High throughputs and information freshness over the internet via transport layer advancements

Abstract

Next-generation applications such as augmented reality, virtual reality, teleoperated driving, and video streaming, along with a wide spectrum of real-time monitoring and actuation applications, are expected to challenge the current Internet on at least two fronts. Many of these applications desire high throughput and reliability at low end-to-end path latencies. To better support them, we must optimize the joint use of the diversity of high link rate wireless access technologies commonly available at end-user devices. In addition, there is a burgeoning class of applications that requires the availability of fresh information (for example, sensor measurements and actuation commands in IoT applications) at the destination. The current Internet treats such applications no differently than it does those that care for throughput. In this thesis, we address the challenges posed by these distinct requirements of high throughput and high freshness via innovations at the transport layer of the networking stack. We address the challenge posed by applications that require high end-to-end throughputs via a novel cross-layer scheduler, QAware, for Multipath TCP (MPTCP). The QAware scheduler uses local queue occupancy information for every access network available on a user device in addition to the typically used end-to-end round-trip delay estimates. This results in a more efficient use of the available interfaces and considerable gains in aggregate throughput compared to other MPTCP schedulers for a varied set of applications and over heterogeneous access networks. For real-time monitoring and remote sensing applications, we address the challenge of enabling freshness, as quantified by the metric of age-of-information (AoI) over an end-to-end Internet path. Specifically, we propose and detail the Age Control Protocol (ACP) and its improved version called ACP+. Both use ACKs to maintain an estimate of the number of unacknowledged packets in the system along with the end-to-end RTT. This, together with an estimate of the time average age of updates is used to determine an ACP source’s update rate. We study the efficacy of ACP and ACP+ using extensive simulations and real-world experiments over the Internet. To gain further insight into age control, we also empirically compare ACP+ with a mix of loss-based, delay-based, and hybrid congestion control algorithms used by TCP. TCP tries to fill the network pipe using estimates of bottleneck rate and baseline RTT, but these estimates may not shed as much light on the age-optimizing update rate. In our experiments over paths in the Internet, ACP+ utilizes only a fraction of the bottleneck link rate for achieving low age. When the path had a wireless access as its first hop that was the bottleneck link, the path beyond the access, with links much faster than the access, was the constraining factor with regards to minimizing age over the end-to-end path. Age being optimized at update rates much lower than typical bottleneck access link rates has interesting consequences for end-to-end flows sharing the wireless access to send updates to the cloud. We experimented with a large number of ACP+ sources (up to 80 sources and fixed physical layer rates of 6, 12 and 24 Mbps) sharing a WiFi access point to send their updates to a cloud server over the Internet. We show that ACP+ allows sources to share the access well. When the wireless access isn’t the constraining factor, given the low age minimizing rate over the end-to-end path, all sources send at the minimizing rate. As the number of sources increases, the resulting congestion over the WiFi access has ACP+ gradually reduce the rate of updates per source in a manner such that the sources together fully utilize the WiFi access link rate. While fully utilizing the access like TCP, ACP+, however, keeps age much lower than TCP congestion control algorithms. Even the packet retry rates because of collisions over WiFi are much lower than for TCP. In fact, TCP algorithms are unsuitable for age control, which we demonstrate using simulations and real-experiments in this thesis. On the other hand, ACP+’s behavior, as determined using controlled simulations, is in line with what would be expected of a good age control strategy enabling sharing of access amongst multiple sources. User devices are expected to support a mix of applications, some of which may care for high throughput and others for the freshness of information. We conclude this thesis with a study on the coexistence of ACP+ and TCP flows sharing an end-to-end Internet path over a WiFi access. In line with expectation, ACP+ flows coexisting with TCP flows remain unaffected when assigned a higher Differentiated Services Code Point (DSCP) priority when all flows originate in the same device. However, the gains from prioritization vanish when the flows are instead sharing a contended wireless access.