slowstart

Slow start is a TCP congestion control mechanism used to probe the available network capacity at the start of a connection or after a loss event. It temporarily limits the rate of data transmission to avoid overwhelming the network, then gradually increases the sending rate as network conditions permit.

During slow start, the sender's congestion window (cwnd) increases exponentially. Starting from an initial cwnd (often 1 Maximum Segment Size, MSS), for each acknowledged segment, cwnd is increased by roughly one MSS, causing cwnd to roughly double each round-trip time. The slow-start phase continues until cwnd reaches a threshold called ssthresh, or until a packet loss is detected, at which point congestion control reduces the window and transitions to a slower growth mode.

Variants and history: Slow start was described in early TCP congestion control schemes, such as Tahoe and Reno, and formalized in RFC 2001 and RFC 5681. In Tahoe, a loss causes cwnd to drop to 1 MSS and ssthresh to cwnd/2. In Reno and later variants, fast recovery shortens the loss handling by keeping the connection in a modified slow start/congestion avoidance mode, often resetting cwnd to ssthresh rather than 1 MSS. Modern TCP implementations use NewReno or other algorithms; they may also adjust initial cwnd, MSS, and use pacing to smooth transmissions.

Because slow start rapidly increases sending rate, it can be sensitive to large bandwidth-delay products; in such networks the initial phase may underutilize or briefly saturate bandwidth. Adjustments like initial cwnd tuning and pacing have been proposed to improve startup performance.