×

Host firmware tuning

New in this release
  • No reference design updates in this release

Description

Configure system level performance. See Configuring host firmware for low latency and high performance for recommended settings.

If Ironic inspection is enabled, the firmware setting values are available from the per-cluster BareMetalHost CR on the hub cluster. You enable Ironic inspection with a label in the spec.clusters.nodes field in the SiteConfig CR that you use to install the cluster. For example:

nodes:
  - hostName: "example-node1.example.com"
    ironicInspect: "enabled"

The telco RAN DU reference SiteConfig does not enable the ironicInspect field by default.

Limits and requirements
  • Hyperthreading must be enabled

Engineering considerations
  • Tune all settings for maximum performance

    You can tune firmware selections for power savings at the expense of performance as required.

Node Tuning Operator

New in this release
  • No reference design updates in this release

Description

You tune the cluster performance by creating a performance profile. Settings that you configure with a performance profile include:

  • Selecting the realtime or non-realtime kernel.

  • Allocating cores to a reserved or isolated cpuset. OKD processes allocated to the management workload partition are pinned to reserved set.

  • Enabling kubelet features (CPU manager, topology manager, and memory manager).

  • Configuring huge pages.

  • Setting additional kernel arguments.

  • Setting per-core power tuning and max CPU frequency.

Limits and requirements

The Node Tuning Operator uses the PerformanceProfile CR to configure the cluster. You need to configure the following settings in the RAN DU profile PerformanceProfile CR:

  • Select reserved and isolated cores and ensure that you allocate at least 4 hyperthreads (equivalent to 2 cores) on Intel 3rd Generation Xeon (Ice Lake) 2.20 GHz CPUs or better with firmware tuned for maximum performance.

  • Set the reserved cpuset to include both hyperthread siblings for each included core. Unreserved cores are available as allocatable CPU for scheduling workloads. Ensure that hyperthread siblings are not split across reserved and isolated cores.

  • Configure reserved and isolated CPUs to include all threads in all cores based on what you have set as reserved and isolated CPUs.

  • Set core 0 of each NUMA node to be included in the reserved CPU set.

  • Set the huge page size to 1G.

You should not add additional workloads to the management partition. Only those pods which are part of the OpenShift management platform should be annotated into the management partition.

Engineering considerations
  • You should use the RT kernel to meet performance requirements.

    You can use the non-RT kernel if required.

  • The number of huge pages that you configure depends on the application workload requirements. Variation in this parameter is expected and allowed.

  • Variation is expected in the configuration of reserved and isolated CPU sets based on selected hardware and additional components in use on the system. Variation must still meet the specified limits.

  • Hardware without IRQ affinity support impacts isolated CPUs. To ensure that pods with guaranteed whole CPU QoS have full use of the allocated CPU, all hardware in the server must support IRQ affinity. For more information, see About support of IRQ affinity setting.

In OKD 4.15, any PerformanceProfile CR configured on the cluster causes the Node Tuning Operator to automatically set all cluster nodes to use cgroup v1.

For more information about cgroups, see Configuring Linux cgroup.

PTP Operator

New in this release
  • No reference design updates in this release

Description

See PTP timing for details of support and configuration of PTP in cluster nodes. The DU node can run in the following modes:

  • As an ordinary clock (OC) synced to a grandmaster clock or boundary clock (T-BC)

  • As a grandmaster clock synced from GPS with support for single or dual card E810 Westport Channel NICs

  • As dual boundary clocks (one per NIC) with support for E810 Westport Channel NICs

    Highly available boundary clocks are not supported.

  • Optional: as a boundary clock for radio units (RUs)

Events and metrics for grandmaster clocks are a Tech Preview feature added in the 4.14 telco RAN DU RDS. For more information see Using the PTP hardware fast event notifications framework.

You can subscribe applications to PTP events that happen on the node where the DU application is running.

Limits and requirements
  • High availability is not supported with dual NIC configurations.

  • Digital Phase-Locked Loop (DPLL) clock synchronization is not supported for E810 Westport Channel NICs.

  • GPS offsets are not reported. Use a default offset of less than or equal to 5.

  • DPLL offsets are not reported. Use a default offset of less than or equal to 5.

Engineering considerations
  • Configurations are provided for ordinary clock, boundary clock, or grandmaster clock

  • PTP fast event notifications uses ConfigMap CRs to store PTP event subscriptions

  • Use Intel E810-XXV-4T Westport Channel NICs for PTP grandmaster clocks with GPS timing, minimum firmware version 4.40

SR-IOV Operator

New in this release
  • No reference design updates in this release

Description

The SR-IOV Operator provisions and configures the SR-IOV CNI and device plugins. Both netdevice (kernel VFs) and vfio (DPDK) devices are supported.

