Flow control in computer networks

Abstract

In this dissertation, we examine the problem of flow control in packet-switched communication networks. Flow control mechanisms are used by transport protocols to control the flow of data on a connection. The objective of this control is to ensure that resources at the intermediate nodes and at the destination are not overwhelmed. We initially consider a deterministic model of a virtual circuit. This model is extended to reflect the effect of cross traffic on data flow. The analysis of this model leads us to the design of an adaptive flow control scheme in which the times to send each packet is decided explicitly. The scheme is called send-time control. The approach used to determine the time to send a packet uses information on transit time, and bottleneck service time of the virtual circuit. Since this information may be either delayed or not available at all, these parameters have to be estimated based on the information available at the source (send and receive times of the packets). We have developed estimation techniques to adapt send-times of packets to the network conditions. We designed and implemented a transport protocol, DTP, which incorporates send-time control scheme. DTP provides end-to-end, connection-oriented, and reliable data transfer service between pairs of processes. DTP uses both selective and cumulative acknowledgements. It uses selective retransmission of lost packets. We have studied network dynamics and conducted experiments on the Internet to better understand the dynamic characteristics of the Internet, and take them into account in the design and implementation of the flow control scheme. We have shown that many commonly held assumptions about the network behavior are not always true. For example, we have noticed that packet losses are not necessarily because of buffer overflow. Based on our network dynamics observations, we have refined the estimation techniques used in send-time control. We have developed heuristics to reflect these observations in the estimation process. These heuristics have been validated using simulation. We have used simulation to compare TCP and DTP performance under a variety of network conditions. DTP performs much better in all the scenarios considered. We have also conducted experiments on the Internet comparing DTP with current protocols and observed a significant improvement in throughput when using DTP.