iRS assigns the GPU resources of the hosts in a cluster to a GPU resource pool for a group of VMs that provide the same service to use. When a VM starts or restarts, CVM allocates and releases GPU resources for VMs based on predefined policies.
GPU and vGPU resource pools are supported:
GPU resource pools are used for GPU passthrough. GPU passthrough uses VT-d to directly present a physical GPU to a VM. GPU passthrough does not decrease the resources and performance of the GPU. The GPU acts as if it were directly driven by the VM.
A vGPU resource pool allocates vGPUs to VMs.
With GPU resource pooling, the system can provide on-demand allocation and reclamation and automated elastic expansion services as follows:
The system allocates an idle GPU automatically to a VM if the VM runs a non-linear editing service. If the host attached to the VM does not have available GPUs, the VM is automatically migrated to a host that has idle GPUs to use the GPUs.
When a VM finishes non-linear editing services, the VM automatically releases the GPUs after shutdown so that other VMs can use the GPUs.