This is a cache of https://docs.okd.io/4.19/hardware_accelerators/rdma-remote-direct-memory-access.html. It is a snapshot of the page at 2025-07-26T21:09:03.391+0000.
R<strong>e</strong>mot<strong>e</strong> Dir<strong>e</strong>ct M<strong>e</strong>mory Acc<strong>e</strong>ss (RDMA) | Hardwar<strong>e</strong> acc<strong>e</strong>l<strong>e</strong>rators | OKD 4.19
  1. Documentation
&times;

NVIDIA GPUDirect Remote Direct Memory Access (RDMA)

NVIDIA GPUDirect Remote Direct Memory Access (RDMA) allows for the memory in one computer to directly access the memory of another computer without needing access through the operating system. This provides the ability to bypass kernel intervention in the process, freeing up resources and greatly reducing the CPU overhead normally needed to process network communications. This is useful for distributing GPU-accelerated workloads across clusters. And because RDMA is so suited toward high bandwidth and low latency applications, this makes it ideal for big data and machine learning applications.

There are currently three configuration methods for NVIDIA GPUDirect RDMA:

Shared device

This method allows for an NVIDIA GPUDirect RDMA device to be shared among multiple pods on the OKD worker node where the device is exposed.

Host device

This method provides direct physical ethernet access on the worker node by creating an additional host network on a pod. A plugin allows the network device to be moved from the host network namespace to the network namespace on the pod.

SR-IOV legacy device

The Single Root IO Virtualization (SR-IOV) method can share a single network device, such as an ethernet adapter, with multiple pods. SR-IOV segments the device, recognized on the host node as a physical function (PF), into multiple virtual functions (VFs). The VF is used like any other network device.

each of these methods can be used across either the NVIDIA GPUDirect RDMA over Converged ethernet (RoCe) or Infiniband infrastructures, providing an aggregate total of six methods of configuration.

NVIDIA GPUDirect RDMA prerequisites

All methods of NVIDIA GPUDirect RDMA configuration require the installation of specific Operators. Use the following steps to install the Operators:

Disabling the IRDMA kernel module

On some systems, including the DellR750xa, the IRDMA kernel module creates problems for the NVIDIA Network Operator when unloading and loading the DOCA drivers. Use the following procedure to disable the module.

Procedure

  1. Generate the following machine configuration file by running the following command:

    $ cat <<eOF > 99-machine-config-blacklist-irdma.yaml
    example output
    apiVersion: machineconfiguration.openshift.io/v1
kind: MachineConfig
metadata:
  labels:
    machineconfiguration.openshift.io/role: worker
  name: 99-worker-blacklist-irdma
spec:
  kernelArguments:
    - "module_blacklist=irdma"

  2. Create the machine configuration on the cluster and wait for the nodes to reboot by running the following command:

    $ oc create -f 99-machine-config-blacklist-irdma.yaml
    example output
    machineconfig.machineconfiguration.openshift.io/99-worker-blacklist-irdma created

  3. Validate in a debug pod on each node that the module has not loaded by running the following command:

    $ oc debug node/nvd-srv-32.nvidia.eng.rdu2.dc.redhat.com
Starting pod/nvd-srv-32nvidiaengrdu2dcredhatcom-debug-btfj2 ...
To use host binaries, run `chroot /host`
Pod IP: 10.6.135.11
If you don't see a command prompt, try pressing enter.
sh-5.1# chroot /host
sh-5.1# lsmod|grep irdma
sh-5.1#

Creating persistent naming rules

In some cases, device names won’t persist following a reboot. For example, on R760xa systems Mellanox devices might be renamed after a reboot. You can avoid this problem by using a MachineConfig to set persistence.

Procedure

  1. Gather the MAC address names from the worker nodes for the node into a file and provide names for the interfaces that need to persist. This example uses the file 70-persistent-net.rules and stashes the details in it.

    $ cat <<eOF > 70-persistent-net.rules
SUBSYSTeM=="net",ACTION=="add",ATTR{address}=="b8:3f:d2:3b:51:28",ATTR{type}=="1",NAMe="ibs2f0"
SUBSYSTeM=="net",ACTION=="add",ATTR{address}=="b8:3f:d2:3b:51:29",ATTR{type}=="1",NAMe="ens8f0np0"
SUBSYSTeM=="net",ACTION=="add",ATTR{address}=="b8:3f:d2:f0:36:d0",ATTR{type}=="1",NAMe="ibs2f0"
SUBSYSTeM=="net",ACTION=="add",ATTR{address}=="b8:3f:d2:f0:36:d1",ATTR{type}=="1",NAMe="ens8f0np0"
eOF

