摘要
多接入边缘计算(MEC)是一种新兴的云计算。低算力的物联网设备可以把计算任务卸载到MEC上处理,以提供更高质量的服务。当MEC的卸载网络在面临大量设备接入时,各设备请求服务时相互竞争的网络连接会发生大量碰撞从而导致MEC卸载网络的性能下降。在Wi-Fi作为MEC的接入点场景中,面对较少数量的设备接入时,802.11协议的退避算法可以合理地设置竞争窗口的值来减轻碰撞所带来的网络吞吐量下降,但默认的退避算法无法有效地应对较多的接入设备或动态变化的网络拓扑。为优化竞争窗口的设置以改善网络性能,提出两种竞争窗口优化的深度强化学习(DRL)方法,将深度Q网络(DQN)与深度确定性策略梯度(DDPG)方法分别用于优化MEC卸载网络竞争窗口大小的设置,以有效应对大量的接入设备和网络拓扑的动态变化。仿真实验结果表明,DRL方法在不同的接入设备数量、静态网络拓扑和动态网络拓扑的条件下,均可稳定网络的吞吐量,且相比于默认的方法有较大的提升,静态条件下相对提升46%,动态条件下相对提升36%,且并没有破坏网络服务的公平性。
Multi-access edge computing(MEC)is an emerging type of cloud computing.Low computing power IoT devices can offload computing tasks to MEC for processing to provide higher quality services.When the offloading network of MEC is faced with a large number of device accesses,there are a lot of collisions among competing network connections when each device requests services,thus leading to performance degradation of the MEC offload network.In the scenario where Wi-Fi is used as the access point for MEC,the 802.11 protocol’s back-off algorithm can reasonably set the value of the contention window(CW)to mitigate the network throughput degradation caused by collisions when facing a smaller number of devices,but the default back-off algorithm cannot effectively cope with a larger number of access devices or dynamically changing network topology.To optimize the setting of the contention window to improve network performance,two deep reinforcement learning(DRL)methods for CW optimization are proposed.The deep Q network(DQN)and deep deterministic policy gradient(DDPG)methods are used to optimize the setting of the contention window size for MEC offloading networks to effectively cope with a large number of access devices and dynamic changes in network topology,respectively.The simulation experimental results show that the DRL method stabilizes the network throughput under different numbers of access devices,static network topologies and dynamic network topologies,and has a large improvement over the default method,with a relative improvement of 46%under static conditions and 36%under dynamic conditions,without destroying the fairness of network services.
作者
詹御
张郭健
彭麟杰
文军
ZHAN Yu;ZHANG Guo-jian;PENG Lin-jie;WEN Jun(School of Information and Software Engineering,University of Electronic Science and Technology of China,Chengdu 610054,China)
出处
《计算机技术与发展》
2022年第6期99-105,共7页
Computer Technology and Development
基金
四川省科技项目(SCITLAB-0013)。
关键词
深度强化学习
多接入边缘计算
物联网
网络优化
竞争窗口
deep reinforcement learning
multi-access edge computing
internet of things
network optimization
contention window
