Network Architecture

9 frameworks, from 1890 to today · Networks · Computer Science

still a working view ended · each leads to its entry below

Overview

Network architecture is the set of large-scale decisions about how a communication network is organized: who decides where traffic goes, how much intelligence sits in the network versus at its edges, how different networks and technologies are joined, and how the whole system is divided into parts that can be built and changed independently. It is the difference between laying cable and designing a system—between asking "how do we move this bit?" and asking "what kind of network should exist at all?" The stakes are enormous because architecture choices outlive hardware: a decision about where to put control logic or how to split responsibilities into layers can shape what is easy to build, secure, and upgrade for decades. The history of network architecture is largely a history of arguments about that division of labor.

Circuit switching and the telephone monopoly

The first global communication network was the telephone system, and its architecture was hierarchical and circuit-switched. A circuit-switched network reserves a dedicated path—a physical circuit—between two callers for the entire duration of a call1. In the telephone system this meant a strict hierarchy: local exchanges fed into toll centers, which fed into regional and national switches, with each level aggregating traffic from the level below. The architecture mirrored the organizational structure of the monopoly telephone administrations that built it. It was extremely good at what it was designed for—reliable, predictable voice service—but its control was centralized, its hierarchy rigid, and setting up a call meant tying up resources end to end whether or not anyone was actually speaking.

Packet switching on both sides of the Atlantic

The decisive break came from packet switching, an idea developed in the 1960s in several places at once. Rather than reserving a circuit, a packet-switched network chops data into small packets, each carrying a destination address, and lets them share links and be forwarded independently2. The intellectual case was made independently by Paul Baran at RAND in the early 1960s (motivated by survivable communications) and Donald Davies at the UK's National Physical Laboratory, who coined the term "packet" around 1965–662. The practical demonstration came with the ARPANET, whose first packet switch was installed at UCLA in 1969. Packet switching changed the architectural question fundamentally: if links are shared and packets are independent, then where does the intelligence go? The ARPANET's answer was to put addressing and forwarding into the network's switches, the Interface Message Processors.

Vendor architectures and the layered idea

As computer networking commercialized in the 1970s, vendors built proprietary architectures to lock customers in. IBM's Systems Network Architecture (SNA), announced in 1974, was the most influential: a complete, hierarchical design for connecting terminals and mainframes, with a master control point and a strict tree of subordinate nodes. Digital Equipment Corporation's DECnet and others offered rival schemes3. These systems worked well within one vendor's world but made it hard to connect different manufacturers' equipment—the "heterogeneous network" problem.

The architectural response was layering: dividing communication into a stack of levels, each providing services to the one above and relying on the one below, so that a layer can be redesigned without disturbing its neighbors. The most influential formalization was the Open Systems Interconnection (OSI) reference model, developed within the International Organization for Standardization from the late 1970s and published in 19843. OSI's seven layers—physical, data link, network, transport, session, presentation, application—became a common vocabulary. On the research side, the ARPANET community developed its own layered model around the Transmission Control Protocol and Internet Protocol (TCP/IP), split in 1978 into a connection-oriented transport layer (TCP) and a connectionless internetwork layer (IP). Where OSI imagined a rich, standardized stack with the network providing many services, TCP/IP imagined a minimal network that simply moved datagrams—independent packets—and pushed complexity to the endpoints.

The end-to-end argument

In 1981, Jerome Saltzer, David Reed, and David Clark published an argument that became the philosophical core of the Internet's architecture4. The end-to-end argument holds that a function should be implemented at the endpoints of communication rather than inside the network whenever the endpoints are the only ones who can implement it correctly and completely. Reliable delivery, for example, must ultimately be checked by the receiving application, since the network cannot know whether data arrived intact at the far end; therefore building reliability into every router is redundant, and the network is better kept simple. This did not claim that in-network functions are never useful—it argued for placing them carefully. The argument convinced a generation of designers because it explained why the Internet could run over any underlying technology and support any application: the network stayed dumb and general, and intelligence lived at the edges.

Federating separate networks

The Internet's second architectural principle was federation. Rather than one network, the Internet is a network of networks: independently administered domains that agree to carry each other's traffic. Early work on internetworking protocols in the 1970s established how packets could cross from one network technology to another. The Border Gateway Protocol (BGP), first specified in 1989 and refined over the following years, became the mechanism by which autonomous systems—separately governed networks—exchange routing information and apply their own policies about whom to carry traffic for. This made global interconnection possible without a central authority, at the cost of a routing system governed by local policy rather than global optimization. The architecture of the Internet thus embedded a political stance: interconnection by voluntary agreement among peers.

