Abstract
Transmission Control Protocol (TCP) is a widely used end-to-end transport protocol on the internet. TCP schemes can be classified into two categories: marking/loss based (e.g., TCP Reno) or delay based (e.g., TCP Vegas, Fast TCP). Delay-based schemes have garnered much attention due to their higher network throughput than loss-based methods. However, the current delay-based TCP versions all suffer from reverse congestion problem. Moreover, when there are multiple bottleneck links or inadequate buffer sizes, delay information alone is insufficient for revealing the incipient network congestion.We start from a theoretical analysis based on Network Calculus (NC) and determine an optimal window size control method. We also show that the basic delay-based TCP congestion avoidance mechanisms can be viewed as different approaches to approximating the optimal NC controller, with each TCP variant making different assumptions in terms of parameter estimation and control implementation strategy. Therefore, the network calculus controller reveals the inherent underlying structure in TCP congestion avoidance methods based on time-variant delay feedback.Furthermore, we outline how the ideal non-causal NC F-model controller can be approximated, resulting in practical congestion avoidance methods with performance gains over other existing delay-based TCP flow control schemes. We propose an initial approximation of the NC controller, a Kalman Filter-based implementation and a practical solution based on a stationary Kalman filter. With an improved bandwidth estimation method, the Kalman Filter-based NC controllers present a marked improvement over all other delay-based TCP schemes.Finally, we consider the design of a congestion control scheme that transcends the two normal categories (marking/loss based or delay based) and instead jointly exploits both delay and marking/loss feedback. We introduce the concept of the normalized queuing delay, which serves as a combined congestion price of delay and marking information. Based on the normalized queuing delay, we go on proposing a solution for dynamically determining the buffer set-point, which is an open problem for various delay-based TCP protocols. We show that the proposed D+M (Delay+Marking) TCP can solve both the dynamic set-point problem and the backward congestion problem. The D+M TCP is able to achieve excellent performance and more versatility than some existing TCP versions as confirmed by comparative ns-2 simulations.