by Douglas Comer, Purdue University
Traditional packet-processing systems use an approach known as demultiplexing to handle incoming packets (refer to  for details). When a packet arrives, protocol software uses the contents of a Type Field in a protocol header to decide how to process the payload in the packet. For example, the Type field in a frame is used to select a Layer 3 module to handle the frame, as Figure 1 illustrates.
Demultiplexing is repeated at each level of the protocol stack. For example, IPv6 uses the Next Header field to select the correct transport layer protocol module, as Figure 2 illustrates.
Modern, high-speed network systems take an entirely different view of packet processing. In place of demultiplexing, they use a technique known as classification . Instead of assuming that a packet proceeds through a protocol stack one layer at a time, they allow processing to cross layers. (In addition to being used by companies such as Cisco and Juniper, classification has been used in Linux  and with network processors by companies such as Intel and Netronome .)
Packet classification is especially pertinent to three key network technologies. First, Ethernet switches use classification instead of demultiplexing when they choose how to forward packets. Second, a router that sends incoming packets over Multiprotocol Label Switching (MPLS) tunnels uses classification to choose the appropriate tunnel. Third, classification provides the basis for Software-Defined Networking (SDN) and the OpenFlow protocol.
Motivation for Classification
To understand the motivation for classification, consider a network system that has protocol software arranged in a traditional layered stack. Packet processing relies on demultiplexing at each layer of the protocol stack. When a frame arrives, protocol software looks at the Type field to learn about the contents of the frame payload. If the frame carries an IP datagram, the payload is sent to the IP protocol module for processing. IP uses the destination address to select a next-hop address. If the datagram is in transit (that is, passing through the router on its way to a destination), IP forwards the datagram by sending it back out one of the interfaces. A datagram reaches TCP only if the datagram is destined for the router itself. TCP then uses the protocol port numbers in the TCP segment to further demultiplex the incoming datagram among multiple application programs.
To understand why traditional layering does not solve all problems, consider MPLS processing. In particular, consider a router at the border between a traditional internet and an MPLS core. Such a router must accept packets that arrive from the traditional internet and choose an MPLS path over which to send the packet. Why is layering pertinent to path selection? In many cases, network managers use transport layer protocol port numbers when choosing a path. For example, suppose a manager wants to send all web traffic down a specific MPLS path. All the web traffic will use TCP port 80, meaning that the selection must examine TCP port numbers.
Unfortunately, in a traditional demultiplexing scheme, a datagram does not reach the transport layer unless the datagram is destined for the local network system. Therefore, protocol software must be reorganized to handle MPLS path selection. We can summarize:
A traditional protocol stack is insufficient for the task of MPLS path selection because path selection often involves transport layer information and a traditional stack will not send transit datagrams to the transport layer.
Classification Instead of Demultiplexing
How should protocol software be structured to handle tasks such as MPLS path selection? The answer lies in the use of classification. A classification system differs from conventional demultiplexing in two ways:
To understand classification, imagine a packet that has been received at a router and placed in memory. Encapsulation means that the packet will have a set of contiguous protocol headers at the beginning. For example, Figure 3 illustrates the headers in a TCP packet (for example, a request sent to a web server) that has arrived over an Ethernet.
Given a packet in memory, how can we quickly determine whether the packet is destined to the web? A simplistic approach simply looks at one field in the headers: the TCP destination port number. However, it could be that the packet is not a TCP packet at all. Maybe the frame is carrying Address Resolution Protocol (ARP) data instead of IP. Or maybe the frame does indeed contain an IP datagram, but instead of TCP the transport layer protocol is the User Datagram Protocol (UDP). To make certain that it is destined for the web, software needs to verify each of the headers: the frame contains an IP datagram, the IP datagram contains a TCP segment, and the TCP segment is destined for the web.
