Logo image
Back
Energy management of multi-component computing platforms under energy constraints
Dissertation   Open access

Energy management of multi-component computing platforms under energy constraints

Rizwana Begum
Doctor of Philosophy (Ph.D.), Drexel University
Apr 2017
DOI:
https://doi.org/10.17918/begr-qb79
pdf
Begum_Rizwana_20176.27 MBDownloadView
PDF Open Access Open Access (License Unspecified)

Abstract

Energy development Power resources--Research Electrical engineering Computer Engineering
Energy management is a problem of all types of computing devices. For example, short battery life is a complaint of today's mobile device users, while high operational costs of cooling and supplying power are the concerns of modern datacenters. It is inevitable that all modern multi-component devices manage energy in order to address these concerns. While, a majority of previous research efforts focus on single components such as CPU and DRAM, only a few optimize the energy consumption of multiple components-without understanding the interactions among these components. Core and memory systems consume a majority of total power consumption in today's computing devices. While Dynamic Voltage and Frequency Scaling (DVFS) for cores is a well established technique to make energy-performance trade-offs, Dynamic Frequency Scaling (DFS) is emerging as a promising technique for memory systems. This dissertation presents a holistic approach that optimizes performance of a system capable of core DVFS and DRAM DFS under energy constraints in a coordinated fashion. The dissertation introduces a relative energy constraint, inefficiency, that is independent of applications and agnostic to devices unlike existing absolute energy constraints. Our cross-component performance and energy models understand the interaction between core and DRAM. The dissertation introduces new relative and adaptive energy management algorithms that use the models to choose optimal frequency settings under given inefficiency budget. In addition to the constraints, the system also needs metrics to provide insights into how well the system is tuning. This dissertation introduces the concept of Power-Agility and a set of novel metrics to measure Power-Agility-ability of the systems to select and transition to the efficient power settings dynamically during the application execution. In the end, the dissertation characterizes and compares the power-agility of the proposed energy management approach to a state-of-the-art server system managing core DVFS and DRAM DFS under performance constraints.

Metrics

40 File views/ downloads
35 Record Views

Details

Logo image