Containerized deployment of micro-services in fog devices: a reinforcement learning-based approach

• Published in: The Journal of supercomputing 2022-04, Vol.78 (5), p.6817-6845
• Main Authors: Nath, Shubha Brata, Chattopadhyay, Subhrendu, Karmakar, Raja, Addya, Sourav Kanti, Chakraborty, Sandip, Ghosh, Soumya K
• Format: Article
• Language: English
• Subjects: Cloud computing; Compilers; Computer networks; Computer Science; Data processing; Electronic devices; Gateways; Internet of Things; Interpreters; Machine learning; Optical character recognition; Optimization; Processor Architectures; Programming Languages; Response time; Switches; Workload; Workloads
• ISSN: 0920-8542; 1573-0484
• DOI: 10.1007/s11227-021-04135-2

The real power of fog computing comes when deployed under a smart environment, where the raw data sensed by the Internet of Things (IoT) devices should not cross the data boundary to preserve the privacy of the environment, yet a fast computation and the processing of the data is required. Devices like home network gateway, WiFi access points or core network switches can work as a fog device in such scenarios as its computing resources can be leveraged by the applications for data processing. However, these devices have their primary workload (like packet forwarding in a router/switch) that is time-varying and often generates spikes in the resource demand when bandwidth-hungry end-user applications, are started. In this paper, we propose pick–test–choose , a dynamic micro-service deployment and execution model that considers such time-varying primary workloads and workload spikes in the fog nodes. The proposed mechanism uses a reinforcement learning mechanism, Bayesian optimization , to decide the target fog node for an application micro-service based on its prior observation of the system’s states. We implement PTC in a testbed setup and evaluate its performance. We observe that PTC performs better than four other baseline models for micro-service offloading in a fog computing framework. In the experiment with an optical character recognition service, the proposed PTC gives average response time in the range of 9.71 sec–50 sec, which is better than Foglets (24.21 sec–80.35 sec), first-fit (16.74 sec–88 sec), best-fit (11.48 sec–57.39 sec) and mobility-based method (12 sec–53 sec). A further scalability study with an emulated setup over Amazon EC2 further confirms the superiority of PTC over other baselines.