Validating cluster tuning for vDU application workloads - Clusters at the network far edge | Scalability and performance

Recommended firmware configuration for vDU cluster hosts
Recommended cluster configurations to run vDU applications
Checking that the recommended cluster configurations are applied

Before you can deploy virtual distributed unit (vDU) applications, you need to tune and configure the cluster host firmware and various other cluster configuration settings. Use the following information to validate the cluster configuration to support vDU workloads.

Additional resources

For more information about single-node OpenShift clusters tuned for vDU application deployments, see Reference configuration for deploying vDUs on single-node OpenShift.

Recommended firmware configuration for vDU cluster hosts

Use the following table as the basis to configure the cluster host firmware for vDU applications running on OKD 4.10.

The following table is a general recommendation for vDU cluster host firmware configuration. Exact firmware settings will depend on your requirements and specific hardware platform. Automatic setting of firmware is not handled by the zero touch provisioning pipeline.

Table 1. Recommended cluster host firmware settings
Firmware setting	Configuration	Description
HyperTransport (HT)	Enabled	HyperTransport (HT) bus is a bus technology developed by AMD. HT provides a high-speed link between the components in the host memory and other system peripherals.
UEFI	Enabled	Enable booting from UEFI for the vDU host.
CPU Power and Performance Policy	Performance	Set CPU Power and Performance Policy to optimize the system for performance over energy efficiency.
Uncore Frequency Scaling	Disabled	Disable Uncore Frequency Scaling to prevent the voltage and frequency of non-core parts of the CPU from being set independently.
Uncore Frequency	Maximum	Sets the non-core parts of the CPU such as cache and memory controller to their maximum possible frequency of operation.
Performance P-limit	Disabled	Disable Performance P-limit to prevent the Uncore frequency coordination of processors.
Enhanced Intel® SpeedStep Tech	Enabled	Enable Enhanced Intel SpeedStep to allow the system to dynamically adjust processor voltage and core frequency that decreases power consumption and heat production in the host.
Intel® Turbo Boost Technology	Enabled	Enable Turbo Boost Technology for Intel-based CPUs to automatically allow processor cores to run faster than the rated operating frequency if they are operating below power, current, and temperature specification limits.
Intel Configurable TDP	Enabled	Enables Thermal Design Power (TDP) for the CPU.
Configurable TDP Level	Level 2	TDP level sets the CPU power consumption required for a particular performance rating. TDP level 2 sets the CPU to the most stable performance level at the cost of power consumption.
Energy Efficient Turbo	Disabled	Disable Energy Efficient Turbo to prevent the processor from using an energy-efficiency based policy.
Hardware P-States	Disabled	Disable `P-states` (performance states) to optimize the operating system and CPU for performance over power consumption.
Package C-State	C0/C1 state	Use C0 or C1 states to set the processor to a fully active state (C0) or to stop CPU internal clocks running in software (C1).
C1E	Disabled	CPU Enhanced Halt (C1E) is a power saving feature in Intel chips. Disabling C1E prevents the operating system from sending a halt command to the CPU when inactive.
Processor C6	Disabled	C6 power-saving is a CPU feature that automatically disables idle CPU cores and cache. Disabling C6 improves system performance.
Sub-NUMA Clustering	Disabled	Sub-NUMA clustering divides the processor cores, cache, and memory into multiple NUMA domains. Disabling this option can increase performance for latency-sensitive workloads.

Enable global SR-IOV and VT-d settings in the firmware for the host. These settings are relevant to bare-metal environments.

Recommended cluster configurations to run vDU applications

Clusters running virtualized distributed unit (vDU) applications require a highly tuned and optimized configuration. The following information describes the various elements that you require to support vDU workloads in OKD 4.10 clusters.

Recommended cluster MachineConfig CRs

The following MachineConfig CRs configure the cluster host:

Table 2. Recommended MachineConfig CRs
CR filename	Description
`02-workload-partitioning.yaml`	Configures workload partitioning for the cluster. Apply this `MachineConfig` CR when you install the cluster.
`MachineConfigSctp.yaml`	Loads the SCTP kernel module. This `MachineConfig` CR is optional and can be omitted if you do not require this kernel module.
`MachineConfigContainerMountNS.yaml`	Configures the container mount namespace and kubelet conf.
`MachineConfigAcceleratedStartup.yaml`	Configures accelerated startup for the cluster.
`06-kdump-master.yaml`, `06-kdump-worker.yaml`	Configures `kdump` for the cluster.

