Redistribute Neighbor

Redistribute neighbor provides a mechanism for IP subnets to span racks without forcing the end hosts to run a routing protocol.

The fundamental premise behind redistribute neighbor is to announce individual host /32 routes in the routed fabric. Other hosts on the fabric can then use this new path to access the hosts in the fabric. If multiple equal-cost paths (ECMP) are available, traffic can load balance across the available paths natively.

The challenge is to accurately compile and update this list of reachable hosts or neighbors. Luckily, existing commonly-deployed protocols are available to solve this problem. Hosts use ARP to resolve MAC addresses when sending to an IPv4 address. A host then builds an ARP cache table of known MAC addresses: IPv4 tuples as they receive or respond to ARP requests.

In the case of a leaf switch, where the default gateway is deployed for hosts within the rack, the ARP cache table contains a list of all hosts that have ARP’d for their default gateway. In many scenarios, this table contains all the layer 3 information that is needed. This is where redistribute neighbor comes in, as it is a mechanism of formatting and syncing this table into the routing protocol.

Redistribute neighbor is distributed as python-rdnbrd.

Target Use Cases and Best Practices

Redistribute neighbor is typically used in these configurations:

• Virtualized clusters
• Hosts with service IP addresses that migrate between racks
• Hosts that are dual connected to two leaf nodes without using proprietary protocols such as MLAG
• Anycast services that need dynamic advertisement from multiple hosts

Follow these guidelines:

• You can connect a host to one or more leafs. Each leaf advertises the /32 it sees in its neighbor table.
• Make sure that a host-bound bridge/VLAN is local to each switch.
• Connect leaf switches with redistribute neighbor enabled directly to the hosts.
• Make sure that IP addressing is non-overlapping, as the host IP addresses are directly advertised into the routed fabric.
• Run redistribute neighbor on Linux-based hosts. NVIDIA has not actively tested other host operating systems.

How It Works

Redistribute neighbor works as follows:

1. The leaf/ToR switches learn about connected hosts when the host sends an ARP request or ARP reply.
2. An entry for the host is added to the kernel neighbor table of each leaf switch.
3. The redistribute neighbor daemon, rdnbrd, monitors the kernel neighbor table and creates a /32 route for each neighbor entry. This /32 route is created in kernel table 10.
4. FRRouting is configured to import routes from kernel table 10.
5. A route-map controls which routes from table 10 are imported.
6. In FRRouting these routes are imported as table routes.
7. BGP, OSPF and so on, are then configured to redistribute the table 10 routes.

Example Configuration

The following example configuration is based on the reference topology created by NVIDIA. Other configurations are possible, based on the use cases outlined above. Here is a diagram of the topology:

Configure the Leafs

The following steps demonstrate how to configure leaf01, but you can follow the same steps on any of the leafs.

The NCLU and vtysh commands save the configuration in the /etc/frr/frr.conf file. The following example uses OSPF as the routing protocol:

frr defaults datacenter
ip import-table 10 route-map REDIST_NEIGHBOR
username cumulus nopassword
!
service integrated-vtysh-config
!
log syslog informational
!
router bgp 65001
 !
 address-family ipv4 unicast
  redistribute table 10
 exit-address-family
!
route-map REDIST_NEIGHBOR permit 10
 match interface swp1
!
route-map REDIST_NEIGHBOR permit 20
 match interface swp2
!
router ospf
 redistribute table 10
!
line vty
!

Configure the Hosts

There are a few possible host configurations that range in complexity. This document only covers the basic use case: dual-connected Linux hosts with static IP addresses assigned.

Configure a Dual-connected Host

Configure a host with the same /32 IP address on its loopback (lo) and uplinks (in this example, eth1 and eth2). This is done so both leaf switches advertise the same /32 regardless of the interface. Cumulus Linux relies on ECMP to load balance across the interfaces southbound, and an equal cost static route (see the configuration below) for load balancing northbound.

