Summary

  • In the 1988 DECbit proposal, a router marked a packet only after an adaptive queue average reached one packet; a sender acted only after collecting a window of returned marks. The bit was the messenger, not the measurement.
  • K. K. Ramakrishnan and Raj Jain paired that evidence with a restrained control rule: reduce the window to seven-eighths when at least half the relevant packets were marked, otherwise add one.

A dashboard light can be binary while the judgment behind it is not. On or off says little unless the observer knows which events were combined, for how long, and what response follows. That was the quiet sophistication of the congestion-avoidance scheme K. K. Ramakrishnan and Raj Jain presented at SIGCOMM in 1988. Its packets carried one congestion indication bit. Its control system carried memory at both ends of the path.

A queue observed across its own rhythm

In “A Binary Feedback Scheme for Congestion Avoidance in Computer Networks with a Connectionless Network Layer”, a congested router sets a bit in a packet travelling toward the destination. A later uncongested router does not erase a mark already made. The destination returns the indication in transport feedback, allowing the source to adjust how many packets it may keep in flight. The paper is jointly authored; the design should not be reduced to a lone-inventor story about Jain.

The marking decision was not based on the queue at a single instant. The authors counted the packet in service as part of queue length and proposed marking when the average reached at least one packet. More important, they defined an adaptive averaging interval: the last complete busy-plus-idle cycle together with the busy portion of the current cycle. The interval therefore followed the resource's own pattern of filling, draining and becoming idle.

The paper considered simpler alternatives. A fixed sampling interval can be too long for a short round trip and too short for a long one. An exponential running average can retain stale history after a burst has passed. Measurements made on unrelated schedules can also make otherwise similar routers report inconsistent conditions. The regeneration-cycle method tried to make the indicator reflect sustained use without letting one brief queue determine the verdict.

The threshold of one packet was an operating rule, not a mystical optimum for every network. Below it, the resource may still have useful capacity; above it, delay begins to accumulate even though throughput cannot exceed the bottleneck's service rate. That logic makes the queue an early sign of offered load outrunning service, while leaving room for the link to remain busy.

A marked packet was not yet an order

The second time filter sat at the source. The endpoint did not halve or trim its window whenever one marked acknowledgement arrived. It waited for a decision interval covering acknowledgements for the previous window and the current window, then examined the returned bits corresponding to the current window. If at least half were set, the source reduced the window to 87.5 percent of its current value. Otherwise it increased the window by one.

Those numbers encode a division of labour. The 50 percent test aggregates path evidence: one anomalous mark need not provoke a cut. Additive increase lets a source probe gradually for newly available capacity. The seven-eighths decrease responds to widespread marking without the violence of repeatedly collapsing the window. The paper says 0.875 was selected partly for ease of implementation and reduced oscillation. A separate 0.8 example illustrates how integer rounding could damage fairness; it is not the chosen algorithm and should not be substituted for it.

The authors evaluated the scheme analytically and through simulations involving multiple users, varied packet sizes, heterogeneous paths, a satellite link, changing service times and a newly arriving source. They reported efficient, adaptive and fair behaviour within those experiments. That evidence is substantial for a design paper, but it is not a measurement of today's Internet and cannot establish universal performance across modern transports, queues or traffic incentives.

From a company design to an Internet survey

RFC 1254, edited by A. Mankin and K. Ramakrishnan in 1991, surveyed gateway congestion-control approaches and described the DECbit mechanism. It records the one-packet average threshold and the seven-eighths response across two windows. Its status is Informational: it maps practice and implementation tasks rather than declaring an Internet Standard.

The RFC also exposes a governance problem. Congestion avoidance depends on endpoints actually responding. A user that ignores feedback may gain short-term throughput while imposing delay or loss on others. Nor can every gateway and host be upgraded simultaneously. Binary feedback therefore joined an engineering mechanism to a transition problem: the network could supply evidence, but shared benefit still required compatible and cooperating receivers of that evidence.

Washington University's current faculty biography calls Jain a co-inventor of DECbit and says it was implemented in various forms in DECnet, OSI, Frame Relay and ATM networks. “Various forms” is an important qualification. It supports the breadth of the idea's influence, not a claim that every later system reproduced the 1988 algorithm exactly. The university's 2017 SIGCOMM award notice likewise places the work within a longer career in performance analysis and congestion control.

Sources and limits

This account relies on the coauthored 1988 paper, the Informational RFC and Washington University biographies. It does not provide a current deployment census, compare DECbit experimentally with present-day congestion controls or establish a direct technical lineage to every later explicit-marking system. Operational implications below are editorial deductions from the documented control structure. The public Washington University portrait grounds the AI editorial likeness; its laboratory background is invented.