Event-Driven Scalability Prediction: Improving Energy Efficiency under Hard
Performance Constraints on Multiprocessors
Dimitris Nikolopoulos
Abstract
The component density and the power and thermal properties of
multicore and multithreaded processors demand energy and temperature
awareness at all layers of the hardware/software
interface. Energy-aware program optimization with hard performance
constraints is an inherently dynamic optimization process, during
which continuous power-performance profiling and rapid detection of
energy-efficient operating points are essential. In this talk I will
discuss the foundations, design and implementation of an event-driven
runtime scalability predictor (EDRSP) for simultaneous energy and
performance optimization on multithreaded and multicore
processors. EDRSP is an optimization tool which combines phase
detection with online analysis of single-phase and cross-phase
power-performance profiles collected from hardware event counters, to
predict operating points of multithreaded code in which power
consumption can be lowered without a performance penalty. I will
present results from applications of EDRSP to energy-efficiency
optimization on real multiprocessors with multithreaded and multicore
processing components.
Biography
Dimitris Nikolopoulos is an Associate Professor of Computer Science at
Virginia Tech since August 2006, after spending four years as an
Assistant Professor of Computer Science at the College of William and
Mary. Prior to assuming tenure-track faculty positions, he held the
positions of Visiting Assistant Professor of the Department of
Electrical and Computer Engineering and Visiting Research Assistant
Professor of the Coordinated Science Laboratory, both at the
University of Illinois, Urbana-Champaign. Nikolopoulos conducts
research in high-end and embedded parallel architectures, with
emphasis on hardware/software support for runtime optimization in the
execution engines and the memory system of processors with multiple
execution contexts. He is a recipient of two early faculty career
development awards from the National Science Foundation (NSF CAREER,
2004) and the United States Department of Energy (DOE ECPI, 2005), as
well as five best paper awards including one from the Supercomputing
(SC) Conference and one from the International Parallel and
Distributed Processing Symposium (IPDPS). Nikolopoulos received his
Doctorate degree in Computer Engineering from the University of Patras
in Greece in 2000. He is a member of the IEEE Computer Society, ACM
SIGARCH and the Technical Chamber of Greece.