Use GPUs with clustered VMs
Applies to: Azure Stack HCI, version 21H2 Preview
This topic provides guidance on how to use graphics processing units (GPUs) with clustered virtual machines (VMs) running the Azure Stack HCI operating system to provide GPU acceleration to workloads in the clustered VMs.
Starting in Azure Stack HCI, version 21H2, you can include GPUs in your Azure Stack HCI cluster to provide GPU acceleration to workloads running in clustered VMs. This topic covers the basic prerequisites of this capability and how to deploy it.
GPU acceleration is provided via Discrete Device Assignment (DDA), also known as GPU pass-through, which allows you to dedicate one or more physical GPUs to a VM. Clustered VMs can take advantage of GPU acceleration, and clustering capabilities such as high availability via failover. Live migrating VMs isn't currently supported, but VMs can be automatically restarted and placed where GPU resources are available in the event of a failure.
To get started, you’ll need an Azure Stack HCI cluster of at least two servers, running Azure Stack HCI, version 21H2. You’ll also need GPUs that are physically installed in every server of the cluster.
The Azure Stack HCI Catalog does not yet indicate GPU compatibility or certification information. Follow your manufacturer's instructions for GPU installation.
This section describes the steps necessary to prepare your cluster servers for GPU usage, assign VMs to clustered GPU resource pools, and test automatic restart.
Prepare the cluster
Prepare the GPUs in each server by installing security mitigation drivers on each server, disabling the GPUs, and dismounting them from the host according to the instructions in Deploy graphics devices using Discrete Device Assignment. Depending on your hardware vendor, you may also need to configure any GPU licensing requirements.
Create a new empty resource pool on each server that will contain the clustered GPU resources. Make sure to provide the same pool name on each server.
In PowerShell, run the following cmdlet as an administrator:
New-VMResourcePool -ResourcePoolType PciExpress -Name "GpuChildPool"
Add the dismounted GPUs from each server to the resource pool that you created in the previous step.
In PowerShell, run the following cmdlets:
$gpu = Get-VMHostAssignableDevice
Add-VMHostAssignableDevice -HostAssignableDevice $gpu -ResourcePoolName "GpuChildPool"
You now have a cluster-wide resource pool (named
GpuChildPool) that is populated with assignable GPUs. The cluster will use this pool to determine VM placement for any started or moved VMs that are assigned to the GPU resource pool.
Assign a VM to a GPU resource pool
First, either create a new VM in your cluster, or find an existing VM.
Prepare the VM for DDA by setting its cache behavior, stop action, and memory-mapped I/O (MMIO) properties according to the instructions in Deploy graphics devices using Discrete Device Assignment.
Configure the cluster VM resource’s default offline action as
In PowerShell, run the following cmdlet:
$vm | Set-ClusterParameter -Name "OfflineAction" -Value 3
Assign the resource pool that you created earlier to the VM. This declares to the cluster that the VM requires an assigned device from the
GpuChildPoolpool when it's either started or moved.
In PowerShell, run the following cmdlet:
$vm | Add-VMAssignableDevice -ResourcePoolName "GpuChildPool"
If you start the VM now, the cluster ensures that it is placed on a server with available GPU resources from this cluster-wide pool. The cluster also assigns the GPU to the VM through DDA, which allows the GPU to be accessed from workloads inside the VM.
You also need to install drivers from your GPU manufacturer inside the VM so that apps in the VM can take advantage of the GPU assigned to them.
Fail over a VM with an assigned GPU
To test the cluster’s ability to keep your GPU workload available, perform a drain operation on the server where the VM is running with an assigned GPU. To drain the server, follow the instructions in Taking an Azure Stack HCI server offline for maintenance. The cluster will restart the VM on another server in the cluster, as long as another server has sufficient available GPU resources in the pool that you created.
For more information, see also: