Configuring Auto Scaling for xGPU Nodes

Prerequisites

• A cluster of v1.27 or later is available.
• CCE AI Suite (NVIDIA GPU) (v2.1.8, v2.7.5 or later), Volcano Scheduler (v1.10.5 or later), and CCE Cluster Autoscaler (v1.27.150, v1.28.78, v1.29.41, or later) have been installed in the cluster.
• The function of automatically adding nodes is enabled when pods in the cluster cannot be scheduled. To do so, go to Settings, click the Auto Scaling tab, enable Auto Scale-out when the load cannot be scheduled under Node Capacity Expansion Conditions.

Notes and Constraints

• If there are workloads that use compute-memory isolation and memory isolation in the same node pool, auto scaling is not supported for the node pool. If node pool auto scaling is enabled, scale-out behavior may become unpredictable.

Step 1: Configure the Node Pool

1. Log in to the CCE console and click the cluster name to access the cluster console. In the navigation pane, choose Nodes.
2. Click Create Node Pool to create an xGPU node pool. For details, see Creating a Node Pool.

   For details about requirements on xGPU nodes, such as the specifications, OS, and runtime, see Preparing Virtualized GPU Resources.

3. After the node pool is created, click Auto Scaling. In the AS Object area, enable Auto Scaling for the target specification and click OK.

Step 2: Configure Heterogeneous Resources

1. In the navigation pane, choose Settings. Then, click the Heterogeneous Resources tab.
2. In the GPU Settings area, locate Node Pool Configurations and select the created node pool.
3. Select a driver that meets GPU virtualization requirements and enable GPU virtualization based on Preparing Virtualized GPU Resources.
   • Clusters of v1.27: GPU virtualization can only be enabled cluster-wide.
     Figure 1 Heterogeneous resource settings for v1.27 clusters
     

   • Clusters of v1.28 or later: GPU virtualization can be enabled node-pool-wide.
     Figure 2 Heterogeneous resource settings for clusters of v1.28 or later
     

4. Click Confirm configuration.

Step 3: Create a GPU Virtualization Workload and Trigger Capacity Expansion

Create a Deployment that uses GPU virtualization resources and requests a number of GPUs exceeding the current upper limit available in the cluster. For details, see Using GPU Virtualization. For example, there is a total of 16 GiB of GPU memory available, with each pod requiring 1 GiB. Then, configure 17 pods, which need a total of 17 GiB of GPU memory.

After a short period of time, you can find GPU node scale-out on the node pool details page.

Helpful Links

To configure workload auto scaling based on GPU metrics, see Configuring Workload Scaling Based on GPU Monitoring Metrics.