Industrial networks are undergoing a radical shift from closed, static OT environments towards open networks that integrate IT and OT. This shift applies IT operation principles to OT environments, such as virtualizing Programmable Logic Controllers and using Artificial Intelligence to increase production and process efficiency. While there is a huge effort to integrate IT principles, this paper demonstrates that IT/OT convergence remains an underexplored area of research, leaving out critical research opportunities for future networking systems. We identify three core challenges: timing constraints, service availability, and changing network traffic characteristics. For each challenge, we provide a concrete use case that demonstrates early findings and opens up new avenues for research within SIGCOMM.

Machine learning models are highly dependent on the quality of their input data. In industrial settings, where video data is transmitted to the model located in the cloud over a network, transmission artifacts (e.g., congestion or losses) can degrade model performance. These performance issues can compromise process quality and result in costly errors. Despite the increasing interest in using Machine Learning (ML) models for video tasks in industrial use cases, existing research has not adequately assessed the impact of individual video artifacts on model performance, nor has it specifically examined the performance of video-based ML models under these conditions.This study aims to fill this gap by considering the impact of degraded data on the model behavior. As a case study, we consider a defect detection scenario where, e.g., a manufacturing robot is monitored for defections using video. To this end, we train a Multiscale Vision Transformer and use an approach that systematically introduces various artifacts, such as bitrate reduction and pixelation. We then assess their impact on the model’s performance accordingly. Hence, we provide the first insights into possible mitigations.

The importance of machine learning (ML) in factories and plants is growing; however, network operators have not yet fully explored the optimization potential that ML applications offer. The literature lacks a detailed analysis of the trade-offs between data quantity and model accuracy in the context of communication demands. This work introduces INOUT OPTIMA, a benchmarking framework for industrial ML applications. INOUT OPTIMA highlights the extensive optimization opportunities ML provides for network planning by analyzing the relationship between data quantity and accuracy. Specifically, it evaluates ML models trained on datasets of varying quality, subjected to controlled data degradation scenarios. This approach allows us to assess how training on degraded data affects inference accuracy and gain knowledge on the behavior of ML applications under degraded input data. The results underline the importance of understanding these trade-offs for factory operators, enabling the design of resilient and efficient ML applications, and providing additional data for factory planners.

Todays factories rely on expansive networks carrying mass amounts of data from sensors, machines, internet of things (IoT) devices, and machine learning (ML) applications. Providing predictable network performance in these factories is a crucial step toward the efficient operation of ML applications and, thus, the factory itself. This paper introduces DIAMOND for designing industrial networks specifically tailored to the realworld requirements of ML applications. DIAMOND analyzes and uses the optimization opportunities lying within the training data of ML applications for network planning. DIAMOND’s first step analyzes ML applications to investigate their accuracy to incoming data rate trade-offs. The method evaluates models trained on data qualities ranging from worst to best under certain degradation scenarios. While data degradation can affect inference performance and accuracy, using it during training leads to more optimization opportunities. The second step uses a mixed integer linear program (MILP) to design network topologies considering available compute resources, latency, and infrastructure costs. This step exploits the newly gained flexibility and robustness of ML models to optimize the network topology and server placement within the infrastructure. The results demonstrate the need for designing efficient networks and avoiding overprovisioning, as purposefully built networks outperform off-the-shelf topologies.

There is no system yet that provides predictable data plane and control plane operations in programmable networks. However, both predictable data plane and control plane operations are needed, e.g., in industrial networks. Particularly there, the operation of the network needs to be planned and, hence, relies on network operations that are deterministic and executed in a timely manner. To fill this gap, this paper proposes our system namedNAGA, which provides data plane deterministic guarantees along with consistent and timely network updates in programmable networks. In order to not rely on specialized hardware,NAGAuses widely-available hardware capabilities such as priority queuing and label-based forwarding. Whereas the real implementation ofNAGAin a P4-based testbed demonstrates that applications receive guaranteed performance in terms of latency and data rate, simulation studies show the ability ofNAGAto be even deployed in large scale scenarios beyond industrial networks, such as wide area and data center networks.

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.

