Evaluating and Reducing Multipath Transport Latency

Abstract: Access to the Internet is a very significant part of everyday life with increasing online services such as news delivery, banking, gaming, audio and high quality movies. Applications require different transport guarantees with some requiring higher bandwidth and others low latency. Upgrading access link capacity does not guarantee faster access to the Internet as it offers higher bandwidth but may not offer low latency. With increasing number of mobile devices supporting more than one access technologies (e.g., WLAN, 3G, 4G,..), there is a need to analyse the impact of using multiple such technologies at the same time. Legacy transport protocols such as TCP or SCTP are only able to connect to one access network at a time to create an end-to-end connection. When more than one access technology is used, there may be a large difference in the data rate offered by each technology. This asymmetry might impact latency sensitive applications by creating out of order delivery. In this thesis, we focus on the latency aspect of multipath transport protocol performance. We consider CMT-SCTP and Multipath TCP as available multipath protocols that were designed for exploiting multiple paths for better throughput and reliability. We consider various real world traffic scenarios such as Video, Gaming and Web traffic to measure end-to-end latency. We perform simulations, emulations and experiments using heterogeneous network settings involving access networks with different bandwidth, delay and loss characteristics. MPTCP performs better in terms of latency than CMT-SCTP and TCP in certain scenarios where available paths are symmetric. However, MPTCP does not perform well in asymmetric scenarios with latency sensitive traffic. This analysis provides insights in to various areas of improvement in MPTCP such as scheduling and loss recovery to achieve low latency. We further focus on packet loss recovery in MPTCP for specific cases of tail losses to reduce latency. Tail losses are the losses that occur at the end of a packet stream. Recovering such losses is of higher significance to latency sensitive applications. We propose a modification to the use of TLP, a mechanism in TCP for tail loss recovery. We evaluate the performance of proposed TLP modification, first using emulations and with real world network experiments. Our results show significant improvements in latency for specific loss scenarios in emulations and up to 50% improvement in experiments.