In the realm of cluster operation, continuously validating and optimizing the configuration requires access to accurate cluster behavioral models. Network Digital Twins (NDTs) have emerged as a paradigm to provide such accurate, live representations of network systems. To capture the live state, NDTs need to anticipate the cluster behavior in a faster than real-time manner. With increasingly complex clusters, classical NDTs relying on detailed handcrafted simulators become too slow to fulfill this task. Leveraging measurements from the actual system demonstrates the potential to create more highlevel, lightweight NDTs that are still fairly accurate. Nonetheless, the degree of abstraction required to create fast and accurate data-driven NDTs is not well understood. To address this, our work investigates the impact of different abstraction levels on modeling accuracy. We develop and compare three Network Digital Twins of a Kubernetes Cluster - a Twin based on a Handcrafted Simulator, a Decentralized Data-driven Twin, abstracting individual system components, and a Centralized Data-driven Twin, abstracting the system as a whole. Our results show that Data-driven Twins improve the performance prediction by 18-53% over the handcrafted one, with the Centralized Twin surpassing the Decentralized Twin in accuracy by 35% and speed by two orders of magnitude.

A significant portion of modern network traffic analysis still relies on human expertise only. To overcome human limitations in light of increases in volume, dynamicity, and overall traffic complexity, modern networks need to autonomously gain an understanding of traffic patterns and present them in an interpretable way. This work presents T-MAW, an approach for Traffic Monitoring and Analysis using Weighted Stochastic Block Models (WSBMs). T-MAW applies WSBMs to network data to create traffic characterizations in human-interpretable form. In addition to the insights gained from the fitted models, T-MAW evaluates unseen traffic against these models to perform anomaly detection. Both, network node behavior characterization and anomaly detection complement human expertise in modern network traffic analysis. As an example, we show how T-MAW can be used to create a behavior-based structured view of network nodes in a real campus network. In the anomaly detection context, we present results for an IP scan attack against the network, as well as from a layer-2 device fault that caused network disruption.

The diversity of deployed applications in data centers leads to a complex traffic mix in the network. Reconfigurable Data Center Networks (RDCNs) have been designed to fulfill the demanding requirements of ever-changing data center traffic. However, they pose new challenges for network traffic engineering, e.g., interference between reconfigurations, transport layer protocols, and congestion control (CC) algorithms. This raises a fundamental research problem: can the current transport layer protocols handle frequent network updates? This paper focuses on TCP and presents a measurement study of TCP performance in RDCNs. In particular, it evaluates diverse traffic mixes combining TCP variants, UDP, and QUIC transport protocols. The quantitative analysis of the measurements shows that migrated TCP flows suffer from frequent reconfigurations. The effect of reconfigurations on the cost, e.g., increased Flow Completion Time (FCT), depending on the traffic mix is modeled with Machine Learning (ML) methods. The availability of such a model will provide insights into the relationship between the reconfiguration settings and the FCT. Our model explains 88% of the variance in the FCT increase under different reconfiguration settings.

Researchers developing new architectures and algorithms for data center networks (DCNs) face the challenge of producing meaningful evaluations of their contributions. Traditional evaluation methods like traffic traces and parametric models can fail to reveal weak spots in DCNs. The concept of adversarial inputs shapes traffic data, making it challenging for a DCN to serve it. Adversarial traffic can provide insight into performance issues of a DCN that might go unnoticed with traces and models. This paper presents AdFAT, a genetic algorithm-based system for automated adversarial input generation for DCNs. While previous work focuses on reordering flow volumes or individual packets, our system uses flow arrival times as the adversarial traffic dimension. By creating adversarial flow arrivals for a demand-oblivious RotorNet topology, we show that AdFAT not only finds traffic that causes 22.64% higher mean flow completion times than traffic with uniform random arrival times but is also sensitive to the inherent periodicities and connection patterns of RotorNet. The results indicate AdFAT can find and exploit temporal and structural properties of dynamic and demand-oblivious topologies in an automated way.

