Differential Shadowing: A Resilience Framework for Extreme-scale, Heterogeneous Environments with Non-Uniform Node Failure Distribution

Achieving resilience in extreme-scale environments, while minimizing energy consumption, is a daunting challenge. At extreme scale, however, the classic checkpoint-restart approach or replication for recovery techniques become inadequate. In this paper, we propose a novel application-aware elastic resilience model, dShadowing, for extreme-scale environments, as an efficient and scalable alternative to checkpointing, pure replication and re-execution. The basic tenet of this model is a dShadow, which is a derivative of its associated main process, whose functional and non-functional attributes are derived to achieve high tolerance to failure, at a minimum energy cost, while closely adhering to QoS requirements. Contrary to current schemes, dShadowing assumes heterogeneous environments, where cores fail independently, but non-identically. The experiment’s results show that dShadowing model can achieve on average over 20% reduction in energy consumption and expected completion time, in comparison to a baseline shadowing model that considers cores fail uniformly. The results also demonstrate the flexibility of the dShadowing model and the ability to tolerate failure at scale adaptively and efficiently.