Skip to main content

Configuring MIG Profiles and Launching VMs with OpenStack

Once MIG mode is enabled on an NVIDIA A100 GPU and instances are created, the next step is to integrate those GPU instances into an OpenStack environment. This allows virtual machines to access specific GPU partitions with hardware-level isolation. This article outlines the steps to configure MIG profiles and deploy VMs with virtual GPUs in OpenStack Nova.

Creating MIG Profiles​

Each MIG profile defines a specific slice of GPU memory and compute resources. On an A100 40GB GPU, a typical profile might be 1g.5gb, which offers one-seventh of the compute and 5GB of memory.

To create MIG instances:

sudo nvidia-smi mig -cgi 19,19,19 -C

This command creates three instances of profile ID 19 (i.e., three 1g.5gb instances). To list available profiles and their IDs:

nvidia-smi mig -lgip

After instance creation, confirm their presence:

nvidia-smi mig -lgi

Each GPU Instance (GI) includes a Compute Instance (CI), and both are required to execute workloads. Use the -lci flag to list compute instances and confirm their allocation.

Mapping MIG Devices for OpenStack​

OpenStack treats MIG instances as mediated devices (mdevs). To expose them to Nova:

Update nova.conf:

[devices]
enabled_mdev_types = nvidia-476

Restart Nova Compute:

sudo systemctl restart nova-compute

Confirm mdev types are detected:

openstack resource provider list

Verify VGPU availability:

openstack allocation candidate list --resource VGPU=1

This output should reflect the MIG-backed resources available to Nova schedulers.

Defining Flavors for GPU-Backed VMs​

Each VM flavor in OpenStack can include GPU traits for resource matching. For example, to create a flavor that requests a single virtual GPU:

openstack flavor create \
  --ram 8192 \
  --disk 40 \
  --vcpus 4 \
  --property resources:VGPU=1 \
  gpu.small

If you want to target a specific MIG profile, additional custom traits can be registered and used with traits: or aggregate_instance_extra_specs.

Launching a VM with a MIG-Backed VGPU​

With flavors and scheduling rules in place, deploy a VM that utilizes one of the available MIG-backed GPU slices:

openstack server create \
  --flavor gpu.small \
  --image <image-id> \
  --network <network-id> \
  --key-name <keypair> \
  gpu-vm-01

Inside the guest VM, install the NVIDIA GRID guest driver, which is required for vGPU device recognition. After installation, validate with:

nvidia-smi

The output will show the assigned MIG device and its available memory and compute capacity.

Managing MIG Devices​

MIG instances can be destroyed and recreated dynamically:

sudo nvidia-smi mig -dci
sudo nvidia-smi mig -dgi

Keep in mind that MIG devices are not persistent across system reboots unless recreated automatically. Use tooling such as mig-parted or a systemd service to manage reinitialization on boot.

Operational Notes​

  • Isolation: MIG provides hardware-level memory and compute isolation. This ensures predictable performance across tenant workloads.

  • Flexibility: You can mix different MIG profiles on a single GPU to balance performance and density.

  • Security: Each MIG instance functions as a discrete device, reducing the risk of cross-tenant interference.

Next Steps​

The next article will focus on Installing and Managing NVIDIA GRID Drivers inside virtual machines and preparing the guest operating system to make full use of the allocated GPU resources.

OpenMetal’s infrastructure is designed to support these advanced virtualization features, enabling organizations to run GPU-powered workloads with full control and without vendor lock-in.