Skip to main content

Troubleshooting

Start here, then follow the link to the focused page.

PageCovers
Jobs will not runPending, failing, cancelled, or slow jobs
Container job failuresEnroot and Pyxis: image imports, mounts, GPUs, NCCL
Cannot connectSSH failures for root and for named users

By Symptom

SymptomGo to
The cluster is stuck in CreatingFirst five checks
The cluster is FailedCluster problems
ssh times out or is refusedCannot connect
Permission denied (publickey)Cannot connect
A job sits in PendingJobs
A job fails instantlyJobs
srun: unrecognized option '--container-image'Containers
A container job cannot see my dataContainers
nvidia-smi shows no GPUs in a containerContainers
A multi-node job hangs at startupContainers
CUDA out of memoryJobs
A node shows XID ErrorNodes and GPU health
A node is stuck UNKNOWNCluster problems
A node keeps getting drainedProlog scripts
GPU utilisation is near zero while a job runsMonitoring
A partition disappeared after scaling downPartitions
A slurm.conf setting was rejectedExtra slurm.conf
An action is missing or greyed outAction availability
My data vanished after a restartWhat persists
A volume will not unmountStorage
A named user cannot log inLogin users

The First Five Checks

Before diving deeper, these five answer most problems.

1. Is the cluster Running?

The status pill on the Details tab. Most tabs and most actions need Running.

2. Are the nodes healthy?

Cluster OverviewNode Health. Grey (UNKNOWN) or red (DOWN) dots mean those nodes are not usable, which makes jobs pend for no visible reason.

sinfo -N -o "%N %t %E"

The %E column shows why a node was drained.

3. Is your storage mounted where you think?

df -h | grep -E '/pfs|/shared'
echo test > /pfs/write-check && rm /pfs/write-check && echo "writable"
The commonest root cause of all

Work written outside a mounted volume — /root, /tmp, /home without per-user logins — is erased whenever the pod is recreated. If output "disappeared", check where it was written before assuming anything is broken.

4. What is Slurm actually saying?

squeue -o "%.18i %.9P %.30j %.8T %.10M %.20R"   # the reason column is the answer
scontrol show job <jobid>
sacct -j <jobid> --format=JobID,State,ExitCode,Elapsed,NodeList

5. What do the logs say?

The Logs tab for the worker your job landed on, plus your own --output file on shared storage.


Cluster Problems

Stuck in Creating

Provisioning normally takes a few minutes.

  1. Wait a little longer — the list refreshes itself while anything is Creating.
  2. Try Actions → Reconfigure Cluster. It is available on a Creating cluster and re-applies the intended state.
  3. If it does not move, note the cluster name and contact support.

Status is Failed

A Failed cluster cannot be repaired into a Running one.

  1. Actions → Clone Cluster — you get the same configuration in a fresh create form.
  2. Adjust anything that might have caused it — most often a plan with no inventory. Try a different GPU type, fewer nodes, or a Private Cluster.
  3. Create the clone, confirm it reaches Running, then Delete the failed record.

A Failed cluster is not billed for compute, so there is no urgency.

A node is stuck UNKNOWN

UNKNOWN means the controller has no report for that node.

  1. Give it a few minutes — normal right after a restart, scale-up or image update.
  2. Check that worker's log on the Logs tab.
  3. Reboot that node from the Nodes tab.
  4. If several nodes are affected, Restart All Workers — but note it ends every running job.

Everything looks fine but nothing schedules

Check whether a partition is the constraint rather than capacity:

sinfo                              # is the partition UP?
scontrol show partition <name> # its nodes, MaxTime, state

A partition in DOWN, DRAIN or INACTIVE accepts nothing. A partition whose nodes were removed by a scale-down stops existing altogether — see Partitions.


FAQs

Does a Slurm Cluster cost more when it is busy? No. You pay for the nodes, per hour, whether they are idle or saturated. Jobs are never billed separately. See Billing.

Can I pause a cluster to stop paying? No. The only way to stop compute charges is to terminate it. Clone it first if you want the configuration back later — a Terminated cluster can still be cloned.

Do I need to recreate the cluster to use a different PyTorch or CUDA version? No. Pass --container-image to srun/sbatch and the job runs in whatever image you name. The cluster's Image is its own Slurm software stack, not your job's environment. See Containers.

Can two teams use different framework versions on one cluster? Yes — that is exactly what container jobs are for. Different jobs can use completely different images at the same time.

Can I add a partition while the cluster is being created? No. A partition names specific nodes, which do not exist yet. Add it from the Cluster Overview tab once the cluster is Running. See Partitions.

Can I edit slurm.conf over SSH? You can, and it will be overwritten. The file is regenerated whenever the cluster changes — a scale, an image update, a volume mount. Use Slurm Configuration so the change persists.

Why was my slurm.conf line rejected? Either the key is platform-managed, or the syntax is invalid — most often a space in the value. See validation messages.

Does scaling down interrupt jobs? Yes — on the removed nodes, and the remaining workers restart too. Drain first. See Scale Cluster.

Will an image update interrupt my jobs? Yes, but gracefully: nodes are drained one at a time, roughly 3–8 minutes each. Long-running jobs are still affected, so schedule it.

Where should checkpoints go? A PFS, SFS or Weka mount. Never a node-local path. See Storage.

How do I see whose job is whose? Enable Login User Management. Without it everyone is root and the User column is meaningless.

Can I cancel a job from the console? No. Use scancel <jobid> over SSH. The Jobs tab is read-only.

Does terminating a cluster delete my data? No. Volumes are unmounted, not deleted — and they keep billing. Delete them separately under Storage when you are done.

Can I get a GPU type that shows "Inventory not available"? Not at that moment. Retry later, or reserve capacity with a Private Cluster.


When to Contact Support

Raise a ticket from Support in the TIR sidebar, or email cloud-platform@e2enetworks.com.

Include the cluster name, the project, and:

ProblemAlso include
Stuck in Creating or TerminatingHow long, and whether Reconfigure was tried
Repeated XID errorsThe XID code, the e2e_node_<id> name, and whether a reboot helped
A node that will not registerThat worker's log from the Logs tab
Fabric performance far below expectationYour NCCL benchmark output and the NCCL_* variables you set
A configuration change that fails to applyThe exact configuration and the error message
Download logs before you restart anything

Restarting recreates the pods and the previous logs are gone. Grab them from the Logs tab first, or you will be asking support to diagnose something with no evidence.


Last updated on September 10, 2026.