The loopback hosts the primary service IP address(es) and to which you can bind services.

Configure the loopback and physical interfaces. Referring back to the topology diagram, server01 is connected to leaf01 via eth1 and to leaf02 via eth2. You should note:

• The loopback IP is assigned to lo, eth1 and eth2.

• The post-up ARPing is used to force the host to ARP as soon as its interface comes up. This allows the leaf to learn about the host as soon as possible.

• The post-up ip route replace is used to install a default route via one or both leaf nodes if both swp1 and swp2 are up.

  ▼ Click to expand

  # The loopback network interface
  auto lo
  iface lo inet loopback
  
  auto lo:1
  iface lo:1
      address 10.1.0.101/32
  
  auto eth1
  iface eth1
      address 10.1.0.101/32
      post-up for i in {1..3}; do arping -q -c 1 -w 0 -i eth1 10.0.0.11; sleep 1; done
      post-up ip route add 0.0.0.0/0 nexthop via 10.0.0.11 dev eth1 onlink nexthop via 10.0.0.12 dev eth2 onlink || true
  
  auto eth2
  iface eth2
      address 10.1.0.101/32
      post-up for i in {1..3}; do arping -q -c 1 -w 0 -i eth2 10.0.0.12; sleep 1; done
      post-up ip route add 0.0.0.0/0 nexthop via 10.0.0.11 dev eth1 onlink nexthop via 10.0.0.12 dev eth2 onlink || true
  ...
  

Install ifplugd

Additionally, install and use ifplugd. ifplugd modifies the behavior of the Linux routing table when an interface undergoes a link transition (carrier up/down). The Linux kernel by default leaves routes up even when the physical interface is unavailable (NO-CARRIER).

After you install ifplugd, edit /etc/default/ifplugd as follows, where eth1 and eth2 are the interface names that your host uses to connect to the leaves.

user@server01:$ sudo nano /etc/default/ifplugd
INTERFACES="eth1 eth2"
HOTPLUG_INTERFACES=""
ARGS="-q -f -u10 -d10 -w -I"
SUSPEND_ACTION="stop"

For full instructions on installing ifplugd on Ubuntu, follow this guide.

Troubleshooting

How do I determine if rdnbrd (the redistribute neighbor daemon) is running?

Run the systemctl status rdnbrd.service command:

cumulus@leaf01:~$ systemctl status rdnbrd.service
* rdnbrd.service - Cumulus Linux Redistribute Neighbor Service
 Loaded: loaded (/lib/systemd/system/rdnbrd.service; enabled)
 Active: active (running) since Wed 2016-05-04 18:29:03 UTC; 1h 13min ago
 Main PID: 1501 (python)
 CGroup: /system.slice/rdnbrd.service
 `-1501 /usr/bin/python /usr/sbin/rdnbrd -d

How do I change the default configuration of rdnbrd?

Edit the /etc/rdnbrd.conf file, then run systemctl restart rdnbrd.service:

cumulus@leaf01:~$ sudo nano /etc/rdnbrd.conf
# syslog logging level CRITICAL, ERROR, WARNING, INFO, or DEBUG
loglevel = INFO

# TX an ARP request to known hosts every keepalive seconds
keepalive = 1

# If a host does not send an ARP reply for holdtime consider the host down
holdtime = 3

# Install /32 routes for each host into this table
route_table = 10

# Uncomment to enable ARP debugs on specific interfaces.
# Note that ARP debugs can be very chatty.
# debug_arp = swp1 swp2 swp3 br1
# If we already know the MAC for a host, unicast the ARP request. This is
# unusual for ARP (why ARP if you know the destination MAC) but we will be
# using ARP as a keepalive mechanism and do not want to broadcast so many ARPs
# if we do not have to. If a host cannot handle a unicasted ARP request, set
#
# Unicasting ARP requests is common practice (in some scenarios) for other
# networking operating systems so it is unlikely that you will need to set
# this to False.
unicast_arp_requests = True
cumulus@leaf01:~$ sudo systemctl restart rdnbrd.service

What is table 10? Why was table 10 chosen?

