| From high-end servers to tiny mobile devices, power consumption has become a major design concern. In the past, power issues were mostly addressed in back-end design stages using design-time worst-case power based approaches, which becomes extremely difficult or simply infeasible in modern system design. In this thesis, we explore power optimization techniques at the system level, and present architectural run-time power optimization techniques for both high-performance, high-power systems and energy-constrained mobile devices, and system-level design automation techniques for low power.;In high-performance networked computer systems, interconnection networks consume a significant portion of total system power. We were the first to address power consumption issues in interconnection networks---presenting architecture-level power optimization and management techniques that can effectively minimize average network run-time power consumption as well as regulate network peak power. Our techniques meet the worst-case power constraint with very low performance impact. Network designers can thus focus on optimizing architectures for the average-case system power and performance. The thermal issue is a critical design concern driven by power. In this thesis, we present an architecture-level thermal modeling technique for on-chip networks. Using this model, we characterize the network thermal impact in modern chip-multiprocessors. We present ThermalHerd, a run-time thermal management technique which can effectively regulate network thermal behavior. ThermalHerd can be extended in the future to address the thermal issue of the entire on-chip systems.;In the wireless scenario, energy is the major concern, which constrains the computation and communication capacity of mobile devices. To tackle this problem, we present a distributed economic subcontracting protocol, called DESP, to dynamically allocate energy resources among mobile devices in an ad-hoc network. On-line bargaining is used to control run-time energy transfer. Mobile devices are modeled as economic agents. Decisionmaking algorithms are proposed for different agents, each of which has a different optimization goal. This technique can fairly and effectively allocate energy resources in both cooperative and competitive network scenarios.;For system-level design automation for low power, we present SLOPES, a multiobjective hardware-software co-synthesis tool for multi-rate, real-time, low power distributed embedded systems. A novel scheduling algorithm is proposed to support dynamically reconfigurable devices. This scheduler uses resource and frame-by-frame reconfiguration information in its efforts to globally minimize the reconfiguration overhead. SLOPES can automatically explore the system design space and simultaneously optimize system performance, power consumption and cost. |
