Data Center Host to ToR Architecture

Bond/Etherchannel is not configured on host to multiple switches (bonds can still occur but only to one switch at a time), so leaf01 and leaf02 see two different MAC addresses.

Configurations

leaf01 Config

auto bridge
iface bridge
  bridge-vlan-aware yes
  bridge-ports swp1 peerlink
  bridge-vids 1-2000
  bridge-stp on
auto bridge.10
iface bridge.10
address 10.1.10.2/24
auto peerlink
iface peerlink
bond-slaves glob swp49-50
auto swp1
iface swp1
mstpctl-portadminedge yes
mstpctl-bpduguard yes

Example Host Config (Ubuntu)

auto eth1
iface eth1 inet manual
auto eth1.10
iface eth1.10 inet manual
auto eth2
iface eth1 inet manual
auto eth2.20
iface eth2.20 inet manual
auto br-10
iface br-10 inet manual
bridge-ports eth1.10 vnet0
auto br-20
iface br-20 inet manual
bridge-ports eth2.20 vnet1

Benefits

• Established technology

  • Interoperability with other vendors

  • Easy configuration for customer

  • Immense documentation from multiple vendors and industry

• Ability to use spanning tree commands

  • mstpctl-portadminedge

  • BPDU guard

• Layer 2 reachability to all VMs

Caveats

• The load balancing mechanism on the host can cause problems. If there is only host pinning to each NIC, there are no problems, but if you are doing a bond, you need to look at an MLAG solution.

• No active-active host links. Some operating systems allow HA (NIC failover), but this still does not utilize all the bandwidth. VMs are using one NIC, not two.

• None (not possible with traditional spanning tree)

• VRR

MLAG (multi-chassis link aggregation) is when both uplinks are utilized at the same time. VRR gives the ability for both spines to act as gateways simultaneously for HA (high availability) and active-active mode (both are being used at the same time).

Configurations

leaf01 Config

auto bridge
iface bridge
  bridge-vlan-aware yes
  bridge-ports host-01 peerlink
  bridge-vids 1-2000
  bridge-stp on
auto bridge.10
iface bridge.10
address 172.16.1.2/24
address-virtual 44:38:39:00:00:10 172.16.1.1/24
auto peerlink
iface peerlink
bond-slaves glob swp49-50
auto peerlink.4094
iface peerlink.4094
address 169.254.1.2
clagd-enable yes
clagd-peer-ip 169.254.1.2
clagd-system-mac 44:38:39:FF:40:94
auto host-01
iface host-01
bond-slaves swp1
clag-id 1
{bond-defaults removed for brevity}

Example Host Config (Ubuntu)

auto bond0
iface bond0 inet manual
  bond-slaves eth0 eth1
  {bond-defaults removed for brevity}
auto bond0.10
iface bond0.10 inet manual
auto vm-br10
iface vm-br10 inet manual
bridge-ports bond0.10 vnet0

Benefits

• 100% of links utilized

Caveats

• More complicated (more moving parts)

• More configuration

• No interoperability between vendors

• ISL (inter-switch link) required

Additional Comments

• Can be done with either the traditional or VLAN-aware bridge driver depending on overall STP needs

• There are a few different solutions including Cisco VPC and Arista MLAG, but none of them interoperate and are very vendor specific

• Cumulus Networks Layer 2 HA validated design guide

• VRR

None

• ToR layer (recommended)

• Spine layer

• Core/edge/exit

The server (physical host) has only has one link to one ToR switch.

Configurations

leaf01 Config

/etc/network/interfaces

auto swp1
iface swp1
  address 172.16.1.1/30

/etc/frr/frr.conf

router ospf
  router-id 10.0.0.11
interface swp1
  ip ospf area 0

leaf02 Config

/etc/network/interfaces

auto swp1
iface swp1
  address 172.16.2.1/30

/etc/frr/frr.conf

router ospf
  router-id 10.0.0.12
interface swp1
  ip ospf area 0

host1 Example Config (Ubuntu)

auto eth1
iface eth1 inet static
  address 172.16.1.2/30
  up ip route add 0.0.0.0/0 nexthop via 172.16.1.1

host2 Example Config (Ubuntu)

auto eth1
iface eth1 inet static
  address 172.16.2.2/30
  up ip route add 0.0.0.0/0 nexthop via 172.16.2.1

Benefits

• Relatively simple network configuration

• No STP

• No MLAG

• No L2 loops

• No crosslink between leafs

• Greater route scaling and flexibility

Caveats

• No redundancy for ToR, upgrades would cause downtime

• Many customers do not have software to support application layer redundancy

Additional Comments

• For additional bandwidth links between host and leaf may be bonded

• No redundancy, uses single ToR as gateway.

• Big Data validated design guide uses single attached ToR

Redistribute neighbor daemon grabs ARP entries dynamically, utilizes redistribute table for FRRouting to grab these dynamic entries and redistribute them into the fabric.

Benefits

• Configuration in FRRouting is simple (route-map + redist table)

• Supported by Cumulus Networks

Caveats

• Silent hosts don't receive traffic (depending on ARP).

• IPv4 only.

• If two VMs are on same L2 domain, they could learn about each other directly rather than utilizing gateway, which causes problems (VM migration for example, or getting their network routed). Put hosts on /32 (no other L2 adjacency).

• VM move does not trigger route withdrawal from original leaf (4 hour timeout).

• Clearing ARP impacts routing. May not be obvious.

• No L2 adjacency between servers without VXLAN.

• Equal cost route installed on server/host/hypervisor to both ToRs to load balance evenly.

