Redução dos efeitos negativos das suspeitas incorretas no algoritmo de consenso de Chandra e Toueg

AUTOR(ES)
FONTE

IBICT - Instituto Brasileiro de Informação em Ciência e Tecnologia

DATA DE PUBLICAÇÃO

06/05/2011

RESUMO

Some protocols in distributed systems, such as atomic broadcast and semi-passive replication, are based on the consensus algorithm proposed by Chandra and Toueg. This algorithm is equipped with an unreliable failure detector. In asynchronous distributed systems, this type of detector can make mistakes by erroneously suspecting a process that is still running. The presence of wrong suspicions degrades significantly the performance of the algorithm and the performance of any protocol that uses it. To reduce this degradation, we propose two new optimizations and an adaptation of the Look-Ahead technique to the algorithm. The first optimization, named Early-Decision, allows to antecipate a decision to the consensus problem. The second optimization, named Additional-Waiting, allows extending the waiting time for messages when it is useful. The Look-Ahead technique helps speed up the execution of consensus when there are processes in different rounds. We present a description of the algorithm that combines these optimizations and prove its correctness. We conducted some simulations to evaluate the effects of optimizations on the performance of the algorithm. In addition, we compared the performance of some consensus algorithms and selected the most efficient, the Paxos algorithm, to be compared with the Chandra and Toueg optimized algorithm. The simulation results show that all optimizations are effective, particularly, when they are combined. In most situations considered, the performance of the Chandra and Toueg optimized algorithm is better than the Paxos algorithm.

ASSUNTO(S)

simulation sistema distribuído assíncrono consenso detector de falhas otimização simulação avaliação de desempenho ciencia da computacao asynchronous distributed system consensus failure detectors optimization performance evaluation

Documentos Relacionados