Summary

This project investigates explores the use of runtime performance, power and scalability prediction techniques for dynamic program adaptation on multi-core and many-core architectures. We develop predictors that combine continuous feedback from a real-time hardware performance monitor with machine learning approaches. The project is part of a broader effort in deploying continuous hardware monitors for autonomous, on-the-fly optimization of concurrent codes. We have applied empirical performance and scalability predictors for core throttling, a technique which enables energy conservation with performance guarantees on multi-core and many-core platforms.

Downloads

As part of this project we have developed a tool for the simplified collection of performance counters on Intel Pentium 4 processors (HT/non-HT) called PACMAN which can be downloaded from here.

People

Matthew Curtis-Maury
Dr. Christos D. Antonopoulos
Prof. Dimitris S. Nikolopoulos

Publications

See [GREENCOM07], [JOATC06], [QEST06], [ICS06], [HPPAC06], [ACMSE07], [FHPM06].