05 The record
Laboratories
Testbeds where this is measured rather than argued about.
Where measurement replaces argument
The history of network congestion is, in large part, a history of people discovering that their intuitions were wrong. Van Jacobson's 1988 congestion-control work at Lawrence Berkeley National Laboratory did not emerge from first principles alone — it came out of watching real traffic collapse on real links, then building the mathematics to explain what the instruments had already shown. The laboratory, literal and institutional, is what turned guesswork into the RFC.
That tradition has continued. Bufferbloat.net, the loose collaborative founded in part by Jim Gettys and Dave Taht, made empirical measurement central to its mission from the start: the bufferbloat problem was demonstrated by running traffic across consumer routers and recording the latency spikes that appeared under load, not by asserting that oversized buffers must cause delay. The Bufferbloat project's documentation collects much of this record, and the methodology — generate a saturating flow, then measure latency in parallel — became the template other groups adopted.
Academic testbeds formalised what informal experiments had sketched. The PlanetLab distributed testbed, deployed across hundreds of university nodes beginning in the early 2000s, let researchers inject traffic into paths that crossed real internet infrastructure rather than simulated hops. Results carried a kind of authority that ns-2 simulations could not quite match, because the queues being filled were the queues actually installed in the network, with all their firmware quirks and undocumented behaviour.
Luigi Rizzo's dummynet, developed in Pisa and later merged into FreeBSD, served a different but complementary role: a controlled, repeatable link emulator that could be tuned to any loss rate, delay or bandwidth constraint a researcher wanted to examine. Repeatability is the point. A measurement taken once, under conditions that cannot be reconstructed, is not a measurement — it is an observation. What distinguishes a laboratory from a lucky packet capture is the ability to run the same scenario again and record the same result, or to vary one parameter and attribute the change in outcome to that parameter alone.
Active queue management schemes — CoDel, FQ-CoDel, PIE — were validated this way, first in software simulation, then on embedded hardware running real traffic, before the IETF moved them toward standardisation. The testbed is the bridge between a paper's claims and a deployed algorithm. Without it, the congestion-control literature would be a long argument with no referee.
