Gestion en ligne pour l'efficacité énergétique des systèmes multi-cœurs/multi-coeurs
Simei Yang · HAL (Le Centre pour la Communication Scientifique Directe) · 2020
Cluster-based multi/many-core platforms represent promising solutions to deliver high computing performance and energy efficiency in modern embedded systems. These platforms often support per-cluster Dynamic Voltage/Frequency Scaling (DVFS), allowing different clusters to change their own v/f levels independently. The increasing application complexity and application dynamism on such platforms arise the need for run-time management. This dissertation focuses on the run-time management of applications on cluster-based multi/many-core systems to improve energy efficiency.Towards the run-time management purpose this dissertation presents different management strategies that estimate the mutual influence between application mapping and cluster v/f configurations to respectively achieve local optimization within a cluster and global optimization in the overall system. The proposed management strategies can achieve near-optimal management solutions with less strategy complexity compared to state-of-the-art strategies. In addition, this dissertation presents a new modelling and simulation approach that allows the evaluation of run-time management strategies in multi/many-core systems to guarantee that system constraints are fully met. The proposed simulation approach is validated using an industrial modelling and simulation framework.