Open Access Dissertation
Date of Award
Electrical and Computer Engineering
Power is a major design concern of today's networked computing systems, from low-power battery-powered mobile and embedded systems to high-power enterprise servers. Embedded systems are required to be power efficiency because most embedded systems are powered by battery with limited capacity. Similar concern of power expenditure rises as well in enterprise server environments due to cooling requirement, power delivery limit, electricity costs as well as environment pollutions.
The power consumption in networked computing systems includes that on circuit board and that for communication. In the context of networked real-time systems, the power dissipation on wireless communication is more significant than that on circuit board. We focus on packet scheduling for wireless real-time systems with renewable energy resources. In such a scenario, it is required to transmit data with higher level of importance periodically. We formulate this packet scheduling problem as an NP-hard reward maximization problem with time and energy constraints. An optimal solution with pseudo polynomial time complexity is presented. In addition, we propose a sub-optimal solution with polynomial time complexity.
Circuit board, especially processor, power consumption is still the major source of system power consumption. We provide a general-purposed, practical and comprehensive power management middleware for networked computing systems to manage circuit board power consumption thus to affect system-level power consumption. It has the functionalities of power and performance monitoring, power management (PM) policy selection and PM control, as well as energy efficiency analysis. This middleware includes an extensible PM policy library. We implemented a prototype of this middleware on Base Band Units (BBUs) with three PM policies enclosed. These policies have been validated on different platforms, such as enterprise servers, virtual environments and BBUs.
In enterprise environments, the power dissipation on circuit board dominates. Regulation on computing resources on board has a significant impact on power consumption. Dynamic Voltage and Frequency Scaling (DVFS) is an effective technique to conserve energy consumption. We investigate system-level power management in order to avoid system failures due to power capacity overload or overheating. This management needs to control the power consumption in an accurate and responsive manner, which cannot be achieve by the existing black-box feedback control. Thus we present a model-predictive feedback controller to regulate processor frequency so that power budget can be satisfied without significant loss on performance.
In addition to providing power guarantee alone, performance with respect to service-level agreements (SLAs) is required to be guaranteed as well. The proliferation of virtualization technology imposes new challenges on power management due to resource sharing. It is hard to achieve optimization in both power and performance on shared infrastructures due to system dynamics. We propose vPnP, a feedback control based coordination approach providing guarantee on application-level performance and underlying physical host power consumption in virtualized environments. This system can adapt gracefully to workload change. The preliminary results show its flexibility to achieve different levels of tradeoffs between power and performance as well as its robustness over a variety of workloads.
It is desirable for improve energy efficiency of systems, such as BBUs, hosting soft-real time applications. We proposed a power management strategy for controlling delay and minimizing power consumption using DVFS. We use the Robbins-Monro (RM) stochastic approximation method to estimate delay quantile. We couple a fuzzy controller with the RM algorithm to scale CPU frequency that will maintain performance within the specified QoS.
Gong, Jiayu, "Qos-aware fine-grained power management in networked computing systems" (2011). Wayne State University Dissertations. 349.