Deep Reinforcement Learning based Compute-Intensive Workload Allocation in Data Centers with High Energy Efficiency

Abstract

Recently the huge amount of energy consumption has become a barrier to the widespread deployment of data centers serving various Internet of Things applications. The reasonable allocation of compute-intensive workloads to physical servers is an efficient way to improve the data center's energy efficiency. Though existing works has proposed some algorithms to manage workloads or virtual machines for energy saving, most of them did not comprehensively consider the high dynamics of server states, and lacked in high scalability in their implementation. In this paper, the Actor Critic based Compute-Intensive Workload Allocation Scheme (AC-CIWAS) is proposed, which can both guarantee the Quality of Service (QoS) of workloads and reduce the computational energy consumption of physical servers. To achieve rational workload allocation, AC-CIWAS captures the dynamic feature of server states continuously, and takes the impact of different workloads on energy consumption into consideration. AC-CIWAS employs the Deep Reinforcement Learning (DRL) based Actor Critic (AC) algorithm to evaluate the expected cumulative return over time, while the cumulative return guides to allocate workloads with high energy efficiency. Simulation results have demonstrated that compared to existing baseline allocation methods, the proposed AC-CIWAS can achieve an approximately 20 percent decrease in server power consumption with QoS guarantee.