Router

A router is a networking device that forwards data packets between networks. For each packet it reads the destination IP address and picks an outgoing interface and a next hop that move the packet closer to its goal. Routers are what join separate local networks into the internet, which runs on the TCP/IP protocol suite, and they sit at nearly every boundary where traffic leaves a local network.

The device matters because modern communication depends on it. Without routers, each network would be isolated, and no single host could reach another across organizational or geographic boundaries. The technology grew out of research networks in the late 1970s and early 1980s and became a commercial product category within a few years. Today the same basic function runs on devices ranging from home Wi-Fi boxes to carrier backbone systems that hold hundreds of thousands of routes. Its ideas also appear in other fields, from graph theory-based path finding to the gating mechanisms of neural networks.

This entry first defines the router and explains how it forwards packets through its routing table. It then covers the protocols that build those tables, the internal hardware that performs forwarding at high speed, and the main device classes. Next it reviews the additional services routers commonly provide, the history of their development, their standards and security issues, and the shift toward software-defined networking. It ends with routing in other technical fields and the etymology of the word.

Definition and role in networks

A router is a device that forwards data packets between networks. What sets it apart from other network equipment is the layer at which it works. A router operates at the network layer (Layer 3 of the OSI reference model), where it reads and acts on IP addresses. A Ethernet switch works one layer lower. It forwards frames by the hardware (MAC) address of each device, and it does so only within a single local segment. A switch keeps a table that maps MAC addresses to ports, and it does not interpret the IP header inside the frame. A router does interpret that header, which lets it join networks that use different address ranges and even different link technologies. The Internet itself is built on this arrangement, using the addressing scheme of TCP/IP.

The router as a gateway

Hosts rarely reach distant networks on their own. A host compares a destination IP address with its own subnet. If the destination lies outside that subnet, the host does not try to find that destination on the local link. It sends the packet to its default gateway, which is the address of a router interface on the local network. The host only needs the link-layer address of that gateway, which it obtains through ARP, and then hands over the packet. From that point the router makes the forwarding decision for the host, selecting the next hop toward the destination. For most end systems, the whole outside world is therefore reached through one configured address.

Broadcast domains

A broadcast domain is the set of devices that receive a link-layer broadcast frame, which is sent to the all-ones MAC address. Ordinary switches flood broadcasts out every port in the same VLAN, so a switched LAN without VLANs forms one large broadcast domain. Routers do not pass broadcast frames from one interface to another. Each router interface therefore marks the edge of its own broadcast domain. The consequences are practical. Broadcasts such as ARP requests and DHCP discovery messages stay within their segment, so they consume less bandwidth on the rest of the network. A broadcast storm or a misbehaving device in one segment is also less likely to affect every other host. Administrators gain clear boundaries for fault isolation and access policy. Layer 3 switches perform the same separation, and VLANs divide broadcast domains at Layer 2, but only routing connects the resulting segments again. A broadcast domain is also not the same thing as an IP subnet, although in common designs one routed VLAN carries one subnet, so the two often coincide.

Router versus routing

The word "router" names a device or a piece of software, while "routing" names the process of choosing a path for packets. Routing is the decision of which neighbor or interface should receive a packet on its way to a destination. A dedicated router is one device that carries out this process, but it is not the only one. General-purpose computers and servers can route traffic when they are configured to do so, and many home devices combine routing with other functions. The process itself depends on tables and protocols that decide which path is best, and those are covered in the sections on forwarding and routing protocols. In this section, the term "router" refers to the device that performs routing at Layer 3 in a network.

Forwarding and the routing table

For every packet, a router runs the same short sequence. It strips the link-layer framing, reads the destination IP address from the network-layer header, and performs a longest-prefix match against its forwarding table. Among all entries whose prefix covers the destination, the entry with the most leading bits in common wins. That entry supplies an outgoing interface and a next-hop address, or, for a directly attached network, the destination host itself. Before transmission the router decrements the time-to-live (TTL) field in the IPv4 header. If the value reaches zero, the packet is discarded, and the router normally returns an ICMP error to the sender. The header checksum is then recalculated, and the destination MAC address is replaced with that of the next hop, so the frame can cross the next segment toward the internet.

Routes in the table come from four kinds of source. Directly connected networks are added automatically from the addresses configured on each interface. Static routes are entered by an administrator; they are predictable but do not react to failures. Dynamically learned routes are installed by routing protocols such as RIP, OSPF, or BGP, which are covered elsewhere in this article. The default route, written 0.0.0.0/0, has a prefix length of zero and therefore matches every address. It is the least specific entry and is consulted only when nothing else matches. A small office, for example, often needs little more than a default route pointing to its upstream provider.

A small worked example shows how these entries interact. Suppose a router holds the following table:

A packet addressed to 10.1.2.57 matches all four entries, because each prefix covers that address. The /24 entry has the most leading bits in common with the destination, so the router sends the packet out eth2 without a further hop. A packet addressed to 10.9.4.4 matches only the /8 entry and the default route. The longer match is the /8, so the packet leaves through eth0 toward 192.168.0.1. The example also shows why the default route is a fallback rather than a competitor: it only wins when no more specific entry applies.

