Deep Reinforcement Learning for DAG-based Concurrent Requests Scheduling in Edge Networks

Abstract

The explosive growth of mobile edge users causes potential pressure for achieving their delay-sensitive requests in edge networks. Moreover, the incoming requests with task-dependency, which can be represented as Directed Acyclic Graphs (DAG), are hard to deal with effectively. In this paper, we intend to mitigate the DAG-based concurrent requests scheduling problem in an online manner. An Markov Decision Process (MDP) model is constructed for the proposed problem, where requests are split into a set of tasks and are assigned to different edge servers in terms of their status. To optimize the scheduling policy in each time slot while minimizing the long term system delay, we propose a Deep Reinforcement Learning (DRL)-based mechanism to promote the scheduling policy and make decision in each step. Extensive experiments are conducted, and evaluation results demonstrate that our proposed DRL technique can effectively improve the long-term performance of scheduling system, compared with other mechanisms.