platform-eng
Kubernetes GPU Scheduling for Isolated Tenant Clusters
Schedule GPU jobs independently for each customer. vMetal provisions capacity, and vCluster gives tenants separate schedulers with dedicated workers through Private Nodes.
Trusted by the fastest-growing AI cloud providers
Problem
Why GPU Scheduling Breaks Down
Namespace-level isolation can't fix the contention built into shared GPU clusters.
Namespace Isolation Is Too Weak
Tenants depend on one API and worker pool. Contention on common workers can affect neighboring GPU workloads.
Dedicated Clusters Are Too Expensive
Provisioning a full physical cluster per tenant multiplies infrastructure cost and slows time to production significantly.
No Scheduler Separation Per Tenant
Quotas and priorities govern one scheduler, but do not give tenants independent scheduling policies and control planes.
Solution
Dedicated Schedulers for Every GPU Tenant
vMetal provisions GPU machines. vCluster gives each tenant its own API, data store and scheduler, with Private Nodes dedicating workers and their networking and storage configuration. vCluster Labs software powers 100K+ GPUs.
Built for GPU Scheduling at Scale
vCluster isolates each layer GPU scheduling touches, from the control plane down to the node and the workload runtime.
Control Plane Isolation
Dedicated Scheduler per GPU Tenant
Each tenant has an independent Kubernetes API, data store and RBAC boundary. Lightweight control-plane hosting reduces dedicated server requirements for GPU scheduling; Private Nodes keep production workers tenant-specific.
- Independent tenant API and data store
- Tenant-scoped scheduling on Private Nodes
- No dedicated control-plane servers per tenant
Node Isolation
Dedicated GPU Nodes Per Tenant
Dedicate GPU workers to each tenant through Private Nodes. This removes cross-tenant workload placement from those workers for GPU scheduling; network bandwidth, storage and GPU partitioning still need explicit policies.
- Dedicated worker nodes per tenant
- Tenant-owned CNI and storage configuration
- Physical servers or supported VMs
AI Workload Integration
Supported AI Workload Templates
Use supported workload templates for GPU scheduling, including the NVIDIA Run:AI partner integration. Configure GPU, network and storage prerequisites, and validate the selected template for Private Nodes before offering it to production customers.
- Native NVIDIA Run:AI catalog
- Validate Private Node configuration
- Release-specific integration prerequisites
Workload Isolation
Kernel-Native Workload Isolation
vNode uses Linux user namespaces and seccomp to restrict workload privileges and system calls. It adds runtime hardening for GPU scheduling without a guest kernel or hypervisor, complementing dedicated workers rather than allocating GPUs.
- Linux user namespaces and seccomp
- No guest kernel or hypervisor
- Helps limit workload breakout
Fleet Operations
Central Management Across GPU Clusters
Coordinate tenant environments for GPU scheduling through one UI, CLI and API. Templates, access policies and quotas provide central operator controls while tenant Kubernetes APIs and dedicated production workers remain separate.
- Unified UI, CLI, and API
- Per-tenant quotas and resource limits
- SSO and RBAC across the fleet
Why vCluster
This isn’t a side project. Behind every vCluster deployment is 5+ years of deep K8s engineering, security hardening, and battle-tested infrastructure work at massive scale.
100K+
GPUs Powered
50+
GPU Clouds & F500s
<45
Days to Launch
30K
GitHub Stars
Get Started in 3 Steps
1
Schedule a Demo
Talk to our team about your stack
2
Deploy vCluster
Deploy vCluster on your infra in minutes
3
Onboard Your Tenants
Go live with a hyperscaler-grade tenant experience in days
FAQs
How does vCluster improve Kubernetes GPU scheduling for isolated tenant clusters?
With Private Nodes, each tenant cluster runs its own Kubernetes scheduler against its own dedicated workers. The API, scheduler and backing store are components of the tenant control plane; they are not necessarily separate pods. Tenant priority and placement policies operate within that cluster. GPU allocation still depends on supported drivers and device resources, while vMetal supplies machine provisioning.
What is the difference between namespace isolation and tenant cluster isolation for GPU scheduling?
Kubernetes supports PriorityClasses, resource quotas and scheduling policies within a common cluster. Those controls do not give each tenant an independent API or scheduler. vCluster with Private Nodes separates tenant control planes and dedicates workers, letting customers manage placement policies within their own environment. GPU driver configuration and capacity planning remain necessary within each tenant's nodes.
Does vCluster support GPU-specific scheduling features like topology awareness?
Tenant clusters expose standard Kubernetes scheduling controls, including node selectors, affinity, topology spread and extended resources. With Private Nodes, customers can configure supported GPU device plugins or DRA drivers on their own workers. Hardware topology and driver capabilities determine which devices can be allocated; Kubernetes conformance alone does not certify a particular GPU configuration or guarantee workload performance.
Can I assign dedicated GPU nodes to specific tenants in vCluster?
Yes. With Private Nodes, the recommended production model, each tenant's GPU nodes join only that tenant's cluster, so no other tenant's workloads can be scheduled onto those servers. You configure this per tenant cluster, which lets GPU cloud operators build service levels on dedicated GPU capacity. For internal dev, test and CI work by trusted teams, Shared Nodes are also available.
How does vCluster prevent GPU resource contention between tenants?
Dedicated Private Nodes remove cross-tenant workload placement from those workers. Bare metal avoids a hypervisor layer, but this is not a throughput or latency guarantee. GPU drivers, workload configuration, network fabrics and storage can still influence performance. vNode adds runtime hardening using the Linux kernel; benchmark the selected stack against the workload and service levels you intend to offer.
Is vCluster production-proven for Kubernetes GPU scheduling at scale?
vCluster Labs software powers 100K+ GPUs and 1M+ CPUs, serving 50+ GPU clouds & Fortune 500s combined. In production, Lintasarta operates 170+ tenant clusters, while Boost Run completed its managed Kubernetes launch in under 45 days from the decision. These named deployments provide scale and launch examples for teams evaluating the platform for their own infrastructure.
Separate Tenant GPU Scheduling and Workers
See how vCluster gives every tenant its own scheduler and dedicated GPU nodes.
