Geographic Distribution of Critical Systems

As seen in Figure 4-6, four servers are deployed within a large network and each is assigned the same anycast IP address. In addition, each server should have a second interface with a unicast address to be used for administration, and possibly other communication (discussed below). A client establishes a connection to an application, which runs on all four servers. Routing protocols deliver the packets between client and server based on a predetermined "metric," such as cost or distance (see Chapter 3 for details on how anycast works with respect to routing protocols).

When the "closest" server is unavailable, routing protocols determine the next "closest" server, and the client then communicates with that server. The client is still using the same anycast address, but is actually communicating with a different server. This mechanism generally provides higher reliability and the client never knows (nor should it care) which server it is communicating with, as long as the application works.

We should note that anycast may be used for almost any application, but generally it works much better with UDP-based applications than TCP-based applications. TCP is a connection-oriented (stateful) protocol and there is generally no way to keep state between the anycast servers such that the sessions from a failed server could be seamlessly transferred to the other anycast servers. If an anycast server fails, or the path to the server changes such that a different server is now the closest, the client would receive a TCP RST (reset), causing it to reestablish the session with the new server.

One mechanism to mitigate the problem with TCP-based applications works with protocols such as LDAP and HTTP (or any protocol supporting application-layer redirection). The client sends the initial TCP-SYN packet to the server, and the server can respond with an application-layer redirect to the unicast address (remember the note about a second interface with a unicast address?). Since this address is unique, a path change would have no effect on TCP, but of course, a server failure would still cause the client to reestablish the connection to a new server.

UDP is a connectionless protocol, so if an anycast server fails, or the path to the server changes such that a different server is now the closest, the client will probably never see any change in communication, or latency, as it would with TCP-based sessions.

Several large service providers utilize anycast, primarily for DNS services. These include Akamai, UltraDNS, and the volunteer root DNS servers (C, F, I, J, K, and M root).

While anycast can provide a high degree of reliability for systems, denial-of-service attacks against an anycast infrastructure may still disrupt service. Until the recent emergence of large botnets , the resources required to sustain a distributed denial-of-service attack against a large anycast infrastructure was beyond the reach of all but the most sophisticated attackers . However, these attacks are becoming more common and are happening with more frequency.