Power- and Performance-constrained Thread Mapping and DVFS on Heterogeneous Multi-Core Systems

Wednesday March 29, 2017
Location: CIC Panther Hollow Room
Time: 4:30PM

Dimitrios Stamoulis (CMU)

Abstract

Modern many-core systems must cope with a wide range of heterogeneity due to either the numerous processing components that are heterogeneous  by design, or due to the different performance requirements of multi-application, multi-threaded workloads. This raises an important question for chip multiprocessor  designers: Can we guarantee per-application performance constraints under  workload and core heterogeneity, while staying within the power budget?

In this talk, I will present our work on an approach for simultaneous thread mapping  and Dynamic Voltage Frequency Scaling (DVFS) on heterogeneous multi-core systems to maximize overall performance, while satisfying the power budget and per-application performance requirements. We formulate this optimization problem as a constrained 0-1  integer linear program (ILP) and we propose a heuristic-based algorithm for efficiently  solving it. Compared with an optimal solver, our method produces results less than 1.5%  away from optimum on average, with four orders of magnitude improvement in runtime. We also show that our method always meets per-application performance requirements,  while agnostic approaches could result in performance bound violations up to 48.1%.

Bio

Dimitrios Stamoulis is a second year PhD student advised by Professor Diana Marculescu. His research is focused on performance optimization for heterogeneous  and dark silicon multi-core systems under power and variability constraints.



Back to the seminar page