Summary

  • Sandy Fraser's virtual-circuit work separated permission to communicate from the bursts that used that permission: a call was requested, accepted and installed before ordinary bytes relied on it.
  • The bargain bought compact forwarding and a common service across unlike devices, but it also made stale mappings, half-open calls, resource accounting and provable teardown part of network correctness.

A path made out of promises

The most revealing packet in Fraser's 1973 patent is not user data. It is an acceptance message. A calling device first sends a connection request on a reserved control channel. The called device may accept or reject it. On acceptance, the information needed to allocate a virtual channel travels back, and switching units install the channel link by link. Only then does communication begin.

That sequence makes a call more than a convenient label. It is a distributed promise. The endpoint has asked for a destination; the far end has consented; intermediate switches have associated local channel identifiers with the next leg; and the network has taken on an obligation to preserve those associations until the call ends. The bytes that follow can be spare because the setup did the verbose work in advance.

Fraser's patent used an especially economical distinction. A path could be virtually allocated—its required characteristics stored—without keeping a physical transmission path busy between bursts. Actual resources were activated as data arrived. This was neither a dedicated telephone circuit nor a succession of entirely self-describing datagrams. It coupled remembered route state with bursty use of shared capacity.

The design also named the end of the promise. A calling or called device could send an end-of-call message; deallocation then proceeded link by link. The patent warns that data still in transit could be lost during that process. Teardown was therefore not clerical tidying. It changed what the network was entitled to forget.

From Spider to Datakit

Fraser's own curriculum vitae places the idea before the product. It records virtual-circuit switching in 1969 and window flow control in 1970, demonstrated in the Spider research network. The Bell Labs historical chronology independently dates the start of Spider's construction to 1970. Fraser then describes designing Datakit in 1975 as a modular network with hardware virtual-circuit switching.

By 1990, according to the CV, Datakit VCS served Bell Laboratories and was used across AT&T, Bell operating companies and customer networks internationally. That reach matters because it turns an elegant call diagram into an operating system of obligations: names must resolve, calls must be admitted or refused, channel state must agree across hops, and failure must leave enough evidence to decide whether a path still exists.

The architecture was also a response to heterogeneity. Digital devices differed in speed, buffering and protocol. Fraser's early system placed programmable interface and switching functions between them, matching transmission and reception characteristics rather than demanding identical endpoints. A short local channel number could stand in for a full destination only because the switch retained the translation that gave the number meaning.

That is the central exchange in any label-switched system. Repeating a global address in every unit spends bandwidth and processing on explicit destination evidence. Installing a mapping once spends control effort and durable state. Neither cost disappears; the architecture chooses where it is paid.

The call processor as an interpreter

A later Bell Labs team—Roy Campbell, Sean Dorward, Arun Iyengar, Charles Kalmanek, Gary Murakami, Ravi Sethi, Ce Kuen Shieh and See-Mong Tan—described how call-processing software evolved around this model. Each line attached to a virtual-circuit switch was managed by a lightweight line process. That process translated device-specific call requests into a uniform internal protocol, while many logical lines could share one physical link.

The point was not merely implementation elegance. Translation created an institutional boundary. Devices could speak in their local conventions; the switch could reason over one internal call language. Adding a new interface did not require every switching component to learn the device's entire dialect. The line process carried the burden of interpretation and supplied the rest of the system with a normalized request.

Plan 9's later network interface demonstrates the same separation from the user's side. Its protocol devices presented similar file-like operations for Datakit and Internet protocols, while a connection server could choose a network and protocol common to the destination. Applications asked to dial or listen without hard-coding every network-specific address scheme. The call remained real, but its mechanics did not have to contaminate every program.

This is a useful correction to the claim that keeping state “in the network” necessarily makes endpoints simple. Complexity moved into named translators, connection servers, admission decisions and switch tables. The achievement was not erasure. It was containment behind interfaces that could be inspected.

A datagram takes a different oath

RFC 791 defines the Internet Protocol as a connectionless, unreliable datagram service. IP itself supplies no end-to-end reliability, sequencing or flow control. Each datagram carries what routing needs without first asking the network to install an end-to-end call.

Putting IP beside Datakit does not produce a historical boxing match. The systems allocate responsibility differently. A connectionless layer avoids the requirement that every intermediate node agree on long-lived per-call state before forwarding. Higher layers and endpoints must recover the continuity the network does not promise. A virtual-circuit network pays setup and state costs so subsequent traffic can use an established treatment.

The useful question is therefore not which word—datagram or circuit—sounds modern. It is which failure the operator can prove. With a datagram, evidence concentrates on reachability, forwarding decisions and endpoint recovery. With a call, evidence must also show whether setup completed, which version of a mapping each hop held, what service was granted and whether teardown reached every holder of state.

XUNET exposes the bill

The XUNET 2 retrospective, written by Charles Kalmanek, Srinivasan Keshav, William Marshall, Stephen Morgan and Robert Restrick, shows the virtual-circuit lineage under higher speed and broader ambition. XUNET 1 had linked research sites with Datakit Virtual Circuit Switches. XUNET 2 used ATM switches, routers and wide-area links as a laboratory for traffic management, switch design and control.

During virtual-circuit setup, users requested a service class. Switch schedulers distinguished circuits, and per-circuit queues and buffer policies protected one user's traffic from another. Connection establishment changed header-translation entries and buffer limits. The call was no longer just a route; it could carry an enforceable resource promise.

That precision had a price. Static per-circuit buffers could consume far more memory than active traffic required. Dynamic allocation reduced memory but demanded signaling and coordination between fast buffer state and the processor making allocation decisions. The authors treated switch auditing, remote maintenance and fault management as architectural work, not accessories.

Here the old call sequence becomes a modern operations lesson. The control plane may believe a mapping exists while the forwarding hardware does not. A retry may create a second call while the first still owns buffers. A controller restart may recover names and routes yet lose the exact grants that made labels safe. The data path can appear healthy and still be carrying someone else's expired promise.

Sandy Fraser's durable contribution

Fraser Research's memorial records a career that extended from Ferranti and Cambridge to Bell Labs, AT&T Labs and the institute he later founded. His CV links awards from ACM, IEEE and the National Academy of Engineering to virtual-circuit switching, flow control and packet-network architecture. Those honors describe an invention; Datakit's more enduring lesson is a method of making network state legible.

The setup request said what was wanted. Acceptance recorded consent. A channel mapping made the decision actionable. Flow control bounded use. End-of-call supplied an instruction to forget. Each stage produced a fact that could, at least in principle, be observed and challenged.

A network built on calls therefore needs more than fast switches. It needs receipts for the life of a promise. Fraser's architecture did not settle the contest between stateful and stateless networks. It made clear that when the network remembers on our behalf, correctness includes knowing exactly what it remembers—and knowing when it has stopped.

Sources