[ES] La complejidad en la gestión de las redes de comunicación tradicionales, así como su poca escalabilidad y flexibilidad, supone un obstáculo para el desarrollo y consolidación de nuevas tecnologías emergentes como es el caso del Internet de las Cosas (Internet of Things), dónde la facilidad para el intercambio y manejo de grandes volúmenes de datos heterogéneos procedentes de sensores es un requisito clave para el correcto funcionamiento del sistema. El Internet de las Cosas se define cómo la interconexión digital de objetos cotidianos dotados de inteligencia (Smart devices) a través de redes de comunicación de datos ya sean públicas (Internet) o privadas. Sin embargo, el Internet de las Cosas no sólo está compuesto por estos dispositivos, toda la infraestructura, plataformas, aplicaciones y servicios que ayudan a los datos a viajar desde los dispositivos origen y hacia sus diferentes destinos, y la gestión de estos también forman parte del denominado Internet de las Cosas. El almacenamiento, análisis, procesado y gestión masiva de dichos datos es lo que se denomina Big Data, y está compuesto de grandes cantidades de datos (massive data) estructurados en diferentes formatos, modelos de datos y protocolos, lo que dificulta su tratamiento y su intercambio a través de las redes de datos convencionales. Ante esta problemática la implementación de redes virtuales definidas por software se presenta como una posible solución para dotar de flexibilidad, escalabilidad y sencillez de gestión a las redes que interconectan estos dispositivos, plataformas y otros elementos IoT, permitiendo una visión global, una gestión centralizada y un desarrollo de servicios a nivel de red específicos para los entornos de Internet de las Cosas. Este proyecto se presenta como una aproximación de estas dos tecnologías y tendrá como objetivo el diseño de una solución donde probar las herramientas de control de redes definidas por software o programables (SDN) y las funciones virtuales de redes (NFV) aplicadas a despliegues de Internet de las Cosas (IoT) de forma que se puedan demostrar sus ventajas e implicaciones y se puedan descubrir nuevas líneas de desarrollo sobre esta base. / [CA] La complexitat en la gestió de les xarxes de comunicació tradicionals, així com la seua poca escalabilitat i flexibilitat, suposa un obstacle per al desenvolupament i consolidació de noves tecnologies emergents com és el cas de la Internet de les Coses (Internet of Things), on la facilitat per a l'intercanvi i maneig de grans volums de dades heterogènies procedents de sensors és un requisit clau per al correcte funcionament del sistema. La Internet de les Coses es defineix com la interconnexió digital d'objectes quotidians dotats d'intel·ligència (Smart devices) a través de xarxes de comunicació de dades ja siguen públiques (Internet) o privades. No obstant això, la Internet de les Coses no sols està compost per aquests dispositius, tota la infraestructura, plataformes, aplicacions i serveis que ajuden les dades a viatjar des dels dispositius d'origen i cap a les seues diferents destinacions, i la gestió d'aquests també formen part de la denominada Internet de les Coses. L'emmagatzematge, anàlisi, processament i gestió massiva d'aquestes dades és el que es denomina Big Data, i està compost de grans quantitats de dades (massive data) estructurats en diferents formats, models de dades i protocols, la qual cosa dificulta el seu tractament i el seu intercanvi a través de les xarxes de dades convencionals. Davant aquesta problemàtica la implementació de xarxes virtuals definides per software es presenta com una possible solució per a dotar de flexibilitat, escalabilitat i senzillesa de gestió a les xarxes que interconnecten aquests dispositius, plataformes i altres elements IoT, permetent una visió global, una gestió centralitzada i un desenvolupament de serveis a nivell de xarxa específics per als entorns d'Internet de les Coses. Aquest projecte es presenta com una aproximació d'aquestes dues tecnologies i tindrà com a objectiu el disseny d'una solució on provar les eines de control de xarxes definides per software o programables (SDN) i les funcions virtuals de xarxes (NFV) aplicades a desplegaments d'Internet de les Coses (IoT) de manera que es puguen demostrar els seus avantatges i implicacions, i es puguen descobrir noves línies de desenvolupament sobre aquesta base. / [EN] Nowadays, the complexity of traditional network administration, together with the lack of scalability and flexibility, has been a challenge for the proper development and integration of new emerging technologies which make use of this network. As an example, we have the so-called Internet of Things (IoT). The principal IoT network requirement that enables the growth of this paradigm is the need to facilitate high data volume exchange and administration, from very heterogeneous sources. The IoT concept is defined