Técnicas de tolerância a falhas aplicadas a redes intra-chip

• Ano de defesa: 2015
• Autor(a) principal: Fochi, Vinicius Morais
• Orientador(a): Moraes, Fernando Gehm
• Tipo de documento: Dissertação
• 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 COMPUTADOR; TOLERÂNCIA A FALHAS (INFORMÁTICA); MULTIPROCESSADORES
• Link de acesso: http://hdl.handle.net/10923/7393

The continuous development of the transistor technology has enabled hundreds of processors to work interconnected by a NoC (network-on-chip). Nanotechnology has enabled the development of complex systems, however, fault vulnerability also increased. The literature presents partial solutions for fault tolerance issues, targeting parts of the system. An important gap in the literature is an integrated method from the router-level fault detection to the correct execution of applications in the MPSoC. The main goal of this dissertation is to present a fault-tolerant method from the physical layer to the transport layer. The MPSoC is modeled at the RTL level using VHDL. This work proposes fault tolerance techniques applied to intra-chip networks. Related work on fault tolerance at a systemic level, router level, link level and routing algorithms are studied. This work presents the research and development of two techniques: (i) protocols to enable the correct communication between task with partial degradation of the link enabling the router to operate even with faulted physical channels; (ii) test recovery method and of the router. This Dissertation considers permanent and transient faults. The HeMPS platform is the reference platform to evaluate the proposed techniques, together with a fault injection campaign where up to five random failures were injected simultaneously at each simulated scenario. Two applications were used to evaluate the proposed techniques, MPEG encoder and a synthetic application, resulting in 2,000 simulated scenarios. The results demonstrated the effectiveness of the proposal, with most scenarios running correctly with routers operating in degraded mode, with an impact on the execution time below 1%, with a router area overhead around 30%.