Website Fingerprinting (WFP) attacks infer the websites or webpages a user is visiting from encrypted traffic. To date, it remains uncertain if WFP can attack many users from a central location in an online scenario. We close this gap with PROFI, a WFP attack that detects websites based on the initial TLS connection from the client to the server using at most the connection’s first 30 packets. PROFI achieves a precision and recall of 86.51% and 85.35% in a closed-world, and 68.90% and 78.71% in an open-world scenario, which is competitive to state-of-the-art (SoA) WFP attacks, while taking a fraction of the time of SoA attacks to classify a webpage. Further, we implement PROFI as a micro service-based prototype and evaluate the attack in an online scenario with real traffic traces. We show that PROFI can monitor up to 100 websites at 10 G, corresponding to up to 424 webpages per second. We also show that PROFI has the potential to interfere with a victim’s webpage access.

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.

Realizing Service Function Chaining with a micro-service-based architecture results in an increased number of computationally cheap Virtual Network Functions (VNFs). Pinning cheap VNFs to dedicated CPU cores can waste resources since not every VNF fully utilizes its core. Thus, cheap VNFs should share CPU cores to improve resource utilization. However, sharing cores can result in degraded performance due to interference between VNFs, even in mildly loaded scenarios. We proposeD2A, a system that combines Neural Combinatorial Optimization, Machine Learning (ML)-based Digital Twins (DTs), and Game Theory to optimize VNF assignments. Measurements in a testbed show thatD2Aincreases throughput by up to 46% and reduces latency by up to 93%, compared to three baseline algorithms. Using an ML-based DT to model VNF interference increases throughput by up to 11%, and reduces latency by up to 90% compared to an analytical model of the system.

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.

Since caches are shared and coherent, a memory access of one process may evict from the cache another process' memory block with an address mapped to the same cache line. This property is exploited by several attacks to form side channels. We show that MAC learning in Software Defined Networks (SDNs) has a similar property in the sense that a MAC address discovered by one network device may be revoked by the discovery of the same address at another switch. This allows us to implement Macchiato, a covert channel for SDNs between any two network devices (including hosts); prior SDN covert channels required at least one malicious switch. We evaluate a prototype implementation of Macchiato and discuss how methods to improve the performance of cache side channels (such as deep neural networks) can also be used in Macchiato.

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.

Programmable networks come with the promise of logically centralized control, in order to optimize the network's routing behavior. However, until now, controllers are heavily involved in network operations to prevent inconsistencies such as blackholes, loops, and congestion. In this paper, we propose the P4Update framework, based on the network programming language P4, to shift the consistency control and most of the routing update logic out of the overloaded and slow control plane. As such P4Update avoids high and unnecessary control plane delays by mainly scheduling and offloading the update process to the data plane.

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.

Providing predictable performance to tenants is mission critical for network hypervisors. As a hypervisor acts as an intermediary between tenants controllers and the physical infrastructure, its resources (e.g., CPU, RAM) should be provisioned and allocated carefully. Initially, we demonstrate that state-of-the-art CPU prediction approaches are not suitable for provisioning network hypervisor CPU resources, since they predict only the mean CPU utilization. However, provisioning the resources with a mean value can significantly degrade the forwarding performance of a network hypervisor. In this article, we present a novel approach which provisions network hypervisor CPU resources efficiently, while avoiding performance degradation. We take three steps to achieve our goal:(i)conducting a profound measurement campaign to determine what is the minimum amount of CPU resources that needs to be allocated to a network hypervisor in order to have no performance degradation;(ii)revealing the key properties of virtual networks that affect the CPU utilization;(iii)designing a precise CPU prediction model. Using randomly generated virtual networks and arbitrary physical topologies, we show that our prediction model exhibits an average relative error of around 4%. Further, our evaluations indicate that provisioning the CPU resources of a network hypervisor based on the proposed prediction model does not degrade the hypervisor forwarding performance. Utilizing our approach, network operators can minimize their resources consumption while still providing predictable and undegraded forwarding performance to tenants.

Network programmability also sneaked into the mobile world leading to the emergence of Software-Defined Radio Access Network (SD-RAN) architectures. Interestingly, while only a small number of prototype architectures exist for SD-RAN, their performance evaluations are unfortunately also limited. Recent evaluations are carried out for small network dimensions of up to 50 devices, while emerging 5G/6G networks envision numbers of devices beyond 5000. Although 5G/6G applications are more stringent with respect to latency guarantees, performance evaluations of such low scale remain questionable. To fill this void, this paper presentsMARC: a novel benchmarking tool for SD-RAN architectures and their controllers. We useMARCto measure, analyze and identify performance implications for two state-of-the-art open source SD-RAN solutions:FlexRANand5G-EmPOWER. We perceive results for monitoring application scenarios considering fully centralized control. For this setting, our findings show that the proposed architectures with a single SD-RAN controller are not scalable and can even lead to unpredictable network operations. Using our tool and based on our insights, we provide and implement design guidelines for the internal working behavior of the existing controllers.

