Power management techniques for conserving energy in multiple system components.

Nevine AbouGhazaleh (CS Grad/Pitt)

PhD Defense

Thursday, February 7, 2008
11:00 a.m. - SENSQ 6106 - Eli Lilly Room

Abstract

Energy consumption is a limiting constraint for both embedded and high performance systems. CPU-core, caches and memory contribute a large fraction of energy consumption in most computing systems. As a result, reducing the energy consumption in these components can significantly reduce the system’s overall energy consumption. However, applying multiple independent power management policies in the system (one for each component) may interfere with each other and in some occasions increase the combined energy consumption.

In this talk, I present three power management techniques that target more than a single component in a system. The focus is on reducing the total energy consumption in processors, caches and memory combinations. First, I present a memory-aware processor power management using collaboration between the OS and the compiler. The technique objectives are: (1) finish the application execution before its deadline and (2) minimize the combined energy consumption in processor and memory. Second, I present an Integrated Dynamic Voltage Scaling (IDVS) techniques for processor and on-chip cache power management in multi-voltage domains systems. IDVS co-ordinates power management decisions across voltage domains rather that being applied in isolation in each domain. Third, I present a Power-Aware Cached DRAM (PA-CDRAM) memory organization for reducing the energy consumption in DRAM memory and off-chip caches. PA-CDRAM exploits the high internal memory bandwidth by bringing the off-chip caches “closer” to the memory, which also improves the overall performance.

The techniques in my talk highlight the importance of designing power management schemes that consider multiple components and their interactions (in terms of power and performance) in the system rather than applying multiple isolated power management polices. This study should lay the foundation for further research in the domain of integrated power management, where a single power manager controls many system components.

Dissertation CO-AdviserS

Dr. Rami Melhem, Department of Computer Science
Dr. Daniel Mosse, Department of Computer Science

Committee Members

Dr. Bruce R. Childers
Dr. Mary Jane Irwin (Penn State)