A dual scheduling framework for task and resource allocation in clouds using deep reinforcement learning

Abstract

Abstract In the three-tier cloud computing marketplace, Software-as-a-Service (SaaS) providers play a dual role. As consumers of Infrastructure-as-a-Service (IaaS) providers, they rent virtual machine (VM) instances from public cloud platforms on a pay-as-you-go basis to deploy their applications. Meanwhile, as service providers for end users, they are responsible for handling changing workloads consisting of user-submitted tasks with some QoS requirements. Under this business model, how to minimize the cost of leasing VM instances while guaranteeing the quality of service for SaaS applications is a challenging issue. In this paper, we present an intelligent dual scheduling framework for SaaS providers. The QoS-aware task allocation scheduler and the workload-aware VM auto-scaling scheduler work together to achieve efficient task scheduling and elastic resource provisioning. Accordingly, we propose a dual scheduling approach based on deep reinforcement learning, which consists of two layers of Deep Q-Network algorithms to realize these functions. We conducted experiments using simulated workloads and real-world workloads under fixed and elastic resource pool conditions, respectively. Experimental results show that our approach is effective in ensuring higher task success rates, improving resource utilization, shortening average task response time, and significantly reducing the cost of renting VM instances compared to other baseline solutions.

Keywords

• Cloud computing
• Task scheduling
• VM Auto-scaling
• Deep reinforcement learning