The recent trend of network softwarization suggests a radical shift in the implementation of traditional network intelligence. In Software Defined Networking (SDN), for instance, the control plane functions of forwarding devices are outsourced to the controller. Softwarized network components are expected to provide uninterrupted service during long periods of time, which makes them prone to the effects ofsoftware aging, a phenomena that has been observed in operational software systems where the failure rate increases or the performance of the software degrades with the elapsed time since the last restart. The effects of software aging in operational networks are typically mitigated bysoftware rejuvenation, i.e., planned restarts cleaning the internal system state in order to prevent or postpone aging-related failures. This article presentsARES, a three-step methodological framework for the management of the effects of software aging in softwarized networks, applied to the case study of open source SDN orchestration platforms. Using ARES, we demonstrate that software aging is a systematic problem that cannot be neglected in network orchestration systems. It stems not only from aging-related bugs and natural aging due to fragmentation, but also from design choices, e.g., when implementing distributed systems. Measurements for Open Network Operating System (ONOS) and OpenDaylight (ODL) demonstrate how “simple” and common networking tasks let network performance degrade rapidly and even lead to crashes: for instance, adding and removing 300 intents per second in ONOS significantly increases the response time by 50% per day and depletes the memory at the rate of 18GB per day. Moreover, we demonstrate a first rejuvenation approach that can mitigate the effects of aging in ONOS.

This paper presents Chameleon, a cloud network providing both predictable latency and high utilization, typically two conflicting goals, especially in multi-tenant datacenters. Chameleon exploits routing flexibilities available in modern communication networks to dynamically adapt toward the demand, and uses network calculus principles along individual paths. More specifically, Chameleon employs source routing on the "queue-level topology", a network abstraction that accounts for the current states of the network queues and, hence, the different delays of different paths. Chameleon is based on a simple greedy algorithm and can be deployed at the edge; it does not require any modifications of network devices. We implement and evaluate Chameleon in simulations and a real testbed. Compared to state-of-the-art, we find that Chameleon can admit and embed significantly, i.e., up to 15 times more flows, improving network utilization while meeting strict latency guarantees.

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.

To meet their increasingly stringent dependability requirements, communication networks need to be predictable, both in terms of correctness and performance. In principle, Software-Defined Networks (SDN) enable such more predictable networks, however, these networks still depend the underlying switches. This paper presents an empirical study of the predictability of SDN switches. Our extensive benchmarking of seven hardware OpenFlow switches from four different manufacturers raises several concerns regarding the dependability of these switches. We uncover several incorrect and unpredictable behaviors and performance issues. In particular, we identify unpredictable behaviors related to the management of flows and buffers, and observe that existing quality-of-service mechanisms, such as priority queuing, introduce unexpected overheads. The latter, in turn, can lead to violations of latency guarantees. Based on our insights, we discuss first solutions toward more predictable architectures.

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.

Virtualization and multi-tenancy are attractive paradigms to improve the utilization of computing infrastructures and hence to reduce costs. In order to provide a high degree of resource sharing without sacrificing predictable cloud application performance, strict performance isolation needs to be ensured. This is non-trivial and requires models which account for all components where applications may interfere: similarly to security, the predictability of cloud application performance can only be as good as the least predictable component in the model. This paper identifies a new source of potential performance interference that has been overlooked so far: the network hypervisor - a critical component in any multi-tenant network. We present a first measurement study of the performance implications of the network hypervisor in Software-Defined Networks (SDNs). For the purpose of our study, we developed a new open-source benchmarking tool for OpenFlow control and data planes. We show that cloud application performance may appear unpredictable if the network hypervisor is not accounted for: the performance does not only depend on the specific hypervisor implementation and workload (e.g., OpenFlow message types), but also on the number of tenants and the size of the network. Hence, our results suggest that hypervisors should be included in our performance models, and their performance benchmarked and compared similarly to other crucial software components such as the SDN controller.

