Section: New Results

Stochastic models for optimizing checkpoint protocol

After our past studied on design of origin checkpoint protocols, we have proposed a new stochastic performance model of the parallel execution in presence of failures. Thanks to this formulation, we are able to optimize several criteria (the time lost due to failure; the expected completion time) by making right decision of the date of each checkpoint. The model is general and it does not take into account the failure distribution law and accept variable checkpoint time estimation, which is important for dynamic parallelism applications.