as the digital interconnection of daily objects endowed with more "intelligence" (Smart devices) through a data communication network either public (Internet) or private. However, this technological trend does not only depend on the "smart devices", but on the whole infrastructure, platforms, frameworks, services, and applications that helps data to travel from the source devices to their different destinations. Also, the handling of the massive volumes of data extracted from those smart devices, their storage, processing, and analysis, known as Big Data, is a key part of this paradigm. This data is gathered from very different sources, and hence, it has diverse data structures and formats. Moreover, it is exchanged using various network protocols (LoRa, CoAp, etc.) which hinder its management and communication through conventional networks, that were not created for such data traffic. Given this problem, several technological approaches have emerged to solve it. Virtual software-defined networking is presented as a possible solution to provide flexibility, scalability, and simplicity of management to the networks that interconnect these devices, platforms, services, and other IoT elements. The virtualization of the network infrastructure, includes an extra layer of abstraction, thus providing a holistic vision of the network and centralizing the administration of its elements and the development of specific network services for IoT deployments. This project is presented as an approximation of these two technological paradigms and will have as the main objective the design of an architectural blueprint and testbed were testing the control tools of software-defined networks (SDN) and the virtualized network functions (NFV) applied to IoT deployments. Thereby, its advantages and implications can be evaluated, and new lines of development can be discovered on this base. / Suárez De Puga García, J. (2022). Especificación e implementación de un sistema de red definida por software con funciones virtuales adaptadas a despliegues de Internet de las cosas [Tesis doctoral]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/181555 / TESIS

In this project, two tests were performed. On the first test, Mininet-WiFi was used to simulate a Software Defined Network to demonstrate Mininet-WiFi’ s ability to be used as the Software Defined Network emulator which can also be integrated to the existing network using a Network Virtualized Function (NVF). A typical organization’s computer network was simulated which consisted of a website hosted on the LAMP (Linux, Apache, MySQL, PHP) virtual machine, and an F5 application delivery controller (ADC) which provided load balancing of requests sent to the web applications. A website page request was sent from the virtual stations inside Mininet-WiFi. The request was received by the application delivery controller, which then used round robin technique to send the request to one of the web servers on the LAMP virtual machine. The web server then returned the requested website to the requesting virtual stations using the simulated virtual network. The significance of these results is that it presents Mininet-WiFi as an emulator, which can be integrated into a real programmable networking environment offering a portable, cost effective and easily deployable testing network, which can be run on a single computer. These results are also beneficial to modern network deployments as the live network devices can also communicate with the testing environment for the data center, cloud and mobile provides. On the second test, a Software Defined Network was created in Mininet using python script. An external interface was added to enable communication with the network outside of Mininet. The amazon web services elastic computing cloud was used to host an OpenDaylight controller. This controller is used as a control plane device for the virtual switch within Mininet. In order to test the network, a webserver hosted on the Emulated Virtual Environment – Next Generation (EVENG) software is connected to Mininet. EVE-NG is the Emulated Virtual Environment for networking. It provides tools to be able to model virtual devices and interconnect them with other virtual or physical devices. The OpenDaylight controller was able to create the flows to facilitate communication between the hosts in Mininet and the webserver in the real-life network. / Electrical and Mining Engineering / M. Tech. (Electrical Engineering)

