Les centres de données verts de dernière génération ont été conçus pour une consommation optimisée et une meilleure qualité du niveau de service SLA. Cependant,ces dernières années, le marché des centres de données augmente rapidement,et la concentration de la puissance de calcul est de plus en plus importante, ce qui fait augmenter les besoins en puissance électrique et refroidissement. Un centre de données est constitué de ressources informatiques, de systèmes de refroidissement et de distribution électrique. De nombreux travaux de recherche se sont intéressés à la réduction de la consommation des centres de données afin d'améliorer le PUE, tout en garantissant le même niveau de service. Certains travaux visent le dimensionnement dynamique des ressources en fonction de la charge afin de réduire le nombre de serveurs démarrés, d'autres cherchent à optimiser le système de refroidissement qui représente un part important de la consommation globale.Dans cette thèse, afin de réduire le PUE, nous étudions la mise en place d'un système autonome d'optimisation globale du refroidissement, qui se base sur des sources de données externes tel que la température extérieure et les prévisions météorologiques, couplé à un module de prédiction de charge informatique globale pour absorber les pics d'activité, pour optimiser les ressources utilisés à un moindre coût, tout en préservant la qualité de service. Afin de garantir un meilleur SLA, nous proposons une architecture distribuée pour déceler les anomalies de fonctionnements complexes en temps réel, en analysant de gros volumes de données provenant des milliers de capteurs du centre de données. Détecter les comportements anormaux au plus tôt, permet de réagir plus vite face aux menaces qui peuvent impacter la qualité de service, avec des boucles de contrôle autonomes qui automatisent l'administration. Nous évaluons les performances de nos contributions sur des données provenant d'un centre de donnée en exploitation hébergeant des applications réelles. / Next-generation green datacenters were designed for optimized consumption and improved quality of service level Service Level Agreement (SLA). However, in recent years, the datacenter market is growing rapidly, and the concentration of the computing power is increasingly important, thereby increasing the electrical power and cooling consumptions. A datacenter consists of computing resources, cooling systems, and power distribution. Many research studies have focused on reducing the consumption of datacenters to improve the PUE, while guaranteeing the same level of service. Some works aims the dynamic sizing of resources according to the load, to reduce the number of started servers, others seek to optimize the cooling system which represents an important part of total consumption. In this thesis, in order to reduce the PUE, we study the design of an autonomous system for global cooling optimization, which is based on external data sources such as the outside temperature and weather forecasting, coupled with an overall IT load prediction module to absorb the peaks of activity, to optimize activere sources at a lower cost while preserving service level quality. To ensure a better SLA, we propose a distributed architecture to detect the complex operation anomalies in real time, by analyzing large data volumes from thousands of sensors deployed in the datacenter. Early identification of abnormal behaviors, allows a better reactivity to deal with threats that may impact the quality of service, with autonomous control loops that automate the administration. We evaluate the performance of our contributions on data collected from an operating datacenter hosting real applications.

Datahallar behöver reservkraft av god kvalitet för att garantera upprätthållandet av dess funktion. Laster i en datahall kommer att generera ström och spänningsövertoner som kan skapa problematik med elkvaliteten. Coromatic är intresserade av att veta mer hur dessa laster påverkar reservkraftsgeneratorn. En mätning utfördes på en datahall under ett funktionsprov. Resultatet blev att halten av THD ökade, främst är det 3:e övertonen som är framträdande. Mätvärdena för THDV ligger under gränsvärdena för SS-EN 50160 och 61000-2-2, men gränsvärdena för 3:e ton ligger långt över. Ingen åtgärd föreslås i nuläget för att hantera problemet. Denna avhandling har gett värdefull information till Coromatic att ta i beaktande vid utförande av nya reservkraftsanläggningar. / A facility full with computers needs backup-power to guarantee the function. Loads in this facility will produce current and voltage harmonics that can pollute and cause trouble with the quality of electricity. Coromatic are interested in knowing more about how these loads can affect the generator. A measurement was performed on a facility when they ran a functional test. The results indicated that THD increased, the third harmonic turned out to be the single harmonic with the highest value. The value is within the boundaries for THDV according to SS-EN 50160 and 61000-2-2, except for the third harmonic. Its value was far too high. No action is propsed to deal with the problem at the moment. This thesis has provided Coromatic with valuable information too consider when they building new systems.

