Improving the performance of TCP applications using network-assisted mechanisms

In the current Internet, IP provides a “best effort”, unreliable packet delivery service. End-to-end reliability is provided by TCP. TCP also implements congestion control mechanisms, whereby data sources reduce their transmission rate after detecting packet loss. In this dissertation, we identify problems with the performance of TCP applications, which are attributed to TCP's reaction to packet loss. We address these problems with the assistance of network level mechanisms, without modifying TCP's operation. First, we consider switched Ethernet Local Area Networks, which are characterized by large link speed mismatches. In this context, we show that an unexpectedly low throughput can result from packet loss. In order to address this problem, we use a hop-by-hop back-pressure mechanism, as specified in the IEEE802.3x standard. We show that this mechanism can improve network performance in some situations, but leads to poor performance in others. We propose a selective scheme based on MAC address and traffic class information, which overcomes these limitations. Then, in the context of the Internet, we use large simulation scenarios and detailed application models to show how congestion-induced packet loss causes unacceptably large delays for interactive TCP applications. We address this problem using service differentiation, in the form of prioritized dropping in network queues. First, we consider giving priority to interactive applications' traffic in the network, and show that this significantly decreases their delays, albeit at the expense of non-interactive ones. Second, we present a marking scheme whereby packets are prioritized based on each connection's TCP window size. We show how this scheme results in good response times for short transfers, which are characteristic of interactive applications, without significantly affecting longer ones. Finally, we consider the future Internet, where TCP applications are expected to share the network with multimedia (e.g., video) applications which use UDP. We show how the user-perceived performance of both types of applications can be degraded as a result of this sharing. We then demonstrate how appropriate TCP marking and video layering can be used in association with prioritized dropping in network queues to obtain excellent performance for both, at times where it would have otherwise been unacceptable.