Observing and Improving the Reliability of Internet Last-mile Links

Title of Document: OBSERVING AND IMPROVING THE RELIABILITY OF INTERNET LAST-MILE LINKS Aaron David Schulman, Doctor of Philosophy, 2013 Directed By: Professor Neil Spring Department of Computer Science People rely on having persistent Internet connectivity from their homes and mobile devices. However, unlike links in the core of the Internet, the links that connect people’s homes and mobile devices, known as “last-mile” links, are not redundant. As a result, the reliability of any given link is of paramount concern: when last-mile links fail, people can be completely disconnected from the Internet. In addition to lacking redundancy, Internet last-mile links are vulnerable to failure. Such links can fail because the cables and equipment that make up last-mile links are exposed to the elements; for example, weather can cause tree limbs to fall on overhead cables, and flooding can destroy underground equipment. They can also fail, eventually, because cellular last-mile links can drain a smartphone’s battery if an application tries to communicate when signal strength is weak. In this dissertation, I defend the following thesis: By building on existing infrastructure, it is possible to (1) observe the reliability of Internet last-mile links across different weather conditions and link types; (2) improve the energy efficiency of cellular Internet last-mile links; and (3) provide an incrementally deployable, energy-efficient Internet last-mile downlink that is highly resilient to weather-related failures. I defend this thesis by designing, implementing, and evaluating systems. First, I study the reliability of last-mile links during weather events. To observe failures of last-mile links, I develop ThunderPing—a system that monitors a geographically diverse set of last-mile links without participation from providers or customers. So far, ThunderPing has collected 4 billion pings from 3.5 million IP addresses over 400 days of probing from PlanetLab hosts. Because pings may fail to solicit a response even when a last-mile link has not failed, losses must be analyzed to determine if they constitute last-mile link failures. Among other challenges I encountered in this project, I found that determining the connectivity state from noisy pings is similar to finding the edges in a noisy picture. As such, I use algorithmic edge detection to find when a host transitions between connectivity states. By matching these connectivity states with weather reports from weather stations at airports, I observe how weather affects last-mile link failure rate and failure duration. Second, I improve the reliability of cellular links by reducing wasted energy. To do so, I develop Bartendr, a system that predicts when a moving smartphone will experience high signal strength. A key challenge is to predict high signal strength without consuming more energy than exploiting it would save. I also develop energy-aware scheduling algorithms for different application workloads—syncing and streaming—based on these predictions. I evaluate the scheduling algorithms with a simulation driven by traces obtained during actual drives. Third, I design a reliable broadcast system that is inexpensive to deploy to many users and is energy-efficient to receive. I adapt reliable FM Radio Data System (RDS) broadcasts to act as an Internet last-mile link. To accomplish this, I design and implement an over-the-air protocol, receiver software, and a hardware bridge for incremental deployment. I implement the full end-to-end system, deploy it on a 3 kW commercial FM radio station in a metropolitan area, and evaluate the loss rate, energy consumption, and synchronization on either a smartphone or on my new hardware bridge. The results indicate that the full end-to-end system can be reliable, a smartphone receiver can sleep between desired broadcasts, and two receivers tend to deliver the same broadcast within about 5 ms. OBSERVING AND IMPROVING THE RELIABILITY OF INTERNET LAST-MILE LINKS