• For host/VM/container mobility, use the same default route on all hosts (such as x.x.x.1) but don't distribute or advertise the .1 on the ToR into the fabric. This allows the VM to use the same gateway no matter which pair of leafs it is cabled to.

• Cumulus Networks blog post introducing redistribute neighbor

Routing on the host means there is a routing application (such as FRRouting) either on the bare metal host (no VMs/containers) or the hypervisor (for example, Ubuntu with KVM). This is highly recommended by the Cumulus Networks Professional Services team.

Benefits

• No requirement for MLAG

• No spanning-tree or layer 2 domain

• No loops

• 3 or more ToRs can be used instead of usual 2

• Host and VM mobility

• Traffic engineering can be used to migrate traffic from one ToR to another for upgrading both hardware and software

Caveats

• Certain hypervisors or host OSes might not support a routing application like FRRouting and will require a virtual router on the hypervisor

• No L2 adjacnecy between servers without VXLAN

• The first hop is still the ToR, just like redistribute neighbor

• A default route can be advertised by all leaf/ToRs for dynamic ECMP paths

• Install the FRRouting package on an Ubuntu server

• Configuring FRRouting

Instead of routing on the hypervisor, each virtual machine utilizes its own routing stack.

Benefits

• In addition to routing on host:

  • Hypervisor/base OS does not need to be able to do routing

  • VMs can be authenticated into routing fabric

Caveats

• All VMs must be capable of routing

• Scale considerations might need to be taken into an account -
  instead of one routing process, there are as many as there are VMs

• No L2 adjacency between servers without VXLAN

• The first hop is still the ToR, just like redistribute neighbor

• Multiple ToRs (2+) can be used

• Install the FRRouting package on an Ubuntu server

• Configuring FRRouting

Virtual router (vRouter) runs as a VM on the hypervisor/host, sends routes to the ToR using BGP or OSPF.

Benefits

In addition to routing on a host:

• Multi-tenancy can work (multiple customers sharing same racks)

• Base OS does not need to be routing capable

Caveats

• ECMP might not work correctly (load balancing to multiple ToRs); Linux kernel in older versions is not capable of ECMP per flow (does it per packet)

• No L2 adjacency between servers without VXLAN

• The gateway would be the vRouter, which has two routes out (two ToRs)

• Multiple vRouters could be used

• Install the FRRouting package on an Ubuntu server

• Configuring FRRouting

In contrast to routing on the host (preferred), this method allows a user to route to the host. The ToRs are the gateway, as with redistribute neighbor, except because there is no daemon running, the networks must be manually configured under the routing process. There is a potential to black hole unless a script is run to remove the routes when the host no longer responds.

Configurations

leaf01 Config

/etc/network/interfaces

auto swp1
iface swp1
  address 172.16.1.1/30

/etc/frr/frr.conf

router ospf
  router-id 10.0.0.11
interface swp1
  ip ospf area 0

leaf02 Config

/etc/network/interfaces

auto swp2
iface swp2
  address 172.16.1.1/30

/etc/frr/frr.conf

router ospf
  router-id 10.0.0.12
interface swp1
  ip ospf area 0

Example Host Config (Ubuntu)

auto lo
iface lo inet loopback
auto lo:1
iface lo:1 inet static
address 172.16.1.2/32
up ip route add 0.0.0.0/0 nexthop via 172.16.1.1 dev eth0 onlink

nexthop via 172.16.1.1 dev eth1 onlink
auto eth1
iface eth2 inet static
address 172.16.1.2/32
auto eth2
iface eth2 inet static
address 172.16.1.2/32

Benefits

• Most benefits of routing on the host

• No requirement for host to run routing

• No requirement for redistribute neighbor

Caveats

• Removing a subnet from one ToR and re-adding it to another (hence, network statements from your router process) is a manual process

• Network team and server team would have to be in sync, or server team controls the ToR, or automation is being used whenever VM migration happens

• When using VMs/containers it is very easy to black hole traffic, as the leafs continue to advertise prefixes even when VM is down

• No L2 adjacency between servers without VXLAN

• The gateways would be the ToRs, exactly like redistribute neighbor with an equal cost route installed

The host runs LACP (Etherchannel/bond) to the pair of ToRs. LNV (Lightweight Network Virtualization) then transports the L2 bridges across an L3 fabric.

Configurations

leaf01 Config

/etc/network/interfaces

auto lo
iface lo inet loopback
  address 10.0.0.11/32
  vxrd-src-ip 10.0.0.11
  vxrd-svcnode-ip 10.10.10.10
  clagd-vxlan-anycast-ip 36.0.0.11
auto vni-10
iface vni-10
vxlan-id 10
vxlan-local-tunnelip 10.0.0.11
auto br-10
iface br-10
bridge-ports swp1 vni-10

leaf02 Config

/etc/network/interfaces

auto lo
iface lo inet loopback
  address 10.0.0.12/32
  Vxrd-src-ip 10.0.0.12
  vxrd-svcnode-ip 10.10.10.10
  clagd-vxlan-anycast-ip 36.0.0.11
auto vni-10
iface vni-10
vxlan-id 10
vxlan-local-tunnelip 10.0.0.12
auto br-10
iface br-10
bridge-ports swp1 vni-10

Benefits

• Layer 2 domain is reduced to the pair of ToRs

• Aggregation layer is all L3 (VLANs do not have to exist on spine switches)

• Greater route scaling and flexibility

• High availability

Caveats

• Needs MLAG (with the same caveats from the MLAG section above) and spanning tree

• VRR

None

• ToR layer or exit leafs

• Cumulus Linux Lightweight Network Virtualization (LNV) documentation