Enable the eBPF dataplane

Big picture

Enable the eBPF dataplane on an existing cluster.

Value

The eBPF dataplane mode has several advantages over standard Linux networking pipeline mode:

  • It scales to higher throughput.

  • It uses less CPU per GBit.

  • It has native support for Kubernetes services (without needing kube-proxy) that:

    • Reduces first packet latency for packets to services.
    • Preserves external client source IP addresses all the way to the pod.
    • Supports DSR (Direct Server Return) for more efficient service routing.
    • Uses less CPU than kube-proxy to keep the dataplane in sync.

To learn more and see performance metrics from our test environment, see the blog, Introducing the Calico eBPF dataplane.

Concepts

eBPF

eBPF (or “extended Berkeley Packet Filter”), is a technology that allows safe mini programs to be attached to various low-level hooks in the Linux kernel. eBPF has a wide variety of uses, including networking, security, and tracing. You’ll see a lot of non-networking projects leveraging eBPF, but for Calico our focus is on networking, and in particular, pushing the networking capabilities of the latest Linux kernels to the limit.

Before you begin

Supported

  • x86-64

  • arm64 (little endian)

  • Kubernetes datastore driver.

  • Distributions:

    • Generic or kubeadm
    • kOps
    • OpenShift
    • EKS
    • AKS with limitations:
    • RKE (RKE2 recommended because it supports disabling kube-proxy)

  • Linux distribution/kernel:

    • Ubuntu 20.04.
    • Red Hat v8.2 with Linux kernel v4.18.0-193 or above (Red Hat have backported the required features to that build).
    • Another supported distribution with Linux kernel v5.3 or above. Kernel v5.8 or above with CO-RE enabled is recommended for better performance.

  • An underlying network fabric that allows VXLAN traffic between hosts. In eBPF mode, VXLAN is used to forward Kubernetes NodePort traffic.

  • IPv6

Limitations:

  • IPIP is not supported (Calico iptables does not support it either). VXLAN is the recommended overlay for eBPF mode.

To enable IPv6 in eBPF mode, see Configure dual stack or IPv6 only. You may be able to run with non-Calico IPAM. eks-cni is known to work.

Not supported

  • Other processor architectures.

  • etcd datastore driver. The etcd datastore driver doesn’t support watching Kubernetes services, which is required for some features in eBPF mode.

  • Distributions:

    • GKE. This is because of an incompatibility with the GKE CNI plugin.

  • Clusters with some eBPF nodes and some standard dataplane and/or Windows nodes.

  • Floating IPs.

  • SCTP (either for policy or services). This is due to lack of kernel support for the SCTP checksum in BPF.

  • Log action in policy rules. This is because the Log action maps to the iptables LOG action and BPF programs cannot access that log.

  • VLAN-based traffic.

Performance

For best pod-to-pod performance, we recommend using an underlying network that doesn’t require Calico to use an overlay. For example:

  • A cluster within a single AWS subnet.
  • A cluster using a compatible cloud provider’s CNI (such as the AWS VPC CNI plugin).
  • An on-prem cluster with BGP peering configured.

If you must use an overlay, we recommend that you use VXLAN, not IPIP. VXLAN has much better performance than IPIP in eBPF mode due to various kernel optimisations.

How to

Verify that your cluster is ready for eBPF mode

This section explains how to make sure your cluster is suitable for eBPF mode.

To check that the kernel on a node is suitable, you can run

  1. uname -rv

The output should look like this:

  1. 5.4.0-42-generic #46-Ubuntu SMP Fri Jul 10 00:24:02 UTC 2020

In this case the kernel version is v5.4, which is suitable.

On Red Hat-derived distributions, you may see something like this:

  1. 4.18.0-193.el8.x86_64 (mockbuild@x86-vm-08.build.eng.bos.redhat.com)

Since the Red Hat kernel is v4.18 with at least build number 193, this kernel is suitable.

Configure Calico to talk directly to the API server

In eBPF mode, Calico implements Kubernetes service networking directly (rather than relying on kube-proxy). Of course, this makes it highly desirable to disable kube-proxy when running in eBPF mode to save resources and avoid confusion over which component is handling services.

To be able to disable kube-proxy, Calico needs to communicate to the API server directly rather than going through kube-proxy. To make that possible, we need to find a persistent, static way to reach the API server. The best way to do that varies by Kubernetes distribution:

  • If you created a cluster manually (for example by using kubeadm) then the right address to use depends on whether you opted for a high-availability cluster with multiple API servers or a simple one-node API server.