Communication networks are the key enabling technology for our digital society. In order to sustain their critical services in the future, communication networks need to flexibly accommodate new requirements and changing contexts due to emerging diverse applications. In contrast to traditional networking technologies, software-oriented networking concepts, such as software-defined networking (SDN) and network function virtualization (NFV), provide ample opportunities for highly flexible network operations, enabling fast and simple adaptation of network resources and flows. This paper identifies the opportunities and challenges of adaptable softwarized networks and introduces a conceptual framework for adaptations in softwarized networks. We first explain how softwarized networks contribute to network adaptability through the functional primitives observation, composition, and control. We review the wide range of options for fine-granular observations as well as fine-granular composition and control provided by SDN and NFV. The multitude of fine-granular “tuning knobs” in adaptable softwarized networks complicates the decision making, which is the main focus of this paper. We propose to enhance the functional primitives observation, composition, and control with data-driven decision making, e.g., machine learning modules, resulting in deep observation, composition, and control. The data-driven decision making modules can learn and react to changes in the environment, e.g., new flow demands, so as to support meaningful decision making for adaptation in softwarized networks. Finally, we make the case for employing the concept of empowerment to realize truly “self-driving” networks.

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.

In software defined networking (SDN), critical control plane functions are offloaded to a software entity known as the SDN controller. Today's SDN controllers are complex software systems, owing to heterogeneity of networks and forwarding devices they support, and are inherently prone to bugs. Our previous work showed that software reliability growth models (SRGM) can model the stochastic nature of bug manifestation process open source SDN controllers. In this paper, we focus on different applications of our SRGM framework crucial for an efficient management of SDN-based networks. We provide guidelines for network operators to decide when the controller software is mature enough to be deployed in operational environment, based on the reliability requirements of network applications, and quantify the marginal benefits of the prolonged testing phase on the software quality. We show how the accuracy of software reliability prediction in the early phase of the software lifecycle can be improved by extrapolating the behavior of previous controller software releases. We also propose software maturity metrics that can be used by operators to discriminate between the competing SDN controller designs, i.e., ONOS and OpenDaylight, when software reliability is a major concern.

Software-Defined Networking (SDN) controllers are network entities that act as strategic control points in an SDN network. Controller placement studies mostly aim at optimizing network performance in terms of control latency, reliability and resilience, given network characteristics that are static. Yet dynamic traffic conditions, if not adapted by the controller placement properly, may cause high end-to-end flow setup time. For reactive controllers, the end-to-end flow setup time of a flow implies the difference between sending time at the source and receiving time at the sink of the first packet in that flow. Therefore, end-to-end flow setup time indicates the amount of time needed to set up forwarding rules in all involved switches and acts as a primary concern in terms of service establishment of network operators. In this paper, we analyze the controller placement for dynamic traffic flows based on a combined controller placement model: controller locations and switch-to-controller assignments are simultaneously optimized for minimum average flow setup time with respect to different traffic conditions inside the network. Linearization method is applied to transform the problem into a Mixed Integer Programming (MIP) problem which can be solved optimally. Two derivatives are also presented for comparison, one optimizing only controller locations and the other optimizing only switch-to-controller assignments. Our simulations cover two real network topologies and we explain the effects of the models have on the flow setup time with respect to dynamic flows. For low flow densities, the controller placement that adapts to flows could reduce the average flow setup time by about 50% compared to the static placement. However, when densities are high, the need of changing controller placement to guarantee flow setup performance is marginal.

This paper is motivated by the emerging vision of an automated and data-driven optimization of communication networks, making it possible to fully exploit the flexibilities offered by modern network technologies and heralding an era of fast and self-adjusting networks. We build upon our recent study of machine-learning approaches to (statically) optimize resource allocations based on the data produced by network algorithms in the past. We take our study a crucial step further by considering dynamic scenarios: scenarios where communication patterns can change over time. In particular, we investigate network algorithms which learn from the traffic distribution (the feature vector), in order to predict global network allocations (a multi-label problem). As a case study, we consider a well-studied fc-median problem arising in Software-Defined Networks, and aim to imitate and speedup existing heuristics as well as to predict good initial solutions for local search algorithms. We compare different machine learning algorithms by simulation and find that neural network can provide the best abstraction, saving up to two-thirds of the algorithm runtime.

Developing models to generate realistic graphs of communication networks often requires a deep understanding and extensive analysis of the underlying network structure. Since deployed communication networks are dynamic, the findings a generator is based on might lose validity. We alleviate the need for extensive analysis of graphs by estimating parameters of a probabilistic model. The model parameters encode the structure of the graph, which is thus learned in an unsupervised fashion. Synthetic graphs can be generated from the model and will have the structure previously inferred. For this, we use the Stochastic-Block-Model (SBM) and the Degree-Corrected-Block-Model (DCBM), a variant allowing for heavy tailed degree distributions. The models originate in the social sciences and separate a graph into groups of nodes. To show the applicability of the models to the task of synthetic graph generation in the domain of communication networks, we use one router level and one IP-to-IP communication graph. We assert the quality of the generated models by evaluating a number of graph features and comparing our results to those obtained with the network generator Orbis. We find our approach to be on par with, or even outperforming Orbis. Furthermore, the models are able to capture large-scale structure in communication networks.