Table size varies enormously with the position of the router. A branch or campus router typically holds from a few entries to a few thousand, many of them static or learned from an internal protocol, and software can search such a table without difficulty. Backbone routers in the default-free zone are in another class. One APNIC measurement set showed about 1,050,000 IPv4 prefixes in the table at the start of 2026, while the CIDR Report counted 979,075 IPv4 prefixes in April 2025. Most of these entries are small blocks: /24, /23 and /22 prefixes together made up 84% of the IPv4 table at the end of 2025. IPv6 adds hundreds of thousands more. The growth crossed a practical threshold in August 2014, when the IPv4 table passed 512,000 entries and some older equipment, whose lookup memory was sized for that limit, ran into trouble. Table size is therefore a hardware constraint as well as a bookkeeping figure.

Routing protocols and path selection

A routing protocol is a set of rules that lets routers share reachability information and work out which path each destination prefix should take. The first design decision is whether routes are entered by hand or learned automatically.

Static and dynamic routing

Static routes are entered by an administrator. They are predictable, easy to audit, and cost almost no processing or bandwidth. However, every change has to be made manually, and a static route has no way to notice that its next hop has failed. Dynamic routing lets routers exchange information and recompute paths when the topology changes, which suits networks with redundant links. The cost is protocol traffic, more complex configuration, and the risk that a faulty update spreads automatically. Most real networks combine the two, using static routes for stub sites and default routes, and dynamic protocols inside and between larger domains.

Distance-vector protocols

Distance-vector protocols choose paths by a metric, usually hop count, and each router sends its routing table to its directly connected neighbors. RIP is the classic example. It sends full table updates every 30 seconds by default, and it computes routes with the Bellman-Ford algorithm, a dynamic programming method. A router knows only what its neighbors report, not the full topology. RIP limits the metric to 15 hops and treats a destination at 16 hops as unreachable. This bound stops endlessly looping routes from growing without limit, but it also restricts RIP to small networks. Split horizon, which forbids advertising a route back out the interface where it was learned, and poison reverse, which advertises failed routes with an infinite metric, reduce loops. Even so, failure news travels one neighbor at a time, so convergence is slow, and the count-to-infinity problem can still occur in some topologies.

Link-state protocols

Link-state protocols, such as OSPF and IS-IS, take a different approach. Each router advertises the state of its own links, and these advertisements are flooded through the area until every router holds an identical link-state database, which is a map of the topology. Each router then runs a shortest-path computation on this map, typically Dijkstra's algorithm, to build its own tree of best routes. Because every router computes from the same picture, convergence is fast and loops are rare. The price is higher CPU and memory use, along with a database that grows large in complex networks. OSPF derives its default link cost from interface bandwidth, so faster links are preferred automatically.

Path-vector protocols and BGP

BGP is a path-vector protocol used between autonomous systems across the Internet. Instead of a single number, routes carry the sequence of autonomous systems they have crossed. A router rejects any path that already contains its own AS, which prevents loops. BGP's path choice is driven mainly by administrative policy, such as local preference values, import and export filters, and attributes set by operators, rather than by a simple shortest-path calculation. For this reason, BGP is called an exterior gateway protocol, while RIP, OSPF, and IS-IS are interior gateway protocols that work within a single AS.

Metrics, administrative distance, and convergence

Each protocol uses its own metric, such as hop count for RIP, bandwidth-derived cost for OSPF, and policy attributes for BGP. These values cannot be compared across protocols, so routers use administrative distance (AD), a local trust ranking. On Cisco devices the default values are 0 for connected routes, 1 for static routes, 20 for external BGP, 90 for internal EIGRP, 110 for OSPF, 115 for IS-IS, 120 for RIP, and 200 for internal BGP. When several sources offer a route to the same prefix, the lowest AD wins, and a value of 255 means the route is not installed. AD is never advertised to other routers, and vendors such as Juniper call the same concept "preference".

Convergence is the time the network needs to agree on a consistent set of routes after a change. During this interval, routers may hold conflicting information, which can produce temporary loops or packets sent into black holes. Protocols reduce these risks with techniques such as split horizon, hold-down timers, triggered updates, and fast link-state flooding, and the choice between them is largely a trade-off between speed of convergence and resource use.

Internal architecture and forwarding hardware

A router is usually organized into two cooperating parts, or planes. The control plane runs the routing protocols and maintains the routing information base (RIB), which can hold several candidate paths to the same destination along with the administrative distance of each source. The data plane works from the forwarding information base (FIB), a reduced copy of the RIB that keeps only the selected route for each prefix and is arranged for fast lookups. Route changes are much rarer than packet arrivals, so the control plane can tolerate slower computation, while the data plane must keep pace with traffic. Where the FIB is built and stored depends on the vendor: in many chassis the route processor computes it and pushes it to the line cards, and some designs keep additional copies, such as a separate control-plane FIB.

