To keep up with the continuous growth in demand, cloud providers spend millions of dollars augmenting the capacity of their wide-area backbones and devote significant effort to efficiently utilizing WAN capacity. A key challenge is striking a good balance between network utilization and availability, as these are inherently at odds; a highly utilized network might not be able to withstand unexpected traffic shifts resulting from link/node failures. We advocate a novel approach to this challenge that draws inspiration from financial risk theory: leverage empirical data to generate a probabilistic model of network failures and maximize bandwidth allocation to network users subject to an operator-specified availability target. Our approach enables network operators to strike the utilization-availability balance that best suits their goals and operational reality. We present TEAVAR (Traffic Engineering Applying Value at Risk), a system that realizes this risk management approach to traffic engineering (TE). We compare TEAVAR to state-of-the-art TE solutions through extensive simulations across many network topologies, failure scenarios, and traffic patterns, including benchmarks extrapolated from Microsoft's WAN. Our results show that with TEAVAR, operators can support up to twice as much throughput as state-of-the-art TE schemes, at the same level of availability.

Recently, much attention has been devoted to the question of whether/when traditional network protocol design, which relies on the application of algorithmic insights by human experts, can be replaced by a data-driven (i.e., machine learning) approach. We explore this question in the context of the arguably most fundamental networking task: routing. Can ideas and techniques from machine learning (ML) be leveraged to automatically generate "good" routing configurations? We focus on the classical setting of intradomain traffic engineering. We observe that this context poses significant challenges for data-driven protocol design. Our preliminary results regarding the power of data-driven routing suggest that applying ML (specifically, deep reinforcement learning) to this context yields high performance and is a promising direction for further research. We outline a research agenda for ML-guided routing.

Recent studies have observed that large data center networks often have a few hotspots while most of the network is underutilized. Consequently, numerous data center network designs have explored the approach of identifying these communication hotspots in real-time and eliminating them by leveraging flexible optical or wireless connections to dynamically alter the network topology. These proposals are based on the premise that statically wired network topologies, which lack the opportunity for such online optimization, are fundamentally inefficient, and must be built at uniform full capacity to handle unpredictably skewed traffic.

Despite extensive efforts to meet ever-growing demands, today's datacenters often exhibit far-from-optimal performance in terms of network utilization, resiliency to failures, cost efficiency, incremental expandability, and more. Consequently, many novel architectures for high-performance datacenters have been proposed. We show that the benefits of state-of-the-art proposals are, in fact, derived from the fact that they are (implicitly) utilizing "expander graphs" (aka expanders) as their network topologies, thus unveiling a unifying theme of these proposals. We observe, however, that these proposals are not optimal with respect to performance, do not scale, or suffer from seemingly insurmountable deployment challenges. We leverage these insights to present Xpander, a novel datacenter architecture that achieves near-optimal performance and provides a tangible alternative to existing datacenter designs. Xpander's design turns ideas from the rich graph-theoretic literature on constructing optimal expanders into an operational reality. We evaluate Xpander via theoretical analyses, extensive simulations, experiments with a network emulator, and an implementation on an SDN-capable network testbed. Our results demonstrate that Xpander significantly outperforms both traditional and proposed datacenter designs. We discuss challenges to real-world deployment and explain how these can be resolved.

Many architectures for high-performance datacenters have been proposed. Surprisingly, recent studies show that datacenter designs with random network topologies outperform more sophisticated designs, achieving near-optimal throughput and bisection bandwidth, high resiliency to failures, incremental expandability, high cost efficiency, and more. Unfortunately, the inherent unstructuredness and unpredictability of random designs pose serious, arguably insurmountable, obstacles to their adoption in practice. Can these guarantees be achieved by well-structured, deterministic datacenters? We provide a surprising affirmative answer. We show, through a combination of theoretical analyses, extensive simulations, and experiments with a network emulator, that any "expander" network topology (as indeed are random graphs) comes with these benefits. We leverage this insight to present Xpander, a novel deterministic datacenter architecture that achieves all of the above desiderata while providing a tangible alternative to existing datacenter designs. We discuss challenges en route to deploying Xpander (including physical layout, cabling costs and complexity, backwards compatibility) and explain how these can be resolved.

Software-defined networking (SDN) is a new paradigm for operating and managing computer networks. SDN enables logically-centralized control over network devices through a "controller" software that operates independently from the network hardware, and can be viewed as the network operating system. Network operators can run both inhouse and third-party SDN programs (often called applications) on top of the controller, e.g., to specify routing and access control policies. SDN opens up the possibility of applying formal methods to prove the correctness of computer networks. Indeed, recently much effort has been invested in applying finite state model checking to check that SDN programs behave correctly. However, in general, scaling these methods to large networks is challenging and, moreover, they cannot guarantee the absence of errors.