L'informatique en nuage (Cloud Computing) a émergé comme un nouveau paradigme pour offrir des ressources informatiques à la demande et pour externaliser des infrastructures logicielles et matérielles. Le Cloud Computing est rapidement et fondamentalement en train de révolutionner la façon dont les services informatiques sont mis à disposition et gérés. Ces services peuvent être demandés à partir d’un ou plusieurs fournisseurs de Cloud d’où le besoin de la mise en réseau entre les composants des services informatiques distribués dans des emplacements géographiquement répartis. Les utilisateurs du Cloud veulent aussi déployer et instancier facilement leurs ressources entre les différentes plateformes hétérogènes de Cloud Computing. Les fournisseurs de Cloud assurent la mise à disposition des ressources de calcul sous forme des machines virtuelles à leurs utilisateurs. Par contre, ces clients veulent aussi la mise en réseau entre leurs ressources virtuelles. En plus, ils veulent non seulement contrôler et gérer leurs applications, mais aussi contrôler la connectivité réseau et déployer des fonctions et des services de réseaux complexes dans leurs infrastructures virtuelles dédiées. Les besoins des utilisateurs avaient évolué au-delà d'avoir une simple machine virtuelle à l'acquisition de ressources et de services virtuels complexes, flexibles, élastiques et intelligents. L'objectif de cette thèse est de permettre le placement et l’instanciation des ressources complexes dans des infrastructures de Cloud distribués tout en permettant aux utilisateurs le contrôle et la gestion de leurs ressources. En plus, notre objectif est d'assurer la convergence entre les services de cloud et de réseau. Pour atteindre cela, nous proposons des algorithmes de mapping d’infrastructures virtuelles dans les centres de données et dans le réseau tout en respectant les exigences des utilisateurs. Avec l'apparition du Cloud Computing, les réseaux traditionnels sont étendus et renforcés avec des réseaux logiciels reposant sur la virtualisation des ressources et des fonctions réseaux. En plus, le nouveau paradigme d'architecture réseau (Software Defined Networks) est particulièrement pertinent car il vise à offrir la programmation du réseau et à découpler, dans un équipement réseau, la partie plan de données de la partie plan de contrôle. Dans ce contexte, la première partie propose des algorithmes optimaux (exacts) et heuristiques de placement pour trouver le meilleur mapping entre les demandes des utilisateurs et les infrastructures sous-jacentes, tout en respectant les exigences exprimées dans les demandes. Cela inclut des contraintes de localisation permettant de placer une partie des ressources virtuelles dans le même nœud physique. Ces contraintes assurent aussi le placement des ressources dans des nœuds distincts. Les algorithmes proposés assurent le placement simultané des nœuds et des liens virtuels sur l’infrastructure physique. Nous avons proposé aussi un algorithme heuristique afin d’accélérer le temps de résolution et de réduire la complexité du problème. L'approche proposée se base sur la technique de décomposition des graphes et la technique de couplage des graphes bipartis. Dans la troisième partie, nous proposons un cadriciel open source (framework) permettant d’assurer la mise en réseau dynamique entre des ressources Cloud distribués et l’instanciation des fonctions réseau dans l’infrastructure virtuelle de l’utilisateur. Ce cadriciel permettra de déployer et d’activer les composants réseaux afin de mettre en place les demandes des utilisateurs. Cette solution se base sur un gestionnaire des ressources réseaux "Cloud Network Gateway Manager" et des passerelles logicielles permettant d’établir la connectivité dynamique et à la demande entre des ressources cloud et réseau. Le CNG-Manager offre le contrôle de la partie réseau et prend en charge le déploiement des fonctions réseau nécessaires dans l'infrastructure virtuelle des utilisateurs / Cloud computing emerged as a new paradigm for on-demand provisioning of IT resources and for infrastructure externalization and is rapidly and fundamentally revolutionizing the way IT is delivered and managed. The resulting incremental Cloud adoption is fostering to some extent cloud providers cooperation and increasing the needs of tenants and the complexity of their demands. Tenants need to network their distributed and geographically spread cloud resources and services. They also want to easily accomplish their deployments and instantiations across heterogeneous cloud platforms. Traditional cloud providers focus on compute resources provisioning and offer mostly virtual machines to tenants and cloud services consumers who actually expect full-fledged (complete) networking of their virtual and dedicated resources. They not only want to control and manage their applications but also control connectivity to easily deploy complex network functions and services in their dedicated virtual infrastructures. The needs of users are thus growing beyond the simple provisioning of virtual machines to the acquisition of complex, flexible, elastic and intelligent virtual resources and services. The goal of this thesis is to enable the provisioning and instantiation of this type of more complex resources while empowering tenants with control