Towards Predicting the Impact of Roll-Forward Failure Recovery for HPC Applications

摘要

The roll-forward recovery schemes on HPC systems implicitly trade off faster time to solution for higher risk: as it usually performs a probabilistic repair, this may cause further failures such as SDCs. It is essential for users to be able to reason about the impact of a particular repair exercised by the scheme. Towards this goal, we identify two research questions aiming to determine the outcome of a repair either at the failure point or at the end of the execution. For the former, we propose a promising hybrid approach that combines machine learning and error propagation analysis techniques.