The hardware of a modular router has four main parts. Line cards carry the physical ports, typically four, eight or sixteen per card. On an ingress line card, a packet passes through layer 2 processing, and a network processor performs the destination lookup to decide which egress card should carry it. Ingress features such as filtering, policing and header rewriting are applied at this stage, and the packet is then held in a buffer until the switching fabric is ready to carry it. The switching fabric connects all line cards to one another. Large systems use crossbar or multistage designs; the Cisco Carrier Routing System, for example, uses a cell-switched, three-stage Benes fabric split into several independent planes. The route processor runs the protocols, builds the tables, distributes them to the line cards, and handles diagnostics and card management. Packet buffers and output queues complete the picture, since the fabric and the egress ports are the places where congestion first appears.

Forwarding lookups can run in software or in hardware. Early routers had a central CPU consult a table held in RAM, which was flexible and cheap but slow and a constant drain on the processor. Modern software forwarding on general-purpose processors relies on compact, cache-aware structures such as tries; its advocates note that software is easier to patch than silicon, while hardware cost, power draw and vendor concentration push prices up. High-end routers generally forward in hardware, using application-specific integrated circuits (ASICs) and ternary content-addressable memory (TCAM). A TCAM holds prioritized key-value entries and can compare a packet against an entire list in a single lookup. Separate TCAMs are often used for the FIB, access control lists and quality-of-service rules. The drawbacks are real: TCAM is expensive, a single high-density device can draw 12 to 15 watts when all entries are searched, capacity is finite, and updates are costly because the entries must stay in priority order.

Performance is measured in several ways. Throughput is expressed in packets per second as well as bits per second, because small packets stress the lookup engine far more than large ones. Latency is the time a packet spends crossing the device, and jitter is the variation in that delay, which matters greatly for voice and video. When the arrival rate on an output port exceeds what the link can carry, packets accumulate in queues. Queue management determines how this congestion is handled: a simple policy drops packets only when buffers are full, while active schemes discard packets earlier and in a controlled way to keep queues short. Buffer depth is a trade-off, since deep buffers reduce loss but add latency and jitter. Achieving line rate across all ports at once is therefore a matter of ASIC design, memory choice and scheduling together.

Device classes and deployments

Routers come in several forms, and the boundaries between them are blurred by price and by the features vendors bundle. At the small end, a wireless router or Wi-Fi router is a device that performs the functions of a router and also includes the functions of a wireless access point. Household products also usually add a switch and a firewall to the routing function, since consumer routers are almost always a combination of both the routing and switching components with a firewall added in for good measure. The modem supplied by the ISP is normally a separate device. Its job is the link to the provider, and it is only suited to linking a single computer to the internet. Address translation is what lets a household work with that limit: the ISP usually provides one public address, and the router lets every laptop, television and phone share it.

Enterprise, campus and branch routers occupy the middle tier, where the typical problem is connecting sites that are far apart. A branch can either continue to use its ISP and use a site to site VPN to communicate with the main office, or it can use a private circuit, such as a leased line or Metro Ethernet, so that it becomes an extension of the main network without using Internet facilities. A common design pairs a leased line with a backup VPN. The VPN is kept idle until the private path fails, and the two routers connected to the leased line must be configured to use dynamic routing (OSPF, BGP, or RIP). Vendors describe the failover as a model of how you could set up your network to automatically fail over to a branch office VPN when the primary link goes down.

Provider edge and core routers are large modular systems built for carrier networks. Cisco's Carrier Routing System is a modular and distributed core router that is generally positioned in the core or edge of a service provider network. Capacity is the main differentiator: a standalone CRS-3 system can handle 2.2 Tbit/s and a multi-chassis system could be designed to handle 322 Tbit/s. Juniper's MX2020, another chassis platform, is listed at 80 Tbps throughput. The roles are also split by position. A provider edge (PE) router faces the customer's router, while the P router which is a provider router that is located in the core of the provider's network.

Specialized and virtual forms cover a wider range. Mid-range carrier products come as compact 1U pizza box form factors as well as chassis with hot-swappable modules and built-in redundancy. Some white-box hardware is rated for harsh outdoor conditions. For example, one unit is able to operate at a wide operational temperature range from -40˚C to 65˚C, which suits remote cabinets. Open designs also separate the hardware from its software. Hardware can be loaded with the Open Network Install Environment (ONIE), which supports the installation of compatible Network Operating System (NOS) software, so the same box can run different network software. Cellular routers, which use mobile networks as the uplink, industrial routers, and wireless mesh nodes that relay traffic to one another fill niches not covered above. Routing software that runs on ordinary servers or in cloud platforms is the fully virtual end of the spectrum.

Additional network functions

Beyond choosing paths, most routers sold for homes, offices and branches also provide services at the boundary between networks. The most consequential is network address translation (NAT). The router rewrites the addresses in outgoing packets so that many private hosts can share one public address, and it keeps a translation table so that replies can be returned to the correct internal host. Port address translation (PAT, also called NAPT) extends this by using port numbers to tell sessions apart, so a single public IPv4 address can serve a whole household or company. NAT was never a permanent fix for the exhaustion of IPv4 addresses. It breaks the end-to-end model of the internet, and each public address has a finite number of ports, so heavy traffic can exhaust them. Inbound connections that match no existing mapping are dropped, which complicates peer-to-peer applications.