With the rapid growth of user traffic, service innovation, and the persistent necessity to reduce costs, today's mobile operators are faced with several challenges. In networking, two concepts have emerged aiming at cost reduction, increase of network scalability and deployment flexibility, namely Network Functions Virtualization (NFV) and Software Defined Networking (SDN). NFV mitigates the dependency on hardware, where mobile network functions are deployed as software virtual network functions on commodity servers at cloud infrastructure, i.e., data centers. SDN provides a programmable and flexible network control by decoupling the mobile network functions into control plane and data plane functions. The design of the next generation mobile network (5G) requires new planning and dimensioning models to achieve a cost optimal design that supports a wide range of traffic demands. We propose three optimization models that aim at minimizing the network load cost as well as data center resources cost by finding the optimal placement of the data centers as well the SDN and NFV mobile network functions. The optimization solutions demonstrate the trade-offs between the different data center deployments, i.e., centralized or distributed, and the different cost factors, i.e., optimal network load cost or data center resources cost. We propose a Pareto optimal multi-objective model that achieves a balance between network and data center cost. Additionally, we use prior inference, based on the solutions of the single objectives, to pre-select data center locations for the multi-objective model that results in reducing the optimization complexity and achieves savings in run time while keeping a minimal optimality gap.

The allocation of physical resources to virtual networks, i.e., the virtual network embedding (VNE), is still an on-going research field due to its problem complexity. While many solutions for the online VNE problem exist, only few have focused on methods that can be generally applied for optimization of online embeddings. In this paper, we propose an admission control based on a Recurrent Neural Network (RNN) to improve the overall system performance for the online VNE problem. Before running a VNE algorithm to embed a virtual network request, the RNN predicts whether the request will be accepted by the VNE algorithm based on the current state of the substrate and the virtual network request (VNR). The RNN prevents VNE algorithms from spending time on VNRs that are either infeasible or that cannot be embedded in acceptable time. In order to train and operate the RNN efficiently, we additionally propose new representations for substrate networks and virtual network requests. The representations are based on topological and network resource features to represent the substrate network and the VNRs with low computational complexity. Via simulations, we show that our admission control reduces the overall computational time for the online VNE problem by up to 91 % while preserving VNE performance on average. Using our new substrate and request representations, the RNN achieves an accuracy ranging between 89 % and 98 % for different VNE algorithms, substrate sizes, and VNR arrival rates.

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.

Virtualization of software defined networks enables tenants to bring their own controller and manage their virtual resources with the full programmability provided by Software Defined Networking (SDN). Distributed SDN hypervisors are proposed to provide an efficient platform for the virtualization of physical SDN networks. They address the issue of scalability that a centralized hypervisor could suffer from. As virtualization provides the possibility to change virtual SDN networks on run time, a hypervisor layer needs efficient mechanisms to dynamically adapt to the changing requirements. Existing proposals provide only a static configuration setup for their distribution of the hypervisor instances. However, in order to satisfy the dynamics of virtual SDN networks, management protocols are needed to support dynamic changes. In this paper, we propose a control path migration protocol for distributed hypervisors. Such protocol is needed to support the dynamic adaptation of the virtualization layer. Our protocol is providing the missing procedure that allows a dynamic change of control connections between virtual SDN networks and the tenants' controllers, respectively. We provide a proof of concept implementation for our proposal. Through measurements in a real testbed setup, we show that our protocol is efficient in terms of control latency overhead and provides transparency to the controllers of the virtual SDN networks.

We present a demonstration of HyperFlex [1], a Software-Defined Networking (SDN) virtualization architecture with flexible hypervisor function allocation guaranteeing control-plane virtualization. Network Virtualization (NV) is expected to overcome the ossification of today's communication networks. SDN is seen as an enabler for programmable network control. In order to fully virtualize software-defined networks, not only the virtualization of the data-plane, but also the virtualization of the control-plane has to be considered. HyperFlex is a virtualization hypervisor that ensures isolated control-plane slices for virtual SDN networks. Control-plane isolation also protects the hypervisor resources from exhaustion. Furthermore, virtualization hypervisors have to be scalable and flexible in order to provide the best possible performance for the virtual software-defined networks. They should be able to adapt to the current state of the network and the divergent demands of virtual SDN networks. HyperFlex distributes the hypervisor functions flexibly and dynamically in order to adapt to the current network state.

