An Energy-efficient Offloading Framework with Predictable Temporal Correctness

As battery-powered embedded devices have limited computational capacity, computation offloading becomes a promising solution that selectively migrates computations to powerful remote severs. The driving problem that motivates this work is to leverage remote resources to facilitate the development of mobile augmented reality (AR) systems. Due to the (soft) timing predictability requirements of many AR-based computations (e.g., object recognition tasks require bounded response times), it is challenging to develop an offloading framework that jointly optimizes the two (somewhat conflicting) goals of achieving timing predictability and energy efficiency. This paper presents a comprehensive offloading and resource management framework for embedded systems, which aims to ensure predictable response time performance while minimizing energy consumption. We develop two offloading algorithms within the framework, which decide the task components that shall be offloaded so that both goals can be achieved simultaneously. We have fully implemented our framework on an Android smartphone platform. An in depth evaluation using representative Android applications and benchmarks demonstrates that our proposed offloading framework dominates existing approaches in term of timing predictability (e.g., ours can support workloads with 100% more required CPU utilization), while effectively reducing energy consumption.