Consumer and small-office routers typically bundle several further services. A built-in DHCP server assigns addresses, default gateway settings and DNS server addresses to client devices. Many routers also forward DNS queries to an upstream resolver or answer them from a local cache. Packet filtering comes in two main forms. Access control lists (ACLs) are stateless: they match header fields such as addresses, protocol and port numbers, and permit or deny each packet on its own, so return traffic must be explicitly allowed. A stateful firewall instead tracks the state of each connection and admits replies to sessions that the inside network initiated. Many modern routers combine both approaches, and filtering is the basis of the broader cybersecurity measures discussed elsewhere in this article.

Quality of service (QoS) addresses congestion, when the offered traffic exceeds the capacity of an outgoing link. A router classifies packets, marks them, places them into queues and schedules transmission, so that latency-sensitive traffic is served first. Voice traffic is a common example: packets can be marked with a differentiated services code point and moved to the front of a priority queue, which reduces delay and jitter during bursts of bulk downloads. Traffic shaping is related but different. It limits the rate at which traffic leaves an interface, for example to match the speed of a remote link or to enforce a bandwidth policy, and it typically buffers excess packets rather than discarding them. QoS is effective only where the router controls the queue. It cannot manage congestion inside another operator's network, and the configuration of consumer devices is often simplified to a few named application categories.

A router can also act as one endpoint of an encrypted virtual-private-network (VPN) tunnel. In this arrangement the device encapsulates and encrypts traffic bound for a remote site, such as a branch office, and decrypts traffic arriving from it, so that the two networks behave as if they were directly connected across the public internet. Encryption and key negotiation consume processor resources, so VPN throughput is an important specification for branch and enterprise models, and many devices offload this work to dedicated hardware. Some deployments also handle multicast traffic, in which one sender delivers the same packets to a group of receivers. On the local network, hosts signal their membership in groups using protocols such as IGMP, and the router uses multicast routing protocols to forward those streams between subnets. Multicast is used in some video distribution and streaming setups, though many networks restrict it to avoid unnecessary load.

History and commercialization

The commercial router grew out of a campus networking problem. In early 1980 Xerox gave Stanford, MIT and Carnegie Mellon access to its Ethernet technology, and Stanford's researchers quickly needed a way to join separate Ethernet segments and link them to the national internet. Xerox Corporation granted Stanford (along with MIT and Carnegie Mellon University) access to their Ethernet technology. The Ethernet design (by Metcalfe and Boggs) was a much more general solution and was adopted eagerly. This immediately led to the need for what became known as "router technology". The task went to Bill Yeager, a software engineer working on the SUMEX-AIM project, who had to write an operating system from scratch. His machine was a DEC PDP-11/05 with 56 kilobytes of memory, which forced very compact code. The resulting system is documented as an early software-driven router that supported both PUP and later IP routing. A second version, built on a board designed by Andreas Bechtolsheim, routed PUP, XNS, Chaosnet, and IP and directly influenced later hardware-based commercial designs.

Cisco Systems and the origin dispute

Cisco Systems was incorporated in December 1984 in Menlo Park, California. Its founders, Leonard Bosack and Sandra Lerner, were a married couple who ran computing groups at Stanford: Bosack was the manager of the computer science department's laboratory, and Lerner oversaw the computers at the graduate school of business. Their first practical work linked two campus LANs, 500 yards across campus. Stanford University later tried to obtain $11 million in licensing fees from the new company, according to one account. The dispute was settled in 1987, when Stanford licensed the router software and two computer boards to Cisco. Cisco shipped its first product, a router for the TCP/IP protocol suite, in 1986.

How much of the commercial product came from Stanford work, and from whom, is disputed. Yeager's supporters point to his code: one biographical source states that the code was licensed by upstart Cisco Systems in 1987 and comprised the core of the first Cisco IOS. Lerner's own account says that they later adapted software used at Stanford into what would become the foundation for Cisco IOS. A Stanford staff perspective, written by Tom Rindfleisch, stresses that all of this work was done by Yeager himself in the SUMEX laboratory and involved no contact with Bosack or Lerner. It also argues that by time Bosack had access to the router source code in 1985, multiple-protocol routers were a relatively mature technology. Bosack and Lerner's departure is also documented: in 1990, Lerner was fired and Bosack quit. The company went public the same year, and by the end of the dot-com bubble in 2000, had a market capitalization of $500 billion, before losing much of that value after 2000.

Standardization in the 1980s

The 1980s set the technical framework that routers later implemented. Ethernet was standardized within the IEEE 802 committee work, and the OSI reference model was published in 1984 as a layered description of network functions. Protocol choices were also shifting: Internet protocols were rolled out in 1982 to replace the ARPANET, and the military portion of the ARPANET was integrated into the Defense Data Network in the early 1980s, and the civilian ARPANET was taken out of service in 1990, by which time NSFNET had become the main national backbone. Routing protocols also changed. In the late 1980s, NSF presided over another seminal design choice: the switch to a routing protocol that replaced EGP with BGP. The older protocol presumed a known pathway for connecting systems, whereas BGP enables fully decentralized routing.

