Research On Edge Offloading Methods For Compute-intensive Tasks

With the rapid development of Artificial Intelligence(AI),Internet of Things(Io T)and Internet of Vehicles(Io V),we have to perform increasingly more resourcehungry and compute-intensive applications on Io T and Io V devices,where the available computing resources are insufficient.How to execute a large number of compute-intensive applications on the edge computing system with limited computing resources is a critical problem.In this paper,we aim to make optimal offloading decisions for compute-intensive tasks in the MEC environment and VEC environment.The main contributions of this thesis include:1.We propose two jointly optimization methods of capacity-competition multiple knapsacks problems for optimizing the offloading problem of computeintensive tasks in task-overflowed situations,named OAKGM and OAMKP.Firstly,we propose an offloading indicator,through iteratively segment which we make a pre-allocation of computing resources of the edge server and then reallocate computing resources on local or on edge,respectively.The proposed methods can not only optimize the cost about time and energy,but also penalize the total amount of overflowed computations for reducing the task pressure in the next workflow.The work is the earliest research about the task-overflowed situations and the jointly optimization methods of capacity-competition multiple knapsacks problems.When the proposed algorithms are compared with other offloading algorithms,they perform better in the utility ratio of computing resources.2.To enhance the generalization ability of OAMKP,we train a deep neural network with an attention mechanism to predict the thresholds for the offloading indicator and calculate the optimal offloading decisions.Firstly,we gain the best thresholds of different workflows of different MEC situations through OAMKP methods as training dataset of the network,and then compute the offloading decisions by the predictions of thresholds.The trained neural network can gain accurately predictions for new MEC situations,which can enhance the generalization ability and reduce the time complexity.Compared with the end-to-end learning method based on U-Net,the proposed deep learning framework performs better under a small sample dataset.3.To make dynamic and efficient offloading decisions for compute-intensive tasks in VEC environment,we propose a dynamic framing offloading algorithm based on deep reinforcement learning method for optimizing the total time of sequential subtasks.Firstly,we segment each task into several sequential subtasks and divided the the process of making the offloading decision of whole tasks in the current road into different offloading frames by the generation time of subtasks.Then regard the offloading frames as the time steps and build an offloading framework based on deep reinforcement learning.This work fills the gap of mobile vehicles sequential subtasks offloading methods in VEC environment.Extensive experimental results demonstrate the effectiveness and superiority of the proposed method when compared with other DRL-based methods and greedy-based methods.