Recommended cluster Operators

The following Operators are required for clusters running vDU applications and are a part of the baseline reference configuration:

Node Tuning Operator (NTO). NTO packages functionality that was previously delivered with the Performance Addon Operator, which is now a part of NTO.
PTP Operator
SR-IOV Network Operator
Red Hat OpenShift Logging Operator
Local Storage Operator

Recommended cluster kernel configuration

Always use the latest supported realtime kernel version in your cluster. You should also ensure that the following configurations are applied in the cluster:

Ensure the following additionalKernelArgs are set in the cluster performance profile:

spec:
  additionalKernelArgs:
  - "idle=poll"
  - "rcupdate.rcu_normal_after_boot=0"
  - "efi=runtime"

Ensure that the performance-patch profile in the Tuned CR configures the correct CPU isolation set that matches the isolated CPU set in the related PerformanceProfile CR, for example:

spec:
  profile:
    - name: performance-patch
      # The 'include' line must match the associated PerformanceProfile name
      # And the cmdline_crash CPU set must match the 'isolated' set in the associated PerformanceProfile
      data: |
        [main]
        summary=Configuration changes profile inherited from performance created tuned
        include=openshift-node-performance-openshift-node-performance-profile
        [bootloader]
        cmdline_crash=nohz_full=2-51,54-103 (1)
        [sysctl]
        kernel.timer_migration=1
        [scheduler]
        group.ice-ptp=0:f:10:*:ice-ptp.*
        [service]
        service.stalld=start,enable
        service.chronyd=stop,disable

1	Listed CPUs depend on the host hardware configuration, specifically the number of available CPUs in the system and the CPU topology.

Checking the realtime kernel version

Always use the latest version of the realtime kernel in your OKD clusters. If you are unsure about the kernel version that is in use in the cluster, you can compare the current realtime kernel version to the release version with the following procedure.

Prerequisites

You have installed the OpenShift CLI (oc).
You are logged in as a user with cluster-admin privileges.
You have installed podman.

Procedure

Run the following command to get the cluster version:

$ OCP_VERSION=$(oc get clusterversion version -o jsonpath='{.status.desired.version}{"\n"}')

Get the release image SHA number:

$ DTK_IMAGE=$(oc adm release info --image-for=driver-toolkit quay.io/openshift-release-dev/ocp-release:$OCP_VERSION-x86_64)

Run the release image container and extract the kernel version that is packaged with cluster’s current release:
```
$ podman run --rm $DTK_IMAGE rpm -qa | grep 'kernel-rt-core-' | sed 's#kernel-rt-core-##'
```
Example output
```
4.18.0-305.49.1.rt7.121.el8_4.x86_64
```
This is the default realtime kernel version that ships with the release.

The realtime kernel is denoted by the string .rt in the kernel version.

Verification

Check that the kernel version listed for the cluster’s current release matches actual realtime kernel that is running in the cluster. Run the following commands to check the running realtime kernel version:

Open a remote shell connection to the cluster node:
```
$ oc debug node/<node_name>
```

Check the realtime kernel version:

sh-4.4# uname -r

Example output

4.18.0-305.49.1.rt7.121.el8_4.x86_64

Checking that the recommended cluster configurations are applied

You can check that clusters are running the correct configuration. The following procedure describes how to check the various configurations that you require to deploy a DU application in OKD 4.10 clusters.

Prerequisites

You have deployed a cluster and tuned it for vDU workloads.
You have installed the OpenShift CLI (oc).
You have logged in as a user with cluster-admin privileges.

Procedure

Check that the default Operator Hub sources are disabled. Run the following command:
```
$ oc get operatorhub cluster -o yaml
```
Example output
```
spec:
    disableAllDefaultSources: true
```

Check that all required CatalogSource resources are annotated for workload partitioning (PreferredDuringScheduling) by running the following command:

$ oc get catalogsource -A -o jsonpath='{range .items[*]}{.metadata.name}{" -- "}{.metadata.annotations.target\.workload\.openshift\.io/management}{"\n"}{end}'

Example output

certified-operators -- {"effect": "PreferredDuringScheduling"}
community-operators -- {"effect": "PreferredDuringScheduling"}
ran-operators (1)
redhat-marketplace -- {"effect": "PreferredDuringScheduling"}
redhat-operators -- {"effect": "PreferredDuringScheduling"}

1	`CatalogSource` resources that are not annotated are also returned. In this example, the `ran-operators` `CatalogSource` resource is not annotated and does not have the `PreferredDuringScheduling` annotation.

In a properly configured vDU cluster, only a single annotated catalog source is listed.

