Packet classification

Routers have evolved from traditional destination-based forwarding devices to what are called packet classification routers. In modern routers, the route and resources allocated to a packet are determined by the destination address as well as other header fields of the packet, such as the source address and TCP/UDP port numbers. Packet classification unifies the forwarding functions required by firewalls, resource reservations, QoS routing, unicast routing, and multicast routing. In classification, the forwarding database of a router consists of a potentially large number of rules on key header fields. A given packet header can match multiple rules. So each rule is given a cost, and the packet is forwarded using the least-cost matching rule. The world has changed significantly since the first edition as follows, but most of the relevant changes are a subset of the changes described at the start of the last chapter on IP lookups. The most relevant changes are the significant use of IPv6 (which complicates packet classification), the increasing use of Software Defined Networks (SDN) and hypervisor switches to do flexible forwarding using packet classification instead of simpler IP lookups, and the emergence of Network Function Virtualization (NFV) which requires software solutions to packet classification. The packet classification problem is motivated in Section 12.1. The classification problem is formulated precisely in Section 12.2, and the metrics used to evaluate rule schemes are described in Section 12.3. Section 12.4 presents simple schemes such as linear search and CAMs. Section 12.5 begins the discussion of more efficient schemes by describing an efficient scheme called grid of tries that works only for rules specifying values of only two fields. Section 12.6 transitions to general rule sets by describing a set of insights into the classification problem, including the use of a geometric viewpoint. Section 12.7 begins the transition to algorithms for the general case with a simple idea to extend 2D schemes. A general approach based on divide-and-conquer is described in Section 12.8. This is followed by three very different examples of algorithms based on divide-and-conquer: simple and aggregated bit vector linear search (Section 12.9), cross-producting (Section 12.10), and RFC, or equivalenced cross-producting (Section 12.11). Section 12.12 presents the most promising of the current algorithmic approaches, an approach based on decision trees. This chapter will continue to exhibit the set of principles introduced in Chapter 3, as well as illustrate three general problem-solving strategies: solving simpler problems first before solving a complex problem, collecting different viewpoints, and exploiting the structure of input data sets.