THESIS
2019
xiii, 115 pages : illustrations ; 30 cm
Abstract
Energy efficiency has become a critical design metric for high-performance systems. Various power management techniques have been proposed for the processor cores such as dynamic voltage and frequency scaling (DVFS), while few solutions consider the power losses suffered on the power delivery system (PDS), even though they have a significant impact on the system overall energy efficiency. With the explosive growth of system complexity and highly dynamic workloads variations, it is also challenging to find the optimal power management policies which can effectively match the power delivery with the power consumption. Besides, process variations (PV) add heterogeneity to systems and make traditional power management methods less effective. In the thesis, firstly, we analytically and comp...[
Read more ]
Energy efficiency has become a critical design metric for high-performance systems. Various power management techniques have been proposed for the processor cores such as dynamic voltage and frequency scaling (DVFS), while few solutions consider the power losses suffered on the power delivery system (PDS), even though they have a significant impact on the system overall energy efficiency. With the explosive growth of system complexity and highly dynamic workloads variations, it is also challenging to find the optimal power management policies which can effectively match the power delivery with the power consumption. Besides, process variations (PV) add heterogeneity to systems and make traditional power management methods less effective. In the thesis, firstly, we analytically and comprehensively study the energy efficiency of different PDS paradigms and power management schemes for the PDS. Then a workload-aware quantized power management scheme is proposed to dynamically improve the energy efficiency of PDS. Finally, we propose a reinforcement learning-based Chip-Specific Power co-Management (CSPM) scheme for PV-aware manycore systems. Both the PDS and the processor cores are jointly adjusted by distributed agents with modular Q-Learning to improve the system overall energy efficiency. The learning agents distributed across power domains not only manage the power states of processor cores but also control the on/off states of on-chip VRs to actively adapts to the workloads variations. By integrating multiple modules in each agent, inter-power domain dependencies are considered for faster convergence and better quality of the policies which benefit the global optimization objective. Weight adaption is adopted to achieve optimal energy efficiency with a desired performance target. System characteristics are naturally included in the learning process to obtain chip-specific policies. Experimental results show that when applied to PV-aware manycore systems with a hybrid PDS constructed by both on- and off-chip voltage regulators, the proposed method on average achieves an effective 60% reduction of the system overall energy-delay-product (EDP) compared to a traditional DVFS approach.
Post a Comment