  2. Convert that file into a base64 string without line breaks and set the output to the variable PeRSIST:

    $ PeRSIST=`cat 70-persistent-net.rules| base64 -w 0`

$ echo $PeRSIST
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

  3. Create a machine configuration and set the base64 encoding in the custom resource file by running the following command:

    $ cat <<eOF > 99-machine-config-udev-network.yaml
    apiVersion: machineconfiguration.openshift.io/v1
kind: MachineConfig
metadata:
   labels:
     machineconfiguration.openshift.io/role: worker
   name: 99-machine-config-udev-network
spec:
   config:
     ignition:
       version: 3.2.0
     storage:
       files:
       - contents:
           source: data:text/plain;base64,$PeRSIST
         filesystem: root
         mode: 420
         path: /etc/udev/rules.d/70-persistent-net.rules

  4. Create the machine configuration on the cluster by running the following command:

    $ oc create -f 99-machine-config-udev-network.yaml
    example output
    machineconfig.machineconfiguration.openshift.io/99-machine-config-udev-network created

  5. Use the get mcp command to view the machine configuration status:

    $ oc get mcp
    example output
    NAMe     CONFIG                                             UPDATeD   UPDATING   DeGRADeD   MACHINeCOUNT   ReADYMACHINeCOUNT   UPDATeDMACHINeCOUNT   DeGRADeDMACHINeCOUNT   AGe
master   rendered-master-9adfe851c2c14d9598eea5ec3df6c187   True      False      False      1              1                   1                     0                      6h21m
worker   rendered-worker-4568f1b174066b4b1a4de794cf538fee   False     True       False      2              0                   0                     0                      6h21m

The nodes will reboot and when the updating field returns to false, you can validate on the nodes by looking at the devices in a debug pod.

Configuring the NFD Operator

The Node Feature Discovery (NFD) Operator manages the detection of hardware features and configuration in an OKD cluster by labeling the nodes with hardware-specific information. NFD labels the host with node-specific attributes, such as PCI cards, kernel, operating system version, and so on.

Prerequisites

  • You have installed the NFD Operator.

Procedure

  1. Validate that the Operator is installed and running by looking at the pods in the openshift-nfd namespace by running the following command:

    $ oc get pods -n openshift-nfd
    example output
    NAMe                                      ReADY   STATUS    ReSTARTS   AGe
nfd-controller-manager-8698c88cdd-t8gbc   2/2     Running   0          2m

  2. With the NFD controller running, generate the NodeFeatureDiscovery instance and add it to the cluster.

    The ClusterServiceVersion specification for NFD Operator provides default values, including the NFD operand image that is part of the Operator payload. Retrieve its value by running the following command:

    $ NFD_OPeRAND_IMAGe=`echo $(oc get csv -n openshift-nfd -o json | jq -r '.items[0].metadata.annotations["alm-examples"]') | jq -r '.[] | select(.kind == "NodeFeatureDiscovery") | .spec.operand.image'`

  3. Optional: Add entries to the default deviceClassWhiteList field, to support more network adapters, such as the NVIDIA BlueField DPUs.

    apiVersion: nfd.openshift.io/v1
kind: NodeFeatureDiscovery
metadata:
  name: nfd-instance
  namespace: openshift-nfd
spec:
  instance: ''
  operand:
    image: '${NFD_OPeRAND_IMAGe}'
    servicePort: 12000
  prunerOnDelete: false
  topologyUpdater: false
  workerConfig:
    configData: |
      core:
        sleepInterval: 60s
      sources:
        pci:
          deviceClassWhitelist:
            - "02"
            - "03"
            - "0200"
            - "0207"
            - "12"
          deviceLabelFields:
            - "vendor"

  4. Create the 'NodeFeatureDiscovery` instance by running the following command:

    $ oc create -f nfd-instance.yaml
    example output
    nodefeaturediscovery.nfd.openshift.io/nfd-instance created

  5. Validate that the instance is up and running by looking at the pods under the openshift-nfd namespace by running the following command:

    $ oc get pods -n openshift-nfd
    example output
    NAMe                                    ReADY   STATUS    ReSTARTS   AGe
nfd-controller-manager-7cb6d656-jcnqb   2/2     Running   0          4m
nfd-gc-7576d64889-s28k9                 1/1     Running   0          21s
nfd-master-b7bcf5cfd-qnrmz              1/1     Running   0          21s
nfd-worker-96pfh                        1/1     Running   0          21s
nfd-worker-b2gkg                        1/1     Running   0          21s
nfd-worker-bd9bk                        1/1     Running   0          21s
nfd-worker-cswf4                        1/1     Running   0          21s
nfd-worker-kp6gg                        1/1     Running   0          21s

  6. Wait a short period of time and then verify that NFD has added labels to the node. The NFD labels are prefixed with feature.node.kubernetes.io, so you can easily filter them.

