La consommation d'énergie des centres de données est un enjeu majeur. Leurs communications internes représentent près du quart de cette consommation. Les technologies de commutation optique ont en principe une meilleure efficacité énergétique que les solutions actuelles. Ce travail porte sur les réseaux optiques en mode paquet pour des centres de données de petite et moyenne taille. Il s'est déroulé dans le cadre du projet EPOC (Energy Proportional and Opportunistic Computing) qui s'intéresse à la réduction de la consommation d'énergie d'un centre de données alimenté partiellement par des sources renouvelables. Une hypothèse clé est l'absence d'un réseau de stockage dédié réduisant ainsi la consommation des interconnexions. Par contre, afin de pouvoir éteindre certains serveurs selon la charge de travail et l'énergie disponible, le débit doit être de l'ordre de 100 Gbit/s. Après un état de l'art des réseaux optiques pour centre de données nous choisissons une solution reposant sur une infrastructure entièrement passive et des émetteurs rapidement accordables en longueur d'onde, proposée récemment dans la littérature (POPI).Nous étudions les limitations dues aux composants (pertes d'insertion, plage d'accord et espacement des canaux). Nous proposons une extension (E-POPI) qui permet d'augmenter le nombre de serveurs connectés en utilisant plusieurs plages de longueurs d'onde. Pour les centres de données de plus grande taille nous proposons un réseau à deux étages (intra- et inter-racks) opérant respectivement dans les bandes C et L, POPI+. La connexion entre étages se fait via une passerelle transparente dans un sens et opaque dans l'autre. Différentes solutions de contrôle des deux étages sont détaillées.Un des éléments essentiels de faisabilité de ces architectures est de concilier la montée en débit avec les pertes du réseau passif d'interconnexion. Les techniques cohérentes des transmissions longue distance ne sont pas actuellement envisageables pour un centre de données. Nous avons donc étudié les formats PAM 4 et 8, par simulation avec différents débits (jusqu'à 112 Gbit/s et récepteurs (PIN, APD et SOA-PIN) et aussi, expérimentalement, à 12 et 18 Gbit/s. Nous avons développé une méthode de compensation des distorsions générées par les différents composants qui procure un compromis entre précision de correction et temps de calcul.Ces résultats nous permettent de déterminer les pertes d'insertion tolérables. Nous les combinons avec les limitations liées à la plage d'accord des émetteurs et à l'encombrement spectral des canaux occupant des fenêtres multiples de 12,5 GHz pour dimensionner les différentes architectures. Les réseaux POPI, E-POPI et POPI+ permettent respectivement la connexion de 48, 99 et 2352 entités à 112 Gbit/s. Nos évaluations tiennent compte d'une possible dispersion des caractéristiques des principaux composants. / Data-center energy consumption is nowadays a major issue. Intra-data-center networking accounts almost for a quarter of the data-center total power consumption. Optical switching technologies could provide higher power efficiency than current solutions based on electrical-packet switching. This work focuses on optical-packet-switched networks for small- and medium-size data centers. It takes part of the EPOC (Energy-Proportional and Opportunistic Computing) project, which main interest consists on reducing the overall power consumption of a data center partially powered by renewable sources. A key assumption is that our data center does not rely on a dedicated storage network, in order to reduce the consumption of those interconnections. In addition, with the aim of being able to turn off some servers according to the workload and the available energy, the bit rate must be close to 100 Gbit/s. We have chosen, after studying the state of the art of data-center interconnects, a purely passive network architecture based on fast-wavelength-tunable transmitters under the name of POPI.We study POPI's limitations due to its components (insertion loss, tuning range and channel spacing). We then propose an extension called E-POPI that allows to increase the number of connected servers by using several transmission bands. For larger data centers, we propose POPI+, a two-stage infrastructure for intra- and inter-rack communications operating in the C and L bands, respectively. The connection between both stages is done via a transparent gateway in one direction and an opaque one in the other. We discuss different control solutions for both stages.The feasibility of these architectures depends on, among other factors, dealing with bit-rate increasing and power losses of a passive interconnect. Coherent long-distance-transmission techniques are not currently suited to data centers. We therefore studied PAM 4 and 8 modulation formats with direct detection. On one hand, by simulation, with different bit rates (up to 112 Gbit/s) and receivers (PIN, APD and SOA-PIN) and, on the other hand, experimentally, at 12 and 18 Gbit/s. We have developed a method for compensating the distortions generated by the different network components. Our method takes into account a good tradeoff between correction accuracy and computation time.Simulation results allow us to determine the amount of insertion loss that may be supported. We then combine these results with the limitations of transmitters-tuning range and channel spacing using multiple of 12.5 GHz slots for dimensioning the proposed architectures. POPI, E-POPI and POPI+ interconnects allow the connection of 48, 99 and 2352 entities, respectively, at 112 Gbit/s. Our assessments take into account a potential dispersion of the characteristics of the main architecture components.

