typhoon/docs/topics/hardware.md
2020-07-25 16:32:08 -07:00

6.1 KiB

Hardware

Typhoon ensures certain networking hardware integrates well with bare-metal Kubernetes.

Ubiquiti

Ubiquiti EdgeRouters and EdgeOS work well with bare-metal Kubernetes clusters. Familiarity with EdgeRouter setup and CLI usage is required.

DHCP

Assign static IPs to clients with known MAC addresses. This is called a static mapping by EdgeOS. Configure the router with the commands based on region inventory.

configure
show service dhcp-server shared-network
set service dhcp-server shared-network-name LAN subnet SUBNET static-mapping NAME mac-address MACADDR
set service dhcp-server shared-network-name LAN subnet SUBNET static-mapping NAME ip-address 10.0.0.20

DNS

Add DNS A records to static IPs as dnsmasq host-records.

configure
set service dns forwarding options host-record=node.example.com,10.0.0.20

Forward *.svc.cluster.local queries to the CoreDNS Kubernetes service IP to allow clients to resolve Kubernetes services.

set service dns forwarding options server=/svc.cluster.local/10.3.0.10
commit-confirm

Restart dnsmasq.

sudo /etc/init.d/dnsmasq restart

PXE

Ubiquiti EdgeRouters can provide a PXE-enabled network boot environment for client machines.

ISC DHCP

With ISC DHCP, add a subnet parameter to the LAN DHCP server to include an ISC DHCP config file.

configure
show service dhcp-server shared-network-name NAME subnet SUBNET
set service dhcp-server shared-network-name NAME subnet SUBNET subnet-parameters "include "/config/scripts/ipxe.conf";"
commit-confirm

Switch to root (i.e. sudo -i) and write the ISC DHCP config /config/scripts/ipxe.conf. iPXE client machines will chainload to matchbox.example.com, while non-iPXE clients will chainload to undionly.kpxe (requires TFTP).

allow bootp;
allow booting;
next-server ADD_ROUTER_IP_HERE;

if exists user-class and option user-class = "iPXE" {
  filename "http://matchbox.example.com/boot.ipxe";
} else {
  filename "undionly.kpxe";
}

dnsmasq

With dnsmasq for DHCP, add options to chainload PXE clients to iPXE undionly.kpxe (requires TFTP), tag iPXE clients, and chainload iPXE clients to matchbox.example.com.

set service dns forwarding options 'dhcp-userclass=set:ipxe,iPXE'
set service dns forwarding options 'pxe-service=tag:#ipxe,x86PC,PXE chainload to iPXE,undionly.kpxe'
set service dns forwarding options 'pxe-service=tag:ipxe,x86PC,iPXE,http://matchbox.example.com/boot.ipxe'

TFTP

Use dnsmasq as a TFTP server to serve undionly.kpxe. Compiling from source with TLS support is strongly recommended. If you use a pre-compiled copy, you must set download_protocol = "http" in your cluster definition (discouraged).

sudo -i
mkdir /config/tftpboot && cd /config/tftpboot
curl http://boot.ipxe.org/undionly.kpxe -o undionly.kpxe

Add dnsmasq command line options to enable the TFTP file server.

configure
show service dns forwarding
set service dns forwarding options enable-tftp
set service dns forwarding options tftp-root=/config/tftpboot
commit-confirm

Routing

Static Routes

Add static route(s) to Kubernetes node(s) that can route to Kubernetes service IPs (default: 10.3.0.0/16). Kubernetes service IPs will become routeable on the LAN.

configure
show protocols static route
set protocols static route 10.3.0.0/16 next-hop NODE_IP
commit-confirm

!!! note Adding multiple next-hop nodes provides equal-cost multi-path (ECMP) routing. EdgeOS v2.0+ is required. The kernel in prior versions used flow-hash to balanced packets, whereas with v2.0, round-robin sessions are used.

BGP

EdgeRouter can exchange routes with other autonomous systems, including a cluster's Calico AS. Peers will exchange podCIDR routes to make individual pods routeable on the LAN.

Define the EdgeRouter AS (if undefined).

configure
show protocols bgp 1
set protocols bgp 1 parameters router-id ROUTER_IP

Peer with node(s) in another AS (eg. Calico default 64512)

set protocols bgp 1 neighbor NODE1_IP remote-as 64512
set protocols bgp 1 neighbor NODE2_IP remote-as 64512
set protocols bgp 1 neighbor NODE3_IP remote-as 64512
commit-confirm

Configure Calico node(s) as to peer with the EdgeRouter.

apiVersion: crd.projectcalico.org/v1
kind: BGPPeer
metadata:
  name: NODE_NAME-to-edgerouter
spec:
  peerIP: ROUTER_IP
  asNumber: 1
  node: NODE_NAME

Or, if every node is to be peered (i.e. full mesh), define a global BGPPeer.

apiVersion: crd.projectcalico.org/v1
kind: BGPPeer
metadata:
  name: global
spec:
  peerIP: ROUTER_IP
  asNumber: 1

If Calico nodes should advertise Kubernetes Service IPs (i.e. ClusterIPs) as well, add a BGPConfiguration.

apiVersion: crd.projectcalico.org/v1
kind: BGPConfiguration
metadata:
  name: default
spec:
  logSeverityScreen: Info
  nodeToNodeMeshEnabled: true
  serviceClusterIPs:
    - cidr: 10.3.0.0/16

Show a summary of peers and exchanged routes.

show ip bgp summary
show ip route bgp

Port Forwarding

Expose the Ingress Controller by adding port-forward rules that DNAT a port on the router's WAN interface to an internal IP and port. By convention, a public Ingress controller is assigned a fixed service IP (e.g. 10.3.0.12).

configure
set port-forward wan-interface eth0
set port-forward lan-interface eth1
set port-forward auto-firewall enable
set port-forward hairpin-nat enable
set port-forward rule 1 description 'ingress http'
set port-forward rule 1 forward-to address 10.3.0.12
set port-forward rule 1 forward-to port 80
set port-forward rule 1 original-port 80
set port-forward rule 1 protocol tcp_udp
set port-forward rule 2 description 'ingress https'
set port-forward rule 2 forward-to address 10.3.0.12
set port-forward rule 2 forward-to port 443
set port-forward rule 2 original-port 443
set port-forward rule 2 protocol tcp_udp
commit-confirm

Web UI

The web UI is often accessible from the LAN on ports 80/443 by default. Edit the ports to 8080 and 4443 to avoid a conflict.

configure
show service gui
set service gui http-port 8080
set service gui https-port 4443
commit-confirm