluster. The dialog is titled Scale pool: <name> and shows the GPU type and Current nodes before anything else.

Two modes, selected by the Scale up / Scale down buttons.

Scale up

  1. ActionsScale ClusterScale up.
  2. Set Total Node Count with the stepper. The value is the new total, not the number to add, and the side text shows the resulting GPU count.
  3. A cost card shows New hourly cost and Change from current.
  4. Click Scale Pool.

If capacity is short you get, instead of the stepper:

Inventory Unavailable — No additional inventory is currently available.

What you can do:

  • Check back later — inventory is refreshed frequently.
  • Contact support for NVIDIA <GPU type> capacity.

When only some of what you asked for is free, the message reads "Only <n> nodes available — you requested <m>." with a link that sets the stepper to the maximum available.

DetailValue
Maximum nodes512
Restart requiredNo — new nodes join; existing jobs keep running
New nodes are placedOn the same rack as your existing nodes, to keep fabric performance consistent
New nodes are not in your custom partitions automatically

A partition holds the nodes you named. After scaling up, edit the partition to include the new nodes — otherwise they are only reachable via the platform all partition. See Manage partitions.

Scale down

  1. ActionsScale ClusterScale down.
  2. Under Select nodes to shut down, click the nodes to remove. A counter shows N selected and a cost card shows the saving.
  3. Click Remove N nodes.
DetailValue
MinimumAt least one node must remain. The last unselected node cannot be chosen, and selecting everything shows "At least 1 node must remain active."
RestartScale-down also restarts the remaining worker pods. The dialog does not say so
Scale-down ends jobs on the removed nodes, and restarts the rest

Nothing checks the Slurm queue before removing a node. Jobs running on a node you remove are lost, and because the remaining workers restart, jobs on those nodes are affected too. Drain the cluster first: scontrol update nodename=<node> state=drain, wait for squeue to clear, then scale down.

A partition can silently stop working

If every node a partition named is removed, that partition stops being written to slurm.conf and shows as not live in the partition table. Repoint it at nodes that still exist. See the scale-down note.

Scale Cluster is hidden on committed plans. The API refuses it with "Scale cluster is not supported for committed cluster".


Restart Actions

Three restart actions with very different blast radius. Each asks a single confirmation.

ActionDialog bodyWhat restartsUse it when
Restart All Workers"Are you sure you want to restart all workers in cluster <name>?"Every worker (slurmd) podNodes are unhealthy, reporting XID errors, or wedged after a hardware fault
Restart Login Service"Are you sure you want to restart the login service for cluster <name>?" Confirm button: RestartThe login node pods onlyYou cannot SSH in, or the login node is out of resources
Restart Cluster"Are you sure you want to restart the entire cluster <name>?"Controller, login node, all workers and the accounting databaseNothing else has worked. This is the heaviest routine action

Restart All Workers is also a button in the details header, labelled Restart Workers. It is disabled unless the cluster is Running.

All three end running jobs on the pods they restart

There is no drain step and no queue check. Restart All Workers and Restart Cluster end every running job. Check the Jobs tab first, and prefer draining and letting jobs finish where you can.

Restart Login Service disconnects open SSH sessions

Anyone logged in is dropped, and anything running in a plain shell there dies. Work you care about should be a Slurm job (sbatch), not a foreground process on the login node — a job survives a login restart, a shell does not.

Escalate in order

Try the narrowest action first: Restart Login Service for access problems → Restart All Workers for node or GPU problems → Restart Cluster only if the controller itself is unhealthy. For a single bad node, use the per-node reboot on the Nodes tab instead of restarting everything.


Reconfigure Cluster

Re-applies the cluster's stored configuration without changing anything. It is a repair action.

  1. ActionsReconfigure Cluster.
  2. Confirm: "Are you sure you want to Reconfigure Cluster <name>?"
DetailValue
Available whenRunning or Creating — one of the few actions offered on a Creating cluster
ChangesNothing in your configuration. It re-pushes what is already stored
When to reach for it

A cluster stuck in Creating, or one whose live state has drifted from what the console shows — a volume that reads as mounted but is not there, a setting that did not take. Reconfigure re-applies the intended state. If it does not help, it has not made anything worse.

This is different from Slurm Configuration, which changes the configuration. Reconfigure just re-applies it.


Clone Cluster

Creates a new cluster pre-filled from this one. Available at any status — including Failed and Terminated, which makes it the standard way to recover from a bad cluster.

