GPUStack Operator

RDMA Networking

The Device Manager publishes usable remote direct memory access (RDMA) endpoints as Kubernetes resources. Kubelet allocates endpoints to containers that request them.

Contents

Overview

Interface discovery

The interface inventory records link state. The device plugin advertises eligible endpoints, and kubelet grants them to containers. Queue admission does not reserve RDMA endpoints.

RDMA discovery and allocation flow

Configuration

RDMA requests

Choose a key and endpoint count in RDMA Operations . Configure an optional selection policy to filter adapters. See Network Topology for inventory and link states.

Operations

  1. Step 1: Inspect capacity. Read the node’s allocatable RDMA keys.
  2. Step 2: Set alignment. Configure the kubelet policy if the workload needs NUMA alignment.
  3. Step 3: Verify the grant. Follow the RDMA walkthrough .

Read the effective policy in Verification . Monitor Pending Pods for endpoint exhaustion in RDMA Operations .

Troubleshooting

For unhealthy endpoints or Pending Pods, use the RDMA troubleshooting table . It links each symptom to a check and recovery action.