Instead of parsing protocol headers, think of the packet as an array of octets in memory. Consider IPv4 as an example. To be an IPv4 datagram, the Ethernet Type field (located in array positions 12 and 13) must contain 0x0800. The IPv4 Protocol field, located at position 23, must contain 6 (the protocol number for TCP). The Destination Port field in the TCP header must contain 80. To know the exact position of the TCP header, we must know the size of the IP header. Therefore, we check the header length octet of the IPv4 header. If the octet contains 0x45, the TCP destination port number will be found in array positions 36 and 37.
As another example, consider classifying Voice over IP (VoIP) traffic that uses the Real-Time Transport Protocol (RTP). Because RTP is not assigned a specific UDP port, vendors use a heuristic to determine whether a given packet carries RTP traffic: check the Ethernet and IP headers to verify that the packet carries UDP, and then examine the octets at a known offset in the RTP packet to verify that the value matches the value used by a known codec.
Observe that all the checks described in the preceding paragraphs require only array lookup. That is, the lookup mechanism treats the packet as an array of octets and merely checks to verify that location X contains value Y, location Z contains value W, and so on—the mechanism does not need to understand any of the protocol headers or the meaning of values. Furthermore, observe that the lookup scheme crosses multiple layers of the protocol stack.
We use the term classifier to describe a mechanism that uses the lookup approach described previously, and we say that the result is a packet classification. In practice, a classification mechanism usually takes a list of classification rules and applies them until a match is found. For example, a manager might specify three rules: send all web traffic to MPLS path 1, send all FTP traffic to MPLS path 2, and send all VPN traffic to MPLS path 3.
Layering When Classification Is Used
If classification crosses protocol layers, how does it relate to traditional layering diagrams? We can think of classification as an extra layer that has been squeezed between Layer 2 and Layer 3. When a packet arrives, the packet passes from a Layer 2 module to the classification module. All packets proceed to the classifier; no demultiplexing occurs before classification. If any of the classification rules matches the packet, the classification layer follows the rule. Otherwise, the packet proceeds up the traditional protocol stack. For example, Figure 4 illustrates layering when classification is used to send some packets across MPLS paths.
Interestingly, a classification layer can subsume all demultiplexing. That is, instead of classifying packets only for MPLS paths, the classifier can be configured with additional rules that check the Type field in a frame for IPv4, IPv6, ARP, Reverse ARP (RARP), and so on.
Classification Hardware and Network Switches
The text in the previous section describes a classification mechanism that is implemented in software—an extra layer is added to a software protocol stack that classifies frames after they arrive at a router. Classification can also be implemented in hardware. In particular, Ethernet switches and other packet-processing hardware devices contain classification hardware that allows packet classification and forwarding to proceed at high speed. The next sections explain hardware classification mechanisms.
We think of network devices, such as switches, as being divided into broad categories by the level of protocol headers they examine and the consequent level of functions they provide:
A Layer 2 Switch examines the Media Access Control (MAC) source address in each incoming frame to learn the MAC address of the computer that is attached to each port. When a switch learns the MAC addresses of all the attached computers, the switch can use the destination MAC address in each frame to make a forwarding decision. If the frame is unicast, the switch sends only one copy of the frame on the port to which the specified computer is attached. For a frame destined to the broadcast or a multicast address, the switch delivers a copy of the frame to all ports.
A VLAN Switch adds one level of virtualization by permitting a manager to assign each port to a specific VLAN. Internally, VLAN switches extend forwarding in a minor way: instead of sending broadcasts and multicasts to all ports on the switch, a VLAN switch consults the VLAN configuration and sends them only to ports on the same VLAN as the source.
A Layer 3 Switch acts like a combination of a VLAN switch and a router. Instead of using only the Ethernet header when forwarding a frame, the switch can look at fields in the IP header. In particular, the switch watches the source IP address in incoming packets to learn the IP address of the computer attached to each switch port. The switch can then use the IP destination address in a packet to forward the packet to its correct destination.
A Layer 4 Device extends the examination of a packet to the transport layer. That is, the device can include the TCP or UDP Source and Destination Port fields when making a forwarding decision.
