Offloading strategy for dependency-aware tasks in MEC system based on deep Q-network

Abstract

Task offloading strategy is one of the promising methods in mobile edge computing (MEC) to reschedule computing resources. Scheduling for computationally intensive tasks was mainly focused on and the dependency restriction between tasks was generally ignored. In this paper, an End-to-Edge collaborative resource allocation model is established. Firstly, in order to improve the unrealistic assumption raised in previous studies that tasks can be offloaded to MEC servers with arbitrary proportion and order, we establish the directed acyclic graphs to describe the dependency relationship within tasks. And then, system time cost including execution time, transmission delay and waiting latency are considered, and the task offloading problem is transformed into an optimization model to minimize time consumption. To manage the non-convex problem, an offloading strategy under dependency constraints based on deep Q-network (DQN) is proposed. Simulations prove that the proposed algorithm can obtain smaller time cost by comparing with other baseline algorithms.