Check that all applicable OKD Operator namespaces are annotated for workload partitioning. This includes all Operators installed with core OKD and the set of additional Operators included in the reference DU tuning configuration. Run the following command:

$ oc get namespaces -A -o jsonpath='{range .items[*]}{.metadata.name}{" -- "}{.metadata.annotations.workload\.openshift\.io/allowed}{"\n"}{end}'

Example output

default --
openshift-apiserver -- management
openshift-apiserver-operator -- management
openshift-authentication -- management
openshift-authentication-operator -- management

Additional Operators must not be annotated for workload partitioning. In the output from the previous command, additional Operators should be listed without any value on the right-hand side of the -- separator.

Check that the ClusterLogging configuration is correct. Run the following commands:

Validate that the appropriate input and output logs are configured:

$ oc get -n openshift-logging ClusterLogForwarder instance -o yaml

Example output

apiVersion: logging.openshift.io/v1
kind: ClusterLogForwarder
metadata:
  creationTimestamp: "2022-07-19T21:51:41Z"
  generation: 1
  name: instance
  namespace: openshift-logging
  resourceVersion: "1030342"
  uid: 8c1a842d-80c5-447a-9150-40350bdf40f0
spec:
  inputs:
  - infrastructure: {}
    name: infra-logs
  outputs:
  - name: kafka-open
    type: kafka
    url: tcp://10.46.55.190:9092/test
  pipelines:
  - inputRefs:
    - audit
    name: audit-logs
    outputRefs:
    - kafka-open
  - inputRefs:
    - infrastructure
    name: infrastructure-logs
    outputRefs:
    - kafka-open
...

Check that the curation schedule is appropriate for your application:

$ oc get -n openshift-logging clusterloggings.logging.openshift.io instance -o yaml

Example output

apiVersion: logging.openshift.io/v1
kind: ClusterLogging
metadata:
  creationTimestamp: "2022-07-07T18:22:56Z"
  generation: 1
  name: instance
  namespace: openshift-logging
  resourceVersion: "235796"
  uid: ef67b9b8-0e65-4a10-88ff-ec06922ea796
spec:
  collection:
    logs:
      fluentd: {}
      type: fluentd
  curation:
    curator:
      schedule: 30 3 * * *
    type: curator
  managementState: Managed
...

Check that the web console is disabled (managementState: Removed) by running the following command:
```
$ oc get consoles.operator.openshift.io cluster -o jsonpath="{ .spec.managementState }"
```
Example output
```
Removed
```

Check that chronyd is disabled on the cluster node by running the following commands:

$ oc debug node/<node_name>

Check the status of chronyd on the node:

sh-4.4# chroot /host

sh-4.4# systemctl status chronyd

Example output

● chronyd.service - NTP client/server
    Loaded: loaded (/usr/lib/systemd/system/chronyd.service; disabled; vendor preset: enabled)
    Active: inactive (dead)
      Docs: man:chronyd(8)
            man:chrony.conf(5)

Check that the PTP interface is successfully synchronized to the primary clock using a remote shell connection to the linuxptp-daemon container and the PTP Management Client (pmc) tool:

Set the $PTP_POD_NAME variable with the name of the linuxptp-daemon pod by running the following command:
```
$ PTP_POD_NAME=$(oc get pods -n openshift-ptp -l app=linuxptp-daemon -o name)
```

Run the following command to check the sync status of the PTP device:

$ oc -n openshift-ptp rsh -c linuxptp-daemon-container ${PTP_POD_NAME} pmc -u -f /var/run/ptp4l.0.config -b 0 'GET PORT_DATA_SET'

Example output

sending: GET PORT_DATA_SET
  3cecef.fffe.7a7020-1 seq 0 RESPONSE MANAGEMENT PORT_DATA_SET
    portIdentity            3cecef.fffe.7a7020-1
    portState               SLAVE
    logMinDelayReqInterval  -4
    peerMeanPathDelay       0
    logAnnounceInterval     1
    announceReceiptTimeout  3
    logSyncInterval         0
    delayMechanism          1
    logMinPdelayReqInterval 0
    versionNumber           2
  3cecef.fffe.7a7020-2 seq 0 RESPONSE MANAGEMENT PORT_DATA_SET
    portIdentity            3cecef.fffe.7a7020-2
    portState               LISTENING
    logMinDelayReqInterval  0
    peerMeanPathDelay       0
    logAnnounceInterval     1
    announceReceiptTimeout  3
    logSyncInterval         0
    delayMechanism          1
    logMinPdelayReqInterval 0
    versionNumber           2

Run the following pmc command to check the PTP clock status:

$ oc -n openshift-ptp rsh -c linuxptp-daemon-container ${PTP_POD_NAME} pmc -u -f /var/run/ptp4l.0.config -b 0 'GET TIME_STATUS_NP'

Example output

sending: GET TIME_STATUS_NP
  3cecef.fffe.7a7020-0 seq 0 RESPONSE MANAGEMENT TIME_STATUS_NP
    master_offset              10 (1)
    ingress_time               1657275432697400530
    cumulativeScaledRateOffset +0.000000000
    scaledLastGmPhaseChange    0
    gmTimeBaseIndicator        0
    lastGmPhaseChange          0x0000'0000000000000000.0000
    gmPresent                  true (2)
    gmIdentity                 3c2c30.ffff.670e00

1	`master_offset` should be between -100 and 100 ns.
2	Indicates that the PTP clock is synchronized to a master, and the local clock is not the grandmaster clock.

