OpenBSD manual page server

Manual Page Search Parameters

CARP(4) Device Drivers Manual CARP(4)

carpCommon Address Redundancy Protocol

pseudo-device carp

The carp interface is a pseudo-device which implements and controls the CARP protocol. carp allows multiple hosts on the same local network to share a set of IP addresses. Its primary purpose is to ensure that these addresses are always available, but in some configurations carp can also provide load balancing functionality.

A carp interface can be created at runtime using the ifconfig carpN create command or by setting up a hostname.if(5) configuration file for netstart(8).

To use carp, the administrator needs to configure at minimum a common virtual host ID (VHID) and virtual host IP address on each machine which is to take part in the virtual group. Additional parameters can also be set on a per-interface basis: advbase and advskew, which are used to control how frequently the host sends advertisements when it is the master for a virtual host, and pass which is used to authenticate carp advertisements. Finally carpdev is used to specify which interface the carp device attaches to. If unspecified, the kernel attempts to set it by looking for another interface with the same subnet. These configurations can be done using ifconfig(8), or through the SIOCSVH ioctl.

carp can also be used in conjunction with ifstated(8) to respond to changes in CARP state; however, for most uses this will not be necessary. See the manual page for ifstated(8) for more information.

Additionally, there are a number of global parameters which can be set using sysctl(8):

net.inet.carp.allow
Accept incoming carp packets. Enabled by default.
net.inet.carp.preempt
Allow virtual hosts to preempt each other. Disabled by default.
net.inet.carp.log
Make carp log state changes, bad packets, and other errors. May be a value between 0 and 7 corresponding with syslog(3) priorities. The default value is 2, which limits logging to changes in CARP state.

LOAD BALANCING

carp provides two mechanisms to load balance incoming traffic over a group of carp hosts: ARP balancing and IP balancing.

Which one to use mainly depends on the network environment carp is being used in. ARP balancing has limited abilities for load balancing the incoming connections between hosts in an Ethernet network. It only works for clients in the local network, because ARP balancing spreads the load by varying ARP replies based on the source MAC address of the host sending the query. Therefore it cannot balance traffic that crosses a router, because the router itself will always be balanced to the same virtual host.

IP balancing is not dependent on ARP and therefore also works for traffic that comes over a router. This method should work in all environments and can also provide more fine grained load balancing than ARP balancing. The downside of IP balancing is that it requires the traffic that is destined towards the load balanced IP addresses to be received by all carp hosts. While this is always the case when connected to a hub, it has to play some tricks in switched networks, which will result in a higher network load.

A rule of thumb might be to use ARP balancing if there are many hosts on the same network segment and to use IP balancing for all other cases.

To configure load balancing one has to specify multiple carp nodes using the carpnodes option. Each node in a load balancing cluster is represented by at least one "vhid:advskew" pair in a comma separated list. carp tries to distribute the incoming network load over all configured carpnodes. The following example creates a load balancing group consisting of three nodes, using vhids 3, 4 and 6:

# ifconfig carp0 carpnodes 3:0,4:0,6:100

The advskew value of the last node is set to 100, so that this node is designated to the BACKUP state. It will only become MASTER if all nodes with a lower advskew value have failed. By varying this value throughout the machines in the cluster it is possible to decide which share of the network load each node receives. Therefore, all carp interfaces in the cluster are configured identically, except for a different advskew value within the carpnodes specification.

See the EXAMPLES section for a practical example of load balancing.

For ARP balancing, one has to configure multiple carpnodes and choose the balancing mode arp.

Once an ARP request is received, the CARP protocol will use a hashing function against the source MAC address in the ARP request to determine which carpnode the request belongs to. If the corresponding carpnode is in master state, the ARP request will be answered, otherwise it will be ignored.

The ARP load balancing has some limitations. Firstly, ARP balancing only works on the local network segment. It cannot balance traffic that crosses a router, because the router itself will always be balanced to the same carpnode. Secondly, ARP load balancing can lead to asymmetric routing of incoming and outgoing traffic, thus combining it with pfsync(4) requires special care, because this can create a race condition between balanced routers and the host they are serving. ARP balancing can be safely used with pfsync if the pf(4) ruleset translates the source address to an unshared address on the outgoing interface using a NAT rule. This requires multiple CARP groups with IP addresses on the outgoing interface, configured so that each host is the master of one group.