Switching Decisions and VLAN Tags
All types of switching hardware described previously use classification. That is, switches operate on packets as if a packet is merely an array of octets, and individual fields in the packet are specified by giving offsets in the array. Thus, instead of demultiplexing packets, a switch treats a packet syntactically by applying a set of classification rules similar to the rules described previously.
Surprisingly, even VLAN processing is handled in a syntactic manner. Instead of merely keeping VLAN information in a separate data structure that holds meta information, the switch inserts an extra field in an incoming packet and places the VLAN number of the packet in the extra field. Because it is just another field, the classifier can reference the VLAN number just like any other header field.
We use the term VLAN Tag to refer to the extra field inserted in a packet. The tag contains the VLAN number that the manager assigned to the port over which the frame arrived. For Ethernet, IEEE standard 802.1Q specifies placing the VLAN Tag field after the MAC Source Address field. Figure 5 illustrates the format.
A VLAN tag is used only internally—after the switch has selected an output port and is ready to transmit the frame, the tag is removed. Thus, when computers send and receive frames, the frames do not contain a VLAN tag.
An exception can be made to the rule: a manager can configure one or more ports on a switch to leave VLAN tags in frames when sending the frame. The purpose is to allow two or more switches to be configured to operate as a single, large switch. That is, the switches can share a set of VLANs—a manager can configure each VLAN to include ports on one or both of the switches.
We can think of hardware in a switch as being divided into three main components: a classifier, a set of units that perform actions, and a management component that controls the overall operation. Figure 6 illustrates the overall organization and the flow of packets.
As black arrows in the figure indicate, the classifier provides the high-speed data path that packets follow. When a packet arrives, the classifier uses the rules that have been configured to choose an action. The management module usually consists of a general-purpose pro-cessor that runs management software. A network administrator can interact with the management module to configure the switch, in which case the management module can create or modify the set of rules the classifier follows.
A network system, such as a switch, must be able to handle two types of traffic: transit traffic and traffic destined for the switch itself. For example, to provide management or routing functions, a switch may have a local TCP/IP protocol stack and packets destined for the switch must be passed to the local stack. Therefore, one of the actions a classifier takes may be "pass packet to the local stack for Demultiplexing".
High-Speed Classification and TCAM
Modern switches can allow each interface to operate at 10 Gbps. At 10 Gbps, a frame takes only 1.2 microseconds to arrive, and a switch usually has many interfaces. A conventional processor cannot handle classification at such speeds, so a question arises: how can a hardware classifier achieve high speed? The answer lies in a hardware technology known as Ternary Content Addressable Memory (TCAM).
TCAM uses parallelism to achieve high speed—instead of testing one field of a packet at a given time, TCAM checks all fields simultaneously. Furthermore, TCAM performs multiple checks at the same time. To understand how TCAM works, think of a packet as a string of bits. We imagine TCAM hardware as having two parts: one part holds the bits from a packet and the other part is an array of values that will be compared to the packet. Entries in the array are known as slots. Figure 7 illustrates the idea.
In the figure, each slot contains two parts. The first part consists of hardware that compares the bits from the packet to the pattern stored in the slot. The second part stores a value that specifies an action to be taken if the pattern matches the packet. If a match occurs, the slot hardware passes the action to the component that checks all the results and announces an answer.
One of the most important details concerns the way TCAM handles multiple matches. In essence, the output circuitry selects one match and ignores the others. That is, if multiple slots each pass an action to the output circuit, the circuit accepts only one and passes the action as the output of the classification. For example, the hardware may choose the lowest slot that matches. In any case, the action that the TCAM announces corresponds to the action from one of the matching slots.
The figure indicates that a slot holds a pattern rather than an exact value. Instead of merely comparing each bit in the pattern to the corresponding bit in the packet, the hardware performs a pattern match. The adjective ternary is used because each bit position in a pattern can have three possible values: a one, a zero, or a "don't care". When a slot compares its pattern to the packet, the hardware checks only the one and zero bits in the pattern—the hardware ignores pattern bits that contain "don't care". Thus, a pattern can specify exact values for some fields in a packet header and omit other fields.
