| The popularity of resource-hungry applications is at record high and is increasing, creating a data consumption boom that is growing at a rapid rate. As a result, resource-constrained networks need to keep up with this demand for data and bandwidth, while guaranteeing accessibility, reliability, and speed to maintain a satisfactory level of end-to-end performance for all users and applications. To achieve this goal, much effort has been put into optimizing for network performance, including optimizing Web applications to adapt to network conditions, designing new transport protocols that better fit modern applications' requirements, and applying in-network management techniques by network operators for better handling of traffic loads. However, many of aforementioned solutions do not go through sufficient evaluation, resulting in poor understanding of their implications or how well they interact with other optimization efforts. This can cause gaps between intended and actual performance of applications across a range of environments. Moreover, some of these approaches disrupt the Internet's openness and neutrality. Further frustrating such scenarios is the lack of visibility into networks, making it very difficult (or impossible) to pinpoint the root causes of poor performance or detect open Internet violations.;This thesis posits that an effective way to address these issues is to create frameworks and tools for rigorous evaluation of approaches for optimizing network performance. By building such tools, we can detect instances of inconsistency between intended and actual network behaviors, diagnose the root causes, and help developing techniques to address shortcomings.;Given the above hypothesis, I first introduce a framework for rigorous evaluation of rapidly evolving transport protocols. Using this framework, I conduct a comprehensive study viii on the QUIC protocol in a wide range of network conditions and environments. In these evaluations, I examine the protocol's fairness, perform head-to-head comparisons with TCP, investigate the effectiveness of in-network proxies on performance, and reason about observed weaknesses and strengths through code instrumentation and inferred state machines.;Next, I present a record-and-replay framework that utilizes the native VPN support of modern operating systems to record and replay traffic from arbitrary applications. Through two measurement studies on traffic differentiation and a zero-rating service by a US operator, I demonstrate how my record-and-replay tool can be used to reveal application-specific in- network policies and shedding light on their implications. My approach for detecting traffic differentiation addresses limitations of prior work by proposing a more accurate statistical test, validating the methodology using real shaping boxes, and enabling testing of arbitrary applications directly from users' computers or mobile devices. My evaluation of a number of mobile Internet service providers revealed instances of differential policies. Finally, I investigated Binge On, a zero-rating service offered by T-Mobile USA, both to characterize its impact on performance as well as understand how the policy is implemented. I discovered a variety of behaviors that have important implications for subscribers, and demonstrate that Binge On's current simple implementation is brittle, potentially inaccurate, and easily subverted by malicious users to free-ride on T-Mobile. |
