Modern AI Factories increasingly need to support heterogeneous workloads, including HPC simulations, AI training, AI inference, cloud-native applications, and traditional virtual machines, on the same infrastructure. Static resource partitioning between Slurm clusters, Kubernetes environments, and virtualized workloads often leads to poor GPU utilization and operational inefficiencies.

To run an AI Factory that is both efficient and secure, administrators have to achieve two goals at once: 

• Maximize resource usage: keep expensive GPUs and InfiniBand fabrics busy by letting any tenant access them when further capacity is needed.
• Isolate tenants: guarantee that teams sharing the same physical infrastructure cannot compromise each other’s workloads and data.

Operators can rely on static partitioning to guarantee tenant isolation, but those rigid hardware islands come at the cost of poor utilization. Overcoming this inefficiency requires pooling resources (i.e. compute nodes, GPUs, high-speed networks) and allocating them to isolated environments dynamically, so that resources are given to one environment when it needs them and reallocated to another when it doesn’t.

OpenNebula Elastic Capacity Management reconciles both goals, effectively eliminating the trade-off between tenant isolation and infrastructure utilization. It lets multiple Slurm and Kubernetes clusters share a common pool of GPU-enabled, InfiniBand-connected servers, dynamically and with strong multi-tenant isolation, so that devices stay busy without tenants’ workloads interfering with each other.

The key technologies and capabilities of OpenNebula that enable Elastic Capacity Management are the following:

• Virtualization: each Slurm or Kubernetes node runs as a virtual machine, so it can be created, destroyed, and moved across the shared pool in minutes. 
• GPU and InfiniBand passthrough: physical GPUs and InfiniBand adapters are attached directly to the virtual machines via PCI passthrough, delivering native accelerator and fabric performance inside a VM.
• Elastic scaling: grow and shrink Slurm and Kubernetes environments dynamically as demand shifts, so idle capacity in one environment can be reclaimed and handed to another.
• Multi-tenancy and isolation: each tenant runs in its own isolated environment, within separate virtual clusters, virtual networks, and access controls, so independent teams share the same physical infrastructure without interfering with each other’s workloads.
• Unified management: provision HPC clusters, Kubernetes environments, and traditional VMs from a single OpenNebula deployment, simplifying operations.

This screencast demonstrates how OpenNebula Elastic Capacity Management enables dynamic resource sharing across multiple Slurm and Kubernetes clusters while preserving tenant isolation and high-performance access to accelerated hardware.

Using an AI Factory testbed provided by CESGA (Galician Supercomputing Center), we show how OpenNebula manages a shared pool of GPU-enabled servers interconnected with NVIDIA InfiniBand, allowing compute nodes, GPUs, and networking resources to be provisioned and reallocated on demand between independent HPC and Kubernetes environments.

The demonstration also includes InfiniBand performance benchmarking, comparing native execution with virtual machines using PCI passthrough. The results demonstrate near-zero virtualization overhead, confirming that OpenNebula can deliver the flexibility of virtualization while preserving the performance required by demanding AI and HPC workloads.

The screencast covers:

• Building a shared AI Factory infrastructure managed by OpenNebula.
• Provisioning independent Slurm and Kubernetes clusters for different tenants.
• Dynamic elastic capacity management across workload environments.
• GPU and InfiniBand passthrough for high-performance virtual machines.
• InfiniBand benchmarking demonstrating near-native performance with virtualization.

Special thanks to CESGA for providing the AI Factory infrastructure used to produce this demonstration

To learn more about how OpenNebula enables AI Factories, explore our AI Factory solution.