Commercial growth in the 1990s

Demand for routers rose sharply as the internet expanded beyond research use. NSFNET traffic illustrates the pace: NSFNET usage grew dramatically, jumping from 85 million packets in January 1988 to 37 billion packets in September 1993. In 1992 the National Science Foundation announced its intention to phase out federal support for the Internet backbone, and encouraged commercial entities to set up private backbones. Federal support for the NSFNET backbone ended on April 30, 1995, and the NSF had earlier funded exchange points in Northern California, Chicago, and New York, where backbone providers could exchange traffic. By the early 1990s a router industry already existed, as there already was an industry building routers consistent with widely employed software protocols.

Exterior routing grew with the internet. BGP version 4 was published in July of 1994, with the first release under RFC 1654. The number of autonomous systems also rose: in the early 1990s the Internet was composed of just a few hundred different Autonomous Systems (ASes) inter-connected on a customer-provider basis, while nowadays there are more than 70,000. Growth at that scale is one reason later sections of this article deal with routing table size, convergence and security.

Standards, interoperability and security

Standards bodies and protocol specifications

Router behavior is defined mostly by open specifications. The IETF publishes these as Requests for Comments (RFCs). The primary function of a BGP speaking system is to exchange network reachability information with other BGP systems. That behavior is set out in RFC 4271, a Standards Track document dated January 2006. OSPF is a link-state routing protocol. Its version 2, specified in RFC 2328 (April 1998), covers IPv4, and IPv6 has its own routing and addressing RFCs. Link-layer behavior for Ethernet ports is a separate matter, handled by the IEEE 802.3 family of standards rather than by the IETF.

Interoperability and vendor lock-in

Because the protocols are published openly, equipment from different manufacturers can usually exchange routes with one another. Vendor documentation makes this visible: a Cisco guide lists the RFCs its platform implements, including the IETF RFCs supported in Cisco NX-OS, and these lists are largely the same ones other vendors support. Operators can therefore mix equipment in one network and avoid tying their whole infrastructure to a single supplier. Proprietary protocols still appear in some sites, but according to a French commentary on RFC 2328, OSPF is by far the most widespread interior gateway protocol, even though RIP remains in use in some places and IS-IS has also been deployed.

Routing security: BGP hijacking and RPKI

The most serious routing-level threat concerns the Internet's exterior routing. BGP provides no native mechanism to authenticate the origin of routing announcements, so any AS can falsely claim to originate a prefix it does not legitimately control, which diverts traffic to its own network. Such prefix hijacking incidents occur regularly. A related technique uses a more specific prefix: by announcing a more-specific prefix, an attacker attracts traffic without removing the legitimate route from the table. Accidental misconfigurations are also common, and the distinction between a deliberate hijack and an unintentional route leak is important in incident analysis.

The main countermeasure is the Resource Public Key Infrastructure. The IETF standardized RPKI to bind IP prefixes to their legitimate origin ASes through cryptographically signed Route Origin Authorizations (ROAs). Border routers then perform Route Origin Validation, and classify each route learned in BGP into one of three states: valid, invalid or unknown. Routes that fail validation can be dropped. The protection has clear limits. ROV checks only the origin and not the rest of the path, so it does not prove that every AS in the path is trustworthy, or that the path is free from a route leak. Path-validation extensions such as BGPsec would close this gap, but BGPsec remains essentially undeployed due to performance overhead and operational complexity. Adoption is also uneven: ROA adoption has grown rapidly in recent years, while ROV deployment has advanced much more slowly.

Device-level risks

Many security problems sit in the devices themselves rather than in the protocols. Default administrator credentials are a persistent weakness. Router defaults often follow predictable patterns, and these default passwords are also weak and usually have fixed rules, such as admin/admin or root/root. A 2025 survey of 3,242 users reported that 81% haven't changed their router's default admin password. Botnets exploit this habit: the Mirai malware family exploits weak default credentials and vulnerabilities to compromise devices, granting attackers remote access. In December 2024, Juniper warned that Session Smart Router products with default passwords were being targeted as part of a malicious campaign that deploys the Mirai botnet malware.

Delayed or absent firmware maintenance compounds the problem. In May 2016, a worm exploited vulnerabilities in Ubiquiti Network routers running outdated firmware, using default login credentials to gain access, and self-propagating to other vulnerable devices. Some vendors never fix their flaws: Quanta decided not to patch the vulnerabilities. A large analysis of embedded firmware found that the authors were able to extract 35,000 RSA private keys, weak password hashes, and hardcoded credentials from images for printers, routers and cameras. Exposed management interfaces widen the attack surface; for example, a Huawei flaw (CVE-2018-7900) allowed credentials information to leak, so attackers could find devices with default passwords through an IoT search engine. Common mitigations include changing default passwords, applying firmware updates promptly, and restricting access to administrative interfaces.

Software-defined networking and virtualization

Software-defined networking (SDN) is not a single product but a design approach. It takes the decision logic that a traditional router or switch keeps on the device and moves it into a controller that can manage many devices at once. The devices keep the fast data-plane work of moving packets, while the controller computes forwarding behaviour for the whole network. The two sides communicate through a southbound interface. In the classic SDN model, the controller also exposes a northbound interface to applications, such as management, security, and orchestration tools.

