Power and energy constraints are now the driving force in all devices from servers to smartphones. This talk starts with quantitative power, performance, and energy measurements on a range of workloads and devices that point to the need for hardware heterogeneity to match software characteristics. However, programming heterogeneous hardware directly is a nightmare. We show how to abstract, choose, and exploit hardware heterogeneity. For interactive server workloads, we promote jobs from slow to fast cores to deliver substantial improvements in throughput and energy compared to homogeneous designs. These results offer some hope that software may survive and perhaps thrive as heterogeneity hardware evolves in the post Denard era.

Speaker Bio
Kathryn S. McKinley is a Principal Researcher at Microsoft and an Endowed Professor of Computer Science at The University of Texas at Austin. She and her collaborators have produced widely used tools: the DaCapo Java Benchmarks, TRIPS Compiler, Hoard memory manager, MMTk garbage collector toolkit, and Immix garbage collector. Her awards include: NSF Career, ASPLOS 2009 Best Paper, 2012 IEEE Top Picks, CACM Research Highlights (2006, 2012), Most Influential OOPSLA Paper from 2002 (awarded 2012), the 2011 ACM SIGPLAN Distinguished Service Award, and the 2012 ACM SIGPLAN Programming Languages Software Award. She has graduated 17 PhD students. She is an IEEE Fellow and ACM Fellow.