This project explores running AI agent workloads on OpenStack infrastructure, leveraging Nova’s compute capabilities to provision and manage virtual machines for autonomous AI agents.

Overview

Agentic workloads are AI-driven processes that autonomously execute tasks, make decisions, and interact with their environment. OpenStack provides the ideal infrastructure layer for running these workloads at scale.

Key Capabilities

  • VM provisioning for isolated agent environments
  • GPU passthrough via vGPU for accelerated inference
  • Dynamic scaling based on workload demand
  • Network isolation between agent instances

Architecture

┌──────────────┐     ┌──────────────┐     ┌──────────────┐
│  Agent API   │────▶│  Nova        │────▶│  Compute     │
│  (Orchestr.) │     │  (Scheduler) │     │  (Hypervisor)│
└──────────────┘     └──────────────┘     └──────────────┘
       │                                         │
       ▼                                         ▼
┌──────────────┐                          ┌──────────────┐
│  Placement   │                          │  Agent VM    │
│  (Resources) │                          │  (Workload)  │
└──────────────┘                          └──────────────┘

Quick Start

  1. Deploy an OpenStack cloud with Nova and Placement services
  2. Configure a flavor with GPU resources for agent workloads: 
    openstack flavor create agent-gpu \
  --vcpus 4 --ram 8192 --disk 40 \
  --property resources:VGPU=1
  3. Launch an agent instance: 
    openstack server create my-agent \
  --flavor agent-gpu \
  --image ubuntu-22.04 \
  --network agent-net

How It Works

The system uses Nova’s scheduling and placement services to match agent workload requirements with available compute resources:

  • Placement API tracks GPU inventory across compute nodes
  • Nova Scheduler selects the best host based on resource availability
  • Live migration enables workload rebalancing without agent downtime