Anycast

The Internet Protocol and other network addressing systems recognize five main addressing methods:

• Unicast addressing uses a one-to-one association between a sender and destination: each destination address uniquely identifies a single receiver endpoint.
• Broadcast uses a one-to-all association; a single datagram from one sender is routed to all of the possibly multiple endpoints associated with the broadcast address. The network automatically replicates datagrams as needed to reach all the recipients within the scope of the broadcast, which is generally an entire network subnet.
• Multicast addressing uses a one-to-many-of-many or many-to-many-of-many association; datagrams are routed simultaneously in a single transmission to many recipients. It differs from broadcast in that the destination address designates a subset, not necessarily all, of the accessible nodes.
• Anycast addressing is a one-to-one-of-many association where datagrams are routed to any single member of a group of potential receivers that are all identified by the same destination address. The routing algorithm selects the single receiver from the group based on least-expensive routing metric. In practice, this means that packets are routed to the topologically-nearest member of an anycast group. Anycasting in the Internet architecture was first described in RFC 1546 [1]
• Geocast refers to the delivery of information to a group of destinations in a network identified by their geographical locations. It is a specialized form of multicast addressing used by some routing protocols for mobile ad hoc networks.

Anycast can be implemented by using Border Gateway Protocol (BGP). Multiple hosts (usually in different geographic areas) are given the same unicast IP address and different routes to the address are announced through BGP. Routers consider these to be alternative routes to the same destination, though they are actually routes to different destinations with the same address. As usual, routers select a route by whatever distance metric is in use (the least cost, least congested, shortest). Selecting a route in this setup amounts to selecting a destination.

The pitfall of this approach is that a connection to an anycast address may fail because the network can change the routing of packets in mid-connection due to congestion or changes in the network, with the result that the destination changes mid-connection, though the new destination is not aware of the connection and is not maintaining the connection state. These conditions are typically referred to as a "PoP switch". With a normal unicast address a routing change would not be a problem, as this merely results in a different route to the same eventual destination. All packets within a connection do not normally need to follow the same path. But when the address is actually an anycast address masquerading as a unicast address (as in this design), a routing change could be a destination change.

Because of the possibility of a "PoP switch" in IPv4, anycast is generally used with connectionless protocols based on UDP, as a way to provide high availability and load balancing for stateless services. For example, one of the well-known applications of IPv4 anycast is the Domain Name System (DNS). This is well-suited to anycast because it is a UDP-based service providing connectionless access to domain name data which is replicated across multiple, geographically-dispersed, stateless servers, with severe demands for availability and scalability.

Because anycast is more error-prone with connection-oriented protocols such as TCP where the destination maintains state, it is less commonly used with these protocols. Some custom IP stacks use proprietary methods to provide healing of stateful protocols when necessary,[2] mitigating problems due to anycast PoP switches. These methods usually involve some form of Network function virtualization[3] in form of packet rewriting as has been demonstrated by AWS HyperPlane,[4][5] and Google Maglev[6] and/or packet encapsulation as laid out by protocols like Generic Routing Encapsulation, IPOP.

Anycast is supported explicitly in IPv6. RFC 4291, which covers IPv6 addressing architecture, reserves Interface Identifier 0 within an IPv6 subnet as the "Subnet Router" anycast address. In addition, RFC 2526 reserves a block of 128 Interface Identifiers within a subnet as anycast addresses.

Most IPv6 routers on the path of an anycast packet through the network will not distinguish it from a unicast packet, but special handling is required from the routers near the destination (that is, within the scope of the anycast address) as they are required to route an anycast packet to the "nearest" interface within that scope which has the proper anycast address, according to whatever measure of distance (hops, cost, etc.) is being used.

The method used in IPv4 of advertising multiple routes in BGP to multiply-assigned unicast addresses also still works in IPv6, and can be used to route packets to the nearest of several geographically dispersed hosts with the same address. This approach, which does not depend on anycast-aware routers, has the same use-cases together with the same problems and limitations as in IPv4.

With the growth of the Internet, network services increasingly have high-availability requirements. As a result, operation of anycast services (RFC 4786) has grown in popularity among network operators.[7]

Anycast allows any operator whose routing information is accepted by an intermediate router to hijack any packets intended for the anycast address. While this at first sight appears insecure, it is no different from the routing of ordinary IP packets, and no more or less secure. As with conventional IP routing, careful filtering of who is and is not allowed to propagate route announcements is crucial to prevent man-in-the-middle or blackhole attacks. The former can also be prevented by encrypting and authenticating messages, such as using Transport Layer Security, while the latter can be frustrated by onion routing.

Anycast is normally highly reliable, as it can provide automatic failover. Anycast applications typically feature external "heartbeat" monitoring of the server's function, and withdraw the route announcement if the server fails. In some cases this is done by the actual servers announcing the anycast prefix to the router over OSPF or another IGP. If the servers die, the router will automatically withdraw the announcement.

"Heartbeat" functionality is important because, if the announcement continues for a failed server, the server will act as a "black hole" for nearby clients; this failure mode is the most serious mode of failure for an anycast system. Even in this event, this kind of failure will only cause a total failure for clients that are closer to this server than any other, and will not cause a global failure.

In denial-of-service attacks, a rogue network host may advertise itself as an anycast server for a vital network service, to provide false information or simply block service.

Anycast methodologies on the Internet may be exploited to distribute DDoS attacks and reduce their effectiveness: As traffic is routed to the closest node, a process over which the attacker has no control, the DDoS traffic flow will be distributed amongst the closest nodes. Thus, not all nodes might be affected. This may be a reason to deploy anycast addressing.[12]

The effectiveness of this technique to divert attacks is questionable, however, because unicast addresses (used for maintenance) can be easy to obtain, at least on IPv6. Now obsoleted RFC 2373 defined that "An anycast address must not be used as the source address of an IPv6 packet." Therefore, pinging an anycast address will return the unicast address of the closest node, since the reply must come from a unicast address. An attacker can then attack individual nodes from any location, bypassing anycast addressing methods. This same method works on some, but not all, IPv4 anycast addresses.[13] RFC 2373 also restricted anycast IPv6 addresses to routers only. However, both of these restrictions were lifted in RFC 4291.

Authentication of anycast transmissions may solve this problem.[14]

Some anycast deployments on the Internet distinguish between local and global nodes to benefit the local community, by addressing local nodes preferentially. An example is the Domain Name System. Local nodes are often announced with the no-export BGP community to prevent hosts from announcing them to their peers, i.e. the announcement is kept in the local area. Where both local and global nodes are deployed, the announcements from global nodes are often AS prepended (i.e. the AS is added a few more times) to make the path longer so that a local node announcement is preferred over a global node announcement.[15]

• Multihoming
• Line hunting, for an equivalent system for telephones

• Best Practices in IPv4 Anycast Routing Tutorial on anycast routing configuration.