RLPAS: Reinforcement Learning-based Proactive Auto-Scaler for Resource Provisioning in Cloud Environment

摘要

Public cloud system offers Infrastructure-as-a-Service (IaaS) to deliver the computational resources on demand. Resource requirements of a cloud environment are always fluctuating because of the dynamic nature of the arriving workload, and traditional reactive scaling techniques are employed to deal with this problem. Automated resource provisioning is an effective methodology for handling workload fluctuations by provisioning the resources on demand. Simple reactive approaches affect the performance of elastic system by over-provisioning the resources that substantially increase the costs whereas under-provisioning leads to starvation. An intelligent resource provisioning mechanism overcomes the stated issues by allocating necessary resources by learning the environment dynamically. In this article, RLPAS (Reinforcement Learning based Proactive Auto-Scaler) algorithm is proposed, and it is based on the existing Reinforcement Learning (RL)-SARSA algorithm that learns the environment in parallel and allocates the resources. The performance of RLPAS algorithm is validated using real workloads, and it outperforms existing auto-scaling approaches in terms of CPU utilization, response time and throughput. Further, it also converges at an optimal time step and proves to be feasible for the extensive range of real cloud applications.