Along with exciting visions for 5G communications and the Tactile Internet, the networking requirement of attaining extremely low end-to-end latency has appeared. While network devices are typically equipped with buffers to counteract packet loss caused by short-lived traffic bursts, the more those buffers get filled, the more delay is added to every packet passing through. In this paper, we develop congestion avoidance methods that harness the power of fully programmable data-planes. The corresponding programmable switches, through languages such as P4, can be programmed to gather and react to important packet meta-data, such as queue load, while the data packets are being processed. In particular, we enable P4 switches to (1) track processing and queuing delays of latency-critical flows and (2) react immediately in the data-plane to congestion by rerouting the affected flows. Through a proof-of-concept implementation in emulation and on real hardware, we demonstrate that a data-plane approach reduces average and maximum delay, as well as jitter, when compared to non-programmable approaches.@en

Throughout the last decades, communication networks have become embedded into almost every aspect of our day-to-day lives (e.g., watching movies, online shopping, sharing moments with friends and family). Moreover, as the support for the transport of audio and video became the norm, new application domains have kept emerging every day. One of these, the Tactile Internet, enables the transport of the sense of touch. Consequently, it allows the end-users to interact with a remote environment in the same way they would if they were present locally. While such applications could revolutionize many industries by enabling users to transport their skills (e.g., surgical skills) across the globe, they pose many new challenges to communication networks, such as the need for very low latency. Yet, providing low latency is fundamentally different from providing high bandwidth, and, as this thesis demonstrates, existing solutions developed for bandwidth-oriented services are not directly applicable to low-latency services. This thesis explores how programmable networks can be used to facilitate emerging low-latency services. Specifically, it combines the advantages of (1) Software-Defined-Networking (SDN), a paradigm in networking that centralizes the control plane, and (2) programmable data planes, which enable an on-the-fly deployment of novel algorithms to the network switches. In particular, this thesis explores what SDN controller tasks are feasible to be offloaded to the data plane, the trade-offs in doing so, and their benefits on low-latency applications. Moreover, it takes advantage of the more fine-grained monitoring possibilities of programmable data planes and incorporates these measurements into the data plane algorithms. As a result, this thesis develops a set of solutions that enable network switches to react to short-term changes in the networking traffic and act independently (or with limited input), improving the Quality of Service (QoS) of low-latency flows. First, we investigate the limitations of programmable switches and ways to overcome them by developing an application to detect heavy hitters (e.g., flows that consume most resources in the network). Next, we explore the concept of network slicing, i.e., reserving part of a physical network for a specific service. We demonstrate that network switches can combine data plane measurements and limited (preconfigured) input from the central controller to enable elasticity, i.e., the ability to automatically scale the assigned network resources based on the flows' requirements with negligible delay. Next, we analyze the co-existence and interactions between flows using different congestion control algorithms and/or having different RTTs. We use this information to develop a data plane algorithm to improve their interactions. Finally, we demonstrate how congestion detection and avoidance can be achieved in the data plane without any assistance from the end-hosts. @en

The Tactile Internet presently constitutes a vision of an Internet over which, in addition to current communications modalities, a sense of touch can be transported. In that case, people would no longer need to be physically near the systems they operate, but could control them remotely. The main problem that needs to be solved to realize the Tactile Internet is summarized by the “1 ms challenge.” If the response time of a system is below 1 ms, the end-user will not be able to tell the difference between controlling a system locally or from another location. This paper offers a summary of the requirements for haptic communications, followed by an overview of challenges in realizing the Tactile Internet. In addition, possible solutions to these challenges are proposed and discussed. For example, the development of the fifth generation mobile communication networks will provide a good foundation upon which a Tactile Internet could be built. This paper also describes the design of a modular testbed needed for testing of a wide variety of haptic system applications.@en

Networks often need to concurrently process mil- lions of flows with varying Quality-of-Service (QoS) require- ments. Doing so by deploying flow-specific rules at network nodes would require significant memory and overhead. In this paper, we take a fundamentally different approach, called P4QoS, by embedding QoS requirements in the packets themselves and leveraging P4-programmable network switches to process the traffic based on them. We illustrate and evaluate our approach with latency as our QoS metric, but our concept can be applied to other metrics as well. Our evaluation, both in software (Mininet) and in hardware (Intel Tofino), shows that P4QoS can satisfy application-specific QoS requirements with negligible memory overhead.@en