To understand TCAM pattern matching, consider a pattern that identifies IP packets. Identifying such packets is easy because an Ethernet frame that carries an IPv4 datagram will have the value 0x0800 in the Ethernet Type field. Furthermore, the Type field occupies a fixed position in the frame: bits 96 through 111. Thus, we can create a pattern that starts with 96 "don't care" bits (to cover the Ethernet destination and source MAC addresses) followed by 16 bits with the binary value 0000100000000000 (the binary equivalent of 0x0800) to cover the Type field. All remaining bit positions in the pattern will be "don't care". Figure 8 illustrates the pattern and example packets.
Although a TCAM hardware slot has one position for each bit, the figure does not display individual bits. Instead, each box corresponds to one octet, and the value in a box is a hexadecimal value that corresponds to 8 bits. We use hexadecimal simply because binary strings are too long to fit into a figure comfortably.
The Size of a TCAM
A question arises: how large is a TCAM? The question can be divided into two important aspects:
A switch can also use patterns to control broadcasting. When a manager configures a VLAN, the switch can add an entry for the VLAN broadcast. For example, if a manager configures VLAN 9, an entry can be added in which the destination address bits are all 1s (that is, the Ethernet broadcast address) and the VLAN tag is 9. The action associated with the entry is "broadcast on VLAN 9".
A Layer 3 switch can learn the IP source address of computers attached to the switch, and can use TCAM to store an entry for each IP address. Similarly, it is possible to create entries that match Layer 4 protocol port numbers (for example, to direct all web traffic to a specific output). SDN technologies allow a manager to place patterns in the classifier to establish paths through a network and direct traffic along the paths. Because such classification rules cross multiple layers of the protocol stack, the potential number of items stored in a TCAM can be large.
TCAM seems like an ideal mechanism because it is both extremely fast and versatile. However, TCAM has two significant drawbacks: cost and heat. The cost is high because TCAM has parallel hardware for each slot and the overall system is designed to operate at high speed. In addition, because it operates in parallel, TCAM consumes much more energy than conventional memory (and generates more heat). Therefore, designers minimize the amount of TCAM to keep costs and power consumption low. A typical switch has 32,000 entries.
Classification-Enabled Generalized Forwarding
Perhaps the most significant advantage of a classification mechanism arises from the generalizations it enables. Because classification examines arbitrary fields in a packet before any demultiplexing occurs, cross-layer combinations are possible. For example, classification can specify that all packets from a given MAC address should be forwarded to a specific output port regardless of the packet contents. In addition, classification can make forwarding decisions depend on combinations of source and destination. An Internet Service Provider (ISP) can choose to forward all packets with IP source address X that are destined for web server W along one path while forwarding packets with IP source address Y that are destined to the same web server along another path.
ISPs need the generality that classification offers to handle traffic engineering that is not usually available in a conventional protocol stack. In particular, classification allows an ISP to offer tiered services in which the path a packet follows depends on a combination of the type of traffic and how much the customer pays.
Classification is a fundamental performance optimization that allows a packet-processing system to cross layers of the protocol stack without demultiplexing. A classifier treats each packet as an array of bits and checks the contents of fields at specific locations in the array.
Classification offers high-speed forwarding for network systems such as Ethernet switches and routers that send packets across MPLS tunnels. To achieve the highest speed, classification can be implemented in hardware; a hardware technology known as TCAM is especially useful because it employs parallelism to perform classification at extremely high speed.
The generalized forwarding capabilities that classification provides allow ISPs to perform traffic engineering. When making a forwarding decision, a classification mechanism can use the source of a packet as well as the destination (for example, to choose a path based on the tier of service to which a customer subscribes).
Material in this article has been taken with permission from Douglas E. Comer, Internetworking With TCP/IP Volume 1: Principles, Protocols, and Architecture, Sixth edition, 2013.