ARP balancing also works for IPv6, but instead of ARP the Neighbor Discovery Protocol (NDP) is used.

IP load balancing works by utilizing the network itself to distribute incoming traffic to all carp nodes in the cluster. Each packet is filtered on the incoming carp interface so that only one node in the cluster accepts the packet. All the other nodes will just silently drop it. The filtering function uses a hash over the source and destination address of the IPv4 or IPv6 packet and compares the result against the state of the carpnode.

IP balancing is activated by setting the balancing mode to ip. This is the recommended default setting. In this mode, carp uses a multicast MAC address, so that a switch sends incoming traffic towards all nodes.

However, there are a few OS and routers that do not accept a multicast MAC address being mapped to a unicast IP. This can be resolved by using one of the following unicast options. For scenarios where a hub is used it is not necessary to use a multicast MAC and it is safe to use the ip-unicast mode. Manageable switches can usually be tricked into forwarding unicast traffic to all cluster nodes ports by configuring them into some sort of monitoring mode. If this is not possible, using the ip-stealth mode is another option, which should work on most switches. In this mode carp never sends packets with its virtual MAC address as source. Stealth mode prevents a switch from learning the virtual MAC address, so that it has to flood the traffic to all its ports. Please note that activating stealth mode on a carp interface that has already been running might not work instantly. As a workaround the VHID of the first carpnode can be changed to a previously unused one, or just wait until the MAC table entry in the switch times out. Some layer 3 switches do port learning based on ARP packets. Therefore the stealth mode cannot hide the virtual MAC address from these kind of devices.

If IP balancing is being used on a firewall, it is recommended to configure the carpnodes in a symmetrical manner. This is achieved by simply using the same carpnodes list on all sides of the firewall. This ensures that packets of one connection will pass in and out on the same host and are not routed asymmetrically.

For most scenarios it is desirable to have a well-defined master, achieved by enabling the preempt option. Enable it on both host A and B:

# sysctl net.inet.carp.preempt=1

Assume that host A is the preferred master and 192.168.1.x/24 is configured on one physical interface and 192.168.2.y/24 on another. This is the setup for host A:

# ifconfig carp0 192.168.1.1 vhid 1
# ifconfig carp1 192.168.2.1 vhid 2

The setup for host B is identical, but it has a higher advskew:

# ifconfig carp0 192.168.1.1 vhid 1 advskew 100
# ifconfig carp1 192.168.2.1 vhid 2 advskew 100

LOAD BALANCING

In order to set up a load balanced virtual host, it is necessary to configure one carpnodes entry for each physical host. In the following example, two physical hosts are configured to provide balancing and failover for the IP address 192.168.1.10.

First the carp interface on Host A is configured. The advskew of 100 on the second carpnode entry means that its advertisements will be sent out slightly less frequently and will therefore become the designated backup.

# ifconfig carp0 192.168.1.10 carpnodes 1:0,2:100 balancing ip

The configuration for host B is identical, except the skew is on the carpnode entry with virtual host 1 rather than virtual host 2.

# ifconfig carp0 192.168.1.10 carpnodes 1:100,2:0 balancing ip

If ARP balancing or a different mode of IP balancing is desired the balancing mode can be adjusted accordingly.

sysctl(3), inet(4), pfsync(4), hostname.if(5), ifconfig(8), ifstated(8), netstart(8), sysctl(8)

The carp device first appeared in OpenBSD 3.5.

If load balancing is used in setups where the carpdev does not share an IP in the same subnet as carp, it is not possible to use the IP of the carp interface for self originated traffic. This is because the return packets are also subject to load balancing and might end up on any other node in the cluster.

If an IPv6 load balanced carp interface is taken down manually, it will accept all incoming packets for its address. This will lead to duplicated packets.

March 12, 2010 OpenBSD-5.1