    $ oc get node -o json | jq '.items[0].metadata.labels | with_entries(select(.key | startswith("feature.node.kubernetes.io")))'
{
  "feature.node.kubernetes.io/cpu-cpuid.ADX": "true",
  "feature.node.kubernetes.io/cpu-cpuid.AeSNI": "true",
  "feature.node.kubernetes.io/cpu-cpuid.AVX": "true",
  "feature.node.kubernetes.io/cpu-cpuid.AVX2": "true",
  "feature.node.kubernetes.io/cpu-cpuid.CeTSS": "true",
  "feature.node.kubernetes.io/cpu-cpuid.CLZeRO": "true",
  "feature.node.kubernetes.io/cpu-cpuid.CMPXCHG8": "true",
  "feature.node.kubernetes.io/cpu-cpuid.CPBOOST": "true",
  "feature.node.kubernetes.io/cpu-cpuid.eFeR_LMSLe_UNS": "true",
  "feature.node.kubernetes.io/cpu-cpuid.FMA3": "true",
  "feature.node.kubernetes.io/cpu-cpuid.FP256": "true",
  "feature.node.kubernetes.io/cpu-cpuid.FSRM": "true",
  "feature.node.kubernetes.io/cpu-cpuid.FXSR": "true",
  "feature.node.kubernetes.io/cpu-cpuid.FXSROPT": "true",
  "feature.node.kubernetes.io/cpu-cpuid.IBPB": "true",
  "feature.node.kubernetes.io/cpu-cpuid.IBRS": "true",
  "feature.node.kubernetes.io/cpu-cpuid.IBRS_PReFeRReD": "true",
  "feature.node.kubernetes.io/cpu-cpuid.IBRS_PROVIDeS_SMP": "true",
  "feature.node.kubernetes.io/cpu-cpuid.IBS": "true",
  "feature.node.kubernetes.io/cpu-cpuid.IBSBRNTRGT": "true",
  "feature.node.kubernetes.io/cpu-cpuid.IBSFeTCHSAM": "true",
  "feature.node.kubernetes.io/cpu-cpuid.IBSFFV": "true",
  "feature.node.kubernetes.io/cpu-cpuid.IBSOPCNT": "true",
  "feature.node.kubernetes.io/cpu-cpuid.IBSOPCNTeXT": "true",
  "feature.node.kubernetes.io/cpu-cpuid.IBSOPSAM": "true",
  "feature.node.kubernetes.io/cpu-cpuid.IBSRDWROPCNT": "true",
  "feature.node.kubernetes.io/cpu-cpuid.IBSRIPINVALIDCHK": "true",
  "feature.node.kubernetes.io/cpu-cpuid.IBS_FeTCH_CTLX": "true",
  "feature.node.kubernetes.io/cpu-cpuid.IBS_OPFUSe": "true",
  "feature.node.kubernetes.io/cpu-cpuid.IBS_PReVeNTHOST": "true",
  "feature.node.kubernetes.io/cpu-cpuid.INT_WBINVD": "true",
  "feature.node.kubernetes.io/cpu-cpuid.INVLPGB": "true",
  "feature.node.kubernetes.io/cpu-cpuid.LAHF": "true",
  "feature.node.kubernetes.io/cpu-cpuid.LBRVIRT": "true",
  "feature.node.kubernetes.io/cpu-cpuid.MCAOVeRFLOW": "true",
  "feature.node.kubernetes.io/cpu-cpuid.MCOMMIT": "true",
  "feature.node.kubernetes.io/cpu-cpuid.MOVBe": "true",
  "feature.node.kubernetes.io/cpu-cpuid.MOVU": "true",
  "feature.node.kubernetes.io/cpu-cpuid.MSRIRC": "true",
  "feature.node.kubernetes.io/cpu-cpuid.MSR_PAGeFLUSH": "true",
  "feature.node.kubernetes.io/cpu-cpuid.NRIPS": "true",
  "feature.node.kubernetes.io/cpu-cpuid.OSXSAVe": "true",
  "feature.node.kubernetes.io/cpu-cpuid.PPIN": "true",
  "feature.node.kubernetes.io/cpu-cpuid.PSFD": "true",
  "feature.node.kubernetes.io/cpu-cpuid.RDPRU": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SeV": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SeV_64BIT": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SeV_ALTeRNATIVe": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SeV_DeBUGSWAP": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SeV_eS": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SeV_ReSTRICTeD": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SeV_SNP": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SHA": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SMe": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SMe_COHeReNT": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SPeC_CTRL_SSBD": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SSe4A": "true",