    • If you opted to set up a high availability cluster then you should use the address of the load balancer that you used in front of your API servers. As noted in the Kubernetes documentation, a load balancer is required for a HA set-up but the precise type of load balancer is not specified.
    • If you opted for a single control plane node then you can use the address of the control plane node itself. However, it’s important that you use a stable address for that node such as a dedicated DNS record, or a static IP address. If you use a dynamic IP address (such as an EC2 private IP) then the address may change when the node is restarted causing Calico to lose connectivity to the API server.

  • kops typically sets up a load balancer of some sort in front of the API server. You should use the FQDN and port of the API load balancer, for example api.internal.<clustername> as the KUBERNETES_SERVICE_HOST below and 443 as the KUBERNETES_SERVICE_PORT.

  • OpenShift requires various DNS records to be created for the cluster; one of these is exactly what we need: api-int.<cluster_name>.<base_domain> should point to the API server or to the load balancer in front of the API server. Use that (filling in the <cluster_name> and <base_domain> as appropriate for your cluster) for the KUBERNETES_SERVICE_HOST below. Openshift uses 6443 for the KUBERNETES_SERVICE_PORT.

  • For AKS and EKS clusters you should use the FQDN of the API server’s load balancer. This can be found with

    1. kubectl cluster-info

    which gives output like the following:

    1. Kubernetes master is running at https://60F939227672BC3D5A1B3EC9744B2B21.gr7.us-west-2.eks.amazonaws.com
    2. ...

    In this example, you would use 60F939227672BC3D5A1B3EC9744B2B21.gr7.us-west-2.eks.amazonaws.com for KUBERNETES_SERVICE_HOST and 443 for KUBERNETES_SERVICE_PORT when creating the config map.

The next step depends on whether you installed Calico using the operator, or a manifest:

  • Operator
  • Manifest

If you installed Calico using the operator, create the following config map in the tigera-operator namespace using the host and port determined above:

  1. kind: ConfigMap
  2. apiVersion: v1
  3. metadata:
  4.   name: kubernetes-services-endpoint
  5.   namespace: tigera-operator
  6. data:
  7.   KUBERNETES_SERVICE_HOST: '<API server host>'
  8.   KUBERNETES_SERVICE_PORT: '<API server port>'

The operator will pick up the change to the config map automatically and do a rolling update of Calico to pass on the change. Confirm that pods restart and then reach the Running state with the following command:

  1. watch kubectl get pods -n calico-system

If you do not see the pods restart then it’s possible that the ConfigMap wasn’t picked up (sometimes Kubernetes is slow to propagate ConfigMaps (see Kubernetes issue #30189)). You can try restarting the operator.

If you installed Calico using a manifest, create the following config map in the kube-system namespace using the host and port determined above:

  1. kind: ConfigMap
  2. apiVersion: v1
  3. metadata:
  4.   name: kubernetes-services-endpoint
  5.   namespace: kube-system
  6. data:
  7.   KUBERNETES_SERVICE_HOST: '<API server host>'
  8.   KUBERNETES_SERVICE_PORT: '<API server port>'

Wait 60s for kubelet to pick up the ConfigMap (see Kubernetes issue #30189); then, restart the Calico pods to pick up the change:

  1. kubectl delete pod -n kube-system -l k8s-app=calico-node
  2. kubectl delete pod -n kube-system -l k8s-app=calico-kube-controllers

And, if using Typha:

  1. kubectl delete pod -n kube-system -l k8s-app=calico-typha

Confirm that pods restart and then reach the Running state with the following command:

  1. watch "kubectl get pods -n kube-system | grep calico"

You can verify that the change was picked up by checking the logs of one of the calico/node pods.

  1. kubectl get po -n kube-system -l k8s-app=calico-node

Should show one or more pods:

  1. NAME                                       READY   STATUS    RESTARTS   AGE
  2. calico-node-d6znw                          1/1     Running   0          48m
  3. ...

Then, to search the logs, choose a pod and run:

  1. kubectl logs -n kube-system <pod name> | grep KUBERNETES_SERVICE_HOST

You should see the following log, with the correct KUBERNETES_SERVICE_... values.

  1. 2020-08-26 12:26:29.025 [INFO][7] daemon.go 182: Kubernetes server override env vars. KUBERNETES_SERVICE_HOST="172.16.101.157" KUBERNETES_SERVICE_PORT="6443"

Configure kube-proxy

In eBPF mode Calico replaces kube-proxy so it wastes resources (and reduces performance) to run both.
This section explains how to disable kube-proxy in some common environments.

