Autoscaling of Internet of Things applications in serverless edge computing

Internet of Things (IoT) services are deployed in edge clouds, which are close to end-users’ devices, when service latency is an important constraint, as often occurs. However, latency control cannot be achieved with ample resource allocation as in core clouds, since in edge environments resources are limited and thus operational efficiency is essential. Serverless computing, an emerging service model for both edge and core clouds, may help in this regard. In fact, autoscaling of serverless functions can provide both reduced latency and efficient resource usage. In this work, we focus on the usage of reinforcement learning to drive resource-based autoscaling of IoT applications on OpenFaaS, the most-adopted open-source serverless platform, running on Kubernetes clusters. We use the Proximal Policy Optimization (PPO) algorithm to dynamically configure the value of the Kubernetes Horizontal Pod Autoscaler, trained on a function trace representative of edge IoT traffic taken from the well-known Azure dataset. We discuss the performance of PPO as a function of service parameters in Kubernetes, obtaining an optimal configuration. With this configuration, we compare with state of the art baselines. Then, we delve into state definition, discussing performance and robustness of PPO with respect to traffic burstiness, as well as comparing its performance with an heuristic, which can be considered representative of the long term behavior of a trained Q-Learning approach. We show that, by limiting the features used in state space definition, it is possible to obtain a simpler yet more robust version of PPO, able to perform well under different burstiness of the input process. The price to pay is a small penalty in terms of resource utilization with base PPO when the burstiness of the input process during the test phase matches the one used for the training phase. In addition, we show that the performance of base PPO can be achieved with the designed heuristic, thus without the need of expensive training, provided that tail latency is not a primary requirement. • Study of the impact of state space variables and action space size in performance of PPO for controlling serverless edge computing autoscaling. • Evaluation of sensitivity of trained models to variations in traffic profile burstiness. • Comparison between performance of PPO variants, an a-priori heuristic, and multiple baselines (static HPA, SAC, predictive non reinforcement learning).