  "feature.node.kubernetes.io/cpu-cpuid.STIBP": "true",
  "feature.node.kubernetes.io/cpu-cpuid.STIBP_ALWAYSON": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SUCCOR": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SVM": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SVMDA": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SVMFBASID": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SVML": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SVMNP": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SVMPF": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SVMPFT": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SYSCALL": "true",
  "feature.node.kubernetes.io/cpu-cpuid.SYSee": "true",
  "feature.node.kubernetes.io/cpu-cpuid.TLB_FLUSH_NeSTeD": "true",
  "feature.node.kubernetes.io/cpu-cpuid.TOPeXT": "true",
  "feature.node.kubernetes.io/cpu-cpuid.TSCRATeMSR": "true",
  "feature.node.kubernetes.io/cpu-cpuid.VAeS": "true",
  "feature.node.kubernetes.io/cpu-cpuid.VMCBCLeAN": "true",
  "feature.node.kubernetes.io/cpu-cpuid.VMPL": "true",
  "feature.node.kubernetes.io/cpu-cpuid.VMSA_ReGPROT": "true",
  "feature.node.kubernetes.io/cpu-cpuid.VPCLMULQDQ": "true",
  "feature.node.kubernetes.io/cpu-cpuid.VTe": "true",
  "feature.node.kubernetes.io/cpu-cpuid.WBNOINVD": "true",
  "feature.node.kubernetes.io/cpu-cpuid.X87": "true",
  "feature.node.kubernetes.io/cpu-cpuid.XGeTBV1": "true",
  "feature.node.kubernetes.io/cpu-cpuid.XSAVe": "true",
  "feature.node.kubernetes.io/cpu-cpuid.XSAVeC": "true",
  "feature.node.kubernetes.io/cpu-cpuid.XSAVeOPT": "true",
  "feature.node.kubernetes.io/cpu-cpuid.XSAVeS": "true",
  "feature.node.kubernetes.io/cpu-hardware_multithreading": "false",
  "feature.node.kubernetes.io/cpu-model.family": "25",
  "feature.node.kubernetes.io/cpu-model.id": "1",
  "feature.node.kubernetes.io/cpu-model.vendor_id": "AMD",
  "feature.node.kubernetes.io/kernel-config.NO_HZ": "true",
  "feature.node.kubernetes.io/kernel-config.NO_HZ_FULL": "true",
  "feature.node.kubernetes.io/kernel-selinux.enabled": "true",
  "feature.node.kubernetes.io/kernel-version.full": "5.14.0-427.35.1.el9_4.x86_64",
  "feature.node.kubernetes.io/kernel-version.major": "5",
  "feature.node.kubernetes.io/kernel-version.minor": "14",
  "feature.node.kubernetes.io/kernel-version.revision": "0",
  "feature.node.kubernetes.io/memory-numa": "true",
  "feature.node.kubernetes.io/network-sriov.capable": "true",
  "feature.node.kubernetes.io/pci-102b.present": "true",
  "feature.node.kubernetes.io/pci-10de.present": "true",
  "feature.node.kubernetes.io/pci-10de.sriov.capable": "true",
  "feature.node.kubernetes.io/pci-15b3.present": "true",
  "feature.node.kubernetes.io/pci-15b3.sriov.capable": "true",
  "feature.node.kubernetes.io/rdma.available": "true",
  "feature.node.kubernetes.io/rdma.capable": "true",
  "feature.node.kubernetes.io/storage-nonrotationaldisk": "true",
  "feature.node.kubernetes.io/system-os_release.ID": "rhcos",
  "feature.node.kubernetes.io/system-os_release.OPeNSHIFT_VeRSION": "4.17",
  "feature.node.kubernetes.io/system-os_release.OSTRee_VeRSION": "417.94.202409121747-0",
  "feature.node.kubernetes.io/system-os_release.RHeL_VeRSION": "9.4",
  "feature.node.kubernetes.io/system-os_release.VeRSION_ID": "4.17",
  "feature.node.kubernetes.io/system-os_release.VeRSION_ID.major": "4",
  "feature.node.kubernetes.io/system-os_release.VeRSION_ID.minor": "17"
}