Middleboxes, overlays, and the erosion of the clean model

By the 1990s the tidy end-to-end picture was under strain. Firewalls, network address translators (NATs), and proxies—collectively called middleboxes—sat in the middle of paths and modified or blocked traffic. NATs, which let many private machines share one public address, became ubiquitous as IPv4 addresses ran short5. Meanwhile overlay networks built virtual topologies on top of the Internet: peer-to-peer systems, content delivery networks, and later virtual private networks and overlay routing services. These developments showed that the "dumb network" was in practice full of intelligent devices, and that architecture was increasingly negotiated between what the protocol designers intended and what operators deployed.

Clean slate and the case for redesign

Frustration with the Internet's ossification—its difficulty in adopting new core protocols—led in the 2000s to the clean-slate movement, most visibly the US National Science Foundation's FIND and GENI programs and the European FIRE initiative. Clean-slate researchers argued that incremental patching could not solve deep problems of security, mobility, and addressing, and that the field needed to design new architectures from scratch, unconstrained by backward compatibility. Projects explored named-data networking (where you request content by name rather than address), new naming and identity schemes, and network virtualization as a way to run many experimental architectures at once. The movement produced important ideas and testbeds, though no wholesale replacement of the Internet.

Software-defined networking

The most consequential architectural shift of the 2000s and 2010s came from a different direction: separating the control plane (the logic that decides where traffic goes) from the data plane (the hardware that forwards packets). The idea grew out of earlier work on programmable networks and active networks in the 1990s, and was crystallized by the OpenFlow protocol, introduced around 2008 by researchers at Stanford and Berkeley. OpenFlow let a centralized controller program the forwarding tables of many switches through a standard interface6. This made network behavior programmable from software, enabled network-wide policies and experiments, and led to commercial controllers and to the wider adoption of the control/data-plane split in data centers and enterprise networks. Programmable data planes, such as the P4 language introduced in 2014, extended the idea to the switches' own packet-processing logic. The architectural question—how much centralized control is safe and scalable—remains one of the field's live debates.

Where the field stands

Today's network architecture descends from these layered strata. The Internet's end-to-end and federation principles still structure the public network, though middleboxes, overlays, and cloud provider networks complicate the picture. Software-defined control and programmable hardware are now standard in data centers and increasingly elsewhere. Open questions remain sharp: how to secure routing and naming without a central authority, how to accommodate mobility and massive numbers of devices, how much intelligence should return to the network for performance and policy, and whether new architectures can ever displace the installed base. The answers will be argued, as they always have been, over the division of labor between the center and the edge.

References

  1. Circuit switching, Wikipedia
  2. Packet switching, Wikipedia
  3. OSI model (History, introduction), Wikipedia
  4. End-to-end principle, Wikipedia
  5. Network address translation (introduction, History), Wikipedia
  6. OpenFlow (introduction, History), Wikipedia

The frameworks of Network Architecture

Each framework is a way of working the field has used: when it held, and the question it answers.

  1. 1890-1980

    Circuit Switching

    How do we reserve a path for a call?

    Fly to its tower in Nousfarer

  2. 1965-Present

    Packet Switching

    How can many users share one link?

    Fly to its tower in Nousfarer

  3. 1974-1990

    Vendor Network Architectures

    How do one vendor's machines all connect?

    Fly to its tower in Nousfarer

  4. 1978-Present

    Layered Reference Models

    How do we divide communication into independent parts?

    Fly to its tower in Nousfarer

  5. 1981-Present

    End-to-end principle

    Where should intelligence sit in a network?

    Fly to its tower in Nousfarer

  6. 1983-Present

    Internetworking and Federation

    How do separate networks join without a ruler?

    Fly to its tower in Nousfarer

  7. 1994-Present

    Middlebox and Overlay Architecture

    What happens when devices sit inside the path?

    Fly to its tower in Nousfarer

  8. 2005-Present

    Clean-Slate Architecture

    Can we design a network from scratch?

    Fly to its tower in Nousfarer

  9. 2008-Present

How the ideas connect

Each arrow runs from an earlier idea to the later one it shaped. The list below says what passed between them.

Circuit SwitchingPacket SwitchingVendor Network ArchitecturesLayered Reference ModelsEnd-to-end principleInternetworking and FederationMiddlebox and Overlay ArchitectureClean-Slate ArchitectureSoftware-Defined Networking

Beyond the field

Where Network Architecture's ideas came from in other fields, and what they led to there: the strongest 12 of 52 links.

A city in Nousfarer: a red-carpeted stair climbing terraces to a crowned tower on an acropolis, more towers round it, forest and snowy mountains beyond.
A city in Nousfarer, the game played in the browser.

Network Architecture is a planet in Nousfarer

Its city has a tower for each of these 9 frameworks. Land, walk in, and climb one concept at a time.

Land on Network Architecture

Related fields

Fields its ideas are linked to

More in Networks

See where Network Architecture sits on the Map of Human Knowledge