Automated Resource Dimensioning in Cloud Using Hybrid Reinforcement Learning

Resource dimensioning refers to the process of determining the amount of resources needed to achieve target KPIs of Virtual Network Functions (VNFs) in a Network Service (NS) in a cost-effective fashion. VNFs in an NS usually have some dependency relationship among themselves. This relationship can either be represented as a chain of VNFs (in the case of Service Function Chain "SFC"), or by more general graph topologies (Virtual Network Function - Forwarding Graph "VNF-FG"). In cloud computing, a similar concept can be found in microservices whereby microservices interact with each other to implement the functionality of the application. The dependencies between the VNFs (or microservices) imply that the performance of a given VNF does not depend only on the amount of resources available to itself, but also on the performance of other VNFs on which it depends. This makes developing accurate solutions for resource dimensioning a challenging task. In this paper, we propose a Hybrid Reinforcement Learning (RL)-based solution to automatically generate resource amounts for VNFs such that they meet the expected performance of the NS with minimal resources. The proposed solution relies both on a simulation and a real cloud environment. We evaluated the performance of the proposed solution in terms of training and inferencing convergence, and inferencing time. The results demonstrate that our training and inferencing algorithms successfully converge to an optimal solution.