Engineering considerations
  • Customer variation on the configuration and number of SriovNetwork and SriovNetworkNodePolicy custom resources (CRs) is expected.

  • IOMMU kernel command line settings are applied with a MachineConfig CR at install time. This ensures that the SriovOperator CR does not cause a reboot of the node when adding them.

Logging

New in this release
  • No reference design updates in this release

Description

Use logging to collect logs from the far edge node for remote analysis. The recommended log collector is Vector.

Engineering considerations
  • Handling logs beyond the infrastructure and audit logs, for example, from the application workload requires additional CPU and network bandwidth based on additional logging rate.

  • As of OKD 4.14, Vector is the reference log collector.

    Use of fluentd in the RAN use model is deprecated.

SRIOV-FEC Operator

New in this release
  • No reference design updates in this release

Description

SRIOV-FEC Operator is an optional 3rd party Certified Operator supporting FEC accelerator hardware.

Limits and requirements
  • Starting with FEC Operator v2.7.0:

    • SecureBoot is supported

    • The vfio driver for the PF requires the usage of vfio-token that is injected into Pods. The VF token can be passed to DPDK by using the EAL parameter --vfio-vf-token.

Engineering considerations
  • The SRIOV-FEC Operator uses CPU cores from the isolated CPU set.

  • You can validate FEC readiness as part of the pre-checks for application deployment, for example, by extending the validation policy.

Local Storage Operator

New in this release
  • No reference design updates in this release

Description

You can create persistent volumes that can be used as PVC resources by applications with the Local Storage Operator. The number and type of PV resources that you create depends on your requirements.

Engineering considerations
  • Create backing storage for PV CRs before creating the PV. This can be a partition, a local volume, LVM volume, or full disk.

  • Refer to the device listing in LocalVolume CRs by the hardware path used to access each device to ensure correct allocation of disks and partitions. Logical names (for example, /dev/sda) are not guaranteed to be consistent across node reboots.

    For more information, see the Fedora 9 documentation on device identifiers.

LVMS Operator

New in this release
  • No reference design updates in this release

New in this release
  • Simplified LVMS deviceSelector logic

  • LVM Storage with ext4 and PV resources

LVMS Operator is an optional component.

Description

The LVMS Operator provides dynamic provisioning of block and file storage. The LVMS Operator creates logical volumes from local devices that can be used as PVC resources by applications. Volume expansion and snapshots are also possible.

The following example configuration creates a vg1 volume group that leverages all available disks on the node except the installation disk:

StorageLVMCluster.yaml
apiVersion: lvm.topolvm.io/v1alpha1
kind: LVMCluster
metadata:
  name: storage-lvmcluster
  namespace: openshift-storage
  annotations:
    ran.openshift.io/ztp-deploy-wave: "10"
spec:
  storage:
    deviceClasses:
    - name: vg1
      thinPoolConfig:
        name: thin-pool-1
        sizePercent: 90
        overprovisionRatio: 10
Limits and requirements
  • Ceph is excluded when used on cluster topologies with fewer than 3 nodes. For example, Ceph is excluded in a single-node OpenShift cluster or single-node OpenShift cluster with a single worker node.

  • In single-node OpenShift clusters, persistent storage must be provided by either LVMS or local storage, not both.

Engineering considerations
  • The LVMS Operator is not the reference storage solution for the DU use case. If you require LVMS Operator for application workloads, the resource use is accounted for against the application cores.

  • Ensure that sufficient disks or partitions are available for storage requirements.

Workload partitioning

New in this release
  • No reference design updates in this release

Description

Workload partitioning pins OpenShift platform and Day 2 Operator pods that are part of the DU profile to the reserved cpuset and removes the reserved CPU from node accounting. This leaves all unreserved CPU cores available for user workloads.

The method of enabling and configuring workload partitioning changed in OKD 4.14.

4.14 and later
  • Configure partitions by setting installation parameters:

    cpuPartitioningMode: AllNodes
  • Configure management partition cores with the reserved CPU set in the PerformanceProfile CR

4.13 and earlier
  • Configure partitions with extra MachineConfiguration CRs applied at install-time

Limits and requirements
  • Namespace and Pod CRs must be annotated to allow the pod to be applied to the management partition

  • Pods with CPU limits cannot be allocated to the partition. This is because mutation can change the pod QoS.

  • For more information about the minimum number of CPUs that can be allocated to the management partition, see Node Tuning Operator.

Engineering considerations
  • Workload Partitioning pins all management pods to reserved cores. A sufficient number of cores must be allocated to the reserved set to account for operating system, management pods, and expected spikes in CPU use that occur when the workload starts, the node reboots, or other system events happen.

Cluster tuning

New in this release
  • No reference design updates in this release

Description

The cluster capabilities feature includes a MachineAPI component which, when excluded, disables the following Operators and their resources in the cluster:

  • openshift/cluster-autoscaler-operator

  • openshift/cluster-control-plane-machine-set-operator

  • openshift/machine-api-operator

Use cluster capabilities to remove the Image Registry Operator.

