Institute of Theoretical and Applied Informatics, Polish Academy of Sciences, 44-100 Gliwice, Poland.
Laboratoire I3S, Université Côte d'Azur, 06103 Nice, France.
Sensors (Basel). 2021 Apr 29;21(9):3105. doi: 10.3390/s21093105.
The short latency required by IoT devices that need to access specific services have led to the development of Fog architectures that can serve as a useful intermediary between IoT systems and the Cloud. However, the massive numbers of IoT devices that are being deployed raise concerns about the power consumption of such systems as the number of IoT devices and Fog servers increase. Thus, in this paper, we describe a software-defined network (SDN)-based control scheme for client-server interaction that constantly measures ongoing client-server response times and estimates network power consumption, in order to select connection paths that minimize a composite goal function, including both QoS and power consumption. The approach using reinforcement learning with neural networks has been implemented in a test-bed and is detailed in this paper. Experiments are presented that show the effectiveness of our proposed system in the presence of a time-varying workload of client-to-service requests, resulting in a reduction of power consumption of approximately 15% for an average response time increase of under 2%.
物联网设备需要访问特定服务，这就要求其具有较短的延迟，因此雾计算架构应运而生，它可以作为物联网系统和云之间的有用中介。然而，随着物联网设备数量的增加，人们开始关注此类系统的功耗问题。因此，在本文中，我们描述了一种基于软件定义网络 (SDN) 的客户端-服务器交互控制方案，该方案可以持续测量客户端-服务器的响应时间并估计网络功耗，从而选择能够最小化包括服务质量和功耗在内的复合目标函数的连接路径。我们使用神经网络的强化学习方法在测试平台上实现了该方案，并在本文中详细介绍了该方案。实验结果表明，在存在时变的客户端到服务请求工作负载的情况下，我们提出的系统是有效的，平均响应时间增加不到 2%，而功耗降低了约 15%。
