A Machine Learning Method For Virtual Machine Placement

Cloud computing has revolutionized the way networks are used,and has made rapid development in recent years.In order to improve the resource utilization of data center in cloud computing and reduce carbon emissions,"green cloud data centers" has been widely studied in recent years.A reasonable virtual machine placement(VMP)strategy is of great significance to reduce energy consumption and improve the performance of data center in cloud computing.In this paper,we establish an integer programming model for the virtual machines placement problem.The optimization goal is to minimize the number of active physical machines.We consider the resource requirements of the virtual machines with CPU and memory,all virtual machines are successfully deployed under the resource constraints of physical machine.The best virtual machine placement strategy can reduce the number of active physical machines,thus improve cloud computing performance and achieve the goal of green cloud data center.We propose a method based on machine learning to solve the problem of virtual machine placement.The main work of this paper is summarized as follows:? Considering the similarity between virtual machine placement and the multiclassification of machine learning,we model the deployment of virtual machines as the process of finding the best category for virtual machines.We construct a multi-classification framework based on the LightGBM model.The virtual machine resource requirement data is generated randomly.We use Cplex to label the training data,and train the model using supervised learning.? In order to improve the scalability of machine learning models,we propose a training method of unified input data dimensions and the extension method based on sliding window optimization.We use the linear dimensionality reduction method PCA and the non-linear dimensionality reduction method t-SNE to dimensionally unify the input of the multi-classification model,and the best dimensionality unification method was selected according to the simulation result.In the test phase,the probability of each category is calculated within a specified window,and the window is slid according to rules.We use this method to solve the large-scale problems that do not lie in the categories of training data.? We adopt the method of active learning to overcome the difficulty in obtaining large-scale virtual machine sequence training data labels.According to the query strategy,we select the data that has a significant effect on model training,and label these "important" data,so as to obtain a machine learning model with good performance by using less training data.