Design and Evaluation of a Suite of Congestion Control Algorithms for the Future Internet

Linear Increase/Multiplicative Decrease (LIMD) has typically been the congestion control paradigm of choice in the traditional Internet with tail-drop routers. In recent years, routers with fair packet dropping policies (e.g., RED) are increasingly being deployed in the Internet. In the context of a network with fair packet dropping routers, we challenge the popular notion that Linear Increase/Linear Decrease (LILD) is an unsuitable paradigm for congestion control, and present two key results: (a) LILD is both eecient and fair in the target environment. In fact, LILD outperforms LIMD in most respects. (b) By maintaining the history of transmission rates and loss prooles, congestion control algorithms can eeectively distinguish between congestion-induced and non-congestion-induced loss. We present the LILD/H algorithm that combines the above features. LILD/H reacts gently to random/channel probe loss, and aggressively to congestion loss, thereby adapting eeectively to the dynamics of the network. Consequently, we believe that LILD/H is a viable congestion control algorithm at the end host or the edge router in the future Internet. 1 Introduction Congestion control algorithms that are deployed in the In-ternet today typically use the linear increase/multiplicative decrease (LIMD) paradigm for adapting the transmission rate of a connection to reeect the available bandwidth (precisely , delay-bandwidth product). Brieey, LIMD periodically adapts the sending rate of a connection by gently increasing the rate upon observing no packet losses in order