The emerging cloud services have motivated a fresh look at the design of data center network infrastructure in multiple layers. To transfer the huge amount of data generated by many data intensive applications, data center network has to be fast, scalable and power efficient. To support flexible and efficient sharing in cloud services, service providers deploy a virtualization layer as part of the data center infrastructure. This thesis explores the design and performance analysis of data center network infrastructure in both physical network and virtualization layer. On the physical network design front, we present a hybrid packet/circuit switched network architecture which uses circuit switched optics to augment traditional packet-switched Ethernet in modern data centers. We show that this technique has substantial potential to improve bisection bandwidth and application performance in a cost-effective manner. To push the adoption of optical circuits in real cloud data centers, we further explore and address the circuit control issues in shared data center environments. On the virtualization layer, we present an analytical study on the network performance of virtualized data centers. Using Amazon EC2 as an experiment platform, we quantify the impact of virtualization on network performance in commercial cloud. Our findings provide valuable insights to both cloud users in moving legacy application into cloud and service providers in improving the virtualization infrastructure to support better cloud services.

Today, data centers host a wide variety of applications which generate a mix of diverse internal data center traffic. In a data center environment 90% of the traffic flows, though they constitute only 10% of the data carried around, are short flows with sizes up to a maximum of 1MB. The rest 10% constitute long flows with sizes in the range of 1MB to 1GB. Throughput matters for the long flows whereas short flows are latency sensitive. This thesis studies various data center transport mechanisms aimed at either improving flow completion time for short flows or throughput for long flows. Thesis puts forth two data center transport mechanisms: (1) for improving flow completion time for short flows (2) for improving throughput for long flows. The first data center transport mechanism proposed in this thesis, FA-DCTCP (Flow Aware DCTCP), is based on Data Center Transmission Control Protocol (DCTCP). DCTCP is a Transmission Control Protocol (TCP) variant for data centers pioneered by Microsoft, which is being deployed widely in data centers today. DCTCP congestion control algorithm treats short flows and long flows equally. This thesis demonstrate that, treating them differently by reducing the congestion window for short flows at a lower rate compared to long flows, at the onset of congestion, 99th percentile of flow completion time for short flows could be improved by up to 32.5%, thereby reducing their tail latency by up to 32.5%. As per data center traffic measurement studies, data center internal traffic often exhibit predefined patterns with respect to the traffic flow mix. The second data center transport mechanism proposed in this thesis shows that, insights into the internal data center traffic composition could be leveraged to achieve better throughput for long flows. The mechanism for the same is implemented by adopting the Software Defined Networking paradigm, which offers the ability to dynamically adapt network configuration parameters based on network observations. The proposed solution achieves up to 22% improvement in long flow throughput, by dynamically adjusting network element’s QoS configurations, based on the observed traffic pattern.

Current telecommunication infrastructure is undergoing significant changes. Such changes involve the type of traffic traveling through the network as well as the requirements imposed by the new traffic mix (e.g. strict delay control and low end-to-end delay). In this new networking scenario, the current infrastructure, which remained almost unchanged for the last several decades, is struggling to adapt, and its limitations in terms of power consumption, scalability, and economical viability have become more evident. In this dissertation we explore the potential advantages of using periodic data structures to handle efficiently bandwidth-intensive transactions, which constitute a significant portion of today's network traffic. We start by implementing an approach that can work as a standalone system aiming to provide the same advantages promised by all-optical approaches such as OBS and OFS. We show that our approach is able to provide similar advantages (e.g. energy efficiency, link utilization, and low computational load for the network hardware) while avoiding the drawbacks (e.g. use of optical buffers, inefficient resource utilization, and costly deployment), using commercially available hardware. Aware of the issues of large scale hardware redeployment, we adapt our approach to work within the current transport network architecture, reusing most of the hardware and protocols that are already in place, offering a more gradual evolutionary path, while retaining the advantages of our standalone system. We then apply our approach to Data Center Networks (DCNs), showing its ability to achieve significant improvements in terms of network performance stability, predictability, performance isolation, agility, and goodput with respect to popular DCN approaches. We also show our approach is able to work in concert with many proposed and deployed DCN architectures, providing DCNs with a simple, efficient, and versatile protocol to handle bandwidth-intensive applications within the DCs. / Graduate

