End-to-end loss discrimination for improved throughput performance in heterogeneous networks
Nguyen Nguyen, En‐hui Yang · 2006
In heterogeneous networks, packet losses arise as a result of both congestion and random transmission errors. In these networks, the Transmission Control Protocol (TCP) performs poorly because it assumes all packet losses are caused by congestion and subsequently throttles its transmission rate unnecessarily. This motivates the need to discriminate between different types of packet loss. In this paper, we propose a novel loss discrimination algorithm. Its design guided by a queueing analysis, our algorithm is based on a unique definition of a customer, Lindley's Equation and normalized least-mean square (LMS) prediction. It is accurate, efficient and can be incorporated into any transport layer protocol that employs a congestion control strategy triggered by packet loss. Our simulation results show that when our algorithm is implemented an extension of TCP Reno, throughput can be more than doubled under certain conditions.