  1. ActionsClone Cluster. You are taken to the create form, pre-filled.
  2. Review every section, adjust what you want, and submit Create Training Cluster.
Carried overNot carried over
Image and versionCluster name — a new one is generated
SSH keys, security groupPartitions — they name the original cluster's nodes
Node count, plan / SKU, private clusterLive state: jobs, node health, attached IPs
Lifecycle script
Storage mounts (PFS, SFS, datasets, Weka) and their paths
Extra slurm.conf and all four prolog/epilog scripts
Cloning does not create anything by itself

It only pre-fills the form. Nothing is provisioned, and nothing is billed, until you submit. The clone reflects the source cluster's current shape, not how it was originally created.

Re-add your partitions after the clone is running

Because partitions cannot be set at create time, a cloned cluster starts with only the platform partitions. Re-create yours from the partition table once it is Running.


Convert to Committed

Switches billing from hourly to a discounted committed plan. Titled Convert to Committed Plan.

You can reserve this training cluster by converting it into a committed service at a discounted price.

  1. ActionsConvert to Committed.

  2. Committed Plan — pick a plan; options read like 30 Days Committed, ₹<price>.

  3. Renewal Policy — what happens when the commitment ends:

    OptionEffect
    Auto RenewalRenews the same committed plan
    Convert to Hourly PlanStarts hourly billing at the listed rate
    Auto DeleteDeletes the cluster at the end of the term
  4. Read the acknowledgement, which is ticked by default, and click Continue.

DetailValue
Restart / downtimeNone. This is purely a billing change — nothing on the cluster changes
Not available forCPU clusters, private-cluster plans, clusters already committed
ReversibleNo, not within the commitment window
The full committed amount is charged upfront and is non-refundable

The acknowledgement text is precise: the whole cost for the commitment period is deducted immediately, you cannot terminate before the term ends, and terminating early refunds nothing for the remaining days. Be sure of the node count and duration before confirming.

Committed clusters cannot be scaled

Once committed, Scale Cluster and a further Convert to Committed are both hidden. Size the cluster before you commit.


Terminate Cluster

Stops the cluster, releases its nodes, and stops compute billing.

  1. ActionsTerminate Cluster (in the red group at the bottom).
  2. Confirm: "Are you sure you want to terminate the cluster <name>?"
  3. Click Terminate.

The status goes to Terminating, then Terminated.

What happensDetail
ComputeNodes are released back to the pool. The Slurm release, its pods, its internal volumes and secrets are destroyed
BillingCompute billing stops
Attached storageUnmounted, not deleted. Your PFS, SFS, dataset and Weka volumes survive — and keep billing
IPsReserved and VPC IPs are detached
Login usersPersistent Login is disabled as part of termination
The recordStays in the list at status Terminated, so you can still Clone or Delete it
Terminate is irreversible and it does not warn you twice

There is no acknowledgement checkbox on this dialog — a single confirmation ends the cluster. Running jobs are lost. Anything not written to a PFS, SFS or Weka mount is gone with the nodes.

Storage keeps billing after termination

Terminating the cluster does not delete your volumes. Review them under Storage in the TIR sidebar and delete what you no longer need. See Billing.

Blocked by a VPC IP that is still attaching

If a VPC IP is mid-attach, terminate is refused until it settles. Wait a moment and retry, or check the Network & Security tab.


Delete Cluster

Removes the cluster record permanently. Labelled Delete Training Cluster on the list page and Delete Cluster on the details page.

  1. ActionsDelete Training Cluster.
  2. The dialog reads "This will permanently delete the training cluster and all associated resources."
  3. Tick "I understand and want to delete the training cluster <name>."
  4. Click Delete.
DetailValue
Available atAny status. On an already-Terminated cluster it is immediate, since teardown is done
After deletionThe cluster leaves the list and no further action on it is possible
Attached storageUnmounted, not deleted — the volumes survive
Delete is final

Once issued, the cluster is gone from the list and every operation on it fails. Clone it first if you might want the same configuration again — a Terminated cluster can still be cloned, a deleted one cannot.

Terminate or delete?

TerminateDelete
Final statusTerminatedDeleted
Still in the listYesNo
Can still be clonedYesNo
Acknowledgement checkboxNoYes
Stops billingYesYes

The usual sequence is terminate, keep the record while you might still want to clone its configuration, and delete later during cleanup.


Per-Node Actions

Rebooting a single node is not in the Actions menu — it lives on the Nodes tab, where you can also see which node is reporting XID errors. Prefer it over Restart All Workers when only one node is unhealthy.


Last updated on September 10, 2026.