Clusters that run kube-proxy with a DaemonSet (such as kubeadm)

For a cluster that runs kube-proxy in a DaemonSet (such as a kubeadm-created cluster), you can disable kube-proxy reversibly by adding a node selector to kube-proxy‘s DaemonSet that matches no nodes, for example:

  1. kubectl patch ds -n kube-system kube-proxy -p '{"spec":{"template":{"spec":{"nodeSelector":{"non-calico": "true"}}}}}'

Then, should you want to start kube-proxy again, you can simply remove the node selector.

Enable the eBPF dataplane - 图1note

This approach is not suitable for AKS with Azure CNI since that platform makes use of the Kubernetes add-on manager. the change will be reverted by the system. For AKS, you should follow Avoiding conflicts with kube-proxy below.

OpenShift

If you are running OpenShift, you can disable kube-proxy as follows:

  1. kubectl patch networks.operator.openshift.io cluster --type merge -p '{"spec":{"deployKubeProxy": false}}'

To re-enable it:

  1. kubectl patch networks.operator.openshift.io cluster --type merge -p '{"spec":{"deployKubeProxy": true}}'

Avoiding conflicts with kube-proxy

If you cannot disable kube-proxy (for example, because it is managed by your Kubernetes distribution), then you must change Felix configuration parameter BPFKubeProxyIptablesCleanupEnabled to false. This can be done with kubectl as follows:

  1. kubectl patch felixconfiguration default --patch='{"spec": {"bpfKubeProxyIptablesCleanupEnabled": false}}'

If both kube-proxy and BPFKubeProxyIptablesCleanupEnabled is enabled then kube-proxy will write its iptables rules and Felix will try to clean them up resulting in iptables flapping between the two.

Enable eBPF mode

The next step depends on whether you installed Calico using the operator, or a manifest:

  • Operator
  • Manifest

To enable eBPF mode, change the spec.calicoNetwork.linuxDataplane parameter in the operator’s Installation resource to `“BPF”.

  1.   kubectl patch installation.operator.tigera.io default --type merge -p '{"spec":{"calicoNetwork":{"linuxDataplane":"BPF"}}}'

Enable the eBPF dataplane - 图2note

The operator rolls out the change with a rolling update (non-disruptive) and then swiftly transitions all nodes to eBPF mode. However, it’s inevitable that some nodes will enter eBPF mode before others. This can disrupt the flow of traffic through node ports.

If you installed Calico using a manifest, change Felix configuration parameter BPFEnabled to true. This can be done with calicoctl, as follows:

  1. calicoctl patch felixconfiguration default --patch='{"spec": {"bpfEnabled": true}}'

When enabling eBPF mode, preexisting connections continue to use the non-BPF datapath; such connections should not be disrupted, but they do not benefit from eBPF mode’s advantages.

Try out DSR mode

Direct return mode skips a hop through the network for traffic to services (such as node ports) from outside the cluster. This reduces latency and CPU overhead but it requires the underlying network to allow nodes to send traffic with each other’s IPs. In AWS, this requires all your nodes to be in the same subnet and for the source/dest check to be disabled.

DSR mode is disabled by default; to enable it, set the BPFExternalServiceMode Felix configuration parameter to "DSR". This can be done with calicoctl:

  1. calicoctl patch felixconfiguration default --patch='{"spec": {"bpfExternalServiceMode": "DSR"}}'

To switch back to tunneled mode, set the configuration parameter to "Tunnel":

  1. calicoctl patch felixconfiguration default --patch='{"spec": {"bpfExternalServiceMode": "Tunnel"}}'

Switching external traffic mode can disrupt in-progress connections.

Reversing the process

To revert to standard Linux networking:

  1. (Depending on whether you installed Calico with the operator or with a manifest) reverse the changes to the operator’s Installation or the FelixConfiguration resource:

    • Operator
    • Manifest
    1. kubectl patch installation.operator.tigera.io default --type merge -p '{"spec":{"calicoNetwork":{"linuxDataplane":"Iptables"}}}'
    1. calicoctl patch felixconfiguration default --patch='{"spec": {"bpfEnabled": false}}'

  2. If you disabled kube-proxy, re-enable it (for example, by removing the node selector added above).

    1. kubectl patch ds -n kube-system kube-proxy --type merge -p '{"spec":{"template":{"spec":{"nodeSelector":{"non-calico": null}}}}}'

  3. Since disabling eBPF mode is disruptive to existing connections, monitor existing workloads to make sure they re-establish any connections that were disrupted by the switch.