Reliability-oriented Resource Management for High-Performance Computing

Reliability is an increasingly pressing issue for High-Performance Computing systems, as failures are a threat to large-scale applications, for which an even single run may incur significant energy and billing costs. Currently, application developers need to address reliability explicitly, by integrating application-specific checkpoint/restore mechanisms. However, the application alone cannot exploit system knowledge, which is not the case for system-wide resource management systems. In this paper, we propose a reliability-oriented policy that can increase significantly component reliability by combining checkpoint/restore mechanisms exploitation and proactive resource management policies.