Value and energy optimizing dynamic resource allocation in many-core HPC systems

The conventional approaches to reduce the energy consumption of high performance computing (HPC) data centers focus on consolidation and dynamic voltage and frequency scaling (DVFS). Most of these approaches consider independent tasks (or jobs) and do not jointly optimize for energy and value. In this paper, we propose DVFS-aware profiling and non-profiling based approaches that use design-time profiling results and perform all the computations at run-time, respectively. The profiling based approach is suitable for the scenarios when the jobs or their structure is known at design-time, otherwise, the non-profiling based approach is more suitable. Both the approaches consider jobs containing dependent tasks and exploit efficient allocation combined with identification of voltage/frequency levels of used system cores to jointly optimize value and energy. Experiments show that the proposed approaches reduce energy consumption by 15% when compared to existing approaches while achieving significant amount of value and reducing percentage of rejected jobs leading to zero value.