A Practical System for Revealing ISP Traffic Differentiation

Traffic differentiation empowers network operators to give different levels of service to different types of network flows, and can be used to manage QoE for end users. Without regulation or accountability, such network management practices could be used to raise the barrier to entry for new technologies, or block them entirely. As a result, the FCC recently passed new rules making differentiation illegal in broadband networks [2]. Further, by breaking end-to-end system design, these practices can have negative side-effects on reachability, reliability, and performance. We are currently implementing a systematic approach to understand and expose traffic differentiation and its impact on the Internet, with the following main contributions. First, we have studied real traffic differentiation devices to understand the range of their behaviors, and have developed a testbed to test the effectiveness of our measurement techniques in a controlled environment. Second, we have designed and deployed systems to detect differentiation efficiently, to operate in both resource-constrained mobile environments and fixed-line settings. With a large-scale deployment that includes incentives for adoption by users and content providers, we will characterize the impact of traffic differentiation on network performance. Last, we are developing a model to understand how to detect traffic shaping implementations in general, and analyze large-scale impacts of differentiation. Key challenges for differentiation detection. Our work addresses three key challenges that limit our understanding of traffic differentiation. First, researchers often struggle to characterize traffic differentiation because the products that implement it are generally closed systems, operators rarely acknowledge the presence of middleboxes in networks, and researchers have limited or no access to experiment with such middleboxes. Second, identifying and understanding differentiation behavior often requires measurements from vantage points inside the network (e.g., from end users and on mobile devices), which limits the coverage and feasibility of any study that uses a fixed testbed (e.g., PlanetLab or lab environments) and necessitates the development of easy-to-deploy systems to detect traffic differentiation. Finally, we lack models of how traffic differentiating products operate on network traffic, making it difficult to reason about interactions between multiple devices on a path. A flexible, realistic testbed for experimenting with differentiation detection. We have studied devices that implement differentiation to understand their behavior, and developed a testbed that allows us to validate our measurement techniques in a controlled environment. A key challenge for detecting differentiation is how to support controlled experiments on network traffic from representative applications. Prior work focused on specific applications (e.g., BitTorrent [1]) and manually generated traffic to match the target application. This approach, however, does not scale to large numbers of applications and may even be infeasible when the target application uses proprietary, or otherwise closed, protocols. To address this, we use a trace record and replay approach [4]. We begin by recording a packet capture of the application. The packet capture is then parsed to extract the application layer messages exchanged. Importantly, we capture the application-level client–server interaction, which does not rely on properties of the underlying network (e.g., packet loss, delay) used for recording. The corresponding messages are then replayed between a client and server running our replay software. Importantly, because our approach reproduces payloads from the application layer, it supports both encrypted (e.g., HTTPS, XMPP) and plaintext traffic. For example, our approach precisely replicates the encrypted stream of data recorded from Netflix traffic, without needing to know the corresponding plaintext. We have implemented this approach for six popular apps, and are developing support to create replay traces for arbitrary user-submitted traffic. We have leveraged two commercial packet-shaping devices that we have acquired to conduct grey-box testing of differentiation, and we are working to acquire others. To broaden the utility of our testbed, we will