The Linux kernel supports multiple routing tables and can utilize 0 through 255 as table IDs; however tables 0, 253, 254 and 255 are reserved, and 1 is usually the first one utilized. Therefore, rdnbrd only allows you to specify 2-252. Cumulus Linux uses table ID 10, however you can set the ID to any value between 2-252. You can see all the tables specified here:

cumulus@leaf01:~$ cat /etc/iproute2/rt_tables
#
# reserved values
#
255 local
254 main
253 default
0 unspec
#
# local
#
#1  inr.ruhep

For more information, refer to Linux route tables or you can read the Ubuntu man pages for ip route.

How do I determine that the /32 redistribute neighbor routes are being advertised to my neighbor?

For BGP, run the NCLU net show bgp neighbor <interface> advertise-routes command or the vtysh show ip bgp neighbor swp51 advertised-routes command. For example:

cumulus@leaf01:~$ net show bgp neighbor swp51 advertise-routes
BGP table version is 5, local router ID is 10.0.0.11
Status codes: s suppressed, d damped, h history, * valid, > best, = multipath,
              i internal, r RIB-failure, S Stale, R Removed
Origin codes: i - IGP, e - EGP, ? - incomplete

    Network         Next Hop            Metric LocPrf Weight Path
*> 10.0.0.11/32     0.0.0.0                  0         32768 i
*> 10.0.0.12/32     ::                                     0 65020 65012 i
*> 10.0.0.21/32     ::                                     0 65020 i
*> 10.0.0.22/32     ::                                     0 65020 i

Total number of prefixes 4

How do I verify that the kernel routing table is being correctly populated?

Use the following workflow to verify that the kernel routing table isbeing populated correctly and that routes are being correctly imported/advertised:

1. Verify that ARP neighbor entries are being populated into the Kernel routing table 10.

   cumulus@leaf01:~$ ip route show table 10
   10.0.1.101 dev swp1 scope link
   

   If these routes are not being generated, verify the following that the rdnbrd daemon is running and check the /etc/rdnbrd.conf file to verify the correct table number is used.

2. Verify that routes are being imported into FRRouting from the kernel routing table 10.

   cumulus@leaf01:~$ sudo vtysh
   leaf01# show ip route table
   Codes: K - kernel route, C - connected, S - static, R - RIP,
           O - OSPF, I - IS-IS, B - BGP, A - Babel, T - Table,
           > - selected route, * - FIB route
   T[10]>* 10.0.1.101/32 [19/0] is directly connected, swp1, 01:25:29
   

   Both the > and * should be present so that table 10 routes are installed as preferred into the routing table. If the routes are not being installed, verify the imported distance of the locally imported kernel routes with the ip import 10 distance X command (where X is not less than the administrative distance of the routing protocol). If the distance is too low, routes learned from the protocol might overwrite the locally imported routes. Also, verify that the routes are in the kernel routing table.

3. Confirm that routes are in the BGP/OSPF database and are being advertised.

   leaf01# show ip bgp
   

Considerations

TCAM Route Scale

This feature adds each ARP entry as a /32 host route into the routing table of all switches within a summarization domain. Take care to keep the number of hosts minus fabric routes under the TCAM size of the switch.

Possible Uneven Traffic Distribution

Linux uses source layer 3 addresses only to do load balancing on most older distributions.

Silent Hosts Never Receive Traffic

Freshly provisioned hosts that have never sent traffic may not ARP for their default gateways. The post-up ARPing in /etc/network/interfaces on the host should take care of this. If the host does not ARP, then rdnbrd on the leaf cannot learn about the host.

Unsupported with EVPN

Redistribute neighbor is unsupported when the BGP EVPN Address Family is enabled. Enabling both redistribute neighbor and EVPN will lead to unreachable IPv4 ARP and IPv6 neighbor entries.