Du fait de la croissance de la demande en ressources de calcul, les architectures de centres de données gagnent en taille et complexité.Dès lors, cette thèse prend du recul par rapport aux architectures réseaux traditionnelles, et montre que fournir des primitives génériques directement à la couche réseau permet d'améliorer l'utilisation des ressources, et de diminuer le trafic réseau et le surcoût administratif.Deux architectures réseaux récentes, Segment Routing (SR) et Bit-Indexed Explicit Replication (BIER), sont utilisées pour construire et analyser des protocoles de couche réseau, afin de fournir trois primitives: (1) mobilité des tâches, (2) distribution fiable de contenu, et (3) équilibre de charge.Premièrement, pour la mobilité des tâches, SR est utilisé pour fournir un service de migration de machine virtuelles sans perte.Cela ouvre l'opportunité d'étudier comment orchestrer le placement et la migration de tâches afin de (i) maximiser le débit inter-tâches, tout en (ii) maximisant le nombre de nouvelles tâches placées, mais (iii) minimisant le nombre de tâches migrées.Deuxièmement, pour la distribution fiable de contenu, BIER est utilisé pour fournir un protocole de multicast fiable, dans lequel les retransmissions de paquets perdus sont ciblés vers l'ensemble précis de destinations n'ayant pas reçu ce packet : ainsi, le surcoût de trafic est minimisé.Pour diminuer la charge sur la source, cette approche est étendue en rendant possible des retransmissions par des pairs locaux, utilisant SR afin de trouver un pair capable de retransmettre.Troisièmement, pour l'équilibre de charge, SR est utilisé pour distribuer des requêtes à travers plusieurs applications candidates, chacune prenant une décision locale pour accepter ou non ces requêtes, fournissant ainsi une meilleure équité de répartition comparé aux approches centralisées.La faisabilité d'une implémentation matérielle de cette approche est étudiée, et une solution (utilisant des canaux cachés pour transporter de façon invisible de l'information vers l'équilibreur) est implémentée pour une carte réseau programmable de dernière génération.Finalement, la possibilité de fournir de l'équilibrage automatique comme service réseau est étudiée : en faisant passer (avec SR) des requêtes à travers une chaîne fixée d'applications, l'équilibrage est initié par la dernière instance, selon son état local. / With the development of demand for computing resources, data center architectures are growing both in scale and in complexity.In this context, this thesis takes a step back as compared to traditional network approaches, and shows that providing generic primitives directly within the network layer is a great way to improve efficiency of resource usage, and decrease network traffic and management overhead.Using recently-introduced network architectures, Segment Routing (SR) and Bit-Indexed Explicit Replication (BIER), network layer protocols are designed and analyzed to provide three high-level functions: (1) task mobility, (2) reliable content distribution and (3) load-balancing.First, task mobility is achieved by using SR to provide a zero-loss virtual machine migration service.This then opens the opportunity for studying how to orchestrate task placement and migration while aiming at (i) maximizing the inter-task throughput, while (ii) maximizing the number of newly-placed tasks, but (iii) minimizing the number of tasks to be migrated.Second, reliable content distribution is achieved by using BIER to provide a reliable multicast protocol, in which retransmissions of lost packets are targeted towards the precise set of destinations having missed that packet, thus incurring a minimal traffic overhead.To decrease the load on the source link, this is then extended to enable retransmissions by local peers from the same group, with SR as a helper to find a suitable retransmission candidate.Third, load-balancing is achieved by way of using SR to distribute queries through several application candidates, each of which taking local decisions as to whether to accept those, thus achieving better fairness as compared to centralized approaches.The feasibility of hardware implementation of this approach is investigated, and a solution using covert channels to transparently convey information to the load-balancer is implemented for a state-of-the-art programmable network card.Finally, the possibility of providing autoscaling as a network service is investigated: by letting queries go through a fixed chain of applications using SR, autoscaling is triggered by the last instance, depending on its local state.