Chapter 7. Scalability and performance optimization

Understand your persistent storage options so that you can optimize your OpenShift Container Platform environment.

The following table summarizes the recommended and configurable storage technologies for the given OpenShift Container Platform cluster application.

The following table summarizes the main directories that OpenShift Container Platform components write data to.

OpenShift Container Platform and Kubernetes are sensitive to disk performance, and faster storage is recommended, particularly for etcd on the control plane nodes.

For production Azure clusters and clusters with intensive workloads, the virtual machine operating system disk for control plane machines should be able to sustain a tested and recommended minimum throughput of 5000 IOPS / 200MBps. This throughput can be provided by having a minimum of 1 TiB Premium SSD (P30). In Azure and Azure Stack Hub, disk performance is directly dependent on SSD disk sizes. To achieve the throughput supported by a Standard_D8s_v3 virtual machine, or other similar machine types, and the target of 5000 IOPS, at least a P30 disk is required.

Host caching must be set to ReadOnly for low latency and high IOPS and throughput when reading data. Reading data from the cache, which is present either in the VM memory or in the local SSD disk, is much faster than reading from the disk, which is in the blob storage.

The OpenShift Container Platform Ingress Controller, or router, is the ingress point for ingress traffic for applications and services that are configured using routes and ingresses.

When evaluating a single HAProxy router performance in terms of HTTP requests handled per second, the performance varies depending on many factors. In particular:

While performance in your specific environment will vary, Red Hat lab tests on a public cloud instance of size 4 vCPU/16GB RAM. A single HAProxy router handling 100 routes terminated by backends serving 1kB static pages is able to handle the following number of transactions per second.

In HTTP keep-alive mode scenarios:

In HTTP close (no keep-alive) scenarios:

The default Ingress Controller configuration was used with the spec.tuningOptions.threadCount field set to 4. Two different endpoint publishing strategies were tested: Load Balancer Service and Host Network. TLS session resumption was used for encrypted routes. With HTTP keep-alive, a single HAProxy router is capable of saturating a 1 Gbit NIC at page sizes as small as 8 kB.

When running on bare metal with modern processors, you can expect roughly twice the performance of the public cloud instance above. This overhead is introduced by the virtualization layer in place on public clouds and holds mostly true for private cloud-based virtualization as well. The following table is a guide to how many applications to use behind the router:

In general, HAProxy can support routes for up to 1000 applications, depending on the technology in use. Ingress Controller performance might be limited by the capabilities and performance of the applications behind it, such as language or static versus dynamic content.

Ingress, or router, sharding should be used to serve more routes towards applications and help horizontally scale the routing tier.

For more information on Ingress sharding, see Configuring Ingress Controller sharding by using route labels and Configuring Ingress Controller sharding by using namespace labels.

You can modify the Ingress Controller deployment using the information provided in Setting Ingress Controller thread count for threads and Ingress Controller configuration parameters for timeouts, and other tuning configurations in the Ingress Controller specification.

Cluster administrators can configure the timeout values for the kubelet’s liveness, readiness, and startup probes for router deployments that are managed by the OpenShift Container Platform Ingress Controller (router). The liveness and readiness probes of the router use the default timeout value of 1 second, which is too brief when networking or runtime performance is severely degraded. Probe timeouts can cause unwanted router restarts that interrupt application connections. The ability to set larger timeout values can reduce the risk of unnecessary and unwanted restarts.

You can update the timeoutSeconds value on the livenessProbe, readinessProbe, and startupProbe parameters of the router container.

The following example demonstrates how you can directly patch the default router deployment to set a 5-second timeout for the liveness and readiness probes:

$ oc -n openshift-ingress patch deploy/router-default --type=strategic --patch='{"spec":{"template":{"spec":{"containers":[{"name":"router","livenessProbe":{"timeoutSeconds":5},"readinessProbe":{"timeoutSeconds":5}}]}}}}'

$ oc -n openshift-ingress describe deploy/router-default | grep -e Liveness: -e Readiness:
    Liveness:   http-get http://:1936/healthz delay=0s timeout=5s period=10s #success=1 #failure=3
    Readiness:  http-get http://:1936/healthz/ready delay=0s timeout=5s period=10s #success=1 #failure=3

When you update a route or an endpoint associated with a route, OpenShift Container Platform router updates the configuration for HAProxy. Then, HAProxy reloads the updated configuration for those changes to take effect. When HAProxy reloads, it generates a new process that handles new connections using the updated configuration.

HAProxy keeps the old process running to handle existing connections until those connections are all closed. When old processes have long-lived connections, these processes can accumulate and consume resources.

The default minimum HAProxy reload interval is five seconds. You can configure an Ingress Controller using its spec.tuningOptions.reloadInterval field to set a longer minimum reload interval.

There are two important maximum transmission units (MTUs): the network interface controller (NIC) MTU and the cluster network MTU.

