Runtime Optimization of Mobile Platforms that Employ Heterogeneous CPUs

What is it about?

State-of-the-art mobile platforms utilize powerful system on chips (SoC) that are composed of a combination of big and little CPU cores. This leads to multiple frequency and core configuration knobs making the optimization of the mobile platforms, such as improving the energy efficiency, challenging. In this paper, we present a comprehensive methodology to choose optimal core and frequency configuration at runtime as a function of workload characteristics. We experimentally validate our approach on a Samsung Exynos chip based platform and achieve high gains in performance per watt compared to the default power management algorithms.

Featured Image

Why is it important?

Dynamically selecting the optimal configuration is a challenging task aggravated by two major factors. First, the design space is large for a runtime evaluation and exploration. Therefore, an exhaustive search is prohibitive due to significant overhead associated with exploration. Second, and more importantly, the optimal choice is a strong function of the workload, which itself varies dynamically. We overcome these challenges, first, by leveraging the use of phase-level offline characterization for a number of benchmarks to find the optimal configurations for each phase. Then, we build classifiers that map the characterized feature data to the optimal configurations. Finally, the classifier is used at runtime to select the optimal configuration for a new application phase. In our experimental evaluations, we observe substantial numerical gains in performance per watt compared to a recently proposed algorithm in literature and the default governors.

Perspectives