Check that the expected master offset value corresponding to the value in /var/run/ptp4l.0.config is found in the linuxptp-daemon-container log:

$ oc logs $PTP_POD_NAME -n openshift-ptp -c linuxptp-daemon-container

Example output

phc2sys[56020.341]: [ptp4l.1.config] CLOCK_REALTIME phc offset  -1731092 s2 freq -1546242 delay    497
ptp4l[56020.390]: [ptp4l.1.config] master offset         -2 s2 freq   -5863 path delay       541
ptp4l[56020.390]: [ptp4l.0.config] master offset         -8 s2 freq  -10699 path delay       533

Check that the SR-IOV configuration is correct by running the following commands:

Check that the disableDrain value in the SriovOperatorConfig resource is set to true:

$ oc get sriovoperatorconfig -n openshift-sriov-network-operator default -o jsonpath="{.spec.disableDrain}{'\n'}"

Example output

true

Check that the SriovNetworkNodeState sync status is Succeeded by running the following command:

$ oc get SriovNetworkNodeStates -n openshift-sriov-network-operator -o jsonpath="{.items[*].status.syncStatus}{'\n'}"

Example output

Succeeded

Verify that the expected number and configuration of virtual functions (Vfs) under each interface configured for SR-IOV is present and correct in the .status.interfaces field. For example:

$ oc get SriovNetworkNodeStates -n openshift-sriov-network-operator -o yaml

Example output

apiVersion: v1
items:
- apiVersion: sriovnetwork.openshift.io/v1
  kind: SriovNetworkNodeState
...
  status:
    interfaces:
    ...
    - Vfs:
      - deviceID: 154c
        driver: vfio-pci
        pciAddress: 0000:3b:0a.0
        vendor: "8086"
        vfID: 0
      - deviceID: 154c
        driver: vfio-pci
        pciAddress: 0000:3b:0a.1
        vendor: "8086"
        vfID: 1
      - deviceID: 154c
        driver: vfio-pci
        pciAddress: 0000:3b:0a.2
        vendor: "8086"
        vfID: 2
      - deviceID: 154c
        driver: vfio-pci
        pciAddress: 0000:3b:0a.3
        vendor: "8086"
        vfID: 3
      - deviceID: 154c
        driver: vfio-pci
        pciAddress: 0000:3b:0a.4
        vendor: "8086"
        vfID: 4
      - deviceID: 154c
        driver: vfio-pci
        pciAddress: 0000:3b:0a.5
        vendor: "8086"
        vfID: 5
      - deviceID: 154c
        driver: vfio-pci
        pciAddress: 0000:3b:0a.6
        vendor: "8086"
        vfID: 6
      - deviceID: 154c
        driver: vfio-pci
        pciAddress: 0000:3b:0a.7
        vendor: "8086"
        vfID: 7

Check that the cluster performance profile is correct. The cpu and hugepages sections will vary depending on your hardware configuration. Run the following command:

$ oc get PerformanceProfile openshift-node-performance-profile -o yaml

Example output

apiVersion: performance.openshift.io/v2
kind: PerformanceProfile
metadata:
  creationTimestamp: "2022-07-19T21:51:31Z"
  finalizers:
  - foreground-deletion
  generation: 1
  name: openshift-node-performance-profile
  resourceVersion: "33558"
  uid: 217958c0-9122-4c62-9d4d-fdc27c31118c
spec:
  additionalKernelArgs:
  - idle=poll
  - rcupdate.rcu_normal_after_boot=0
  - efi=runtime
  cpu:
    isolated: 2-51,54-103
    reserved: 0-1,52-53
  hugepages:
    defaultHugepagesSize: 1G
    pages:
    - count: 32
      size: 1G
  machineConfigPoolSelector:
    pools.operator.machineconfiguration.openshift.io/master: ""
  net:
    userLevelNetworking: true
  nodeSelector:
    node-role.kubernetes.io/master: ""
  numa:
    topologyPolicy: restricted
  realTimeKernel:
    enabled: true
status:
  conditions:
  - lastHeartbeatTime: "2022-07-19T21:51:31Z"
    lastTransitionTime: "2022-07-19T21:51:31Z"
    status: "True"
    type: Available
  - lastHeartbeatTime: "2022-07-19T21:51:31Z"
    lastTransitionTime: "2022-07-19T21:51:31Z"
    status: "True"
    type: Upgradeable
  - lastHeartbeatTime: "2022-07-19T21:51:31Z"
    lastTransitionTime: "2022-07-19T21:51:31Z"
    status: "False"
    type: Progressing
  - lastHeartbeatTime: "2022-07-19T21:51:31Z"
    lastTransitionTime: "2022-07-19T21:51:31Z"
    status: "False"
    type: Degraded
  runtimeClass: performance-openshift-node-performance-profile
  tuned: openshift-cluster-node-tuning-operator/openshift-node-performance-openshift-node-performance-profile

