Additive increase early adaptive decrease mechanism for TCP congestion control

Abstract

Due to the fundamental end-to-end design principle, today TCP/IP congestion control algorithm implements an additive increase multiplicative decrease (AIMD) probing algorithm. We propose a new mechanism that we call additive increase early adaptive decrease (AIEAD). The AIEAD algorithm estimates both the available bandwidth and the queue backlog in a end-to-end fashion: the backlog estimate aims at bounding queue lengths, detecting congestion before overflow and discriminating congestion losses from losses that are due to unreliable links; the bandwidth estimate aims at adaptively setting control windows after congestion detection by taking into account the used bandwidth at the time of congestion (early adaptive decrease phase). An implementation of the AIEAD paradigm, which we call New Westwood (NW), is tested and compared with Reno, and Vegas TCP using the ns-2 simulator. Simulations show that NW significantly improves fairness and goodput with respect to both Reno and Vegas, whereas in a last-hop wireless lossy link scenario NW significantly improves the goodput.