Energy-efficient allocation of computing node slots in HPC clusters through parameter learning and hybrid genetic fuzzy system modeling

Abstract

Decision-making mechanisms for online allocation of computer node slots in HPC clusters are commonly based on simple knowledge-based systems comprised of individual sets of if---then rules. In contrast with previous works where these rules were designed using expert knowledge, two different types of evolutionary learning algorithms are compared in this paper. In the first case, some of the numerical parameters defining a human-designed knowledge base are tuned. In the second case, a genetic fuzzy system evolves a partial rule set that, after being combined with some expert rules, conforms the most appropriate knowledge base for a given load scenario. In both cases, the proposed approaches optimize the quality of service and the number of node reconfigurations along with the energy consumption. An experimental study has been made using actual workloads from the Scientific Modeling Cluster at Oviedo University, and statistical evidence was found supporting the adoption of the new learning system.