Analysis and evaluation of distributed checkpoint algorithms to avoid rollback propagation

Checkpointing is a very well known mechanism to achieve fault tolerance. In distributed applications where processes can checkpoint independently of each other, a local checkpoint is useful for fault tolerance purposes only if it belongs to at least one consistent global checkpoint. In this case, execution can be restarted from it without needing to rollback the execution in the past. The paper exploits a theoretical framework that facilitates the definition and analysis of distributed checkpoint algorithms to avoid rollback propagation. Several distributed algorithms are presented which avoid rollback propagation by forcing additional checkpoints in processes. The effectiveness of the algorithms is evaluated in several testbed applications, showing their limited capability of bounding the number of additional checkpoints