Cluster orchestrators such as Kubernetes (K8s) provide many knobs that cloud administrators can tune to conFigure their system. However, different configurations lead to different levels of performance, which additionally depend on the application. Hence, finding exactly the best configuration for a given system can be a difficult task. A particularly innovative approach to evaluate configurations and optimize desired performance metrics is the use of Digital Twins (DT). To achieve good results in short time, the models of the cloud network functions underlying the DT must be minimally complex but highly accurate. Developing such models requires detailed knowledge about the system components and their interactions. We believe that a data-driven paradigm can capture the actual behavior of a network function (NF) deployed in the cluster, while decoupling it from internal feedback loops. In this paper, we analyze the HTTP load balancing function as an example of an NF and explore the data-driven paradigm to learn its behavior in a K8s cluster deployment. We develop, implement, and evaluate two approaches to learn the behavior of a state-of-the-art load balancer and show that Machine Learning has the potential to enhance the way we model NF behaviors.

Traffic Engineering (TE) mechanisms in data center networks make distributed forwarding decisions based on the global network state. Thus, new TE mechanisms require the design and implementation of effective information exchange and efficient decentralized algorithms to compute forwarding decisions, which is challenging and time-intensive. To automate and simplify this process, we proposeMistill.Mistilldistills the forwarding behavior of TE policies from exemplary forwarding decisions into a Neural Network.Mistilllearns (i) how to encode local state into update messages, (ii) which network devices must exchange updates, and (iii) how to map the exchanged updates into forwarding decisions. We demonstrate the abilities ofMistillby learning three TE policies, verifying their performance in simulations on synthetic and real-world traffic patterns, and by showing that the learned policies generalize to unseen traffic patterns. We implementMistillas a proof-of-concept and show thatMistillreacts on average within 1.3ms to changes in the network.

Today’s data centers are hosting various applications under the same roof. The diversity among deployed applications leads to a complex traffic mix in Data Center Networks (DCNs). Reconfigurable Data Center Networks (RD-CNs) have been designed to fulfill the demanding requirements of ever-changing data center traffic. However, they pose new challenges for network traffic engineering, e.g., interference between reconfigurations and congestion control (CC). This raises a fundamental research problem: can the current transport layer protocols handle frequent network updates?; This paper focuses on the Transmission Control Protocol (TCP) and presents a measurement study of TCP variants in RDCNs. The quantitative analysis of the measurements shows that migrated flows suffer from frequent reconfigurations. The effect of reconfigurations on the cost, e.g. increased Flow Completion Time (FCT), depending on the traffic mix is modeled with Machine Learning (ML) methods. The availability of such a model will provide insights into the relationship between the reconfiguration settings and the FCT. Our model explains 88% of the variance in the FCT increase under different reconfiguration settings.

With the advent of 6G and its mission-critical and tactile Internet applications running in a virtualized environment on the same physical infrastructure, even the shortest service disruptions have severe consequences for thousands of users. Therefore, the network hypervisors, which enable such virtualization, should tolerate failures or be able to adapt to sudden traffic fluctuations instantaneously, i.e., should be well-prepared for such unpredictable environmental changes. In this paper, we propose a latency-aware dual hypervisor placement and control path design method, which protects against single-link and hypervisor failures and is ready for unknown future changes. We prove that finding the minimum number of hypervisors is not only NP-hard, but also hard to approximate. We propose optimal and heuristic algorithms to solve the problem. We conduct thorough simulations to demonstrate the efficiency of our method on real-world optical topologies, and show that with an appropriately selected representative set of possible future requests, we are not only able to approach the maximum possible acceptance ratio but also able to mitigate the need of frequent hypervisor migrations for most realistic latency constraints.

