Comparison of different strategies for Time and Energy Efficient Offloading for Mobile Edge Computing

Abstract

Every day, the number of wireless devices, and IoT applications increases, which require extensive computational resources. Therefore, it is possible to mitigate the lack of computational resources in wireless devices by using Mobile Edge Computing (MEC). MEC is a modern technology that brings the capabilities of Cloud Computing at the edge of a mobile network to perform computationally intensive tasks, which reduces the delay and prevents end to end communication with the remote Cloud. This paper proposed a task offloading model for multiple-device, multiple-task MEC system, the model is formulated as an optimization problem with the objective of reducing time of computation and energy consumption. However, the complexity rapidly increases as more devices are added to the system, thus the proposed problem is solved by introducing five strategies which are full local computing, full offloading computing, random offloading, Q learning, Deep Q network, and a distributed DNN, which are compared with the optimal offloading strategy. The results (4 devices with 3 tasks for each device) show that the total cost in terms of time and energy consumption in Q learning, DQN and, Distributed DNN algorithms is near to the optimal offloading strategy, furthermore, these strategies reduce the total cost up to 63.7% when compared to full local strategy, also up to 21.8% when compared to full edge strategy. However, the learning speed of distributed DNN is faster than Deep Q Network, when number of devices increases. In addition, adistributed DNN generates the offloading decision (in 4 milliseconds) faster than DQN algorithm (in 8 milliseconds).