  7. Confirm there is a network device that is discovered:

    $ oc describe node | grep -e 'Roles|pci' | grep pci-15b3
                    feature.node.kubernetes.io/pci-15b3.present=true
                    feature.node.kubernetes.io/pci-15b3.sriov.capable=true
                    feature.node.kubernetes.io/pci-15b3.present=true
                    feature.node.kubernetes.io/pci-15b3.sriov.capable=true

Configuring the SR-IOV Operator

Single root I/O virtualization (SR-IOV) enhances the performance of NVIDIA GPUDirect RDMA by providing sharing across multiple pods from a single device.

Prerequisites

  • You have installed the SR-IOV Operator.

Procedure

  1. Validate that the Operator is installed and running by looking at the pods in the openshift-sriov-network-operator namespace by running the following command:

    $ oc get pods -n openshift-sriov-network-operator
    example output
    NAMe                                      ReADY   STATUS    ReSTARTS   AGe
sriov-network-operator-7cb6c49868-89486   1/1     Running   0          22s

  2. For the default SriovOperatorConfig CR to work with the MLNX_OFeD container, run this command to update the following values:

    apiVersion: sriovnetwork.openshift.io/v1
kind: SriovOperatorConfig
metadata:
  name: default
  namespace: openshift-sriov-network-operator
spec:
  enableInjector: true
  enableOperatorWebhook: true
  logLevel: 2

  3. Create the resource on the cluster by running the following command:

    $ oc create -f sriov-operator-config.yaml
    example output
    sriovoperatorconfig.sriovnetwork.openshift.io/default created

  4. Patch the sriov-operator so the MOFeD container can work with it by running the following command:

    $ oc patch sriovoperatorconfig default   --type=merge -n openshift-sriov-network-operator   --patch '{ "spec": { "configDaemonNodeSelector": { "network.nvidia.com/operator.mofed.wait": "false", "node-role.kubernetes.io/worker": "", "feature.node.kubernetes.io/pci-15b3.sriov.capable": "true" } } }'
    example output
    sriovoperatorconfig.sriovnetwork.openshift.io/default patched

Configuring the NVIDIA network Operator

The NVIDIA network Operator manages NVIDIA networking resources and networking related components such as drivers and device plugins to enable NVIDIA GPUDirect RDMA workloads.

Prerequisites

  • You have installed the NVIDIA network Operator.

Procedure

  1. Validate that the network Operator is installed and running by confirming the controller is running in the nvidia-network-operator namespace by running the following command:

    $ oc get pods -n nvidia-network-operator
    example output
    NAMe                                                          ReADY   STATUS             ReSTARTS        AGe
nvidia-network-operator-controller-manager-6f7d6956cd-fw5wg   1/1     Running            0                5m

  2. With the Operator running, create the NicClusterPolicy custom resource file. The device you choose depends on your system configuration. In this example, the Infiniband interface ibs2f0 is hard coded and is used as the shared NVIDIA GPUDirect RDMA device.

    apiVersion: mellanox.com/v1alpha1
kind: NicClusterPolicy
metadata:
  name: nic-cluster-policy
spec:
  nicFeatureDiscovery:
    image: nic-feature-discovery
    repository: ghcr.io/mellanox
    version: v0.0.1
  docaTelemetryService:
    image: doca_telemetry
    repository: nvcr.io/nvidia/doca
    version: 1.16.5-doca2.6.0-host
  rdmaSharedDevicePlugin:
    config: |
      {
        "configList": [
          {
            "resourceName": "rdma_shared_device_ib",
            "rdmaHcaMax": 63,
            "selectors": {
              "ifNames": ["ibs2f0"]
            }
          },
          {
            "resourceName": "rdma_shared_device_eth",
            "rdmaHcaMax": 63,
            "selectors": {
              "ifNames": ["ens8f0np0"]
            }
          }
        ]
      }
    image: k8s-rdma-shared-dev-plugin
    repository: ghcr.io/mellanox
    version: v1.5.1
  secondaryNetwork:
    ipoib:
      image: ipoib-cni
      repository: ghcr.io/mellanox
      version: v1.2.0
  nvIpam:
    enableWebhook: false
    image: nvidia-k8s-ipam
    repository: ghcr.io/mellanox
    version: v0.2.0
  ofedDriver:
    readinessProbe:
      initialDelaySeconds: 10
      periodSeconds: 30
    forcePrecompiled: false
    terminationGracePeriodSeconds: 300
    livenessProbe:
      initialDelaySeconds: 30
      periodSeconds: 30
    upgradePolicy:
      autoUpgrade: true
      drain:
        deleteemptyDir: true
        enable: true
        force: true
        timeoutSeconds: 300
        podSelector: ''
      maxParallelUpgrades: 1
      safeLoad: false
      waitForCompletion:
        timeoutSeconds: 0
    startupProbe:
      initialDelaySeconds: 10
      periodSeconds: 20
    image: doca-driver
    repository: nvcr.io/nvidia/mellanox
    version: 24.10-0.7.0.0-0
    env:
    - name: UNLOAD_STORAGe_MODULeS
      value: "true"
    - name: ReSTORe_DRIVeR_ON_POD_TeRMINATION
      value: "true"
    - name: CReATe_IFNAMeS_UDeV
      value: "true"

