Alauda GPU Management

Architecture

Alauda GPU Management manages the GPU host stack through a single ClusterPolicy custom resource. After the operator is installed, you create one ClusterPolicy; the operator reconciles the managed components onto the GPU nodes and keeps them in the declared state.

TOC

What the operator manages for youThe driver-mode forkComponents not managed by this operator

What the operator manages for you

Managed componentWhat it does for the userDefault
DriverMakes the NVIDIA kernel modules and user-space driver available on each GPU node, according to the selected driver mode.Enabled (precompiled), or disabled in preinstalled-driver mode
Container toolkitConfigures the containerd runtime and CDI so that GPU workloads can access the device through the nvidia RuntimeClass.Enabled
Device pluginAdvertises GPUs to Kubernetes as the nvidia.com/gpu schedulable resource.Enabled
GPU Feature DiscoveryLabels nodes with GPU model, driver version, and related facts used for scheduling.Enabled
DCGM-ExporterExposes GPU metrics for ACP monitoring and dashboards.Enabled
Driver managerCoordinates driver rollout, node drain, and upgrade sequencing.Enabled with the driver

You do not install or upgrade these components individually. They are declared in one ClusterPolicy and reconciled together, which is the difference from the previous model where each capability was a separate cluster plugin.

The driver-mode fork

The one decision that changes how the stack is deployed is the driver mode. It determines whether the operator delivers the driver and where the driver comes from:

  • Precompiled driver (immutable OS) — the operator stages a precompiled driver image so an immutable host such as SUSE Linux Micro stays unchanged. The matching driver image is selected for the node kernel.
  • Preinstalled host driver — the driver already exists on the host; the operator sets driver.enabled=false and manages only the runtime and Kubernetes-facing components.
  • Upstream driver (mutable OS) — the operator pulls the NVIDIA driver image and builds modules on the node; this mode is configurable but relies on NVIDIA-hosted images.

The driver mode is chosen once, before creating the ClusterPolicy. See Configure the driver mode for the mechanics of each mode and Install for the decision table.

Components not managed by this operator

Some capabilities are related to GPUs but are not part of this product's default scope:

  • GPU sharing, memory/compute partitioning, and HAMi scheduling are provided by HAMi. This operator provides the driver and runtime base that HAMi builds on. See the HAMi documentation.
  • MIG partitioning is available in the operator but off by default and depends on a user-supplied mig-manager image and MIG-capable hardware. See MIG resources.
  • vGPU, KubeVirt GPU passthrough, confidential computing, and GPUDirect Storage are feature-gated and not part of the default delivery. Their images are not included in the product package.