Combining Explicit Flow Measurements and Network Programmability for Network Management
Abstract
The Internet has evolved into an integral part of modern society and is essential for many applications. Therefore, ensuring its robustness is vital. Yet, managing “the Internet” remains a complex challenge, as it comprises thousands of autonomous and diverse networks. To facilitate common management solutions, the constituent networks have gradually and subtly converged around a core set of protocols, essentially disallowing deviations in the process. While this ossification helps network management strengthen the Internet’s reliability, it also comes at a cost: innovation needed to adapt the Internet to contemporary requirements is significantly hindered. Aiming to counter this trend, many Internet giants jointly standardized QUIC, a protocol designed to foster innovation and combat ossification. With these traits, QUIC revitalizes the Internet but also disrupts a well-established ecosystem, complicating network management and calling many management concepts into question. In this dissertation, we focus on two resulting key challenges for network management. First, QUIC flows expose little information, so operators require novel solutions to assess the performance of QUIC traffic in their networks. Second, QUIC is implemented in user space, leading to more diverse protocol behaviour, which needs to be accommodated by more adaptive and flexible network structures. Our work shows that two complementary technologies — explicit flow measurements (EFM) and network programmability — are well-suited for efficiently addressing these challenges and represent promising key enablers for managing a more diverse Internet. Using a combination of theoretical considerations and practical simulation, testbed, and Internet experiments, we first find that EFM estimations indeed reasonably reflect the performance of QUIC traffic and that the one EFM scheme standardized as an actual QUIC feature — the latency spin bit — is already used by real web traffic. We then show that we can monitor the spin bit at line rate on programmable hardware and that the EFM monitoring output can be used to assess the network state systematically with network tomography. Finally, we focus on how EFM and network programmability can help cope with diverse protocol behaviour. In particular, we demonstrate that programmable hardware is a promising platform for congestion management algorithms, although specific constraints require careful designs. Additionally, we conceptualize two novel congestion management solutions that leverage EFM to (i) identify misbehaving flows and (ii) shield benign flows from misbehaving traffic, thus incentivizing using EFM and proper congestion control. In summary, this dissertation highlights the transformative potential of combining EFM with network programmability and provides key input for standardization and operators. With EFM, operators regain visibility into crucial performance metrics, and they can systematically assess their networks and deploy novel solutions that help face evolving protocol dynamics and motivate using mechanisms for the common good. Complementarily, network programmability represents a new vehicle for network management, enabling faster development and deployment by avoiding costly hardware upgrades, thus increasing the adaptiveness and flexibility of solutions in high-speed networks. Together, EFM and network programmability pave the way for effective network management in a revitalized and increasingly complex Internet.
