Examination Committee

Prof Koen Jacques Ferdinand BLANCKAERT, CIVL/HKUST (Chairperson)
Prof Jiang XU, ECE/HKUST (Thesis Supervisor)
Prof Marilyn C WOLF, School of Electrical and Computer Engineering, Georgia Institute of Technology (External Examiner)
Prof Qiang XU, Department of Computer Science and Engineering, The Chinese University of Hong Kong (External Examiner)
Prof Tim CHENG, ECE/HKUST
Prof Howard LUONG, ECE/HKUST
Prof Kai CHEN, CSE/HKUST

Abstract

Energy efficiency has become a critical criterion for the design of multicore systems today. System-level power management are widely used to determine and schedule appropriate power modes for each core by taking advantage of the control knobs provided by low power techniques, such as power-gating and dynamic voltage and frequency scaling (DVFS). A robust and efficient power manager must be resilient and adaptive to the non-idealities, variabilities and uncertainties from the environment, the hardware and the running workloads. This thesis investigates different design concerns in system-level power management for multicore systems, and proposes several self-adaptive runtime approaches to improve the system energy-efficiency in a reliable and efficient way.
 
Power gating is widely used to reduce leakage power in multicore systems. However, the power-mode transition will introduce serious power/ground (P/G) noise, which could be a severe reliability threat to system power integrity. At the same time, the increasing degree of process variation also brings uncertainties to the P/G noise problem. In this thesis, we analyze the characteristics of P/G noise in the presence of process variation, and propose a hardware-software collaborated runtime approach to adaptively protect PUs from P/G noise. The proposed adaptive method achieves a comparable reliability to the most conservative static method while incurring much lower performance and energy overhead.
 
DVFS has also been widely employed in commercial multicore systems to improve energy-efficiency. Recently, on-line learning based control algorithms show great potential in dynamic power management with adaptation to runtime conditions and system behavior patterns. However, such method is usually treated as non-scalable to systems with large number of cores. To address this problem, a modular reinforcement learning based online DVFS control scheme is developed to make the system able to adaptively select globally optimized operating points with incurring only reasonably polynomial overhead to the core count.