BareStax runs your NVIDIA GPU cluster from the installed server to the job queue
BareStax installs, expands, updates, and keeps NVIDIA GPU clusters running in your data center. It manages the path from an installed Ubuntu host to a schedulable BareStax Core node, with the same operating model on connected and fully air-gapped networks.
- Cluster summary
- Cluster
- NodeProfile
- Node
- Onboarding
- Fabric
- StorageTarget
- DriverStack
- Tenant
- Queue
- Jobs
- Usage report
- Operation
- Approvals
- Release and image
- Policies
- Events
- Alarms
- Audit
- Runbooks
| Awaiting approval drain-gpu-03 | |
| Drain · gpu-03 · 0/1 | 11:07 |
| user:operator-1 → — | |
- E-PROV-007 Configuration drift that requires downtime Show impact
- E-NODE-008 Node unreachable
| NAME | PROFILE | PHASE | READY | CORE | CODE |
|---|---|---|---|---|---|
| gpu-01 | hgx-h200 | Ready | True | partial | |
| gpu-02 | hgx-h200 | Ready | Unknown | drained | E-NODE-008 |
| gpu-03 | hgx-h200 | Ready | True | busy | E-PROV-007 |
| gpu-04 | hgx-h200 | Ready | True | busy | |
| gpu-05 | hgx-b300 | Ready | True | busy | |
| gpu-06 | hgx-b300 | Verifying | False | — | |
| gpu-07 | hgx-b300 | Provisioning | False | — | |
| gpu-08 | pcie-l40s | Ready | True | idle |
E-NODE-008 · 10 min
- E-NODE-008 Node unreachable
RB-NODE-008
Your team performs the physical handoff
-
Rack the servers
Connect the GPU nodes, head node, storage, and cluster network in your data center.
-
Install the operating system
Keep your own Ubuntu installation. BareStax works above it and does not erase the disk.
-
Run one command and give one approval
The node reports its hardware, receives the matching profile, installs a tested software set, and passes cluster checks. BareStax handles the rest.
Declarative control reaches below the scheduler
No layers. Just compute.
Kubernetes applies desired state and reconciliation to workloads it manages. BareStax applies that operating model to layers Kubernetes does not manage: system settings, kernel, NVIDIA drivers, CUDA, NVLink, NVSwitch, InfiniBand, BMC, and job scheduling.
There is no Kubernetes or virtualization layer between the job and the GPU. BareStax Core assigns the full GPU directly to the job. BareStax classifies drift by whether correction needs downtime; disruptive changes wait for approved maintenance.
One control model covers the cluster lifecycle
-
Day 0
Install the control plane, connect identity, define network and storage targets, and test the alert path.
barestax-server init -
Day 1
One command and one approval start inventory, profile selection, stack installation, and hardware validation before BareStax Core accepts work.
barestaxd joinbarestaxctl node approve -
Day 2
Drift that needs no downtime is fixed in place; anything that does waits for an approved maintenance window. A faulty node takes no new work while its running job continues. Recovery follows policy.
barestaxctl impact node gpu-03 -
Growth
New servers use the same join path, with separate queues for different GPU classes in one pool.
barestaxd join -
Updates
Signed releases start on one canary node and continue within the disruption budget while running jobs remain protected.
barestax-server upgrade --offline -
Removal
Drain and remove a node without losing active work. A returning node receives a new identity.
barestaxctl node drainbarestaxctl node remove
Six operating functions share one control model
- Declarative stack management Profiles keep the kernel, driver, CUDA libraries, fabric software, runtime, and system settings in tested sets. Details
- BareStax Core and job execution BareStax writes the BareStax Core configuration itself; queues follow GPU classes, and accounting is ready when the first job lands. Details
- Health and recovery Faults stop new work, preserve the active job, and expose impact before recovery. Details
- Air-gapped releases The connected-side bundler prepares a release for transfer into the isolated network. Details
- Security and audit mTLS, cluster PKI, and four-eyes approval protect cluster operations; audit details remain on the Security page. Details
- Console and portal Operators move from alerts to impact and approved operations in the console, while researchers use the matching portal for jobs, sessions, files, and queue reasons. Details
Every fault is tied to operational impact
A failed node is not an isolated red status. The impact view identifies affected jobs, tenants, and services, separates verified impact from possible impact, and states the age of its data.
The path runs from alert to impact and controlled operation; a broken live stream marks the view as stale.
| JOBID | TENANT | USER | STATE | GPU | IMPACT |
|---|---|---|---|---|---|
| 4521 | research-a | researcher-1 | RUNNING | 4 | Verified |
| 4533 | research-a | researcher-2 | RUNNING | 4 | Verified |
- Verified: 1 node, 2 jobs, 1 tenant
- E-PROV-007 · RB-PROV-007
Air-gapped operation uses the same release process
Runtime stays inside the institution's network. Signed bundles cross by approved media or data diode, then deploy by canary within the disruption budget. Deployment details
Errors explain themselves in English and Turkish
English and Turkish cover the console, portal, CLI messages, errors, and runbooks; every coded error points to its offline runbook.
The first session starts with what is in the rack
GPU classes, fabric, and storage set the agenda. We work from your hardware, not a slide deck.