Three-Dimensional (3D) integration of circuits based on die and wafer stacking using through-silicon-via is a critical technology in enabling "more-than-Moore", i.e. functional integration of devices beyond pure scaling ("more Moore"). In particular, the scaling from multi-core to many-core architecture is an excellent candidate for such integration. 3D systems design follows is a challenging and a complex design process involving integration of heterogeneous technologies. It is also expensive to prototype because the 3D industrial ecosystem is not yet complete and ready for low-cost mass production. Networks-on-Chip (NoCs) efficiently facilitates the communication of massively integrated cores on 3D many-core architecture. In this thesis scalability and performance issues of NoCs are explored in terms of architecture, organization and functionality of many-core systems. First, we evaluate on-chip network performance in massively integrated many-core architecture when network size grows. We propose link and channel models to analyze the network traffic and hence the performance. We develop a NoC simulation framework to evaluate the performance of a deflection routing network as the architecture scales up to 1000 cores. We propose and perform comparative analysis of 3D processor-memory model configurations in scalable many-core architectures. Second, we investigate how the deflection routing NoCs can be designed to maximize the benefit of the fast TSVs through clock pumping techniques. We propose multi-rate models for inter-layer communication. We quantify the performance benefit through cycle-accurate simulations for various configurations of 3D architectures. Finally, the complexity of massively integrated many-core architecture by itself brings a multitude of design challenges such as high-cost of prototyping, increasing complexity of the technology, irregularity of the communication network, and lack of reliable simulation models. We formulate a zero-load average distance model that accurately predicts the performance of deflection routing networks in the absence of data flow by capturing the average distance of a packet with spatial and temporal probability distributions of traffic. The thesis research goals are to explore the design space of vertical integration for many-core applications, and to provide solutions to 3D technology challenges through architectural innovations. We believe the research findings presented in the thesis work contribute in addressing few of the many challenges to the field of combined research in many-core architectural design and 3D integration technology. / <p>QC 20151221</p>

Orientadores: Edson Moschim, Felipe Rudge Barbosa / Tese (doutorado) - Universidade Estadual de Campinas, Faculdade de Engenharia Eletrica e de Computação / Made available in DSpace on 2018-08-07T06:09:06Z (GMT). No. of bitstreams: 1 Nascimento_LuizHenriqueBonanido_D.pdf: 2297657 bytes, checksum: b04918af9ef98f4cae4bd6c3d6d4c0ca (MD5) Previous issue date: 2006 / Resumo: Hoje em dia, como uma multiplicidade muito grande de tecnologias converge para as redes fotônicas, a demanda por vazão cresce. Os protocolos para acesso múltiplo e simultâneo às portas de entrada da rede, bem como a abordagem do formato dos pacotes fotônicos e a arquitetura das redes, são considerados como alternativas tecnológicas para a melhora do desempenho das redes fotônicas. Deste modo, apresenta-se aqui uma proposta para a arquitetura de redes fotônicas,utilizando topologias em malha otimizadas para que os pacotes atinjam seus respectivos nós de destino no menor número de hops possível, com nós de rede 2x2. Além disso, utiliza-se o roteamento por deflexão para a resolução de contenda entre pacotes ópticos, bem como buffers eletrônicos na entrada da rede óptica para evitar problemas de perdas de inserção. Além disso, a estrutura de rede permite hierarquização e também se propõe um modelo para o funcionamento do nó que faz a ligação entre hierarquias diferentes dentro da rede óptica. Esta arquitetura de rede é então avaliada quanto à sua capacidade de prover requisitos para QoS em termos de perdas de pacotes, atrasos e inversão de ordem de pacotes, para a qual é proposta uma heurística para uma avaliação quantitativa. Os resultados mostram que uma arquitetura nos moldes apresentados é capaz de prover ganhos de desempenho, oferecendo condições para a garantia de QoS, provida por protocolos de rede em nível superior, até cerca de 70% de carga / Abstract: Nowadays, as a multiplicity of technologies converges upon the optical network, the demand for throughput increases. The protocols for multiple and simultaneous access to network ports, as well as the optical packet approach should be considered as technological alternatives to optical network performance upgrade. Thus, we present here a new proposal to the architecture of optical packet networks, using mesh topologies optimized in such a way that the packets reach their respective destination nodes with the smaller possible number of hops, and using 2x2 network nodes. Besides, we use deflection routing to solve the contention between optical packets, as well as electronical buffers at the optical network input to prevent problems with insertion losses. The structure of the network also allows hierarchization, performing the link between different hierarchies on the optical network. The performance of the proposed architecture is evaluated by the capacity of providing requirements of QoS, in terms of packet losses, delays and the amount of out of order packets, for which is presented a heuristic to get a quantitative estimation. The results show that the proposed architecture is able to provide performance gains, giving conditions to guarantee some levels of QoS, which is done by the specific protocols of superior network layers, up to 70% of network load / Doutorado / Telecomunicações e Telemática / Doutor em Engenharia Elétrica

[pt] Redes Manhattan Street (MS) têm sido descritas como um arranjo linear bidimensional de nós, semelhante à configuração de ruas e avenidas de Manhattan. O roteamento por deflexão é implementado encaminhando os pacotes que atingem um determinado nó a uma de suas saídas de forma síncrona ou assíncrona. O principal objetivo deste trabalho consiste na simulação e análise de redes totalmente ópticas configuradas segundo a topologia MS. O roteamento por deflexão e o assincronismo são considerados, para evitar complexidade eletrônica e armazenamento de pacotes no domínio óptico. Serão apresentadas as características das redes MS, suas propriedades estruturais e os parâmetros utilizados para analisar seu desempenho. Uma metodologia analítica dedicada a obtenção teórica destes parâmetros será introduzida. Serão apresentados alguns conceitos básicos sobre simulação de redes;diversas simulações da rede proposta utilizando os protocolos UDP e TCP; uma descrição do software que foi utilizado para realizar as simulações; uma comparação entre os resultados obtidos através da simulação e os obtidos através da metodologia analítica; e uma análise do efeito da latência na vazão do protocolo TCP. / [en] Manhattan Street (MS) Networks are bidimensional linear node sets arranged as the avenues and streets of Manhattan. The simulation and analysis of all-optical MS networks is the central target of this paper. In order to avoid using complex electronics and/or optical domain buffers, the deflection routing and the asynchronism are taken into account in the analysis. Deflection routing is performed by conveying incoming packets towards one of the two outputs. The characteristics of MS Networks are presented, along with their structural properties and the parameters used for performance analysis. An analytical methodology for the theoretical obtaining of these parameters is described. Some basic concepts on network simulation are discussed. Several simulations of the proposed network are presented, using both UDP and TCP protocols, and the software used for simulations is also described. The obtained results are compared and discussed with respect to the previously described analytic methodologies. Finally, the effect of network latency on the TCP-protocol throughput is assessed.