Energy Efficiency Dynamic Modeling for Cloud Computing Based On Deep Learning

Abstract

Automatic decision-making strategies, for example re- inforcement learning (RL), have been employed to (partially) solve the resource allocation problem adaptively from the cloud calculating system. However, high dimensions are exhibited by a complete cloud resource allocation frame in action and state spaces, which prohibit the usefulness of traditional RL methods. Higher power consumption has become one of the concerns in designing and control of cloud computing systems, which degrades system reliability and raises cooling cost. An effective dynamic power management (DPM) policy must minimize power consumption while keeping performance degradation in a acceptable degree. Thus, a joint virtual machine (VM) resource allocation and power management platform is essential to the overall cloud computing platform. Novel solution framework is imperative to tackle the dimensions in state and action spaces. In this paper, we suggest a novel hierarchical framework for solving power management problem and the resource allocation in cloud computing methods. The suggested hierarchical framework comprises a grade for VM resource allocation into a local tier for power management of servers that are and the servers. The emerging profound reinforcement learning (DRL) technique, which can deal with complex control problems with large state space, is adopted to solve the global tier problem. Additional an autoencoder along with a novel weight sharing structure are adopted to handle the condition space and accelerate the rate. On the other hand, the local tier of server power managements comprises an LSTM based workload predictor plus a model-free RL established energy manager, working in a dispersed manner. Experiment results using real Google cluster traces demonstrate that our proposed frame significantly conserves energy usage and the energy consumption than the when. Meanwhile, the proposed framework can achieve the best trade-off between latency and power/energy ingestion.