[Colloquium] Imes/MS Presentation/Feb 20, 2015

[Margaret Jaffey]

This is an announcement of Connor Imes's MS Presentation.

------------------------------------------------------------------------------
Date:  Friday, February 20, 2015

Time:  2:30 PM

Place:  Ryerson 251

M.S. Candidate:  Connor Imes

M.S. Paper Title: Managing Diversity in Performance and Energy
Characteristics on Embedded Systems

Abstract:
Minimizing energy under timing constraints is a common task for
embedded systems, which are often limited by available energy sources.
Solving this problem even for a single system is hard, and current
solutions are often designed for specific platforms and applications.
Developing general techniques is difficult because of the variety in
performance, power, and energy characteristics of different systems
and applications. While heuristic solutions have often sufficed in the
past, the increase in configurability and heterogeneity of modern
systems means that no single heuristic will perform well on all
platforms, or even for all applications on a single platform. A
portable solution to the problem is needed. This thesis first
identifies two embedded systems that expose different
performance/power and latency/energy tradeoff spaces and demonstrates
that each requires a different heuristic to achieve acceptable energy
consumption under timing constraints. Both systems use hardware
designs that are common in today’s cutting-edge devices, thus
establishing the problem as a genuine concern for system and software
designers. The first is a Sony Vaio tablet, a homogeneous multi-core
system with a mobile Intel Haswell processor; the second is an ODROID
development platform, a single-ISA heterogeneous multi-core system
with an ARM big.LITTLE processor. We show that the Vaio requires a
race-to-sleep heuristic to achieve acceptable energy efficiency while
the ODROID requires a never-idle heuristic, and that choosing the
wrong strategy can decrease energy efficiency to only 25% of optimal
(i.e., consumes 4× the optimal energy). To address this problem, we
present POET, a portable library and runtime that manages system
resources to achieve nearoptimal energy consumption while meeting soft
real-time performance constraints. We show that POET meets performance
targets with small error while consuming, on average, only 1.9% more
energy on the Vaio compared to a dynamic optimal oracle, and 3.8% more
on the ODROID. POET is designed to facilitate writing portable,
energy-efficient applications by providing a simple and extensible API
to users. It is released as open-source software under the BSD
License.

Connor's advisor is Prof. Henry Hoffmann

Login to the Computer Science Department website for details:
 https://www.cs.uchicago.edu/phd/ms_announcements#ckimes

=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Margaret P. Jaffey            margaret at cs.uchicago.edu
Department of Computer Science
Student Support Rep (Ry 156)               (773) 702-6011
The University of Chicago      http://www.cs.uchicago.edu
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=