CPU settings are dependent on the number of cores available on the server and should align with workload partitioning settings. hugepages configuration is server and application dependent.

Check that the PerformanceProfile was successfully applied to the cluster by running the following command:

$ oc get performanceprofile openshift-node-performance-profile -o jsonpath="{range .status.conditions[*]}{ @.type }{' -- '}{@.status}{'\n'}{end}"

Example output

Available -- True
Upgradeable -- True
Progressing -- False
Degraded -- False

Check the Tuned performance patch settings by running the following command:

$ oc get tuneds.tuned.openshift.io -n openshift-cluster-node-tuning-operator performance-patch -o yaml

Example output

apiVersion: tuned.openshift.io/v1
kind: Tuned
metadata:
  creationTimestamp: "2022-07-18T10:33:52Z"
  generation: 1
  name: performance-patch
  namespace: openshift-cluster-node-tuning-operator
  resourceVersion: "34024"
  uid: f9799811-f744-4179-bf00-32d4436c08fd
spec:
  profile:
  - data: |
      [main]
      summary=Configuration changes profile inherited from performance created tuned
      include=openshift-node-performance-openshift-node-performance-profile
      [bootloader]
      cmdline_crash=nohz_full=2-23,26-47 (1)
      [sysctl]
      kernel.timer_migration=1
      [scheduler]
      group.ice-ptp=0:f:10:*:ice-ptp.*
      [service]
      service.stalld=start,enable
      service.chronyd=stop,disable
    name: performance-patch
  recommend:
  - machineConfigLabels:
      machineconfiguration.openshift.io/role: master
    priority: 19
    profile: performance-patch

1	The cpu list in `cmdline=nohz_full=` will vary based on your hardware configuration.

Check that cluster networking diagnostics are disabled by running the following command:

$ oc get networks.operator.openshift.io cluster -o jsonpath='{.spec.disableNetworkDiagnostics}'

Example output

true

Check that the Kubelet housekeeping interval is tuned to slower rate. This is set in the containerMountNS machine config. Run the following command:

$ oc describe machineconfig container-mount-namespace-and-kubelet-conf-master | grep OPENSHIFT_MAX_HOUSEKEEPING_INTERVAL_DURATION

Example output

Environment="OPENSHIFT_MAX_HOUSEKEEPING_INTERVAL_DURATION=60s"

Check that Grafana and alertManagerMain are disabled and that the Prometheus retention period is set to 24h by running the following command:
```
$ oc get configmap cluster-monitoring-config -n openshift-monitoring -o jsonpath="{ .data.config\.yaml }"
```
Example output
```
grafana:
  enabled: false
alertmanagerMain:
  enabled: false
prometheusK8s:
   retention: 24h
```
1. Use the following commands to verify that Grafana and alertManagerMain routes are not found in the cluster:
  $ oc get route -n openshift-monitoring alertmanager-main
  $ oc get route -n openshift-monitoring grafana
  Both queries should return Error from server (NotFound) messages.
Check that there is a minimum of 4 CPUs allocated as reserved for each of the PerformanceProfile, Tuned performance-patch, workload partitioning, and kernel command line arguments by running the following command:
```
$ oc get performanceprofile -o jsonpath="{ .items[0].spec.cpu.reserved }"
```
Example output
```
0-1,52-53
```
Depending on your workload requirements, you might require additional reserved CPUs to be allocated.

Validating single-node OpenShift cluster tuning for vDU application workloads

Recommended firmware configuration for vDU cluster hosts

Recommended cluster configurations to run vDU applications

Recommended cluster MachineConfig CRs

Recommended cluster Operators

Recommended cluster kernel configuration

Checking the realtime kernel version

Checking that the recommended cluster configurations are applied