Summary

  • Hans-Werner Braun’s NSFNET work joined operational authority to an unusually public measurement practice: monthly performance reporting, traffic matrices and flow statistics made growth visible outside the machine room.
  • The later loss of comparable backbone visibility shows that measurement is not an automatic by-product of scale. It is a governance choice that must survive changes in ownership, architecture and commercial incentives.

In September 1991, a picture of NSFNET traffic looked almost like weather moving across a continent. Links thickened under load; the national backbone could be read as a living system rather than an invisible utility. The image was useful because it made a claim that managers, funders and connected institutions could inspect: demand was not merely rising in the abstract. It was accumulating along particular paths in a network with finite capacity.

That legibility was part of the infrastructure. It did not emerge from a single dashboard or a single person. The National Science Foundation funded a public research network; Merit Network coordinated the service; IBM and MCI supplied engineering and communications capacity; ANS later operated the backbone; state and regional networks carried traffic to campuses. Measurement researchers, including k. claffy and George Polyzos, developed ways to interpret what the instruments recorded and to explain what they missed.

Braun’s place in this collective effort matters because he worked where operation, architecture and evidence met. After helping run the interim 56-kilobit-per-second NSFNET backbone, he became a co-principal investigator in the Merit-led proposal for the T1 network. The Internet Hall of Fame’s account describes him installing Dave Mills’s Fuzzball software and, in practice, running the early backbone. The next system had a severe deadline: the T1 backbone was expected to be operational by 30 June 1988. The partnership met that date, then developed a T3 prototype as traffic continued to rise.

The numbers explain why speed upgrades cannot be reduced to technological enthusiasm. The interim network was already overloaded by the middle of 1987. NSFNET traffic later reached 19 billion packets in September 1992, according to Merit’s final report, with growth then averaging roughly 11 per cent a month. By April 1995, near the backbone’s retirement, the service was carrying almost 100 billion packets a month. These were measurements of the instrumented NSFNET service, not a census of the whole Internet. Even with that boundary, they converted complaints about congestion and anecdotes about adoption into a shared operating record.

The reporting obligation was consequential. Under Merit’s agreement with NSF, monthly reports on network performance were required. Merit and IBM assembled source-to-destination traffic matrices and estimates of application use; performance and flow statistics from the NSFNET era were publicly available. An overloaded link could therefore be discussed as more than a local inconvenience. A capacity request could be compared with a trend. A routing change could be examined against an observed path. A funder could see whether the service it supported was becoming essential.

Braun’s RFC work shows the companion idea: visibility is valuable only when responsibility is intelligible. RFC 1093 described routing among administrative domains and the filtering needed at their boundaries. RFC 1222 documented the evolution of the NSFNET backbone and the separation between interior and exterior routing responsibilities. RFC 1104 addressed policy routing, accounting and monitoring in a wider architectural discussion. Together they reveal a network that was not simply one technical object. It was a federation of operators whose authority had to be delimited before evidence could support a decision.

Measurement also had hard limits. CAIDA’s work on long-term NSFNET growth warned that the available instrumentation had not been designed for traffic forecasting. Collection points changed; application classifications could mislead; a topology transition could break a time series even while the network continued to operate. When traffic migrated after the NSFNET backbone was retired, researchers lost visibility at precisely the moment the commercial Internet was expanding. A smooth chart could therefore conceal a discontinuity in what was being counted.

This is where Braun’s 1998 assessment becomes more than nostalgia. In an SDSC announcement for the ARTS traffic-analysis system, he argued that commercial Internet providers had shown little inclination to publish backbone performance statistics comparable to those available in the NSFNET period. The statement should remain attributed to him; it does not prove that every provider concealed every metric. It does identify an incentive change. A publicly funded backbone had a reason, and a contractual duty, to make performance observable.

A competitive provider could regard the same data as a security exposure, a commercial secret or a source of reputational risk.

The governance lesson is not that raw flows should be opened indiscriminately. Network data can expose users, customers, vulnerabilities and business relationships. The better principle is bounded transparency: publish aggregated, privacy-safe and methodologically documented indicators; retain enough continuity to distinguish real change from an instrumentation change; and state who controls the collectors, definitions and release schedule. Independent scrutiny requires provenance as much as volume.

Seen this way, Braun’s contribution is less about predicting demand than about making decisions contestable. Statistics did not automatically choose T1 or T3, and an impressive growth curve did not settle procurement, routing or funding policy. But a common evidence base narrowed the room for convenient denial. It allowed operators, researchers and sponsors to argue about the same network.

That capacity remains scarce. Today, many critical digital systems expose a service-status page while withholding the longer series needed to judge concentration, congestion or resilience. The NSFNET experience suggests a sharper test for public-interest infrastructure: not whether it publishes some numbers, but whether outsiders can understand the scope, continuity and institutional custody of those numbers. A backbone that can be measured only by its owner can be managed. A backbone whose measurement rules are durable and inspectable can be governed.

Sources