Software-Defined Networking (SDN) is an emerging trend in networking that offers a number of advantages such as smoother network management over traditional networks. By decoupling the control and data planes from network elements, a huge amount of new opportunities arise, especially in network virtualization. In cloud datacenters, where virtualization plays a fundamental role, SDN presents itself as the perfect candidate to ease infrastructure management and to ensure correct operation. Even if the original SDN ideology advocates openness of source and interfaces, multiple networking vendors offer their own proprietary solutions. In this work, an open-source SDN solution, named Tungsten Fabric, will be deployed in a virtualized datacenter and a number of SDN-related use-cases will be examined. The main goal of this work is to determine whether Tungsten Fabric can deliver the same set of use-cases as a proprietary solution from Juniper, named Contrail Cloud. Finally, this work will give some guidelines on whether open-source SDN is the right candidate for Ericsson.

A interferência de desempenho é um desafio bem conhecido em redes de datacenter (DCNs), permanecendo um tema constante de discussão na literatura. Diversos estudos concluíram que a largura de banda disponível para o envio e recebimento de dados entre máquinas virtuais (VMs) pode variar por um fator superior a cinco, resultando em desempenho baixo e imprevisível para as aplicações. Trabalhos na literatura têm proposto técnicas que resultam em subutilização de recursos, introduzem sobrecarga de gerenciamento ou consideram somente recursos de rede. Nesta tese, são apresentadas três propostas para lidar com a interferência de desempenho em DCNs: IoNCloud, Predictor e Packer. O IoNCloud está baseado na observação que diferentes aplicações não possuem pico de damanda de banda ao mesmo tempo. Portanto, ele busca prover desempenho previsível e garantido enquanto minimiza a subutilização dos recursos de rede. Isso é alcançado por meio (a) do agrupamento de aplicações (de acordo com os seus requisitos temporais de banda) em redes virtuais (VNs); e (b) da alocação dessas VNs no substrato físico. Apesar de alcançar os seus objetivos, ele não provê conservação de trabalho entre VNs, o que limita a utilização de recursos ociosos. Nesse contexto, o Predictor, uma evolução do IoNCloud, programa dinamicamente a rede em DCNs baseadas em redes definidas por software (SDN) e utiliza dois novos algoritmos para prover garantias de desempenho de rede com conservação de trabalho. Além disso, ele foi projetado para ser escalável, considerando o número de regras em tabelas de fluxo e o tempo de instalação das regras para um novo fluxo em DCNs com milhões de fluxos ativos. Apesar dos benefícios, o IoNCloud e o Predictor consideram apenas os recursos de rede no processo de alocação de aplicações na infraestrutura física. Isso leva à fragmentação de outros tipos de recursos e, consequentemente, resulta em um menor número de aplicações sendo alocadas. O Packer, em contraste, busca prover desempenho de rede previsível e garantido e minimizar a fragmentação de diferentes tipos de recursos. Estendendo a observação feita ao IoNCloud, a observação-chave é que as aplicações têm demandas complementares ao longo do tempo