Limits and requirements
  • Cluster capabilities are not available for installer-provisioned installation methods.

  • You must apply all platform tuning configurations. The following table lists the required platform tuning configurations:

    Table 1. Cluster capabilities configurations
    Feature Description

    Remove optional cluster capabilities

    Reduce the OKD footprint by disabling optional cluster Operators on single-node OpenShift clusters only.

    • Remove all optional Operators except the Marketplace and Node Tuning Operators.

    Configure cluster monitoring

    Configure the monitoring stack for reduced footprint by doing the following:

    • Disable the local alertmanager and telemeter components.

    • If you use RHACM observability, the CR must be augmented with appropriate additionalAlertManagerConfigs CRs to forward alerts to the hub cluster.

    • Reduce the Prometheus retention period to 24h.

      The RHACM hub cluster aggregates managed cluster metrics.

    Disable networking diagnostics

    Disable networking diagnostics for single-node OpenShift because they are not required.

    Configure a single OperatorHub catalog source

    Configure the cluster to use a single catalog source that contains only the Operators required for a RAN DU deployment. Each catalog source increases the CPU use on the cluster. Using a single CatalogSource fits within the platform CPU budget.

Machine configuration

New in this release
  • No reference design updates in this release

Limits and requirements
  • The CRI-O wipe disable MachineConfig assumes that images on disk are static other than during scheduled maintenance in defined maintenance windows. To ensure the images are static, do not set the pod imagePullPolicy field to Always.

    Table 2. Machine configuration options
    Feature Description

    Container runtime

    Sets the container runtime to crun for all node roles.

    kubelet config and container mount hiding

    Reduces the frequency of kubelet housekeeping and eviction monitoring to reduce CPU usage. Create a container mount namespace, visible to kubelet and CRI-O, to reduce system mount scanning resource usage.

    SCTP

    Optional configuration (enabled by default) Enables SCTP. SCTP is required by RAN applications but disabled by default in FCOS.

    kdump

    Optional configuration (enabled by default) Enables kdump to capture debug information when a kernel panic occurs.

    CRI-O wipe disable

    Disables automatic wiping of the CRI-O image cache after unclean shutdown.

    SR-IOV-related kernel arguments

    Includes additional SR-IOV related arguments in the kernel command line.

    RCU Normal systemd service

    Sets rcu_normal after the system is fully started.

    One-shot time sync

    Runs a one-time system time synchronization job for control plane or worker nodes.

Reference design deployment components

The following sections describe the various OKD components and configurations that you use to configure the hub cluster with Red Hat Advanced Cluster Management (RHACM).

Red Hat Advanced Cluster Management (RHACM)

New in this release
  • No reference design updates in this release

Description

RHACM provides Multi Cluster Engine (MCE) installation and ongoing lifecycle management functionality for deployed clusters. You declaratively specify configurations and upgrades with Policy CRs and apply the policies to clusters with the RHACM policy controller as managed by Topology Aware Lifecycle Manager.

  • GitOps Zero Touch Provisioning (ZTP) uses the MCE feature of RHACM

  • Configuration, upgrades, and cluster status are managed with the RHACM policy controller

During installation RHACM can apply labels to individual nodes as configured in the SiteConfig custom resource (CR).

Limits and requirements
  • A single hub cluster supports up to 3500 deployed single-node OpenShift clusters with 5 Policy CRs bound to each cluster.

Engineering considerations
  • Use RHACM policy hub-side templating to better scale cluster configuration. You can significantly reduce the number of policies by using a single group policy or small number of general group policies where the group and per-cluster values are substituted into templates.

  • Cluster specific configuration: managed clusters typically have some number of configuration values that are specific to the individual cluster. These configurations should be managed using RHACM policy hub-side templating with values pulled from ConfigMap CRs based on the cluster name.

  • To save CPU resources on managed clusters, policies that apply static configurations should be unbound from managed clusters after GitOps ZTP installation of the cluster. For more information, see Release a persistent volume.

Topology Aware Lifecycle Manager (TALM)

New in this release
  • No reference design updates in this release

Description
Managed updates

TALM is an Operator that runs only on the hub cluster for managing how changes (including cluster and Operator upgrades, configuration, and so on) are rolled out to the network. TALM does the following:

  • Progressively applies updates to fleets of clusters in user-configurable batches by using Policy CRs.

  • Adds ztp-done labels or other user configurable labels on a per-cluster basis

Precaching for single-node OpenShift clusters

TALM supports optional precaching of OKD, OLM Operator, and additional user images to single-node OpenShift clusters before initiating an upgrade.