  3. Create the NicClusterPolicy custom resource on the cluster by running the following command:

    $ oc create -f network-sharedrdma-nic-cluster-policy.yaml
    example output
    nicclusterpolicy.mellanox.com/nic-cluster-policy created

  4. Validate the NicClusterPolicy by running the following command in the DOCA/MOFeD container:

    $ oc get pods -n nvidia-network-operator
    example output
    NAMe                                                          ReADY   STATUS    ReSTARTS   AGe
doca-telemetry-service-hwj65                                  1/1     Running   2          160m
kube-ipoib-cni-ds-fsn8g                                       1/1     Running   2          160m
mofed-rhcos4.16-9b5ddf4c6-ds-ct2h5                            2/2     Running   4          160m
nic-feature-discovery-ds-dtksz                                1/1     Running   2          160m
nv-ipam-controller-854585f594-c5jpp                           1/1     Running   2          160m
nv-ipam-controller-854585f594-xrnp5                           1/1     Running   2          160m
nv-ipam-node-xqttl                                            1/1     Running   2          160m
nvidia-network-operator-controller-manager-5798b564cd-5cq99   1/1     Running   2         5d23h
rdma-shared-dp-ds-p9vvg                                       1/1     Running   0          85m

  5. rsh into the mofed container to check the status by running the following command:

    $ MOFeD_POD=$(oc get pods -n nvidia-network-operator -o name | grep mofed)
$ oc rsh -n nvidia-network-operator -c mofed-container ${MOFeD_POD}
sh-5.1# ofed_info -s
    example output
    OFeD-internal-24.07-0.6.1:
    sh-5.1# ibdev2netdev -v
    example output
    0000:0d:00.0 mlx5_0 (MT41692 - 900-9D3B4-00eN-eA0) BlueField-3 e-series SuperNIC 400Gbe/NDR single port QSFP112, PCIe Gen5.0 x16 FHHL, Crypto enabled, 16GB DDR5, BMC, Tall Bracket                                                       fw 32.42.1000 port 1 (ACTIVe) ==> ibs2f0 (Up)
0000:a0:00.0 mlx5_1 (MT41692 - 900-9D3B4-00eN-eA0) BlueField-3 e-series SuperNIC 400Gbe/NDR single port QSFP112, PCIe Gen5.0 x16 FHHL, Crypto enabled, 16GB DDR5, BMC, Tall Bracket                                                       fw 32.42.1000 port 1 (ACTIVe) ==> ens8f0np0 (Up)

  6. Create a IPoIBNetwork custom resource file:

    apiVersion: mellanox.com/v1alpha1
kind: IPoIBNetwork
metadata:
  name: example-ipoibnetwork
spec:
  ipam: |
    {
      "type": "whereabouts",
      "range": "192.168.6.225/28",
      "exclude": [
       "192.168.6.229/30",
       "192.168.6.236/32"
      ]
    }
  master: ibs2f0
  networkNamespace: default

  7. Create the IPoIBNetwork resource on the cluster by running the following command:

    $ oc create -f ipoib-network.yaml
    example output
    ipoibnetwork.mellanox.com/example-ipoibnetwork created

  8. Create a MacvlanNetwork custom resource file for your other interface:

    apiVersion: mellanox.com/v1alpha1
kind: MacvlanNetwork
metadata:
  name: rdmashared-net
spec:
  networkNamespace: default
  master: ens8f0np0
  mode: bridge
  mtu: 1500
  ipam: '{"type": "whereabouts", "range": "192.168.2.0/24", "gateway": "192.168.2.1"}'

  9. Create the resource on the cluster by running the following command:

    $ oc create -f macvlan-network.yaml
    example output
    macvlannetwork.mellanox.com/rdmashared-net created

Configuring the GPU Operator

The GPU Operator automates the management of the NVIDIA drivers, device plugins for GPUs, the NVIDIA Container Toolkit, and other components required for GPU provisioning.