The concept of network slicing enables operators to provision multiple virtual networks on top of a single (shared) physical infrastructure. Adding elasticity to slicing, i.e., the ability to on-demand provision/release dedicated network resources, improves resource utilization. However, efficiently allocating and scaling slice resources, while maintaining state consistency, is challenging. Especially with P4-programmab1e network devices that process packets at Tbps speeds, controller-driven scaling of network functions would be too time-consuming, and data-plane scaling is needed. In this paper, we address this need, by developing a custom scaling protocol and framework that can consistently, with negligible delay, scale network slices and functions transparently to the slice end-users. We compare, via emulation and experiments on programmable hardware, our approach to state-of-the-art scaling techniques and demonstrate significant slice resource utilization improvements and scaling duration reductions.@en

Congestion control algorithms are usually developed in isolation without thoroughly investigating their co-existence and interactions with other protocols and/or congestion control algorithms. As a result, flows using different algorithms and/or having different Round-Trip Times may overpower each other, resulting in unfair resource distribution, with a subset of the flows usually claiming most of the capacity. To solve the aforementioned problem, we make use of pro- grammable switches and the network programming language P4 to enforce fairness from within the network itself, instead of from the congestion control algorithms ran at the end-points. Our solution P4air continuously monitors the properties of all flows that pass through a switch and groups them based on the behavior of the congestion control algorithms used. Furthermore, for each group, it applies appropriate measures to suppress the aggressive flows and boost smaller flows. Our experiments, using modern programmable hardware (Barefoot Tofino switch), demonstrate significant performance gains for P4air in terms of fairness compared to state-of-the-art solutions.@en

Congestion control algorithms are crucial in achieving high utilization while preventing overloading the network. Over the years, many different congestion control algorithms have been developed, each trying to improve over others in specific situations. However, their interactions and co-existence has, to date, not been thoroughly evaluated, which is the focus of this paper. Through head-to-head comparisons of loss-based, delay-based and hybrid types of congestion control algorithms, we reveal that fairness in resources claimed is often not achieved, especially when flows sharing a link have different round-trip times or belong to different groups. @en

F1ows that have exceeded a given percentage of the last sliding window of N packets, denoted as heavy-hitter flows, require special handling, since they may disrupt the service of other flows or may be indicative of malicious traffic. However, even when equipped with a programmable switch, it is unclear how to detect heavy hitters on a per-packet basis, while obeying the stringent switch memory access rates. For instance, existing solutions, such as HashPipe, cannot detect heavy hitters without halving the line rate and do not support sliding windows. To the best of our knowledge, this paper is the first to present heavy-hitter detection solutions that provide per-packet granularity at line-rate performance. We realize this by introducing (1) Modulo sketching, a novel counting algorithm that reuses counters and limits the impact of smaller flows beyond early processing stages; and (2) Sequential Zeroing, a new approach to extending interval-based schemes to sliding window measurements. Our solutions are extensively evaluated, both via simulations and experiments on a Netronome SmartNIC, and demonstrate significant performance gains over the state-of-theart.@en

“Tactile internet” refers to a network that can support real-time interactions between human operators and remote cyber-physical systems as if they were near to each other. For this, the network should support ultra-low latency communication, often referred to as the 1ms challenge. However, we observe that network requirements, such as latency and band- width, of tactile internet based cyber-physical systems or Tactile Cyber-Physical Systems (TCPS) are not static: they severely fluctuate over time. Therefore, for TCPS, static provisioning of network resources is sub-optimal. For optimal utilization of network resources, we propose a mechanism to, per TCPS flow, dynamically create, destroy and switch network slices, based on the network resources needed at that time. Our solution consists of two main components. First, we develop a clustering algorithm to determine the slices and their specifications required to support a TCPS flow. Second, we leverage Software-Defined Networking (SDN) and P4-programmable switches to enable on- the-fly provisioning and switching of these slices.@en