Adaptive runtime framework for energy-efficient fault tolerance in large-scale distributed and high-performance computing systems

Large-scale parallel and distributed systems are becoming increasingly vulnerable to failures and energy efficiency is one of the most important limitations in the high-performance computing environment. This paper introduces a dynamic framework for energy-efficient reliability management which is based on lightweight monitoring, hierarchical fault recovery, and dynamic decision making in terms of checkpointing, replication, and task migration. The proposed framework is a dynamic, reliable-energy balanced model through the use of three complementary algorithms of ERARO, HLARO, and RACIO in a closed-loop running system. Experimental validation of cluster-based MPI workload reveals that the framework minimizes energy use, restricts execution overhead and retains a high recovery efficiency which provides a viable and scalable solution to the current HPC and distributed systems.