OpenFlow is the southbound protocol most closely associated with SDN. The Open Networking Foundation standardized it in 2011. It gives software access to the flow tables that tell a switch or router how to handle traffic. When a device receives a flow it has no rule for, it can send the packet to the controller and wait for instructions. Devices also report link and port changes and collect statistics that the controller reads. Other southbound options exist, including ForCES, OvSDB, and NETCONF, but OpenFlow has been the most widely adopted open standard. Many commercial products mix standard interfaces with vendor-specific ones.

A related movement, network function virtualization (NFV), targets the hardware appliances that perform network functions. The idea was pushed by a group of telecommunications operators, whose 2012 white paper led to the creation of an ETSI Industry Specification Group. NFV replaces dedicated boxes, such as firewalls, load balancers, and some routing functions, with virtualized network functions (VNFs) that run as software on commodity servers. These servers are usually managed by a hypervisor, and the functions can be chained into service paths. In practice, a new service then means installing software rather than buying equipment. Routing and firewall functions also run as virtual machines or containers in cloud platforms, and most telecom vendors now sell 5G core products built on ETSI NFV standards.

White-box hardware and disaggregation take a different route. Instead of buying a vendor's integrated product, the operator separates the switch hardware from the network operating system (NOS). White-box switches use merchant switching silicon from several chip makers, and the buyer chooses an NOS, often a Linux-based system such as SONiC. The Switch Abstraction Interface, developed under the Open Compute Project, provides a common layer so that one NOS can run on silicon from different vendors. Industry usage distinguishes bare-metal switches, which ship with no software beyond a boot loader, from white-box switches, which come pre-loaded by a vendor, and from brite-box switches, which are branded products sold more cheaply. The main claimed benefits are lower cost, reduced vendor lock-in, and the freedom to change software without changing hardware. The costs include integration work, the need to find support across two suppliers, and the question of who is responsible when a fault crosses the hardware–software boundary.

SDN and virtualization raise well-known design debates. A logically centralized controller can become a single point of failure: if it fails and no standby exists, the whole network may stop being managed, and in some designs traffic flows are interrupted. Common mitigations include running an active-standby controller pair with synchronized state, or a physically distributed set of controllers that appears to be one logical unit. Newer OpenFlow versions allow a switch to connect to several controllers with defined roles, such as master and equal. Scalability is a second concern, because the control channel must carry flow requests and must stay available and secure. Open-source controllers such as OpenDaylight and ONOS are used in research and industry, but many operators have adopted SDN gradually, running classical protocols alongside controller-driven parts of the network. Adoption has been strongest in data-centre fabrics, large carrier and cloud environments, and campus or enterprise deployments with a clear management need. Wholesale replacement of traditional routing with centrally programmed networks has progressed more slowly than early predictions suggested.

Routing in other fields

The word "routing" also names operations in areas far from packet forwarding. The most active of these is machine learning, where a learned gate decides which sub-networks process each piece of input. Circuit design uses the same idea of choosing paths through a graph, and woodworking uses a separate word that only looks the same.

Mixture-of-experts models

In a mixture-of-experts (MoE) layer, several expert sub-networks sit behind a gating network, which is often called the router. For each token, the gate scores every expert, and the token is sent only to the best-scoring few. In the simplest scheme, top-k routing, the gate keeps the k highest scores. Other designs reverse the direction: under expert-choice routing, each expert selects the tokens it will handle, which improves expert utilization and coherence, according to one 2026 survey of the field. Transformer-based language models typically place an MoE layer where a dense feed-forward block would otherwise be. The Switch Transformer sends each token to one expert (k=1), while GShard uses two (k=2) .

Sparse routing lets a model hold far more parameters than it spends computation on for any single token. Only a small share of the experts runs on each token, so capacity grows much faster than the cost per token. The method has a known failure mode. Left alone, the gate tends to favor a few experts, and those experts then train faster and get picked even more often; researchers call this routing collapse. The usual countermeasure is an auxiliary loss term that pushes tokens to be spread roughly evenly across experts. These techniques appear in large language models and are a core part of how they scale.

Circuit and PCB routing

In electronic design automation, routing means laying out the copper traces that connect component pins on a printed circuit board or an integrated-circuit. The problem is naturally a shortest-path problem on a grid graph. C. Y. Lee's 1961 maze algorithm is an early example: it expands a numbered wavefront from one pin until it reaches the target, then traces back the shortest route, avoiding obstacles. Many later tools restrict wires to horizontal and vertical segments, so the length of a wire is its Manhattan length, the sum of its horizontal and vertical extents. Using alternating horizontal and vertical layers, known as X-Y routing, leaves the routing surface less blocked and suits automation well.

Modern autorouters are iterative. In the PathFinder approach, which was first developed for FPGAs, every edge and node counts as a shared resource. The first pass routes all connections greedily, and later passes raise the cost of congested edges and re-route the worst offenders until no resource is overused. Placement is handled separately and often uses half-perimeter wire length as a proxy for how easy the board is to route. Designers also deliberately lengthen some traces with serpentine patterns to match timing, a technique traced to early 1980s autorouting work.