para múltiplos recursos. Desse modo, o Packer utiliza (i) uma nova abstração para especificar os requisitos temporais das aplicações, denominada TI-MRA (Time- Interleaved Multi-Resource Abstraction); e (ii) uma nova estratégia de alocação de recursos. As avaliações realizadas mostram os benefícios e as sobrecargas do IoNCloud, do Predictor e do Packer. Em particular, os três esquemas proveem desempenho de rede previsível e garantido; o Predictor reduz o número de regras OpenFlow em switches e o tempo de instalação dessas regras para novos fluxos; e o Packer minimiza a fragmentação de múltiplos tipos de recursos. / Performance interference has been a well-known problem in datacenter networks (DCNs) and one that remains a constant topic of discussion in the literature. Several measurement studies concluded that throughput achieved by virtual machines (VMs) in current datacenters can vary by a factor of five or more, leading to poor and unpredictable overall application performance. Recent efforts have proposed techniques that present some shortcomings, such as underutilization of resources, significant management overhead or negligence of non-network resources. In this thesis, we introduce three proposals that address performance interference in DCNs: IoNCloud, Predictor and Packer. IoNCloud leverages the key observation that temporal bandwidth demands of cloud applications do not peak at exactly the same time. Therefore, it seeks to provide predictable and guaranteed performance while minimizing network underutilization by (a) grouping applications in virtual networks (VNs) according to their temporal network usage and need of isolation; and (b) allocating these VNs on the cloud substrate. Despite achieving its objective, IoNCloud does not provide work-conserving sharing among VNs, which limits utilization of idle resources. Predictor, an evolution over IoNCloud, dynamically programs the network in Software-Defined Networking (SDN)-based DCNs and uses two novel algorithms to provide network guarantees with work-conserving sharing. Furthermore, Predictor is designed with scalability in mind, taking into consideration the number of entries required in flow tables and flow setup time in DCNs with high turnover and millions of active flows. IoNCloud and Predictor neglect resources other than the network at allocation time. This leads to fragmentation of non-network resources and, consequently, results in less applications being allocated in the infrastructure. Packer, in contrast, aims at providing predictable and guaranteed network performance while minimizing overall multi-resource fragmentation. Extending the observation presented for IoNCloud, the key insight for Packer is that applications have complementary demands across time for multiple resources. To enable multi-resource allocation, we devise (i) a new abstraction for specifying temporal application requirements (called Time-Interleaved Multi-Resource Abstraction – TI-MRA); and (ii) a new allocation strategy. We evaluated IoNCloud, Predictor and Packer, showing their benefits and overheads. In particular, all of them provide predictable and guaranteed network performance; Predictor reduces flow table size in switches and flow setup time; and Packer minimizes multi-resource fragmentation.

