Enhancing early scheduling in parallel state machine replication

• Ano de defesa: 2020
• Autor(a) principal: Batista, Eliã Rafael de Lima
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Pontifícia Universidade Católica do Rio Grande do Sul; Escola Politécnica; Brasil; PUCRS; Programa de Pós-Graduação em Ciência da Computação
• Palavras-chave em Português: Escalonamento antecipado; Replicação máquina de estados; Roubo de trabalho; Espera ocupada; Early Scheduling; State machine replication; Work-stealing; Busy-wait; CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO
• Link de acesso: http://tede2.pucrs.br/tede2/handle/tede/9165

State machine replication is a well-established approach to provide fault tolerance. One of its key assumptions is that eplicas must execute operations deterministically and thus serially. However, to benefit from multi-core servers, some techniques were built to allow concurrent execution of operations in state machine replication. Invariably, such techniques are based on the fact that independent operations (which do not share common state or do not update shared state) can execute in parallel. Early Scheduling is a promising technique that trades concurrency for expeditious scheduling decisions. In this technique, requests are grouped in classes and a fixed subset of threads is assigned to each class, respecting request dependencies. Early scheduling has been shown to provide significant performance improvements in state machine replication. Although early scheduling shows performance gains, it restricts concurrency according to the class definitions and class-to-threads mapping. We evaluate the impact of these restrictions imposed by the early scheduling technique. Out of the observations, we propose and implement improvements to the basic technique, present their results and compare the resulting system to more classic approaches. Our main contribution concern the use of alternative synchronization techniques, such as busy wait adaptations and application of work-stealing techniques within the former Early Scheduling algorithm.