Defining the Checkpoint Interval for Uncoordinated Checkpointing Protocols

Parallel applications running on large computers suffer from the absence of a reliable environment. Fault tolerance proposals, in general, rely on rollback-recovery strategies supported by checkpoint and/or message logging. There are well-defined models that address the optimum checkpoint interval for coordinated checkpointing. Nevertheless, there is a lack of models concerning uncoordinated checkpointing combined with message logging. First we present a model designed for serial applications or coordinated checkpointing-based solutions. Our contribution is the extension of this model to a scenario based on uncoordinated checkpointing combined with message logging. We introduce two key points to minimise the fault tolerance overhead for parallel applications. The first is the use of a factor to represent the dependency relation between processes. The second is the use a specific checkpoint intervals for each process. Experiments show that our model performs as well as previous studies for serial applications or coordinated checkpointing. While running parallel applications using uncoordinated checkpointing combined with message logging, our checkpoint interval model effectively minimises the overhead introduced by the fault tolerance tasks. Moreover, the overhead prediction error is smaller than 5% for all applications tested.