A transit platform must prove Network Operator wired InfiniBand/RDMA correctly. What validation should operators run?
Select an answer to reveal the explanation.
Short Explanation
You can't prove RDMA wiring from host counters or cluster-Service reachability alone. Run a sample RDMA pod so you actually confirm the Network Operator wired InfiniBand into Kubernetes workloads.
Full Explanation
A transit platform proving that the Network Operator wired InfiniBand/RDMA correctly should test a sample RDMA pod to confirm InfiniBand/RDMA usability from Kubernetes workloads. A workload-level experiment validates the full path: scheduling against RDMA resources, attaching the secondary InfiniBand network, and exercising RDMA communication. For municipal transit ML, that practical proof is what justifies promoting fabric-attached jobs to production. Running an iperf TCP test over the cluster pod overlay to measure Ethernet throughput fails because Ethernet overlay throughput says nothing about InfiniBand/RDMA fabric attachment or RDMA usability from Kubernetes workloads. Checking InfiniBand interface counters on hosts with ibstat to verify link state fails because host link state alone does not prove the Network Operator wired RDMA into Kubernetes workloads or that pods can use the fabric. Verifying Kubernetes Service ClusterIP reachability between namespaces using a test pod fails conceptually: Service connectivity tests general cluster networking, not InfiniBand/RDMA enablement or fabric attachment by the operator. Exam caveat: A sample RDMA pod test should be representative but minimal—enough to prove attach and RDMA function without consuming the entire transit GPU pool. Failure of the sample must drive operator/network diagnosis, not a Kubernetes Service ClusterIP reachability check between namespaces, which does not prove InfiniBand/RDMA wiring. Operational check: Deploy a documented canary RDMA pod or Job on the transit platform, confirm it receives the expected InfiniBand/RDMA interface and passes a basic RDMA connectivity or bandwidth check, capture logs and resource allocations, and only then sign off Network Operator wiring as verified.