tklauser/eks-kube-proxy-free.md

Based on:

NOTE

Currently does not work with Cilium 1.8.0-rc3 due to cilium/cilium#12078 Thus use cilium:latest for now.

Setup instructions

Spin up EKS cluster using eksctl:

$ eksctl create cluster --name test-cluster --without-nodegroup
$ kubectl -n kube-system delete daemonset aws-node

Install Cilium in order to be able to create node group:

$ helm install cilium ./cilium \
  --namespace kube-system \
  --set global.eni=true \
  --set config.ipam=eni \
  --set global.egressMasqueradeInterfaces=eth0 \
  --set global.tunnel=disabled \
  --set global.nodeinit.enabled=true

Create a node group. We need SSH access in order to be able to install a newer kernel and additional tools. We also need to select a larger instance type in order to be able to run XDP on the provided ENA device. Use nodegroup-config.yaml for convenience:

$ cat nodegroup-config.yaml
apiVersion: eksctl.io/v1alpha5
kind: ClusterConfig

metadata:
  name: test-cluster
  region: us-west-2

nodeGroups:
  - name: ng-1
    instanceType: m5n.xlarge
    desiredCapacity: 2
    ssh:
      allow: true
$ eksctl create nodegroup -f nodegroup-config.yaml

Get external IP of nodes:

$ EXT_IPS=$(kubectl get no -o jsonpath='{$.items[*].status.addresses[?(@.type=="ExternalIP")].address }{"\n"}' | tr ' ' '\n')

Install a newer kernel on each node:

bash:

$ for ip in $EXT_IPS ; do ssh ec2-user@$ip "sudo amazon-linux-extras install -y kernel-ng && sudo reboot"; done

zsh:

$ for ip in ${(f)EXT_IPS} ; do ssh ec2-user@$ip "sudo amazon-linux-extras install -y kernel-ng && sudo reboot"; done

Wait until nodes come back up and check their kernel version (should say 5.4.38-17.76.amzn2.x86_64 or similar):

bash:

$ for ip in $EXT_IPS ; do ssh ec2-user@$ip "uname -r"; done

zsh:

$ for ip in ${(f)EXT_IPS} ; do ssh ec2-user@$ip "uname -r"; done

Install ethtool on nodes in order to be able to configure ENA NIC queues:

bash:

$ for ip in $EXT_IPS ; do ssh ec2-user@$ip "sudo yum install -y ethtool"; done

zsh:

$ for ip in ${(f)EXT_IPS} ; do ssh ec2-user@$ip "sudo yum install -y ethtool"; done

Since there was a bug regarding XDP in earlier versions of the ena driver, make sure to use at least version 2.2.8:

bash:

$ for ip in $EXT_IPS ; do ssh ec2-user@$ip "sudo ethtool -i eth0"; done

zsh:

$ for ip in ${(f)EXT_IPS} ; do ssh ec2-user@$ip "sudo ethtool -i eth0"; done

If the reported driver version is < 2.2.8, a new version of the ena driver needs to be built and installed according to the instructions given here. For kernel-ng version 5.4.38-17.76.amzn2.x86_64 a prebuilt rpm of ena driver version 2.2.9g can be dowloaded here.

Set MTU on the netdevs (default is 9001, for XDP the ENA driver allows max. 3818):

bash:

$ for ip in $EXT_IPS ; do ssh ec2-user@$ip "sudo ip link set dev eth0 mtu 3818"; done

zsh:

$ for ip in ${(f)EXT_IPS} ; do ssh ec2-user@$ip "sudo ip link set dev eth0 mtu 3818"; done

Get the current NIC queues configured on the netdevs.

bash:

$ for ip in $EXT_IPS ; do ssh ec2-user@$ip "sudo ethtool -l eth0"; done

zsh:

$ for ip in ${(f)EXT_IPS} ; do ssh ec2-user@$ip "sudo ethtool -l eth0"; done

The output is expected to look something like this:

Channel parameters for eth0:
Pre-set maximums:
RX:             0
TX:             0
Other:          0
Combined:       4
Current hardware settings:
RX:             0
TX:             0
Other:          0
Combined:       4

Set NIC queues on the netdevs to be able to use XDP (use 1/2 of the value given for Combined above):

bash:

$ for ip in $EXT_IPS ; do ssh ec2-user@$ip "sudo ethtool -L eth0 combined 2"; done

zsh:

$ for ip in ${(f)EXT_IPS} ; do ssh ec2-user@$ip "sudo ethtool -L eth0 combined 2"; done

Get the K8S API server IP and port:

$ export API_SERVER_IP=$(kubectl get ep kubernetes -o jsonpath='{$.subsets[0].addresses[0].ip}')
$ export API_SERVER_PORT=443

Delete kube-proxy daemonset:

$ kubectl -n kube-system delete daemonset kube-proxy

Configure Cilium for kube-proxy free mode with NodePort XDP acceleration:

$ helm upgrade cilium ./cilium \
	--namespace kube-system \
	--reuse-values \
	--set global.autoDirectNodeRoutes=true \
	--set global.kubeProxyReplacement=strict \
	--set global.nodePort.acceleration=native \
	--set global.nodePort.mode=snat \
	--set global.k8sServiceHost=$API_SERVER_IP \
	--set global.k8sServicePort=$API_SERVER_PORT

Restart the Cilium daemonset:

$ kubectl rollout restart -n kube-system ds/cilium

	apiVersion: eksctl.io/v1alpha5
	kind: ClusterConfig

	metadata:
	name: test-cluster
	region: us-west-2

	nodeGroups:
	- name: ng-1
	instanceType: m5n.xlarge
	desiredCapacity: 2
	ssh:
	allow: true