Validate GPUDirect RDMA Performance with DMA-BUF

💡 Quick Answer: Deploy two pods with mellanox/cuda-perftest, run ib_write_bw --use_cuda=0 --use_cuda_dmabuf between them, and verify throughput reaches near line-rate (80–95+ Gbps for 100G NICs).

After configuring GPUDirect RDMA, validate that GPU-to-GPU transfers over the network achieve expected throughput using the ib_write_bw benchmark with DMA-BUF.

Step 1 — Get Network Interface Name

kubectl exec -it -n network-operator mofed-ubuntu22.04-ds-xxxxx -- ibdev2netdev

Example output:

mlx5_0 port 1 ==> ens64np1 (Up)

Step 2 — Deploy Test Pods

Create pods requesting GPU and RDMA resources on two different nodes:

apiVersion: v1
kind: Pod
metadata:
  name: rdma-gpu-pod-1
  annotations:
    k8s.v1.cni.cncf.io/networks: rdma-test-network
spec:
  nodeSelector:
    kubernetes.io/hostname: gpu-node-1
  restartPolicy: OnFailure
  containers:
  - image: mellanox/cuda-perftest
    name: rdma-gpu-test
    securityContext:
      capabilities:
        add: ["IPC_LOCK"]
    resources:
      limits:
        nvidia.com/gpu: 1
        rdma/rdma_shared_device_a: 1
      requests:
        nvidia.com/gpu: 1
        rdma/rdma_shared_device_a: 1

Create a matching pod for the second node (rdma-gpu-pod-2 on gpu-node-2).

kubectl apply -f rdma-gpu-pod-1.yaml -f rdma-gpu-pod-2.yaml
kubectl get pods -o wide

Step 3 — Run the Benchmark

Start the server on pod 1:

kubectl exec -it rdma-gpu-pod-1 -- ib_write_bw --use_cuda=0 --use_cuda_dmabuf \
  -d mlx5_0 -a -F --report_gbits -q 1

Run the client on pod 2 (replace IP with pod 1 address):

kubectl exec -it rdma-gpu-pod-2 -- ib_write_bw -n 5000 --use_cuda=0 --use_cuda_dmabuf \
  -d mlx5_0 -a -F --report_gbits -q 1 <pod-1-ip>

Step 4 — Interpret Results

Expected output (100G NIC):

#bytes     #iterations    BW peak[Gb/sec]    BW average[Gb/sec]   MsgRate[Mpps]
65536      5000             92.39              92.38               0.176196
131072     5000             92.42              92.41               0.088131
1048576    5000             92.40              92.40               0.011015
8388608    5000             92.39              92.39               0.001377

Performance targets:

• 100G NIC — expect 90–95 Gbps for large messages
• 200G NIC — expect 180–195 Gbps for large messages
• Small messages — lower throughput is normal due to latency overhead

Step 5 — Validate DMA-BUF Path

The --use_cuda_dmabuf flag confirms the DMA-BUF path. If it falls back to legacy nvidia-peermem, you will see errors or warnings in the output.

Also verify with NCCL:

NCCL_DEBUG=INFO NCCL_NET_GDR_LEVEL=5 all_reduce_test

Look for GPUDirect RDMA DMA-BUF enabled and no peer memory fallback.

Cleanup

kubectl delete pod rdma-gpu-pod-1 rdma-gpu-pod-2

Why This Matters

Benchmarking confirms that GPUDirect RDMA is functioning at the hardware level. Without validation, misconfigurations can silently degrade multi-node training throughput by falling back to CPU-staged transfers.