This document outlines the networking requirements for installing and operating GKE on Bare Metal.
External network requirements
GKE on Bare Metal requires an internet connection for operational purposes. GKE on Bare Metal retrieves cluster components from Container Registry, and the cluster is registered with Connect.
You can connect to Google by using the public internet through HTTPS, a virtual private network (VPN), or a Dedicated Interconnect connection.
If the machines you are using for your admin workstation and cluster nodes use a proxy server to access the internet, your proxy server must allow some specific connections. For details, see the prerequisites section of Install behind a proxy.
Internal network requirements
GKE on Bare Metal can work with Layer 2 or Layer 3 connectivity between cluster nodes. The load balancer nodes can be the control plane nodes or a dedicated set of nodes. For more information, see Choosing and configuring load balancers.
When you use Bundled Layer 2 load balancing with
MetalLB (spec.loadBalancer.mode: bundled
and spec.loadBalancer.type: layer2), load balancer nodes require Layer 2
adjacency. The Layer 2 adjacency requirement applies whether you run the load
balancer on control plane nodes or in a dedicated set of load balancing nodes.
Bundled load balancing with BGP supports
Layer 3 protocol, so strict Layer 2 adjacency isn't required.
The requirements for load balancer machines are as follows:
- For Bundled Layer 2 load balancing, all load balancers for a given cluster are in the same Layer 2 domain. Control plane nodes must also be in the same Layer 2 domain.
- For Bundled Layer 2 load balancing, all virtual IP addresses (VIPs) must be in the load balancer machine subnet and routable to the gateway of the subnet.
- Users are responsible for allowing ingress load balancer traffic.
Pod networking
GKE on Bare Metal 1.7.0 and later versions allows you to configure up to 250 pods per node. Kubernetes assigns a Classless Inter-Domain Routing (CIDR) block to each node so that each pod can have a unique IP address. The size of the CIDR block corresponds to the maximum number of pods per node. The following table lists the size of the CIDR block that Kubernetes assigns to each node based on the configured maximum pods per node:
| Maximum pods per node | CIDR block per node | Number of IP addresses |
|---|---|---|
| 32 | /26 | 64 |
| 33 – 64 | /25 | 128 |
| 65 – 128 | /24 | 256 |
| 129 - 250 | /23 | 512 |
Running 250 pods per node requires Kubernetes to reserve a /23 CIDR block
for each node. Assuming that your cluster uses the default value of /16
for the clusterNetwork.pods.cidrBlocks field, your cluster has a
limit of (2(23-16))=128 nodes. If you intend to grow the
cluster beyond this limit, you can either increase the value of
clusterNetwork.pods.cidrBlocks or decrease the value of
nodeConfig.podDensity.maxPodsPerNode. This method had a few disadvantages.
Single user cluster deployment with high availability
The following diagram illustrates a number of key networking concepts for GDCV for Bare Metal in one possible network configuration.
Consider the following information to meet the network requirements:
- The control plane nodes run the load balancers, and they all have Layer 2 connectivity, while other connections, including worker nodes, only require Layer 3 connectivity.
- Configuration files define IP addresses for worker node pools.
Configuration files also define VIPs for the following
purposes:
- Services
- Ingress
- Control plane access through the Kubernetes API
- You require a connection to Google Cloud.
Port usage
This section shows how UDP and TCP ports are used on cluster and load balancer nodes.
Control plane nodes
| Protocol | Direction | Port range | Purpose | Used by |
|---|---|---|---|---|
| UDP | Inbound | 6081 | GENEVE Encapsulation | Self |
| TCP | Inbound | 22 | Provisioning and updates of admin cluster nodes | Admin workstation |
| TCP | Inbound | 6444 | Kubernetes API server | All |
| TCP | Inbound | 2379 - 2381 | etcd server client API | kube-apiserver and etcd |
| TCP | Inbound | 2382 - 2384 | etcd-events server client API | kube-apiserver and etcd-events |
| TCP | Inbound | 10250 | kubelet API | Self and control plane |
| TCP | Inbound | 10251 | kube-scheduler | Self |
| TCP | Inbound | 10252 | kube-controller-manager | Self |
| TCP | Inbound | 10256 | Node health check | All |
| TCP | Both | 4240 | CNI health check | All |
Worker nodes
| Protocol | Direction | Port range | Purpose | Used by |
|---|---|---|---|---|
| TCP | Inbound | 22 | Provisioning and updates of user cluster nodes | Admin cluster nodes |
| UDP | Inbound | 6081 | GENEVE Encapsulation | Self |
| TCP | Inbound | 10250 | kubelet API | Self and control plane |
| TCP | Inbound | 10256 | Node health check | All |
| TCP | Inbound | 30000 - 32767 | NodePort services | Self |
| TCP | Both | 4240 | CNI health check | All |
Load balancer nodes
| Protocol | Direction | Port range | Purpose | Used by |
|---|---|---|---|---|
| TCP | Inbound | 22 | Provisioning and updates of user cluster nodes | Admin cluster nodes |
| UDP | Inbound | 6081 | GENEVE Encapsulation | Self |
| TCP | Inbound | 443* | Cluster management | All |
| TCP | Both | 4240 | CNI health check | All |
| TCP | Inbound | 7946 | Metal LB health check | load balancer nodes |
| UDP | Inbound | 7946 | Metal LB health check | load balancer nodes |
| TCP | Inbound | 10256 | Node health check | All |
* This port can be configured in the cluster config, using the
controlPlaneLBPort field.
