Charmed HPC Slurm Cluster (Juju)

The slurm-juju deployment application deploys a complete Charmed HPC Slurm cluster — the vantage-slurm-charm-operators Topology‑A bundle — into an existing Juju model, then wires the single Vantage secret the cluster needs. Unlike slurm-multipass (a single VM) or the LXD/Juju extension apps, slurm-juju brings up the full control plane, compute, accounting, identity, shared storage, and observability as Juju applications.

slurm-juju does no Juju provisioning. It assumes you already have a bootstrapped controller and a model; it only runs juju deploy plus the post‑deploy secret steps. It targets the controller and model you name in --options.

What it deploys

A single juju deploy of the bundle brings up ~18 applications and their relations, including:

Role	Applications
Control plane	slurmctld, slurmdbd, slurmrestd, sackd (login)
Compute	slurmd (scale with juju add-unit slurmd -n <N>)
Accounting	mysql
Identity	sssd (SSSD against Vantage LDAP, on every Slurm node)
Vantage integration	vantage-agent (auth proxy + tunnel + OIDC/JWKS)
Shared storage	nfs-data / nfs-srv / nfs-home servers + fs-data / fs-srv / fs-home mounts
Observability	monitoring (Prometheus + Grafana), node-exporter, slurm-job-exporter
Containers	apptainer

The applications above are deployed under these short names; each pulls its corresponding vantage-* charm from Charmhub (e.g. the slurmctld application uses the vantage-slurmctld charm).

Charms are pulled from Charmhub on the edge channel — nothing is built locally. See the upstream deployment guide and architecture for the full topology.

Prerequisites

slurm-juju deploys into infrastructure you provide. Before you start you need:

• The juju CLI installed (it is a strictly‑confined snap — see Troubleshooting):

  sudo snap install juju
  juju version

• A bootstrapped controller and an existing model on a cloud that supports VM constraints (the bundle constrains Slurm/NFS units to virt-type=virtual-machine). For a local sandbox on LXD:

  sudo snap install lxd --channel latest/stable
  sudo lxd init --auto
  juju bootstrap localhost local        # controller named "local"
  juju add-model v8x-1                   # model named "v8x-1"

• Charmhub reachable from the machine running v8x (charms resolve from edge).
• v8x installed and authenticated, and an on_prem cloud account.

slurm-juju never provisions Juju

It will not bootstrap a controller or create a model. If the controller or model named in --options does not exist, the deploy fails fast — create them first with juju bootstrap / juju add-model.

Create the cloud account

v8x cloud account create local-juju --provider on_prem
v8x cloud account list --provider on_prem

Deploy Slurm

Point slurm-juju at the controller and model you bootstrapped:

v8x cluster create charmed-hpc-dev \
  --cloud-account local-juju \
  --app slurm-juju \
  --options controller=local,model=v8x-1

The --options flag for slurm-juju is a comma‑separated list. Both keys are required:

Key	Required	Purpose
controller	yes	Name of an existing Juju controller (valid juju name: lowercase letters, digits, hyphens)
model	yes	Name of an existing Juju model on that controller

Together they form the juju -m controller:model target for every command.

Roadmap

A slurmd-units=<N> option to scale compute at deploy time is planned but not yet implemented. For now, scale after deploy with juju add-unit slurmd -n <N>.

What happens under the hood

slurm-juju reproduces the upstream deployment runbook automatically:

1. Verifies the juju binary is on PATH (aborts with install guidance if not).

2. Renders the bundle. The only value injected is the sssd application's ldap-uri, derived from your active profile's vantage_url (the first hostname label is swapped for openldap), e.g. https://app.dev.vantagecompute.ai → ldaps://openldap.dev.vantagecompute.ai:636.

3. Deploys the bundle into the target model: juju deploy -m controller:model <bundle>.

4. Creates and wires the vantage-cluster secret — the one required post‑deploy step. v8x does this for you (you never run juju add-secret by hand). It carries five keys, all sourced from the Vantage cluster + your settings:

   Secret key	Source
   vantage-url	your profile's vantage_url
   client-id	cluster OIDC client id
   client-secret	cluster OIDC client secret
   org-id	your organization id
   ldap-bind-password	cluster SSSD bind password

   The secret is granted to vantage-agent and sssd, and vantage-secret=<secret-id> is set on both. From there the agent configures itself and SSSD renders sssd.conf; slurmrestd, Grafana, and Prometheus targets are wired by relation (no extra config).

Inspect the deployment

v8x cluster get charmed-hpc-dev
v8x app deployment list

# Watch the cluster converge (Slurm installs from a tarball; first convergence takes a few minutes)
juju status -m local:v8x-1 --watch 5s

# Confirm Slurm sees the compute node
juju exec -m local:v8x-1 --unit slurmctld/0 -- sinfo

Access Slurm

# SSH to the login node
juju ssh -m local:v8x-1 sackd/0

# Submit a quick job
srun -p slurmd hostname

Delete the deployment

Removal tears down the bundle's applications and the vantage-cluster secret, but leaves the Juju model intact (it does not destroy the model it deployed into):

v8x app deployment list                       # find the deployment id
v8x app deployment delete <deployment-id> --force

Or remove via the cluster record:

v8x cluster delete charmed-hpc-dev --cluster-type slurm --app slurm-juju --force

Troubleshooting

juju version
juju models                                   # confirm the controller/model exist
juju status -m <controller>:<model>
juju debug-log -m <controller>:<model>

Juju Required / juju CLI not found — install the binary: sudo snap install juju.

controller/model not found — slurm-juju does not create them. Run juju bootstrap and juju add-model, then re‑run with matching --options controller=,model=.

warning

no charm was found at "<path>" or permission denied on deploy

This is Juju snap confinement, not a v8x bug. The juju CLI is a strictly‑confined snap: it cannot read the host /tmp (the snap has a private /tmp) nor hidden dot‑directories under your home (e.g. ~/.v8x, ~/.cache). v8x already works around this by writing the temporary bundle to a non‑hidden directory under $HOME. If you hand juju deploy a path yourself, keep it out of /tmp and out of dot‑dirs.

Charms fail to resolve — confirm Charmhub is reachable and the vantage-* charms are published on edge. The bundle pins charms to the edge channel.

The secret looks half‑applied after a failed deploy — juju add-secret is not idempotent. Clear it before retrying, or remove the deployment (which also removes the secret):

juju remove-secret -m <controller>:<model> vantage-cluster

Re‑run any deploy with -v for verbose output:

v8x cluster create charmed-hpc-dev --cloud-account local-juju --app slurm-juju \
  --options controller=local,model=v8x-1 -v

Next steps

• Usage Examples
• Troubleshooting
• vantage-slurm-charm-operators documentation