HANSEL: Adaptive horizontal scaling of microservices using Bi-LSTM

Abstract

With the rapid development of 5G network, business scenarios such as intelligent service and new retail are becoming more and more popular. The demand for more flexible and scalable real-time data processing, in particular, the AI-related data processing has also increased in edge computing. Therefore, how to meet such business development has become a major challenge. Focusing on this requirement, microservice architecture, proposed and developed by some big cloud computing companies’ platform, such as Google Kubernetes platform, has gradually become a mainstream technology solution in edge computing.However, many microservices used in edge computing cannot achieve an even time distribution, which is random or sudden. Kubernetes built-in Horizontal POD Autoscaling (HPA) is unable to well handle the change of microservice load, which inevitably leads to the waste of system resources and affects the SLA of microservice.To solve this issue, this paper proposes a HANSEL system based on Kubernetes platform, which can optimize the horizontal elastic scaling policy of Kubernetes by accurately predicting the load of microservices based on the Bi-LSTM load prediction algorithm with attention mechanism. Furthermore, active elastic scaling is realized through reinforcement learning method, and we design a hybrid elastic scaling mechanism through combining reactive and active methods, so as to construct an elastic scaling system for automatic scheduling of working nodes. Our experimental results show that HANSEL system can improve the system resource utilization by about 20% when meeting the microservice SLA of edge computing.