Self-Hosted Cluster Deployment and Operations#

This guide provides Platform Engineers and site reliability engineers (SREs) with a structured path for deploying and operating a self-hosted NVCF stack on Amazon Elastic Kubernetes Service (EKS), Azure Kubernetes Service (AKS), or an on-premises Kubernetes cluster for running Omniverse workloads. For a Kubernetes cluster you operate yourself, see the On-Premises Deployment Guide.

Tip

You can automate most of this deployment with Agentic AI Skills, which plan and apply the self-hosted stack (infrastructure, NVCF, caches, storage, and validation). See Agentic AI Skills and the streaming-self-hosted repository.

Overview#

A self-hosted NVCF deployment runs on Kubernetes (EKS, AKS, or an on-premises cluster you operate) and consists of:

  • Control Plane: API services, invocation handling, and workload orchestration

  • GPU Nodes: Worker nodes with NVIDIA GPUs managed by the NVIDIA Cluster Agent (NVCA)

  • Container Registry: Amazon Elastic Container Registry (ECR) or Azure Container Registry (ACR) for storing your containerized Cloud Functions

High-Level Steps#

1. Self-Hosted Cluster Deployment#

Complete these steps to deploy the core NVCF platform:

Step

Description

Documentation

1a

Access and mirror artifacts - Obtain NVCF images and Helm charts from NGC and mirror them to your registry

Self-Hosted Image Mirroring

1b

Provision Kubernetes infrastructure - Create or select the Kubernetes cluster and cloud resources for the control plane and GPU workers

CSP End-to-End Example

1c

Deploy core services - Install infrastructure components (NATS, Cassandra, OpenBao) and control plane services

Helmfile Installation

2. Enhancements#

Add these components based on your workload requirements:

Component

Description

Caching Services

Accelerate deployments and improve performance with the container cache, the Omniverse Content Cache (OVCC), and the Omniverse Derived Cache (OVDC), which also serves as the shader cache. See Caches.

Low Latency Streaming (LLS)

Enable WebRTC streaming for Omniverse Streaming workloads. Required for real-time interactive applications. See Low Latency Streaming.

3. Integration#

Connect NVCF with your existing infrastructure:

Area

Description

Storage

Configure persistent storage for your Omniverse assets. See Storage API documentation.

Observability

Set up monitoring, logging, and tracing for your NVCF deployment. See Observability Configuration.

4. Validation#

Verify your deployment functions correctly:

Activity

Description

Cluster Validation

Confirm control plane services are healthy and GPU nodes are registered. See NVCF Troubleshooting / FAQ.

Performance Testing

Validate Omniverse Streaming performance.

Deployment Checklist#

Use this checklist to track your progress:

Phase

Task

Prerequisites

Obtain NGC access and API keys

Prerequisites

Set up AWS or Azure account with required permissions

Prerequisites

Install kubectl and Helm

Artifacts

Mirror NVCF images to ECR or ACR

Artifacts

Download self-hosted stack and compute-plane stack packages

Cluster

Provision Kubernetes cluster and required cloud resources

Cluster

Configure networking and security groups

Core

Deploy infrastructure components

Core

Deploy control plane services

Core

Deploy NVCA operator

Enhancements

Deploy caching services

Enhancements

Deploy LLS for streaming

Integration

Configure storage

Integration

Set up observability

Validation

Verify cluster health

Validation

Run performance tests

Next Steps#

After completing the deployment process, your cluster is ready to host NVIDIA Cloud Functions. Share these resources with your development team: