Mapeamento e adaptação de rotas de comunicação em redes em chip

• Ano de defesa: 2010
• Autor(a) principal: Moreno, Edson Ifarraguirre
• Orientador(a): Calazans, Ney Laert Vilar
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Pontifícia Universidade Católica do Rio Grande do Sul; Porto Alegre
• Palavras-chave em Português: INFORMÁTICA; ARQUITETURA DE REDES; ROTEAMENTO - REDES DE COMPUTADORES; MULTIPROCESSADORES; REDES DE COMUNICAÇÃO DE COMPUTADORES
• Link de acesso: http://hdl.handle.net/10923/1470

The constant evolution of market needs requires the availability of computing systems with ever- growing performance. Increases in operating frequencies and instruction level parallelism in microprocessors are not alone sufficient anymore to guarantee performance scaling for such systems. A way to achieve scaling performance has been the adoption of Multiprocessor Systems on a Chip (MPSoCs), which allow distributing application computation costs along a set of processing elements in the MPSoC. The increase in the number of MPSoC processing elements as technology advances into the deep submicron domain is a clear trend. To interconnect such elements it is necessary to employ more efficient communication infrastructures in what concerns electrical characteristics, facility of adoption by designers and performance. Networks on Chip (NoCs) or on chip networks are a clear trend in this sense. In the same way computation performance is expected to increase, so is the performance of communication among processing elements in future MPSoCs. Obviously, NoCs may still suffer from congestion, which degrades the communication quality due, for example, to the increase of latency while delivering messages. The use of adaptive routing algorithms in NoCs allows altogether to solve or at least to alleviate congestion scenarios, but adaptation decisions usually employ only the instantaneous state of the communication network and are based on local information. The problem with this kind of approach is the unpredictability of latency in delivering network packets since the path followed by each packet depends on the state of the network and on the rules adopted by the routing algorithm as well. Additionally, deviations from a route considered congested can take packets to regions with even higher traffic. This thesis proposes two communication infrastructures that allow an improved degree of predictability and are thus more useful to fulfill application communication requirements. Both infrastructures employ source routing strategies. The first, called Hermes-SR NoC, explores the mapping of communication routes at design time. Initial results demonstrate this infrastructure displays gains when compared to the Hermes NoC, a network without congestion solving mechanisms, which uses an XY deterministic routing algorithm. The second infrastructure, called MoNoC (Monitored NoC), explores resources that enable dynamic route adaptation to take place. These resources include special network interfaces, monitors e network probes. Experimental results achieved with MoNoC showed significant reductions for application latency. In both cases, the adoption of adaptive routing algorithms as a base to compute routes enables to turn traffic away from congestion points in the network, which naturally increases latency and packet delivery predictability.