Woodworking and CNC routers

A woodworking router is a high-speed handheld cutter. It is used to trim, shape, and hollow out wood and other materials, and it is typically run with profile bits or in a router table. Its name comes from a separate English verb meaning to hollow out with a gouging action, so it is etymologically unrelated to the word "route." A CNC router replaces the hand with a gantry that moves a router spindle along the X, Y and Z axes under computer control, following toolpaths generated by CAD/CAM software. Compared with metal milling machines, CNC routers emphasize high spindle speed and large sheet capacity, while mills rely on high torque and rigidity for hard metals. None of these machines forwards data, so they have no connection to networking hardware beyond the shared name.

Etymology and pronunciation

"Router" is an agent noun formed from "route," which entered English from Old French and ultimately derives from a Latin phrase meaning a broken or cut-through road. The noun "route" is pronounced in two common ways. Speakers in North America usually say it to rhyme with "out," while many British speakers say it to rhyme with "boot," and the device noun "router" follows the same split. The verb "to route," meaning to send something along a chosen path, is pronounced to match the noun in each variety. Its use as a verb in networking and logistics is more common than the verb form of "router," which is rarely used.

References

1. What is Broadcast Domain in Networking? - PyNet Labspynetlabs.com
2. Broadcast domainen.wikipedia.org
3. Gateway addressen.wikipedia.org
4. Broadcast domain explainedgeek-university.com
5. Exploring Broadcast Domains: Definition & Impactvteams.com
6. Layer 2 Domainlearningnetwork.cisco.com
7. BGP in 2025blog.apnic.net
8. BGP Best Current Practices ISP/IXP Workshops 1 Last updated 4th April 2025nsrc.org
9. Longest Prefix Match - an overviewsciencedirect.com
10. Network switch memory interface configurationimage-ppubs.uspto.gov
11. What will happen when the routing table hits 1024kprefixx.net
12. The Evolving Landscape Of BGP: Past Five Years, Present State, And Future Directions • PALLAS DIGITALpall.as
13. Understand Administrative Distance - Ciscocisco.com
14. Administrative Distances for Routes - Introductiondocs.defenseorchestrator.com
15. Administrative distanceen.wikipedia.org
16. What Is the Routing Information Protocol (RIP)? - JumpCloudjumpcloud.com
17. RIP and RIPng Overviewjuniper.net
18. Routing Information Protocol (RIP) - GeeksforGeeksgeeksforgeeks.org
19. Cisco Secure Firewall Management Center Device Configuration Guide, 7.6 - RIP [Cisco Secure Firewall Management Center] - Ciscocisco.com
20. Dynamic Routing Protocols: Distance Vector vs Link Statepluralsight.com
21. Link State vs Distance Vector Routing Protocolsorhanergun.net
22. Routing: Distance Vector vs. Link Statebaeldung.com
23. Link-State Routing vs Distance Vector Routing vs Path Vector Routingccna-classes.com
24. What is a Routing Information Base (RIB)? - JumpCloudjumpcloud.com
25. What is a Forwarding Information Base (FIB)? - JumpCloudjumpcloud.com
26. Network distributed programmable forwarding plane packet processorimage-ppubs.uspto.gov
27. Custom routing information bases for network devicesimage-ppubs.uspto.gov
28. Forwarding information base entry priorityimage-ppubs.uspto.gov
29. Switching and Routing - Part 1. Router Architecturenetmanias.com
30. Route Processor - an overviewsciencedirect.com
31. Issues and trends in router design - Cornell: Computer Sciencecs.cornell.edu
32. Carrier Routing Systemen.wikipedia.org
33. Cisco CRS Carrier Routing System 16-Slot Line Card Chassis System Description - Switch Fabric [Cisco Carrier Routing System] - Ciscocisco.com
34. Router and switch architectureimage-ppubs.uspto.gov
35. Mutual table for enhancing software and hardware switched packet flow forwardingimage-ppubs.uspto.gov
36. Forwarding Hardware - Networking Notesnull.53bits.co.uk
37. Method and apparatus for performing network routing with use of power efficient TCAM-based forwarding engine architecturesimage-ppubs.uspto.gov
38. Cache-aware data structures for packet forwarding tables on general purpose CPUsarxiv.org
39. GOTO action in network traffic policiesimage-ppubs.uspto.gov
40. Wireless routeren.wikipedia.org
41. 0% found this document useful (0 votes)scribd.com
42. Solved: Setting up a Branch Office - Cisco Communitycommunity.cisco.com
43. Configure a Branch Office VPN for Failover from a Leased Linewatchguard.com
44. Use a Branch Office VPN for Failover From a Private Network Link — Configuration Examplewatchguard.com
45. HPE Networking routers: High-performance, programmable, and sustainable routershpe.com
46. PE-CE Lessons - NetworkLessons.comnetworklessons.com
47. TJ1400 Aggregation & Edge Routertejasnetworks.com
