Asynchronous Hybrid Deep Learning (AHDL): A Deep Learning Based Resource Mapping in DVFS Enabled Mobile MPSoCs

Abstract

Mapping resources to tasks accurately in order to gain performance, energy efficiency, reduction in peak temperature, etc. on an embedded/Edge device is a big challenge. Machine learning has proven to be effective in learning heuristics based resource mapping approaches, but its success is bound by the quality of feature extraction. Additionally, feature extraction in such approaches not just requires expert domain knowledge and human effort, but at the same time requires the application (tasks) to be profiled for such processes. Therefore, the efficacy of such resource mapping methodologies depends on expertise, skills, profiling time and architecture of the system. In this paper, we propose a novel methodology, Asynchronous Hybrid Deep Learning (AHDL), which sets a new paradigm of using Deep Learning approaches to map resources to application (tasks). In our approach, we leverage task profiling methodologies to achieve accurate mapping in order to achieve greater reward from the system, but at the same time is able to allocate resources to unprofiled application (tasks) at the same time without the need of manual feature extraction by domain experts. Our proposed methodology is able to achieve competitive results in comparison with the state-of- the-art without the usual associated challenges such as manual feature extraction.