Energy-efficient Incremental Offloading of Neural Network Computations in Mobile Edge Computing

Deep Neural Network (DNN) has shown remarkable success in Computer Vision and Augmented Reality. However, battery-powered devices still cannot afford to run state-of-the-art DNNs. Mobile Edge Computing (MEC) is a promising approach to run the DNNs on energy-constrained mobile devices. It uploads the DNN model partitions of the devices to the nearest edge servers on demand, and then offloads DNN computations to the servers to save the energy of the devices. Nevertheless, the existing all-at-once computation offloading faces two great challenges. The first one is how to find the most energy-efficient model partition scheme under different wireless network bandwidths in MEC. The second challenge is how to reduce the time and energy cost of the devices waiting for the servers, since uploading all DNN layers of the optimal partition often lakes time. To meet these challenges, we propose the following solution. First, we build regression-based energy consumption prediction models by profiling the energy consumption of mobile devices under varied wireless network bandwidths. Then, we present an algorithm that finds the most energy-efficient DNN partition scheme based on the established prediction models and performs incremental computation offloading upon the completion of uploading each DNN partition. The experimental results show that our solution improves energy efficiency compared to the current all-at-once approach. Under the 100 Mbps bandwidth, when the model uploading takes 1/3 of the total uploading time, the proposed solution can reduce the energy consumption by around 40%.