Network Virtualization (NV) and Software-Defined Networking (SDN) are both expected to increase the flexibility and programmability of today's communication networks. Combining both approaches may even be a further step towards increasing the efficiency of network resource utilization. Multiple solutions for virtualizing SDN networks have already been proposed, however, they are either implemented in software or they require special network hardware. We propose HyperFlex, an SDN hypervisor architecture that relies on the decomposition of the hypervisor into functions that are essential for virtualizing SDN networks. The hypervisor functions can be flexibly executed in software or hosted on SDN network elements. Furthermore, existing hypervisor solutions focus on data-plane virtualization mechanisms and neglect the virtualization of the control-plane of SDN networks. HyperFlex provides control-plane virtualization by adding a control-plane isolation function, either in software or on network elements. The isolation function ensures that the resources of the control-plane are shared correctly between each virtual SDN network while it also protects the hypervisor resources from resource exhaustion.

Video content and, in particular, YouTube's content account for the largest amount of today's Internet traffic. However, little is known about the behavior of video streaming services for different kinds of network environments and under varying network conditions. Due to network operators' lack of knowledge about the transmitted content, network resources may not be optimally used in general. Thus, we propose a dyadic measurement system composed of application, i.e., client-based and network-based monitoring for YouTube's video traffic. Using our proposed monitoring methodology, we analyze the behavior of YouTube's HTTP-based adaptive video streaming mechanisms. In detail, we quantify via experimental measurements on real network traffic YouTube's behavior for different videos under static and varying network conditions. Our measurement results show that in case of varying network conditions, YouTube demands different video qualities in parallel in order to adapt to the network situation. However, this behavior can result in up to 33 % of redundant network traffic, i.e., downloaded video content of different quality levels for the same play time. Due to our findings, network operators should try to optimize the allocation of network resources for video content in a way that avoids varying network conditions, resulting in less waste of network resources.

In this paper, we demonstrate an implementation of a Network Services Abstraction Layer (NSAL) on top of the network control and management plane. Furthermore, we introduce a unified data model for both Software Defined Networking (SDN) and legacy devices that allows managing and configuring both networks in a unified way in order to achieve Quality of Service (QoS) for time-critical tasks (e.g. VoIP). Due to the unified data model, network operators are able to manage their network through one interface. We demonstrate a use case by implementing a VoIP scheduling application on top of the NSAL and evaluate VoIP call quality in a distributed heterogeneous network.

Software-defined Networking (SDN) is a promising and powerful concept to introduce new dimensions of flexibility and adaptability in today's communication networks. In particular, the realization of Quality of Service (QoS) concepts becomes possible in a flexible and dynamic manner with SDN. Although concepts of QoS are well researched, they were not realized in communication networks due to high implementation complexity and realization costs. Using SDN to realize QoS mechanisms enables emerging concepts, such as application-aware resource management solutions. These emerging concepts demand latency or data rate guarantees for end-user applications, e.g., video streaming or gaming. However, the impact of dynamic QoS management on network traffic has not been studied in detail yet. This paper provides a first study of the impact on dynamic QoS mechanisms and their realizations for OpenFlow-enabled SDN switches. Although SDN and, in particular, OpenFlow as one dominant realization claim to provide a standardized interface to control network traffic, our measurement results show a noticeable diversity for different OpenFlow switches. In detail, our investigations reveal a severe impact on the performance of TCP-based network traffic among different switches. These observations of switch diversity may provide SDN application developers insights when realizing QoS concepts in an SDN-based network.

Today's Internet does not provide an exchange of information between applications and networks, which may result in poor application performance. Concepts such as application-aware networking or network-aware application programming try to overcome these limitations. The introduction of Software-Defined Networking (SDN) opens a path towards the realization of an enhanced interaction between networks and applications. Hence, a more dynamic and demand-based allocation of network resources to heterogeneous applications can be realized. The implementation of the resource management action, however, may have an impact on the data transport and application quality. This paper summarizes resource management mechanisms provided by current SDN approaches based on OpenFlow and exemplary evaluates implementation prospects and challenges.