Research On Prediction-Based Migration Of Virtual Machines In Cloud Environments

With the rapid development of digitalization and informatization,various industries have accumulated massive amounts of data,and the demand for high-performance computility by enterprises has exploded.Cloud computing services,as a collection of emerging technologies and a successful business model,fits perfectly into the urgent need for arithmetic power and storage of related enterprises.However,as the number and size of cloud data centers continue to increase,inefficient resource management methods make the energy waste in cloud data centers more and more serious,and new resource management methods for cloud environments need to be studied to improve the resource utilization of server clusters and reduce the energy consumption in cloud data centers.Among all resource management approaches applied to cloud data centers,dynamic virtual machine migration technology can effectively utilize the physical resources inside cloud data centers to achieve the effects of reducing energy consumption,guaranteeing user service quality,and achieving server load balancing,etc.Therefore,the research in this paper focuses on dynamic virtual machine migration problem.The traditional dynamic VM migration process can be roughly divided into three phases:host state detection,VM selection,and VM placement,while this paper adopts a proactive resource allocation strategy of prediction before migration,which first predicts the future load of VMs using a load prediction model and then takes targeted migration operations for VMs based on the prediction results.Therefore,this paper adds a load prediction phase before the traditional dynamic VM migration process,and proposes corresponding improved algorithms for the load prediction phase,host state detection phase and VM placement phase,respectively.In the load prediction phase,this paper proposes a dual-channel hybrid load prediction model HSG＿CBLSTM based on deep learning,which uses SG filter to divide the original load data into two parts and captures the low-frequency features and high-frequency features using two prediction channels respectively to jointly predict the future load.This approach not only reduces the prediction pressure of a single channel,but also compensates for the lost detail information of the filtered data,and employs both convolutional neural network and bi-directional long short-term memory network in the prediction process to improve the model prediction accuracy.In the host state detection phase,this paper proposes MARLOT,an overload detection algorithm based on multi-agent reinforcement learning,which models the overload detection process as a reinforcement learning model and utilizes the self-learning mechanism of the agents in reinforcement learning to learn the respective most suitable threshold determination policy for each physical host based on the feedback rewards from the environment,thereby reducing the energy consumption and service violations in the cloud data center.In the virtual machine placement phase,this paper proposes a predictive capacity-based VM placement algorithm PCB,which first predicts the future VM load situation using the HSG＿CBLSTM prediction model,and takes into account the stability of the future host load and the increase in energy consumption in the VM placement process,reduces the possibility of service violations by introducing penalty factors,and uses greedy thinking to select physical hosts with higher future resource utilization,which reduces the overall energy consumption of the system and decreases the load balance degree.In terms of load prediction,this paper uses real-world datasets from Alibaba and Google to validate the prediction accuracy of the HSG＿CBLSTM prediction model,and the experimental results show that the prediction accuracy of the HSG＿CBLSTM model outperforms that of the single-channel model,the statistical model and the compared deep learning model.In terms of dynamic VM migration algorithms,the experiments use the Cloud Sim cloud simulation platform to verify the effectiveness of the MARLOT overload detection algorithm and the PCB VM placement algorithm using the Planet Lab dataset,respectively.Finally,the two are combined with the HSG＿CBLSTM prediction model to implement the prediction-based dynamic VM migration algorithm MARLOT-PCB and compare it with the mainstream VM migration algorithms.The experimental results show that compared with the comparison algorithms,the MARLOT-PCB algorithm proposed in this paper can achieve lower energy consumption level and service violation rate,and at the same time can guarantee a low load balancing degree.The above experimental results show that the MARLOT-PCB dynamic VM migration algorithm proposed in this paper can effectively reduce the overall energy consumption and service violation rate of the cloud data center,while ensuring the load balancing among servers.