Configuring Fault-Tolerant Servers for Best Performance

When using primary-backup replication, one checkpoints the primary\´s state to reduce the failover time to a backup, upon failure of the primary. A trade-off is involved: by frequent checkpointing the response time for requests during "no-failure" intervals is increased. On the other hand, when checkpointing frequency is low, failover takes longer. In this paper we provide a theoretical ground for computing the optimal checkpointing interval that minimizes the average response time of requests, given parameters such as load, failure rate and service rates. We use queuing theory for modelling the support for failover in a system where a logging server records client update operations. The novelty of the model is inclusion of the backlog processing time for requests accumulated during failover. Also, our model considers the waiting times in the queue of the logging server as well as the application server