"Power consumption is becoming a performance limiting factor in the development of high-end computing systems. Processors containing multiple/many cores are being hailed as a solution to the power crisis, but support for effectively managing parallelism -- in terms of both performance and power consumption -- on these architectures is far from mature. We present a performance prediction model for identifying energy-efficient operating points of concurrency at program phase boundaries in multithreaded scientific applications. We do so via statistical analysis of hardware event counters," said Curtis-Maury.
        

"Additionally, we present a prediction-driven runtime optimization system, ACTOR, that throttles concurrency to reduce power consumption and set performance at the knee of the scalability curve of each parallel execution phase. ACTOR successfully identifies and exploits program phases where poor scalability results in performance losses through the use of more processing elements, providing simultaneous reductions in execution time and power consumption in many cases. ACTOR reduces the overhead of searching the optimization space for power-performance efficiency, and achieves substantial performance and power savings in real parallel applications," Curtis-Maury explained.

 "I very much enjoyed my visit to Cornell. It was a wonderful opportunity both to increase general awareness of my research as well as to interact with peer researchers at a top university and see what kinds of projects they are working on," said Curtis-Maury.

"Curtis-Maury's research enables us to model and better understand emerging many-core processors," said Nikolopoulos, his advisor. "Our ongoing collaborations with CSL at Cornell University present new directions and opportunities in applying machine-learning techniques to understanding how programs scale on modern parallel architectures."

Matthew Curtis-Maury is finishing his Ph.D. with Dr. Dimitrios S. Nikolopoulos in the Department of Computer Science at Virginia Tech. He will graduate in May 2008. He received a B.S. from Wake Forest University in 2003 and an M.S. from The College of William and Mary in 2005, both in Computer Science. His research interests include energy-aware high performance computing and statistical performance prediction of parallel applications. Curtis-Maury has published 11 papers in high performance computing, and he won the best paper award at the International Workshop on OpenMP in 2005.