Prerequisites

  • You have installed the GPU Operator.

Procedure

  1. Check that the Operator pod is running to look at the pods under the namespace by running the following command:

    $ oc get pods -n nvidia-gpu-operator
    example output
    NAMe                          ReADY   STATUS    ReSTARTS   AGe
gpu-operator-b4cb7d74-zxpwq   1/1     Running   0          32s

  2. Create a GPU cluster policy custom resource file similar to the following example:

    apiVersion: nvidia.com/v1
kind: ClusterPolicy
metadata:
  name: gpu-cluster-policy
spec:
  vgpuDeviceManager:
    config:
      default: default
    enabled: true
  migManager:
    config:
      default: all-disabled
      name: default-mig-parted-config
    enabled: true
  operator:
    defaultRuntime: crio
    initContainer: {}
    runtimeClass: nvidia
    use_ocp_driver_toolkit: true
  dcgm:
    enabled: true
  gfd:
    enabled: true
  dcgmexporter:
    config:
      name: ''
    serviceMonitor:
      enabled: true
    enabled: true
  cdi:
    default: false
    enabled: false
  driver:
    licensingConfig:
      nlsenabled: true
      configMapName: ''
    certConfig:
      name: ''
    rdma:
      enabled: false
    kernelModuleConfig:
      name: ''
    upgradePolicy:
      autoUpgrade: true
      drain:
        deleteemptyDir: false
        enable: false
        force: false
        timeoutSeconds: 300
      maxParallelUpgrades: 1
      maxUnavailable: 25%
      podDeletion:
        deleteemptyDir: false
        force: false
        timeoutSeconds: 300
      waitForCompletion:
        timeoutSeconds: 0
    repoConfig:
      configMapName: ''
    virtualTopology:
      config: ''
    enabled: true
    useNvidiaDriverCRD: false
    useOpenKernelModules: true
  devicePlugin:
    config:
      name: ''
      default: ''
    mps:
      root: /run/nvidia/mps
    enabled: true
  gdrcopy:
    enabled: true
  kataManager:
    config:
      artifactsDir: /opt/nvidia-gpu-operator/artifacts/runtimeclasses
  mig:
    strategy: single
  sandboxDevicePlugin:
    enabled: true
  validator:
    plugin:
      env:
        - name: WITH_WORKLOAD
          value: 'false'
  nodeStatusexporter:
    enabled: true
  daemonsets:
    rollingUpdate:
      maxUnavailable: '1'
    updateStrategy: RollingUpdate
  sandboxWorkloads:
    defaultWorkload: container
    enabled: false
  gds:
    enabled: true
    image: nvidia-fs
    version: 2.20.5
    repository: nvcr.io/nvidia/cloud-native
  vgpuManager:
    enabled: false
  vfioManager:
    enabled: true
  toolkit:
    installDir: /usr/local/nvidia
    enabled: true

  3. When the GPU ClusterPolicy custom resource has generated, create the resource on the cluster by running the following command:

    $ oc create -f gpu-cluster-policy.yaml
    example output
    clusterpolicy.nvidia.com/gpu-cluster-policy created

  4. Validate that the Operator is installed and running by running the following command:

    $ oc get pods -n nvidia-gpu-operator
    example output
    NAMe                                                  ReADY   STATUS      ReSTARTS   AGe
gpu-feature-discovery-d5ngn                           1/1     Running     0          3m20s
gpu-feature-discovery-z42rx                           1/1     Running     0          3m23s
gpu-operator-6bb4d4b4c5-njh78                         1/1     Running     0          4m35s
nvidia-container-toolkit-daemonset-bkh8l              1/1     Running     0          3m20s
nvidia-container-toolkit-daemonset-c4hzm              1/1     Running     0          3m23s
nvidia-cuda-validator-4blvg                           0/1     Completed   0          106s
nvidia-cuda-validator-tw8sl                           0/1     Completed   0          112s
nvidia-dcgm-exporter-rrw4g                            1/1     Running     0          3m20s
nvidia-dcgm-exporter-xc78t                            1/1     Running     0          3m23s
nvidia-dcgm-nvxpf                                     1/1     Running     0          3m20s
nvidia-dcgm-snj4j                                     1/1     Running     0          3m23s
nvidia-device-plugin-daemonset-fk2xz                  1/1     Running     0          3m23s
nvidia-device-plugin-daemonset-wq87j                  1/1     Running     0          3m20s
nvidia-driver-daemonset-416.94.202410211619-0-ngrjg   4/4     Running     0          3m58s
nvidia-driver-daemonset-416.94.202410211619-0-tm4x6   4/4     Running     0          3m58s
nvidia-node-status-exporter-jlzxh                     1/1     Running     0          3m57s
nvidia-node-status-exporter-zjffs                     1/1     Running     0          3m57s
nvidia-operator-validator-l49hx                       1/1     Running     0          3m20s
nvidia-operator-validator-n44nn                       1/1     Running     0          3m23s

  5. Optional: When you have verified the pods are running, remote shell into the NVIDIA driver daemonset pod and confirm that the NVIDIA modules are loaded. Specifically, ensure the nvidia_peermem is loaded.