Submitted by Marta Toyoda (1144061@mackenzie.br) on 2018-09-03T21:27:11Z No. of bitstreams: 2 ISRAEL AVANSI MARQUES.pdf: 3183643 bytes, checksum: 04b9d57a4cfc26900497e82d58eecbf5 (MD5) license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) / Approved for entry into archive by Paola Damato (repositorio@mackenzie.br) on 2018-09-27T18:42:40Z (GMT) No. of bitstreams: 2 ISRAEL AVANSI MARQUES.pdf: 3183643 bytes, checksum: 04b9d57a4cfc26900497e82d58eecbf5 (MD5) license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) / Made available in DSpace on 2018-09-27T18:42:40Z (GMT). No. of bitstreams: 2 ISRAEL AVANSI MARQUES.pdf: 3183643 bytes, checksum: 04b9d57a4cfc26900497e82d58eecbf5 (MD5) license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) Previous issue date: 2018-06-14 / Conselho Nacional de Desenvolvimento Científico e Tecnológico / Fundação de Amparo a Pesquisa do Estado de São Paulo / Fundo Mackenzie de Pesquisa / Two-dimensional materials herald the coming of advances in many areas of Engineering, thanks to the peculiar physicochemical properties obtained when one isolates a single layer of minerals such as graphite and black phosphorus, for example. In the field of optical communications, graphene, composed by only one layer obtained from graphite, has been conquering prominence in applications regarding electro-optic modulators. Graphene exhibits potential to diminish the size and energy consumption of integrated photonic devices operating in the C-band of telecommunications (1550 nm), as well as to increase their transmission rates. The use of graphene on integrated photonics devices has been proposed and demonstrated with encouraging results concerning binary amplitude modulation. By means of computational simulations on COMSOL MultiPhysics v5.3 software, graphene-based electro-optical modulators for PAM-4 multilevel amplitude modulation with no requirement of digital-to-analog conversion (DAC) were designed in this work. The PAM-4 format allows a two-fold increase compared to the binary format without escalating the electro-optical bandwidth. Hence, modulation speeds higher than 200 Gbit/s are calculated from the fashioned devices with electro-optical bandwidths of 83,3 GHz. Recently, two-dimensional black phosphorus has also been explored for applications in modulators operating on the mid-infrared regime. Simulations were realized evaluating the use of black phosphorus on the 1550 nm wavelength regarding amplitude modulation for performance confrontation with graphene. It is shown that when compared with monolayer graphene, black phosphorus would display superior functioning if employed starting from a thickness corresponding to 12 layers. / Os materiais bidimensionais prometem avanços em diversas áreas da engenharia, graças às propriedades físico-químicas peculiares obtidas pela separação de uma única camada de minerais como por exemplo, o grafite e o fósforo negro. No campo das comunicações ópticas, o grafeno, composto por uma única camada atômica obtida a partir do grafite, tem angariado destaque em aplicações concernentes à modulação eletro-óptica. O mesmo mostra-se promissor para reduzir o tamanho, o consumo de energia e aumentar a taxa de transmissão de dados de dispositivos fotônicos integrados operando na banda C das telecomunicações (1550 nm). A utilização do grafeno em dispositivos fotônicos integrados tem sido proposta e demonstrada na plataforma da fotônica do silício com resultados promissores na modulação de amplitude binária. Por meio de simulações computacionais no software COMSOL MultiPhysics v5.3, foram projetados nesse trabalho moduladores eletro-ópticos baseados em grafeno para modulação de amplitude multinível PAM-4 sem necessidade de conversão digital-analógica (DAC). O formato PAM-4 permite dobrar a taxa de bits em relação ao formato binário sem aumento da banda eletro-óptica, dobrando assim a eficiência espectral. Logo, taxas de modulação maiores que 200 Gbit/s são calculadas nos dispositivos projetados com bandas eletro-ópticas de 83,3 GHz. Recentemente, o fósforo negro bidimensional também tem sido explorado para aplicações em moduladores operando no infravermelho médio. Simulações foram realizadas avaliando o uso de fósforo negro no comprimento de onda de 1550 nm para modulação de amplitude para confrontação de desempenho com o grafeno. Demonstra-se que quando comparado ao grafeno monocamada, o fósforo negro apresentaria uma performance superior se empregado a partir de uma espessura correspondente a 12 camadas.