DFHR: A Design Framework for HPC Reliability

This paper proposed a design framework for HPC reliability (DFHR) which consists of reliability prediction, failure hotspots (FHs) exploring, reliability program planning, and reliability plan implementation. In DFHR, 1) a concept of failure hotspot is proposed which provides targets for reliability engineering; 2) a unified process of reliability design is included in the framework with a constant respect to reliability technique costs; 3) the engineering practicability of the framework is achieved by hierarchy reliability planning and reasonable approximation; 4) highly precise reliability prediction methods are included in the framework. Based on the practice results of DFHR in the reliability design of a high performance computer, we demonstrate that DFHR is suitable for the cost-effective reliability design of HPC systems.