TCP: Sliding Window and Congestion Control

Everything below this promises nothing. The wire drops bits, the shared medium collides, the network reorders and discards without telling anyone. TCP is where somebody finally promises that the bytes you sent are the bytes that arrive, in order, exactly once — and it can only make that promise by assuming every layer beneath it has already failed. The reliability part is straightforward and was in Cerf and Kahn's 1974 paper: number every byte, acknowledge what has arrived, retransmit what has not, and keep several segments in flight rather than waiting for each. The window has to be at least the bandwidth-delay product, because throughput is window divided by round-trip time, and a protocol that waits for each acknowledgement delivers the same throughput on a ten-gigabit link as on a hundred-megabit one. The hard part arrived thirteen years later. In October 1986 the link between Lawrence Berkeley Laboratory and the Berkeley campus, four hundred metres apart, fell from 32 kilobits per second to 40 BITS per second. Nothing had broken. Every sender, finding packets lost, retransmitted; the retransmissions caused more loss; and the network spent its entire capacity carrying copies of data it had already carried. It is blueprint 7's collapse curve, on a whole internet. Van Jacobson's answer in 1988 is one of the most consequential few hundred lines ever written. A sender must treat loss as a SIGNAL rather than as bad luck: probe upward until something drops, then halve. Nobody is told the capacity of anything. Every flow discovers it, continuously, and the additive-increase multiplicative-decrease pairing is the only simple rule that drives competing flows toward an equal share. FULLY MEASURABLE. Capture a real transfer, plot the window, and find the sawtooth on your own machine.