48. Aggregation Routerspine-networks.com
49. COR550 - Edgecore Networksedge-core.com
50. Network address translationen.wikipedia.org
51. Prioritize your traffic: understanding quality of service (QoS) on your routertotaldefense.com
52. What Is Router Quality of Service (QoS) and Why Should You Care?sip.us
53. NAT vs. PAT: Key Differences Explainedninjaone.com
54. What is Port Address Translation (PAT)? Use Cases, Advantages and Limitations - zenarmor.comzenarmor.com
55. Access Control Lists (ACLs): Types, Placement & Examplesittsystems.com
56. Transit Access Control Lists: Filtering at Your Edge - Ciscocisco.com
57. TCR Origin of Cisco Systemsweb.stanford.edu
58. Inside the Invention of the Stanford Router That Inspired Ciscootl.stanford.edu
59. The Origins of the Stanford University SUMEX-AIM Router Technology: A Historical and Technical Overview - 300000119 - CHMcomputerhistory.org
60. Cisco Systems Incencyclopedia.com
61. FinancialContent - The Network Weaves: A History of Cisco Systemsmarkets.financialcontent.com
62. Dr. William (Bill) Yeagerithistory.org
63. Sandy Lerneren.wikipedia.org
64. Ciscoen.wikipedia.org
65. Telephony communication via varied redundant networksimage-ppubs.uspto.gov
66. The Pre-commercial Internet under DARPAnber.org
67. The History of Border Gateway Protocolmedium.com
68. Method and system to reduce a number of border gateway protocol neighbors crossed to reach target autonomous systemsimage-ppubs.uspto.gov
69. Leonard Bosacken.wikipedia.org
70. Method and system to reduce a number of border gateway protocol neighbors crossed to reach target autonomous systemsimage-ppubs.uspto.gov
71. RFC 4271 on A Border Gateway Protocol 4 (BGP-4)mailarchive.ietf.org
72. RFC 2328 - OSPF Version 2datatracker.ietf.org
73. IETF RFCs Supported by Cisco NX-OS Unicast Featurescisco.com
74. RFC 2328 : OSPF Version 2 St´ephane Bortzmeyer <stephane+blog@bortzmeyer.org>bortzmeyer.org
75. TORCH: Characterizing Invalid Route Filtering via Tunnelled Observationarxiv.org
76. BGP Security: Route Hijacking and How to Detect It - DEV Communitydev.to
77. Stalloris: RPKI Downgrade Attackarxiv.org
78. RPKI and ROAs: Preventing BGP Route Hijacksobelinf.com
79. Cyber security authentication method for displayless networking deviceimage-ppubs.uspto.gov
80. 81% of Router Users at Risk: Default Passwords Expose Devices to Hackersblogs.npav.net
81. Decoding Router Vulnerabilities Exploited by Mirai: Insights from SonicWall’s Honeypot Datasonicwall.com
82. Juniper Warns of Mirai Botnet Targeting SSR Devices with Default Passwordsthehackernews.com
83. Botnetscyber.nj.gov
84. IDIoT: Securing the Internet of Things like it's 1994arxiv.org
85. Huawei Router Flaw Leaks Default Credential Statusthreatpost.com
86. BGP Hijacking - LogicMonitorlogicmonitor.com
87. OpenFlow - an overviewsciencedirect.com
88. Software-Defined Networking: A Comprehensive Surveyarxiv.org
89. A Survey on Software Defined Networking: Architecture for Next Generation Networkarxiv.org
90. Software-Defined Networking: State of the Art and Research Challengesarxiv.org
91. (PDF) Distributed SDN Control: Survey, Taxonomy and Challengesresearchgate.net
92. ETSI White Paper No. 54 Evolving NFV towards the next decadeetsi.org
93. ETSI NFV: A Decade of Transformation - LF Networkinglfnetworking.org
94. Network Service Orchestration: A Surveyarxiv.org
95. What Is Network Functions Virtualization (NFV)?ibm.com
96. Switch Abstraction Interface (SAI) - Breaking the Network Aggregationdesign-reuse.com
97. Everything You Should Know About White Box Switch - Asterfusion Data Technologiescloudswit.ch
98. The White Box Switch: Part 1nxp.com
99. Area-Efficient In-Memory Computing for Mixture-of-Experts via Multiplexing and Cachingarxiv.org
100. Understanding Mixture of Experts (MoE) Neural Networksintuitionlabs.ai
101. Mixture-of-Experts (MoE) LLMs - by Cameron R. Wolfe, Ph.D.cameronrwolfe.substack.com
102. Mixture of Experts in Large Language Modelsarxiv.org
103. What Is Mixture of Experts (MoE) and How It Works?nvidia.com
104. Token-Based Mixture-of-Experts Modelsemergentmind.com
105. Mixture of Experts Explainedhuggingface.co
106. What is Mixture of Experts (MoE)? How it Works and Use Cases - Zilliz Learnzilliz.com
107. GitHub - bbenchoff/OrthoRoute: OrthoRoute: A GPU-accelerated PCB autorouter for KiCad · GitHubgithub.com
108. PCB ROUTERS AND ROUTING METHODSspeedingedge.com
109. Iterative method for establishing connections and resulting productimage-ppubs.uspto.gov
110. System, method, and computer program product for optimization-based printed circuit board designimage-ppubs.uspto.gov
111. PCB Routing Methodsterratel.eu
112. CNC routeren.wikipedia.org
113. Router (woodworking)en.wikipedia.org
114. CNC Router vs Milling Machine: Which Is Right for Your Shop? - Wurth Machinerywurthmachinery.com
115. Best CNC Routers for Woodworking - STYLECNCstylecnc.com