Increasing quality of service using transport layer coding over parallel heterogeneous networks

Abstract

A shift in the way the Internet is accessed and the type of traffic generated has been underway for the last decade and will continue into the foreseeable future. Furthermore, these shifts are expected to place considerable strain on the ability of existing transport layer protocols to meet users' quality of service demands. Internet traffic is being generated by an increasing number of inexpensive, low-power, mobile devices resulting in an increased incidence of packet losses, unreliable network connections, and path instability. In addition, the majority of traffic, such as streaming video, is fast becoming more delay sensitive driving changes throughout the network. This thesis explores methods to improve end-to-end performance through the use of network and erasure coding techniques applied at the transport layer. These techniques will help provide path diversity and connection resiliency in the form of multi-path transport, in addition to increasing throughput and decreasing in-order delivery delay in the presence of network disruptions and packet erasures. The potential gains of a network coded transport layer are first introduced using two protocols: Coded TCP (CTCP) and Multi-Path TCP with Network Coding (MPTCP/NC). These two approaches help illustrate different methods to implement network coding to help overcome packet losses and diversify network connections over multiple paths. Analytic and experimental results are provided that show a network coded transport layer can significantly improve throughput in challenged networks; but more importantly, these results hint at the possible gains for application layer quality of service. Observing the performance of HTTP requests and streaming video, the potential to decrease the in-order delivery delay is highlighted. Based off of the observations outlined above, the second half of the thesis explores different code constructions that reduce in-order delivery delay in multiple path networks. A generation-based systematic code construction, similar to the one used in CTCP, is modeled and analyzed. Of particular note is the inherent trade-off between increasing the generation size to reduce the probability of decode errors resulting in retransmissions and keeping the generation size small enough so that coding delay is minimized. Furthermore, the trade-off between delay and efficiency is explored to help quantify the cost of reducing delay so that a user's quality of service constraint is met. Finally, a multiple path streaming code construction that removes the artificial constraints imposed by partitioning packets into generations is explored. Comparisons between the two constructions help highlight the importance of feedback in reducing delay, as well as the advantages and disadvantages of one type of code construction over the other.