Due to the imperative need to reduce the management costs, operators multiplex several concurrent applications in large datacenters. However, uncontrolled resource sharing between co-hosted applications often results in performance degradation problems, thus creating violations of service level agreements (SLAs) for service providers. Therefore, in order to meet per-application SLAs, per-application performance modeling for dynamic resource allocation in shared resource environments has recently become promising. We introduce Chorus, an interactive performance modeling framework for building application performance models incrementally and on the fly. It can be used to support complex, multi-tier resource allocation, and/or what-if performance inquiry in modern datacenters, such as Clouds. Chorus consists of (i) a declarative high-level language for providing semantic model guidelines, such as model templates, model functions, or sampling guidelines, from a sysadmin or a performance analyst, as model approximations to be learned or refined experimentally, (ii) a runtime engine for iteratively collecting experimental performance samples, validating and refining performance models. Chorus efficiently builds accurate models online, reuses and adjusts archival models over time, and combines them into an ensemble of models. We perform an experimental evaluation on a multi-tier server platform, using several industry- standard benchmarks. Our results show that Chorus is a flexible modeling framework and knowledge base for validating, extending and reusing existing models while adapting to new situations.

Due to the imperative need to reduce the management costs, operators multiplex several concurrent applications in large datacenters. However, uncontrolled resource sharing between co-hosted applications often results in performance degradation problems, thus creating violations of service level agreements (SLAs) for service providers. Therefore, in order to meet per-application SLAs, per-application performance modeling for dynamic resource allocation in shared resource environments has recently become promising. We introduce Chorus, an interactive performance modeling framework for building application performance models incrementally and on the fly. It can be used to support complex, multi-tier resource allocation, and/or what-if performance inquiry in modern datacenters, such as Clouds. Chorus consists of (i) a declarative high-level language for providing semantic model guidelines, such as model templates, model functions, or sampling guidelines, from a sysadmin or a performance analyst, as model approximations to be learned or refined experimentally, (ii) a runtime engine for iteratively collecting experimental performance samples, validating and refining performance models. Chorus efficiently builds accurate models online, reuses and adjusts archival models over time, and combines them into an ensemble of models. We perform an experimental evaluation on a multi-tier server platform, using several industry- standard benchmarks. Our results show that Chorus is a flexible modeling framework and knowledge base for validating, extending and reusing existing models while adapting to new situations.

Machine virtualization has become a cornerstone of modern datacenters. It enables server consolidation as a means to reduce costs and increase efficiencies. The communication endpoints within the datacenter are now virtual machines (VMs), not physical servers. Consequently, the datacenter network now extends into the server and last hop switching occurs inside the server. Today, thanks to increasing core counts on processors, server VM densities are on the rise. This trend is placing enormous pressure on the network I/O subsystem and the last hop virtual switch to support efficient communication, both internal and external to the server. But the current state-of-the-art solutions fall short of these requirements. This thesis presents new architectures and mechanisms for the network subsystem in virtualized servers to build efficient virtualization platforms. Specifically, there are three primary contributions in this thesis. First, it presents a new mechanism to reduce memory sharing overheads in driver domain-based I/O architectures. The key idea is to enable a guest operating system to reuse its I/O buffers that are shared with a driver domain. Second, it describes Hyper-Switch, a highly streamlined, efficient, and scalable software-based virtual switching architecture, specifically for hypervisors that support driver domains. The Hyper-Switch combines the best of the existing architectures by hosting the device drivers in a driver domain to isolate any faults and placing the virtual switch in the hypervisor to perform efficient packet switching. Further, the Hyper-Switch implements several optimizations, such as virtual machine state-aware batching, preemptive copying, and dynamic offloading of packet processing to idle CPU cores, to enable efficient packet processing, better utilization of the available CPU resources, and higher concurrency. This architecture eliminates the memory sharing overheads associated with driver domains. Third, this thesis proposes an alternate virtual switching architecture, called sNICh, which explores the idea of server/switch integration. The sNICh is a combined network interface card (NIC) and datacenter switching accelerator. This takes the Hyper-Switch architecture one step further. It offloads the data plane of the switch to the network device, eliminating driver domains entirely.

Modern datacenters require a network with high cross-section bandwidth, fine-grained security, support for virtualization, and simple management that can scale to hundreds of thousands of hosts at low cost. This thesis first presents the firmware for Rain Man, a novel datacenter network architecture that meets these requirements, and then performs a general scalability study of the design space. The firmware for Rain Man, a scalable Software-Defined Networking architecture, employs novel algorithms and uses previously unused forwarding hardware. This allows Rain Man to scale at high performance to networks of forty thousand hosts on arbitrary network topologies. In the general scalability study of the design space of SDN architectures, this thesis identifies three different architectural dimensions common among the networks: source versus hop-by-hop routing, the granularity at which flows are routed, and arbitrary versus restrictive routing and finds that a source-routed, host-pair granularity network with arbitrary routes is the most scalable.