and management capabilities and to enable the convergence of cloud and network services. To reach these goals, the thesis proposes mapping algorithms for optimized in-data center and in-network resources hosting according to the tenants' virtual infrastructures requests. In parallel to the apparition of cloud services, traditional networks are being extended and enhanced with software networks relying on the virtualization of network resources and functions especially through network resources and functions virtualization. Software Defined Networks are especially relevant as they decouple network control and data forwarding and provide the needed network programmability and system and network management capabilities. In such a context, the first part proposes optimal (exact) and heuristic placement algorithms to find the best mapping between the tenants' requests and the hosting infrastructures while respecting the objectives expressed in the demands. This includes localization constraints to place some of the virtual resources and services in the same host and to distribute other resources in distinct hosts. The proposed algorithms achieve simultaneous node (host) and link (connection) mappings. A heuristic algorithm is proposed to address the poor scalability and high complexity of the exact solution(s). The heuristic scales much better and is several orders of magnitude more efficient in terms of convergence time towards near optimal and optimal solutions. This is achieved by reducing complexity of the mapping process using topological patterns to map virtual graph requests to physical graphs representing respectively the tenants' requests and the providers' physical infrastructures. The proposed approach relies on graph decomposition into topology patterns and bipartite graphs matching techniques. The third part propose an open source Cloud Networking framework to achieve cloud and network resources provisioning and instantiation in order to respectively host and activate the tenants' virtual resources and services. This framework enables and facilitates dynamic networking of distributed cloud services and applications. This solution relies on a Cloud Network Gateway Manager and gateways to establish dynamic connectivity between cloud and network resources. The CNG-Manager provides the application networking control and supports the deployment of the needed underlying network functions in the tenant desired infrastructure (or slice since the physical infrastructure is shared by multiple tenants with each tenant receiving a dedicated and isolated portion/share of the physical resources)

Graphics Processing Units (GPUs) are emerging as the most popular accelerator for many applications, powering the core of Machine Learning applications and many computing-intensive workloads. GPUs have typically been consideredas accelerators, with Central Processing Units (CPUs) in charge of the mainapplication logic, data movement, and network connectivity. In these architectures,input and output data of network-based GPU-accelerated application typically traverse the CPU, and the Operating System network stack multiple times, getting copied across the system main memory. These increase application latency and require expensive CPU cycles, reducing the power efficiency of systems, and increasing the overall response times. These inefficiencies become of higher importance in latency-bounded deployments, or with high throughput, where copy times could easily inflate the response time of modern GPUs. The main contribution of this dissertation is towards a GPU-centric network architecture, allowing GPUs to initiate network transfers without the intervention of CPUs. We focus on commodity hardware, using NVIDIA GPUs and Remote Direct Memory Access over Converged Ethernet (RoCE) to realize this architecture, removing the need of highly homogeneous clusters and ad-hoc designed network architecture, as it is required by many other similar approaches. By porting some rdma-core posting routines to GPU runtime, we can saturate a 100-Gbps link without any CPU cycle, reducing the overall system response time, while increasing the power efficiency and improving the application throughput.The second contribution concerns the analysis of Clockwork, a State-of-The-Art inference serving system, showing the limitations imposed by controller-centric, CPU-mediated architectures. We then propose an alternative architecture to this system based on an RDMA transport, and we study some performance gains that such a system would introduce. An integral component of an inference system is to account and track user flows,and distribute them across multiple worker nodes. Our third contribution aims to understand the challenges of Connection Tracking applications running at 100Gbps, in the context of a Stateful Load Balancer running on commodity hardware. / <p>QC 20240315</p>