Lecture 7: Congestion control

Last lecture, we looked at the phenonmenon of congestion collapse, i.e., what happens when you pick a very large window size in the sliding window protocol. The problem of congestion collapse on the Internet was first documented by John Nagle in 1984 [2]. It was also observed by Van Jacobson in 1986 [8]. In both cases, the congestion collapse was the result of a large number of packets being prematurely retransmitted. The network was still doing work and heavily utilized, but not all of it was contributing to useful transport-layer throughput because the receiver was receiving duplicate copies of the same packet. Figure 3 of Van Jacobson’s paper [8] illustrates this problem beautifully. Quoting the caption of Figure 3 from that paper: “The dashed line shows the 20 KBps bandwidth available for this connection. Only 35% of this bandwidth was used; the rest was wasted on retransmits.” The solution to the Internet’s congestion collapse problem was a congestion-control algorithm. This algorithm was first implemented on end hosts by Van Jacobson in the BSD Operating System around 1986-87 and published in 1988 [8]. This algorithm is today called AIMD (Additive Increase Multiple Increase); the reason for this will become apparent soon. AIMD allows a sliding window sender to adaptively find a window size equal to the BDP of the network. This is unlike what we have done so far, where we hardcoded a fixed window size into the sliding window protocol. AIMD was first developed by Ramakrishnan and Jain at Digital Equipment Corporation (DEC) in an algorithm that later came to be called DECbit [9]. However, Ramakrishnan and Jain’s evaluation was confined to simulated networks and required changes to the routers—something that was quite antithetic to the Internet’s ethic of minimalism in the routers. Van Jacobson then adapted the same algorithm for use in the Internet without modifying the routers and evaluated it on a real network. DECbit was eventually adopted and standardized as the Explicit Congestion Notification (ECN) mechanism in routers in 2001. But it took a while, unlike Van Jacobson’s approach, which was implemented in every Internet end host by the early 1990s. This is another example of keeping things simple on the Internet: by minimizing the number of entities that need to be modified, Van Jacobson’s approach was more readily adopted. We’ll discuss AIMD in more detail now. There are two main parts to this AIMD algorithm: