Research On Energy-aware Virtual Machine Dynamic Consolidation

With the rapid development of cloud computing,the upsurge of data and the large demand for computing capabilities by users may result in excessive use of resources in cloud data centers,which not only causes the raising of energy consumption,but also leads to the reduction of Quality of Service(QoS)of cloud vendors and the increasing of Service Level Agreement Violation(SLAV).Currently,many studies have proved that dynamic consolidation of Virtual Machines(VMs)is one of the effective technologies to reduce energy consumption and improve QoS.Based on Deep Reinforcement Learning(DRL)technology,this thesis proposes an Adaptive DRL-based Virtual Machine Consolidation(ADVMC)method to reduce energy consumption in data center and ensure QoS.The dynamic VM consolidation mainly includes VM selection and VM placement.The main work of this thesis can be summarized into the following two parts:(1)Influence Coefficient-based Virtual Machine Selection PolicyMost of the existing VM selection strategies only consider the relationship between VMs,but do not consider the impact of each VM on the host overload.Aiming at the overloaded host,this thesis takes into account the influence of the VMs on the host on the basis of considering the usage of resources in the system.Firstly,an Influence Coefficient(IC)is constructed according to resource utilization to measure the impact of a VM on producing host overload;then an Influence Coefficient-based VM Selection(ICVMS)policy is proposed.It preferentially selects a VM with the greatest impact on producing host overload for migration,and finds the root cause of host overload quickly and accurately which can improve the accuracy of migration and enable the workload of overloaded hosts to return to a normal state quickly.The experimental results show that ICVMS proposed in this thesis can effectively reduce the energy consumption of the data center,help to optimize the QoS of users,and reduce SLAV.(2)Prediction Aware DRL-based Virtual Machine Placement PolicyAt present,a large number of researches regard the VM placement problem as a markov decision problem,and use Reinforcement Learning(RL)to solve it.This kind of algorithm essentially finds the best VM placement strategy adaptively by constructing a reward mechanism.However,due to the increasing number of hosts and VMs in cloud data centers,the existing RL-based VM placement algorithms generally suffer from the problem of huge state-action space,resulting in slow convergence.In addition,the random exploration strategy will randomly select actions in the initial stage of VM placement,resulting in extra energy consumption.In response to these problems,this thesis proposes a state prediction model based on Long Short-Term Memory(LSTM),which provides more reasonable data center environment states for training the DRL model to accelerate convergence.On this basis,a Prediction Aware DRL-based VM Placement(PADRL)algorithm is proposed.The reward function is built by energy consumption and SLAV jointly to find a suitable host for the VM to be migrated and adaptively obtain the most optimal placement strategy.Finally,combining the VM selection strategy and the VM placement strategy,ADVMC is constructed.The experimental results show that,compared with other VM consolidation algorithms,the VM consolidation method ADVMC proposed in this thesis has a significant effect on reducing energy consumption and ensuring SLA.