  • A PreCachingConfig custom resource is available for specifying optional pre-caching configurations. For example:

    apiVersion: ran.openshift.io/v1alpha1
    kind: PreCachingConfig
    metadata:
      name: example-config
      namespace: example-ns
    spec:
      additionalImages:
        - quay.io/foobar/application1@sha256:3d5800990dee7cd4727d3fe238a97e2d2976d3808fc925ada29c559a47e2e
        - quay.io/foobar/application2@sha256:3d5800123dee7cd4727d3fe238a97e2d2976d3808fc925ada29c559a47adf
        - quay.io/foobar/applicationN@sha256:4fe1334adfafadsf987123adfffdaf1243340adfafdedga0991234afdadfs
      spaceRequired: 45 GiB (1)
      overrides:
        preCacheImage: quay.io/test_images/pre-cache:latest
        platformImage: quay.io/openshift-release-dev/ocp-release@sha256:3d5800990dee7cd4727d3fe238a97e2d2976d3808fc925ada29c559a47e2e
      operatorsIndexes:
        - registry.example.com:5000/custom-redhat-operators:1.0.0
      operatorsPackagesAndChannels:
        - local-storage-operator: stable
        - ptp-operator: stable
        - sriov-network-operator: stable
      excludePrecachePatterns: (2)
        - aws
        - vsphere
    1 Configurable space-required parameter allows you to validate before and after pre-caching storage space
    2 Configurable filtering allows exclusion of unused images
Backup and restore for single-node OpenShift

TALM supports taking a snapshot of the cluster operating system and configuration to a dedicated partition on a local disk. A restore script is provided that returns the cluster to the backed up state.

Limits and requirements
  • TALM supports concurrent cluster deployment in batches of 400

  • Precaching and backup features are for single-node OpenShift clusters only.

Engineering considerations
  • The PreCachingConfig CR is optional and does not need to be created if you just wants to precache platform related (OpenShift and OLM Operator) images. The PreCachingConfig CR must be applied before referencing it in the ClusterGroupUpgrade CR.

  • Create a recovery partition during installation if you opt to use the TALM backup and restore feature.

GitOps and GitOps ZTP plugins

New in this release
  • No reference design updates in this release

Description

GitOps and GitOps ZTP plugins provide a GitOps-based infrastructure for managing cluster deployment and configuration. Cluster definitions and configurations are maintained as a declarative state in Git. ZTP plugins provide support for generating installation CRs from the SiteConfig CR and automatic wrapping of configuration CRs in policies based on PolicyGenTemplate CRs.

You can deploy and manage multiple versions of OKD on managed clusters using the baseline reference configuration CRs. You can also use custom CRs alongside the baseline CRs.

Limits
  • 300 SiteConfig CRs per ArgoCD application. You can use multiple applications to achieve the maximum number of clusters supported by a single hub cluster.

  • Content in the /source-crs folder in Git overrides content provided in the GitOps ZTP plugin container. Git takes precedence in the search path.

  • Add the /source-crs folder in the same directory as the kustomization.yaml file, which includes the PolicyGenTemplate as a generator.

    Alternative locations for the /source-crs directory are not supported in this context.

Engineering considerations
  • To avoid confusion or unintentional overwriting of files when updating content, use unique and distinguishable names for user-provided CRs in the /source-crs folder and extra manifests in Git.

  • The SiteConfig CR allows multiple extra-manifest paths. When files with the same name are found in multiple directory paths, the last file found takes precedence. This allows you to put the full set of version-specific Day 0 manifests (extra-manifests) in Git and reference them from the SiteConfig CR. With this feature, you can deploy multiple OKD versions to managed clusters simultaneously.

  • The extraManifestPath field of the SiteConfig CR is deprecated from OKD 4.15 and later. Use the new extraManifests.searchPaths field instead.

Agent-based installer

New in this release
  • No reference design updates in this release

Description

Agent-based installer (ABI) provides installation capabilities without centralized infrastructure. The installation program creates an ISO image that you mount to the server. When the server boots it installs OKD and supplied extra manifests.

You can also use ABI to install OKD clusters without a hub cluster. An image registry is still required when you use ABI in this manner.

Agent-based installer (ABI) is an optional component.

Limits and requirements
  • You can supply a limited set of additional manifests at installation time.

  • You must include MachineConfiguration CRs that are required by the RAN DU use case.

Engineering considerations
  • ABI provides a baseline OKD installation.

  • You install Day 2 Operators and the remainder of the RAN DU use case configurations after installation.

Additional components

Bare Metal Event Relay

The Bare Metal Event Relay is an optional Operator that runs exclusively on the managed spoke cluster. It relays Redfish hardware events to cluster applications.

The Bare Metal Event Relay is not included in the RAN DU use model reference configuration and is an optional feature. If you want to use the Bare Metal Event Relay, assign additional CPU resources from the application CPU budget.