The NIC MTU is only configured at the time of OpenShift Container Platform installation. The MTU must be less than or equal to the maximum supported value of the NIC of your network. If you are optimizing for throughput, choose the largest possible value. If you are optimizing for lowest latency, choose a lower value.

The OpenShift SDN network plugin overlay MTU must be less than the NIC MTU by 50 bytes at a minimum. This accounts for the SDN overlay header. So, on a normal ethernet network, this should be set to 1450. On a jumbo frame ethernet network, this should be set to 8950. These values should be set automatically by the Cluster Network Operator based on the NIC’s configured MTU. Therefore, cluster administrators do not typically update these values. Amazon Web Services (AWS) and bare-metal environments support jumbo frame ethernet networks. This setting will help throughput, especially with transmission control protocol (TCP).

For OVN and Geneve, the MTU must be less than the NIC MTU by 100 bytes at a minimum.

When installing large clusters or scaling the cluster to larger node counts, set the cluster network cidr accordingly in your install-config.yaml file before you install the cluster:

networking:
  clusterNetwork:
  - cidr: 10.128.0.0/14
    hostPrefix: 23
  machineNetwork:
  - cidr: 10.0.0.0/16
  networkType: OVNKubernetes
  serviceNetwork:
  - 172.30.0.0/16

The default cluster network cidr 10.128.0.0/14 cannot be used if the cluster size is more than 500 nodes. It must be set to 10.128.0.0/12 or 10.128.0.0/10 to get to larger node counts beyond 500 nodes.

Because encrypting and decrypting node hosts uses CPU power, performance is affected both in throughput and CPU usage on the nodes when encryption is enabled, regardless of the IP security system being used.

IPSec encrypts traffic at the IP payload level, before it hits the NIC, protecting fields that would otherwise be used for NIC offloading. This means that some NIC acceleration features might not be usable when IPSec is enabled and will lead to decreased throughput and increased CPU usage.

Mount namespaces are used to isolate mount points so that processes in different namespaces cannot view each others' files. Encapsulation is the process of moving Kubernetes mount namespaces to an alternative location where they will not be constantly scanned by the host operating system.

The host operating system uses systemd to constantly scan all mount namespaces: both the standard Linux mounts and the numerous mounts that Kubernetes uses to operate. The current implementation of kubelet and CRI-O both use the top-level namespace for all container runtime and kubelet mount points. However, encapsulating these container-specific mount points in a private namespace reduces systemd overhead with no difference in functionality. Using a separate mount namespace for both CRI-O and kubelet can encapsulate container-specific mounts from any systemd or other host operating system interaction.

This ability to potentially achieve major CPU optimization is now available to all OpenShift Container Platform administrators. Encapsulation can also improve security by storing Kubernetes-specific mount points in a location safe from inspection by unprivileged users.

The following diagrams illustrate a Kubernetes installation before and after encapsulation. Both scenarios show example containers which have mount propagation settings of bidirectional, host-to-container, and none.

Here we see systemd, host operating system processes, kubelet, and the container runtime sharing a single mount namespace.

The following diagram illustrates the system state after encapsulation.

You can configure mount namespace encapsulation so that a cluster runs with less resource overhead.

These commands return the mount namespaces associated with systemd, kubelet, and the container runtime. In OpenShift Container Platform, the container runtime is CRI-O.

Encapsulation is in effect if systemd is in a different mount namespace to kubelet and CRI-O as in the above example. Encapsulation is not in effect if all three processes are in the same mount namespace.

You can inspect Kubernetes-specific mount points in the cluster host operating system for debugging or auditing purposes by using the kubensenter script that is available in Red Hat Enterprise Linux CoreOS (RHCOS).

SSH shell sessions to the cluster host are in the default namespace. To inspect Kubernetes-specific mount points in an SSH shell prompt, you need to run the kubensenter script as root. The kubensenter script is aware of the state of the mount encapsulation, and is safe to run even if encapsulation is not enabled.

If the encapsulation feature is not enabled, the kubensenter findmnt and findmnt commands return the same output, regardless of whether they are run in an oc debug session or in an SSH shell prompt.

Any monitoring tool that relies on the ability to run in the host operating system and have visibility of mount points created by kubelet, CRI-O, or containers themselves, must enter the container mount namespace to see these mount points. The kubensenter script that is provided with OpenShift Container Platform executes another command inside the Kubernetes mount point and can be used to adapt any existing tools.

The kubensenter script is aware of the state of the mount encapsulation feature status, and is safe to run even if encapsulation is not enabled. In that case the script executes the provided command in the default mount namespace.

For example, if a systemd service needs to run inside the new Kubernetes mount namespace, edit the service file and use the ExecStart= command line with kubensenter.

[Unit]
Description=Example service
[Service]
ExecStart=/usr/bin/kubensenter /path/to/original/command arg1 arg2