Packet switching is a method of moving digital information through a network by dividing communications into packets and forwarding those packets through shared network resources. It is the fundamental communication model underlying the modern Internet and most contemporary data networks. Rather than reserving an entire end-to-end communication channel exclusively for one conversation, packet-switched networks allow packets belonging to many communications to share links and forwarding equipment.
A packet normally carries addressing and control information that enables intermediate routers or switches to determine how to handle it. The network can therefore multiplex traffic from many users over the same physical infrastructure. Packet switching is particularly effective for computer communications because data traffic is usually bursty: a device may transmit substantial information for a short period and then remain relatively idle. Shared capacity can consequently be used by other traffic when one communication is not actively sending data.
Packet switching developed as an alternative to the traditional circuit-switching model used extensively by telephone networks. In circuit switching, a dedicated logical or physical circuit is established for a communication and resources are reserved for the duration of the session. This can provide predictable characteristics, but reserved capacity may remain unused during periods of silence. Packet switching instead statistically multiplexes traffic, allowing packets from independent flows to occupy transmission capacity as they become available.
Early packet-network research included Paul Baran's work on distributed communications and the ARPA-funded networking research that led to ARPANET; Donald Davies independently developed and popularised the term “packet switching” in the 1960s. Modern packet networking evolved through many standards and technologies, with IP eventually becoming the dominant internetworking protocol suite for the global Internet.
There are different forms of packet switching. In datagram networks, each packet is treated as an independently routable unit and normally contains enough addressing information to allow routers to forward it without maintaining a dedicated connection state for the entire communication. IP is the principal example. In virtual-circuit approaches, a logical path or forwarding state can be established before data transmission, and packets associated with the virtual circuit can then use that state.
The two models differ in how forwarding state, routing decisions, failure recovery and traffic management are handled. Packet switching also differs from message switching, where a complete message may be stored and forwarded as a unit rather than being divided into smaller packets. Packet-switched systems generally use buffering because multiple packets can arrive at a device faster than an outgoing link can transmit them, creating queues.
Routers are central to packet switching because they make forwarding decisions between different IP networks. A router examines relevant packet information, particularly the destination address, consults its forwarding information and sends the packet towards an appropriate next hop. The route does not normally represent a physically dedicated path in the circuit-switching sense. Routing protocols and routing tables determine the available paths, while forwarding hardware moves individual packets at high speed.
At the local network level, switches typically forward Ethernet frames according to link-layer addresses, while routers connect separate IP networks. The distinction between switching and routing is therefore important even though everyday descriptions sometimes call all network forwarding “switching”. Packet switching is a system-wide principle involving shared resources, packetisation, forwarding and statistical multiplexing rather than merely a particular type of Ethernet device.
Packet switching introduces variable network behaviour. Each packet can experience transmission delay, propagation delay, processing delay and queuing delay. If a router's outgoing interface becomes heavily loaded, packets may wait in a queue before transmission. When buffers become full, subsequent packets can be discarded. Packets can also take different routes if routing changes, producing different delays and potentially causing out-of-order delivery. TCP is designed to cope with many of these conditions by using sequencing, acknowledgements, retransmission and congestion-control algorithms. Real-time applications may instead use UDP and application-level mechanisms because retransmitting an old packet can be less useful than receiving newer information promptly. Packet switching therefore provides flexibility and efficient resource sharing, but it does not inherently guarantee constant latency, bandwidth, ordering or delivery.
Statistical multiplexing is one of packet switching's principal advantages. Suppose many users share a link whose capacity is substantially greater than the average traffic generated by any individual user. Each user's packets occupy the link when they need it, allowing the network to support aggregate demand without dedicating a fixed portion of the link to each user. This efficiency is particularly valuable for Internet traffic, where usage patterns vary significantly over time.
The trade-off is that simultaneous demand can exceed available capacity, creating congestion. Network operators therefore use queue management, routing, traffic engineering, quality-of-service mechanisms, capacity planning and congestion-control systems to manage shared resources. The network's ability to share capacity dynamically is both the central strength of packet switching and a source of some of its characteristic performance problems.
Packet switching supports a broad range of network architectures, from small home networks to global Internet backbones and large cloud data centres. Technologies such as Ethernet, Wi-Fi, IP, MPLS and various carrier networking systems implement or support packet-oriented communication in different ways. Modern networks may also combine packet switching with mechanisms that provide traffic classes, resource reservations, engineered paths or service guarantees. Software-defined networking can separate aspects of control and forwarding, while modern data-centre networks use high-speed packet switching to interconnect servers and storage systems. The underlying principle remains that information is carried in discrete units across shared infrastructure rather than requiring a permanently dedicated end-to-end circuit for every communication.
Packet switching should not be confused with the mere act of splitting a file into pieces. Packet switching specifically concerns how discrete network units share communication infrastructure and are forwarded through a network. It also does not mean that every packet is routed independently in every practical network; protocols and network designs can maintain forwarding state or engineered paths. Another misconception is that packet switching is inherently unreliable.
Packet-switched IP networks provide best-effort delivery, but higher-level protocols can provide strong reliability, ordering and error recovery. Conversely, packet switching does not guarantee that TCP itself eliminates congestion: TCP actively responds to congestion signals and packet loss but cannot create additional physical capacity. The practical behaviour of a packet-switched network therefore results from interactions among link capacity, routing, queues, transport protocols, application traffic and network-management mechanisms.