ISP networks have become a critical infrastructure in our society. Traffic in these networks is growing and is increasingly dominated by a small number of large CDNs connecting at multiple locations. Simultaneously, the networks are becoming more flexible, in terms of routing, CDN user mapping, and also regarding the IP topology: emerging optical technologies allow to flexibly reconfigure the network. This paper studies the potential gains of these reconfiguration flexibilities. The idea is to make CDN-ISP infrastructure demand-aware, that is, to re-optimize it towards the changing end-user demands over time. We present an optimization framework and conduct an extensive evaluation using data from a large European ISP. We find that such a reconfigurable infrastructure has indeed a high potential: by leveraging spatial and diurnal traffic patterns, the efficiency of ISP networks and CDNs is improved significantly. Specifically, the required backbone capacity is reduced by 15% while reducing path lengths by 30%, on average and during the critical peak hour. Moreover, such infrastructures can leverage re-optimizations during specific events, like the COVID-19 pandemic, and under link failures. We optimistically assume a cooperative environment of ISPs and CDNs, and we conclude by discussing trends that foster the identified benefits in practice.

In order to meet the increasingly stringent throughput and latency requirements in datacenter networks, several innovative network architectures based on reconfigurable optical topologies have been proposed. Examples include demand-oblivious reconfigurable topologies such as RotorNet (SIGCOMM 2017), Opera (NSDI 2020), and Sirius (SIGCOMM 2021), as well as demand-aware topologies such as ProjecToR (SIGCOMM 2016). All these architectures feature attractive performance properties using specific prototypes. However, reproducing these experiments is often difficult due to missing hardware and publicly available software. This paper presents a flexible framework for reconfigurable networks based on off-the-shelf hardware, which supports experimentation and reproducibility at a small scale. We describe how our framework, ExReC, can be instantiated with different configurations, allowing us to emulate existing architectures and to study their trade-offs. Finally, we demonstrate the application of our approach to different use cases and workloads, including distributed machine learning training.

High computational demands of complex deep learning models led to workload distribution across multiple machines. Many frameworks for distributed machine learning (DML) have been developed and are employed in practice for orchestrating workload distribution. In this paper, we analyze and compare network behaviors of three widely used state-of-the-art DML frameworks. The study reveals that traffic can largely vary across the frameworks. While some frameworks exhibit well predictable patterns, others are less structured. We further explore whether and how it is possible to relate the network traffic to the DML jobs' attributes, and present a multiple linear regression model accordingly. Our results can inform the networking community about traffic characteristics and contribute toward the generation of realistic DML traffic for simulation studies.

Communication networks are evolving towards a more adaptive and reconfigurable nature due to the evergrowing demands they face. A framework for measuring network flexibility has been proposed recently, but the cost of rendering communication networks more flexible has not yet been mathematically modeled. As new technologies such as software-defined networking (SDN), network function virtualization (NFV), or network virtualization (NV) emerge to provide network flexibility, a way to estimate and compare the cost of different implementation options is needed. In this paper, we present a comprehensive model of the cost of a flexible network that takes into account its transient and stationary phases. This allows network researchers and operators to not only qualitatively argue about their new flexible network solutions, but also to analyze their cost for the first time in a quantitative way.

In the field of networking research, increased flexibility of new system architecture proposals, protocols, or algorithms is often stated to be a competitive advantage over its existing counterparts. However, this advantage is usually claimed only on an argumentative level and neither formally supported nor thoroughly investigated due to the lack of a unified flexibility framework. As we will show in this paper, the flexibility achieved by a system implementation can be measured, which consequently can be used to make different networking solutions quantitatively comparable with each other. The idea behind our mathematical model is to relate network flexibility to the achievable subset of the set of all possible demand changes, and to use measure theory to quantify it. As increased flexibility might come with additional system complexity and cost, our framework provides a cost model which measures how expensive it is to operate a flexible system. The introduced flexibility framework contains different normalization strategies to provide intuitive meaning to the network flexibility value as well, and also provides guidelines for generating demand changes with (non-)uniform demand utilities. Finally, our network flexibility framework is applied on two different use-cases, and the benefits of a quantitative flexibility analysis compared to pure intuitive arguments are demonstrated.

