GIT-CERCS-05-23
Balasubramanian Seshasayee, Karsten Schwan,
Energy-Efficient Device Scheduling through Contextual Timeouts
Handheld and embedded hardware platforms are operating with an increasing
number of internal and external devices, potentially increasing energy
consumption and more importantly, motivating the need for energy
management techniques for peripheral devices. This paper presents a
platform-wide, system-level approach to dynamic energy management, termed
contextual timeouts.
The approach exploits the fact that most current peripheral devices
support the ability to switch to a low power mode when not in use and
automatically resuming operation upon use. The approach utilizes the energy
savings derived from such device suspensions, considering that the device
suspend/resume actions themselves consume power and have associated latencies.
Contextual timeouts do not require programmer involvement. Instead, dynamic
instrumentation is used to automatically capture and monitor the contexts
(i.e., the execution points) at which programs make the service requests that
cause device usage. From such dynamic monitoring data, system-level algorithms
predict future request times and manage devices to best meet program
needs under predicted behaviors. Adaptive methods for dynamic workload
characterization coupled with runtime techniques for request prediction
result in experimentally obtained energy savings of up to 50% over an
aggressive timeout-based regime on a Linux-based iPAQ PDA.