TCP for High-Speed Networks

Abstract

In response to a series of collapses due to congestion on the Internet in the mid-’80s, congestion control was added to the transmission control protocol (TCP) (Jacobson, 1988), thus allowing individual connections to control the amount of traffic they inject into the network. This control involves regulating the size of the congestion window (cwnd) to impose a limit on the size of the transmission window. In the most deployed TCP variant on the Internet, TCP Reno (Allman, Floyd, & Partridge, 2002), changes in congestion window size are driven by the loss of segments. Congestion window size is increased by 1/cwnd for each acknowledgement (ack) received, and reduced to half for the loss of a segment in a pattern known as additive increase multiplicative decrease (AIMD). Although this congestion control mechanism was derived at a time when the line speed was of the order of 56 kbs, it has performed remarkably well given that the speed, size, load, and connectivity of the Internet have increased by approximately six orders of magnitude in the past 15 years. However, the AIMD pattern of window growth seriously limits efficienct operation of TCP-Reno over high-capacity links, so that the transport layer is the network bottleneck. This text explains the major challenges involved in using TCP for high-speed networks and briefly describes some of the variations of TCP designed to overcome these challenges.