A predictable network performance is mission critical for many applications and yet hard to provide due to difficulties in modeling the behavior of the increasingly complex network equipment. This paper studies the problem of providing deterministic latency guarantees in small networks based on low-capacity hardware (e.g., in-cabin and industrial networks): such networks are of increasing importance, need to meet stringent performance requirements, but have hardly been explored so far. Our main contribution is the design, implementation, and evaluation of Loko, a system which provides predictable latency guarantees in programmable networks using low-cost hardware. Loko relies on a novel measurement-based methodology and uses deterministic network calculus to derive a reliable performance model of a given switch. To this end, we also show that state-of-the-art models in the literature like QJump and Silo fall short to model the behavior of such switches, due to incorrect architectural and performance assumptions. As a case study, we implement Loko for the Zodiac FX switch. Our experiments are encouraging: we find that the derived models are indeed accurate, allowing Loko to provide deterministic end-to-end guarantees with low-cost programmable devices.

Large content providers, known as hyper-giants, are responsible for sending the majority of the content traffic to consumers. These hyper-giants operate highly distributed infrastructures to cope with the ever-increasing demand for online content. To achieve commercial-grade performance of Web applications, enhanced end-user experience, improved reliability, and scaled network capacity, hyper-giants are increasingly interconnecting with eyeball networks at multiple locations. This poses new challenges for both (1) the eyeball networks having to perform complex inbound traffic engineering, and (2) hyper-giants having to map end-user requests to appropriate servers.

Existing virtual network admission control algorithms targeting high utilization of data center infrastructure are computationally expensive or provide poor performance. In particular, existing algorithms have in common that they are oblivious to the past, i.e., requests are handled in a fire-and-forget manner, not taking into account information from previously solved instances. This can be inefficient and misses out on a basic optimization opportunity: as for any network optimization algorithm that faces repeating problem instances, it may be beneficial to learn from network states and the outcome of acceptance decisions of the past. In this paper, we propose Ismael, a machine learning framework for predicting the acceptance of virtual clusters, one of the most common virtual network abstractions in data centers. Ismael can be configured with, and learn from, different existing algorithms by combining fixed-size feature representations for graphs with a convolutional neural network or a fully connected deep neural network. We report on extensive simulations, which demonstrate that it is possible to mimic existing, computationally intensive admission control algorithms with an accuracy of up to 94 %, while significantly reducing runtime.

Network virtualization enables increasingly diverse network services to cohabit and share a given physical infrastructure and its resources, with the possibility to rely on different network architectures and protocols optimized towards specific requirements. In order to ensure a predictable performance despite shared resources, network virtualization requires a strict performance isolation and hence, resource reservations. Moreover, the creation of virtual networks should be fast and efficient. The underlying NP-hard algorithmic problem is known as the Virtual Network Embedding (VNE) problem and has been studied intensively over the last years. This paper presents NeuroViNE, a novel approach to speed up and improve a wide range of existing VNE algorithms: NeuroViNE is based on a search space reduction mechanism and preprocesses a problem instance by extracting relevant subgraphs, i.e., good combinations of substrate nodes and links. These subgraphs can then be fed to an existing algorithm for faster and more resource-efficient embeddings. NeuroViNE relies on a Hopfield network, and its performance benefits are investigated in simulations for random networks, real substrate networks, and data center networks.

Software defined networking (SDN) network hypervisors provide the functionalities needed for virtualizing software-defined networks. Hypervisors sit logically between the multiple virtual SDN networks (vSDNs), which reside on the underlying physical SDN network infrastructure, and the corresponding tenant (vSDN) controllers. Different SDN network hypervisor architectures have mainly been explored through proof-of-concept implementations. We fundamentally advance SDN network hypervisor research by conducting a model-based analysis of SDN hypervisor architectures. Specifically, we introduce mixed integer programming formulations for four different SDN network hypervisor architectures. Our model formulations can also optimize the placement of multi-controller switches in virtualized OpenFlow-enabled SDN networks. We employ our models to quantitatively examine the optimal placement of the hypervisor instances. We compare the control plane latencies of the different SDN hypervisor architectures and quantify the cost of virtualization, i.e., the latency overhead due to virtualizing SDN networks via hypervisors. For generalization, we quantify how the hypervisor architectures behave for different network topologies. Our model formulations and the insights drawn from our evaluations inform network operators about the trade-offs of the different hypervisor architectures and help choosing an architecture according to operator demands.