Metcalfe: Collisions and Random Retry

What the paper explains

Metcalfe and Boggs describe stations deferring while they hear a transmission. Propagation time still allows simultaneous starts; a detected collision leads to abandonment and a random retry interval. They also distinguish packet transport from the additional precautions applications require for reliable communication.

Carrier detection defers a new transmission when another is heard; interference detection identifies damage during an attempt. The paper treats these as different mechanisms. Listening locally cannot reveal a signal that has not yet propagated to that location.

Sources for this section: [1]

An original probability exercise

For this deliberately slotted toy, there are exactly two backlogged stations. Each independently transmits with probability p in a slot. Exactly one transmission succeeds; two collide; if neither transmits, the slot is idle. Assume equal slot durations, no carrier sensing, no propagation detail and fresh independent choices each slot. Compare p = 1/2 with p = 1/4.

Sources for this section: [1]

Reference answer

Exactly one transmits with probability p(1-p)+(1-p)p = 2 p(1-p). At p = 1/2 this is 1/2; collision and idle probabilities are each 1/4. At p = 1/4, success is 3/8, collision is 1/16 and idle is 9/16. Lowering the attempt probability reduces collisions but can waste more slots through idleness. A lower collision count alone is not the same as higher useful throughput.

Sources for this section: [1]

Do not mistake the toy for Ethernet

The real design listens, detects interference and adapts retransmission timing. Our equal-slot model does none of those things; it isolates independent contention. It cannot predict measured Ethernet utilisation, latency or a modern switched network's behaviour. The exercise explains a trade-off without claiming to reproduce the paper's later performance analysis.

Likewise, a successful slot in this toy means exactly one sender attempted. It does not test a packet checksum or confirm that a destination application accepted data. Keeping these layers distinct prevents a probability of contention success from becoming an unsupported end-to-end reliability guarantee.

Sources for this section: [1]