DeepMRA: An Efficient Microservices Resource Allocation Framework with Deep Reinforcement Learning in the Cloud

The rapid growth of cloud computing has precipitated a paradigm shift in application service deployment, transitioning predominantly from monolithic to microservices architectures. This shift to microservices brings new, complex challenges in managing cloud resources. Traditional cloud resource allocation methods struggle with microservices' unique challenges, like complex inter-service dependencies and the need to balance Quality of Service (QoS) with cost efficiency. Recognizing these challenges in cloud resource management, our study proposes an innovative approach to dynamically allocate resources for cloud microservices with Deep Reinforcement Learning (DRL). Specifically, we introduce DeepMRA, an efficient microservices resource allocation framework with Deep Reinforcement Learning in the Cloud, with multiple agents navigating the complexities arising from varying workloads. We propose a performance predictor to forecast application performance, guiding the training of agents in DRL. Due to the shortcomings of traditional performance data collection methods in the context of microservices, we developed the Parallel and Asynchronous Uncertainty-Directed Sampling (PAUDS) algorithm. This algorithm is specifically designed to optimize data collection processes, ensuring a robust dataset for building a reliable performance predictor. Extensive experiments conducted with microservice-based applications indicate that the proposed method reduces resource consumption while upholding QoS requirements under varying workloads.