On the optimum recovery of distributed programs

In a previous paper (1992) the authors have presented a checkpointing algorithm for distributed applications. That algorithm is based on a non-blocking coordinated global checkpoint of the distributed program. However, the associated rollback algorithm does not provide the best results, since in most cases it forces the rollback of all the processes. This paper presents two algorithms for roll-back-recovery that minimize the number of processes which need to roll back. One of the algorithms is oriented to systems that use message logging, while the other is more general and can be used in those systems that only rely an a coordinated checkpoint and do not log messages. We will show that our proposal achieves the optimum results In the minimization of the number of processes that have to roll back