Aggressive TCP Slow Start

TCP Slow Start is a mechanism that is intended to allow a data transfer to start at a low bandwidth and quickly reach the peak bandwidth that it can use, subject to congestion control constraints that are enforced by intermediate network nodes. The current version is defined by RFC 5681 [2]. In general, the mechanism used for congestion control in TCP networks is that a congested router will randomly drop a packet as a signal to the sending system that it is using too much bandwidth and needs to slow down. Sending systems control the bandwidth that they are using by controlling the number of packets that can be in flight at once (the congestion window), linearly increasing this number until a packet is dropped, at which point it is halved. At steady state, TCP connections exhibit a sawtooth pattern where the bandwidth repeatedly slowly increases to some peak, at which point it drops and starts increasing again. TCP Slow Start is meant to address the issue that connection must start with a pretty small window in order to support the slowest network connections, and it can take a very long time for the connection to ramp up from the initial congestion window size to the steady state congestion window size. During TCP Slow Start, the congestion window size increases exponentially rather than linearly, which in general allows the connection to reach the steady state bandwidth quickly. In practice, in addition to performing slow start at the beginning of a transfer, most modern TCP implementations also perform slow start periodically during a transfer in order to adjust to changing congestion conditions throughout the transfer. While TCP slow start works well in general, a data transfer must not experience dropped packets during slow start, because if it does the transfer will move from exponential increase to linear increase very quickly. If this occurs, the transfer might 25 0 5 10 15 20 10