Multi-cluster port requirements
In a multi-cluster configuration, added clusters must include the following ports to communicate with the admin cluster.
| Protocol | Direction | Port range | Purpose | Used by |
|---|---|---|---|---|
| TCP | Inbound | 22 | Provisioning and updates of cluster nodes | All nodes |
| TCP | Inbound | 443* | Kubernetes API server for added cluster | Control plane and load balancer nodes |
* This port can be configured in the cluster config, using the
controlPlaneLBPort field.
Configure firewalld ports
You are not required to disable firewalld to run GKE on Bare Metal on Red
Hat Enterprise Linux (RHEL) or CentOS. To use firewalld, you must open the UDP
and TCP ports used by control plane, worker, and load balancer nodes as described in
Port usage on this page. The following example configurations
show how you can open ports with firewall-cmd, the firewalld command-line
utility. You should run the commands as the root user.
Control plane node example configuration
The following block of commands shows an example of how you can open the needed ports on servers running control plane nodes:
firewall-cmd --permanent --zone=public --add-port=22/tcp
firewall-cmd --permanent --zone=public --add-port=4240/tcp
firewall-cmd --permanent --zone=public --add-port=6444/tcp
firewall-cmd --permanent --zone=public --add-port=6081/udp
firewall-cmd --permanent --zone=public --add-port=10250-10252/tcp
firewall-cmd --permanent --zone=public --add-port=10256/tcp
firewall-cmd --permanent --zone=public --add-port=2379-2380/tcp
firewall-cmd --permanent --zone=public --add-port=443/tcp
firewall-cmd --permanent --zone=public --add-port=30000-32767/tcp
firewall-cmd --permanent --new-zone=k8s-pods
firewall-cmd --permanent --zone=k8s-pods --add-source PODS_CIDR
firewall-cmd --permanent --zone=k8s-pods --set-target=ACCEPT
firewall-cmd --reload
Replace PODS_CIDR with the CIDR blocks reserved for your
pods configured in the clusterNetwork.pods.cidrBlocks field. The default CIDR
block for pods is 192.168.0.0/16.
Worker node example configuration
The following block of commands shows an example of how you can open the needed ports on servers running worker nodes:
firewall-cmd --permanent --zone=public --add-port=22/tcp
firewall-cmd --permanent --zone=public --add-port=4240/tcp
firewall-cmd --permanent --zone=public --add-port=6444/tcp
firewall-cmd --permanent --zone=public --add-port=6081/udp
firewall-cmd --permanent --zone=public --add-port=10250/tcp
firewall-cmd --permanent --zone=public --add-port=10256/tcp
firewall-cmd --permanent --zone=public --add-port=443/tcp
firewall-cmd --permanent --zone=public --add-port=30000-32767/tcp
firewall-cmd --permanent --new-zone=k8s-pods
firewall-cmd --permanent --zone=k8s-pods --add-source PODS_CIDR
firewall-cmd --permanent --zone=k8s-pods --set-target=ACCEPT
firewall-cmd --reload
Replace PODS_CIDR with the CIDR blocks reserved for your
pods configured in the clusterNetwork.pods.cidrBlocks field. The default CIDR
block for pods is 192.168.0.0/16.
Load balancer node example configuration
The following block of commands shows an example of how you can open the needed ports on servers running load balancer nodes:
firewall-cmd --permanent --zone=public --add-port=22/tcp
firewall-cmd --permanent --zone=public --add-port=4240/tcp
firewall-cmd --permanent --zone=public --add-port=6444/tcp
firewall-cmd --permanent --zone=public --add-port=7946/tcp
firewall-cmd --permanent --zone=public --add-port=7946/udp
firewall-cmd --permanent --zone=public --add-port=6081/udp
firewall-cmd --permanent --zone=public --add-port=10250/tcp
firewall-cmd --permanent --zone=public --add-port=10256/tcp
firewall-cmd --permanent --zone=public --add-port=443/tcp
firewall-cmd --permanent --zone=public --add-port=30000-32767/tcp
firewall-cmd --permanent --new-zone=k8s-pods
firewall-cmd --permanent --zone=k8s-pods --add-source PODS_CIDR
firewall-cmd --permanent --zone=k8s-pods --set-target=ACCEPT
firewall-cmd --reload
Replace PODS_CIDR with the CIDR blocks reserved for your
pods configured in the clusterNetwork.pods.cidrBlocks field. The default CIDR
block for pods is 192.168.0.0/16.
Confirm your port configuration
To verify your port configuration, use the following steps on control plane, worker, and load balancer nodes:
Run the following Network Mapper command to see what ports are open:
nmap localhostRun the following commands to get your firewalld configuration settings:
firewall-cmd --zone=public --list-all-policies firewall-cmd --zone=public --list-ports firewall-cmd --zone=public --list-services firewall-cmd --zone=k8s-pods --list-all-policies firewall-cmd --zone=k8s-pods --list-ports firewall-cmd --zone=k8s-pods --list-servicesIf necessary, rerun the commands from the preceding sections to configure your nodes properly. You may need to run the commands as the root user.
Known issue for firewalld
When running GKE on Bare Metal with firewalld enabled on either CentOS or
Red Hat Enterprise Linux (RHEL), changes to firewalld can remove the Cilium
iptables chains on the host network. The iptables chains are added by the
anetd Pod when it's started. The loss of the Cilium iptables chains causes
the Pod on the Node to lose network connectivity outside of the Node.
Changes to firewalld that remove the iptables chains include, but aren't
limited to:
Restarting
firewalld, usingsystemctlReloading
firewalldwith the command line client (firewall-cmd --reload)
To apply firewalld changes without removing iptables chains, restart anetd
on the Node:
Locate and delete the
anetdPod with the following commands to restartanetd:kubectl get pods -n kube-system kubectl delete pods -n kube-system ANETD_XYZReplace ANETD_XYZ with the name of the
anetdPod.