    $ oc rsh -n nvidia-gpu-operator $(oc -n nvidia-gpu-operator get pod -o name -l app.kubernetes.io/component=nvidia-driver)
sh-4.4# lsmod|grep nvidia
    example output
    nvidia_fs             327680  0
nvidia_peermem         24576  0
nvidia_modeset       1507328  0
video                  73728  1 nvidia_modeset
nvidia_uvm           6889472  8
nvidia               8810496  43 nvidia_uvm,nvidia_peermem,nvidia_fs,gdrdrv,nvidia_modeset
ib_uverbs             217088  3 nvidia_peermem,rdma_ucm,mlx5_ib
drm                   741376  5 drm_kms_helper,drm_shmem_helper,nvidia,mgag200

  6. Optional: Run the nvidia-smi utility to show the details about the driver and the hardware:

sh-4.4# nvidia-smi

+ .example output

Wed Nov  6 22:03:53 2024
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.90.07              Driver Version: 550.90.07      CUDA Version: 12.4     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. eCC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA A40                     On  |   00000000:61:00.0 Off |                    0 |
|  0%   37C    P0             88W /  300W |       1MiB /  46068MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
|   1  NVIDIA A40                     On  |   00000000:e1:00.0 Off |                    0 |
|  0%   28C    P8             29W /  300W |       1MiB /  46068MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI        PID   Type   Process name                              GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|  No running processes found                                                             |
+-----------------------------------------------------------------------------------------+

  1. While you are still in the driver pod, set the GPU clock to maximum using the nvidia-smi command:

    $ oc rsh -n nvidia-gpu-operator nvidia-driver-daemonset-416.94.202410172137-0-ndhzc
sh-4.4# nvidia-smi -i 0 -lgc $(nvidia-smi -i 0 --query-supported-clocks=graphics --format=csv,noheader,nounits | sort -h | tail -n 1)
    example output
    GPU clocks set to "(gpuClkMin 1740, gpuClkMax 1740)" for GPU 00000000:61:00.0
All done.
    sh-4.4# nvidia-smi -i 1 -lgc $(nvidia-smi -i 1 --query-supported-clocks=graphics --format=csv,noheader,nounits | sort -h | tail -n 1)
    example output
    GPU clocks set to "(gpuClkMin 1740, gpuClkMax 1740)" for GPU 00000000:e1:00.0
All done.

  2. Validate the resource is available from a node describe perspective by running the following command:

    $ oc describe node -l node-role.kubernetes.io/worker=| grep -e 'Capacity:|Allocatable:' -A9
    example output
    Capacity:
  cpu:                          128
  ephemeral-storage:            1561525616Ki
  hugepages-1Gi:                0
  hugepages-2Mi:                0
  memory:                       263596712Ki
  nvidia.com/gpu:               2
  pods:                         250
  rdma/rdma_shared_device_eth:  63
  rdma/rdma_shared_device_ib:   63
Allocatable:
  cpu:                          127500m
  ephemeral-storage:            1438028263499
  hugepages-1Gi:                0
  hugepages-2Mi:                0
  memory:                       262445736Ki
  nvidia.com/gpu:               2
  pods:                         250
  rdma/rdma_shared_device_eth:  63
  rdma/rdma_shared_device_ib:   63
--
Capacity:
  cpu:                          128
  ephemeral-storage:            1561525616Ki
  hugepages-1Gi:                0
  hugepages-2Mi:                0
  memory:                       263596672Ki
  nvidia.com/gpu:               2
  pods:                         250
  rdma/rdma_shared_device_eth:  63
  rdma/rdma_shared_device_ib:   63
Allocatable:
  cpu:                          127500m
  ephemeral-storage:            1438028263499
  hugepages-1Gi:                0
  hugepages-2Mi:                0
  memory:                       262445696Ki
  nvidia.com/gpu:               2
  pods:                         250
  rdma/rdma_shared_device_eth:  63
  rdma/rdma_shared_device_ib:   63