The emergence of fog computing has shown promising results for reducing network latency and congestion in the cloud. In this environment, effective autoscaling to handle spatio-temporal load bursts in geographically distributed and resource-constrained fog nodes has become a timely problem. A typical strategy for autoscaling is based on centralized monitoring of the fog nodes and the deployed service instances. However, centralized collection and analysis of the metrics for autoscaling can become infeasible with the increasing number of fog nodes. Moreover, the dynamic and fluctuating characteristics of the fog nodes make effective autoscaling challenging in fog computing environments. In this work, we propose HiDRA, a Hierarchical Decentralized Autoscaler that scales and places microservice instances based on multi-agent reinforcement learning. In HiDRA, agents are divided into scaling and placement agents that collaborate with each other to effectively handle spatio-temporal load bursts in fog computing. These Deep Q-Network-based autoscaling agents are trained solely based on their regional observations at runtime, eliminating the need for a centralized collection of metrics. We evaluated HiDRA in multiple simulated fog environments created using a real-world dataset. The environments were divided into three levels of sparsity, each consisting of 20, 15, and 10 initial instances and unstable nodes. The result shows that comparatively by ratio against the baseline, HiDRA increased the average request success rate by 10.7%, 16.4%, and 36.7% and reduced the number of created instances by 12.3%, 15.9%, and 16.8% in environments with 20, 15, and 10 initial instances and unstable nodes, respectively.

While the emergence of fog computing has shown promising results for reducing the network latency and congestion in the cloud, effective autoscaling to handle spatio-temporal bursts of load has become a timely problem. In this work, we propose HiDRA, a Hierarchical Decentralized Reactive Autoscaler that scales and places microservice instances based on multi-agent reinforcement learning. In HiDRA, the agents are divided into scaling and placement agents that collaborate with each other to effectively handle spatio-temporal bursts of load in fog computing. These hierarchical agents are trained solely based on their regional observations at runtime, eliminating the need for a centralized collection of metrics. We evaluated HiDRA in 20 simulated fog environments to show that HiDRA reduces the created number of instances by 36.70% while resulting in similar scaling performance in terms of response success rate.

With the increasing number of Web of Things devices, the network and processing delays in the cloud have also increased. As a solution, fog computing has emerged, placing computational resources closer to the user to lower the communication overhead and congestion in the cloud. In fog computing systems, microservices are deployed as containers, which require an orchestration tool like Kubernetes to support service discovery, placement, and recovery. A key challenge in the orchestration of microservices is automatically scaling the microservices in case of an unpredictable burst of load. In cloud computing, a centralized autoscaler can monitor the deployed microservice instances and make scaling actions based on the monitored metric values. However, monitoring an increasing number of microservices in fog computing can cause excessive network overhead and thereby delay the time to scaling action. We propose DESA, a fully DEcentralized Self-adaptive Autoscaler through which microservice instances make their own scaling decisions, cloning or terminating themselves through self-monitoring. We evaluate DESA in a simulated fog computing environment with different numbers of fog nodes. Furthermore, we conduct a case study with the 1998 World Cup website access log, examining DESA’s performance in a realistic scenario. The results show that DESA successfully reduces the scaling reaction time in large-scale fog computing systems compared to the centralized approach. Moreover, DESA resulted in a similar maximum number of instances and lower average CPU utilization during bursts of load.