Paskutiniu metu duomenų centrai teikia skirtingas tinklo, informacijos talpinimo, terminalinių serverių ir kitas paslaugas. Paslaugų kokybė priklauso ne tik nuo specifikuojamų SLA parametrų, bet ir nuo vartotojo poreikių. Taigi buvo sudaryta duomenų centrų paslaugų įvertinimo metodika, skirta palyginti paslaugų kokybę tiek vartotojams, tiek ir paslaugų teikėjams. Siūloma taikyti serverių bei tinklo sujungimų rezervacijas, norint užtikrinti aukštą paslaugų pateikiamumą. Dvi populiariausios duomenų centrų paslaugos buvo ištirtos: „thin-client“ (terminalinių serverių) architektūra, ir Web paslaugos. Web paslaugos kokybės įvertinimo bandymai buvo atlikti, norint parodyti, kad WRT kokybės parametras yra labiau informatyvus nei įprasti duomenų perdavimo tinklo parametrai. „Thin-client“ architektūros analizė parodė, kad ji gali būti naudojama LAN ir WAN tinkluose, ir gali būti labai ekonomiškai naudinga. Tinklo ir serverių resursų stebėjimas turi būti vykdomas, norint teikti aukštos kokybės paslaugas - taigi duomenų centrų paslaugų teikėjai privalo naudoti reikiamą aparatūrinę ir programinę įrangą periodiškiems matavimams atlikti, bei saugoti ir apdoroti gautus rezultatus. / Nowadays datacenters provide different network, hosting, terminal desktop and other services, and it has become very important to be able to evaluate the quality of these services. SLA (Service Level Agreement) metrics define the quality of service, but it also depends on the needs of the particular user. So the evaluation methodology was created in order to compare the quality of services, both: for the user, and for the datacenter service provider. A service provider has to implement the server architecture with network connection and server redundancy in order to assure high service availability. Two most popular datacenter services were analyzed: “thin-client” architecture and web services. The web service quality evaluation experiments have been made to show that web oriented performance metric "Web Response Time" (WRT) is more informative for web service quality assurance, comparing to the traditional latency and loss metrics in the network environment. “Thin-client” architecture analysis show that it can be used both in LAN and WAN networks, and it can be economically useful to outsource IT systems to datacenters by using this technology. Network and server resource performance monitoring has to be done in order to supply high quality services - so datacenter service providers have to use monitoring hardware and software to make periodical measurements and collect, store and process the results.

<p>The soaring demands for always-on and fast-response online services have driven modern datacenter networks to undergo tremendous growth. These networks often rely on scale-out designs with large numbers of commodity switches to reach immense capacity while keeping capital expenses under check. Today, datacenter network operators spend tremendous time and efforts on two key challenges: 1) how to efficiently utilize the bandwidth connecting host pairs and 2) how to promptly handle network failures with minimal disruptions to the hosted services.</p><p>To resolve the first challenge, we propose solutions in both network layer and transport layer. In the network layer solution, We advocate to design practical datacenter architectures for easy operation, i.e., an architecture should be reliable, capable of improving bisection bandwidth, scalable and debugging-friendly. By strictly following these four guidelines, We propose DARD, a Distributed Adaptive Routing architecture for Datacenter networks. DARD allows each end host to reallocate traffic from overloaded paths to underloaded paths without central coordination. We use congestion game theory to show that DARD converges to a Nash equilibrium in finite steps and its gap to the optimal flow allocation is bounded in the order of 1/logL, with L being the number of links. We use a testbed implementation and simulations to show that DARD can achieve a close-to-optimal flow allocation with small control overhead in practice.</p><p>In the transport layer solution, We propose Explicit Multipath Congestion Control Protocol (MPXCP), which achieves four desirable properties: fast convergence, efficiency, being fair to flows with different RTTs and negligible queue size. Intensive ns-2 simulation shows that MPXCP can quickly converge to efficiency and fairness without building up queues despite different delay-bandwidth products.</p><p>To resolve the second challenge, recent research efforts have focused on automatic failure localization. Yet, resolving failures still requires significant human interventions, resulting in prolonged failure recovery time. Unlike previous work, we propose NetPilot, a system aims to quickly mitigate rather than resolve failures. NetPilot mitigates failures in much the same way operators do -- by deactivating or restarting suspected offending components. NetPilot circumvents the need for knowing the exact root cause of a failure by taking an intelligent trial-and-error approach. The core of NetPilot is comprised of an Impact Estimator that helps guard against overly disruptive mitigation actions and a failure-specific mitigation planner that minimizes the number of trials. We demonstrate that NetPilot can effectively mitigate several types of critical failures commonly encountered in production datacenter networks.</p> / Dissertation